CN117371812A - Aircraft group collaborative decision generation method, system and equipment - Google Patents

Abstract

The invention discloses an aircraft group collaborative decision-making generation method, system and equipment, and relates to the technical field of aircraft. By using the preset density peak clustering algorithm to calculate event pseudo-labels, and training the initial event extraction model based on the event pseudo-label set, the target event extraction model is obtained. The target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively and perform event comparison to obtain the solution comparison results. When the plan comparison results are inconsistent, the optimal instruction sequence is selected based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic policy gradient algorithm to obtain the collaborative decision-making of the aircraft group. When the plan comparison results are consistent, the aircraft group plan will be used as the aircraft group collaborative decision-making. Fine-tune the initial event extraction model based on event pseudo-labels, and then extract events to quickly grasp the overall situation of the flight environment and formulate corresponding plans to deal with emergencies.

Description

A method, system and device for collaborative decision-making generation of aircraft groups

Technical field

The present invention relates to the field of aircraft technology, and in particular to methods, systems and equipment for generating collaborative decision-making for aircraft groups.

Background technique

The role of aircraft is to carry out flight missions inside or outside the atmosphere, including transportation, reconnaissance, military, scientific research and other fields. In transportation, aircraft can be used for long-distance travel and cargo transportation, greatly shortening distances and improving transportation efficiency. In scientific research, aircraft are widely used in various scientific research, such as astronomy, meteorology, geology and other fields. Among them, an aircraft group is a group composed of multiple aircraft that work together to complete specific tasks, such as reconnaissance and monitoring, rescue and search, logistics and transportation, etc. Swarms of aircraft enable more efficient, precise operations and, in some cases, greater capabilities and flexibility.

The flight control and stability requirements of a group of high-speed flying aircraft are higher, because when performing a mission, multiple high-speed flying aircraft need to effectively ensure coordinated operation and avoid mutual interference. This requires real-time, efficient communication and collaborative work among aircraft groups to ensure cooperation and coordination between individual aircraft.

However, aircraft will face more complex challenges in the aerodynamic environment, such as airflow turbulence, aerodynamic instability, etc. These factors may cause aircraft attitude control to be difficult, unstable or even out of control. However, the existing cooperative decision-making of aircraft groups is usually set in advance and cannot be adjusted according to emergencies. As a result, the failure of a single aircraft in a high-speed flying aircraft group may affect the normal completion of the operation and even cause damage to the normal operation of other aircraft. interference, causing huge losses.

Contents of the invention

The present invention provides a method, system and equipment for generating collaborative decision-making of an aircraft group, which solves the problem that the existing collaborative decision-making of an aircraft group cannot be adjusted according to emergencies and is prone to affect the normal completion of operations due to the failure of a single aircraft in the aircraft group. technical problem.

The invention provides an aircraft group collaborative decision-making generation method, including:

When receiving the emergency situation data of the aircraft group, a preset density peak clustering algorithm is used to calculate the event pseudo-label corresponding to the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station, Get the event pseudo-label set;

Train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model;

The target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively and perform event comparison to obtain solution comparison results;

When the solution comparison results are inconsistent, the optimal instruction sequence is selected based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the aircraft group corresponding to the emergency situation data. Collaborative decision-making solutions;

When the plan comparison results are consistent, the aircraft group plan in the aircraft group data set is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

Optionally, when receiving emergency situation data of an aircraft group, a preset density peak clustering algorithm is used to calculate the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station. Corresponding event pseudo-labels, the steps to obtain the event pseudo-label set include:

When receiving the emergency situation data of the aircraft group, obtain the aircraft group data set generated by the aircraft group based on the emergency situation data and the ground control station data set generated by the ground control station;

Extract the feature vectors of the aircraft group data set and the ground control station data set through a preset word skip model to obtain a feature vector set;

Calculate the density corresponding to each point in the feature vector set using a preset density calculation formula to obtain the density;

The preset density calculation formula is:

Among them, ρ(xi) represents the density of point x _i ; k represents the number of neighbors of a point, that is, using _the Euclidean distance as the metric, the top k points closest to a point; Represents the Euclidean distance between point x _i and point x _j , u represents the dimension of a feature vector concentration point, that is, the number of features, g is a subscript, used to index different features; KNN(xi ₎ represents the A set of the first k points with the smallest Euclidean distance from point x _i ; i represents the i-th point x _i in the data set; j represents the j-th point x _j in the data set;

Substitute the density into the preset relative density calculation formula to calculate the relative density of the point to obtain the relative density;

The preset relative density calculation formula is:

Among them, r _ρ ( _xi ) represents the relative density of point x _i ; ρ ( _xi ) represents the density of point _xi ; ρ (x _j ) represents the density of point x _j ; k represents the number of neighbors of a point, that is Using Euclidean distance as the metric, the top k points closest to a point; KNN(xi ₎ represents the set of the top k points with the smallest Euclidean distance from point x _i in the data set; i represents the i-th point in the data set point x _i ;j represents the jth point x _j in the data set;

Select points where the relative density is greater than the density average corresponding to the relative density to obtain multiple cluster centers;

According to the cluster center, the feature vector set is divided using a preset domain calculation formula to obtain multiple sets;

The calculation formula of the preset domain is:

D _mn (x _p )={x _q |x _q ∈MN(x _p )∨(x _m ∈MN(x _p ))∧x _q ∈MN(x _m ))};

In the formula, D _mn (x _p ) represents the domain of point x _p ; MN (x _p ) represents the set of all neighbors of point x _p ; MN (x _m ) represents the set of all neighbors of point x _m ; x _p represents the feature The p-th point in the vector set; x _m represents the m-th point in the feature vector set; x _q represents the q-th point in the feature vector set;

According to the labels corresponding to the cluster centers, corresponding labels are set for the points corresponding to each set to obtain an event pseudo-label set.

Optionally, the step of training an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model includes:

Input the event pseudo-label set into the initial event extraction model to obtain the initial loss function value;

Fine-tune the initial event extraction model according to the initial loss function value to obtain an intermediate event extraction model;

Input the event pseudo-label set into the intermediate event extraction model to obtain the target loss function value;

Determine whether the target loss function value is a preset function value;

If so, use the intermediate event extraction model corresponding to the target loss function value as the target event extraction model;

If not, use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain an initial loss function value.

Optionally, the step of extracting events from the aircraft group data set and the ground control station data set respectively through the target event extraction model and performing event comparison to obtain solution comparison results includes:

Extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model to obtain aircraft group events and ground events;

Compare the aircraft group event and the ground event according to the trigger word category and parameter similarity to obtain an event comparison result;

Using all the event comparison results, a plan comparison result is constructed.

Optionally, when the plan comparison results are inconsistent, the optimal instruction sequence is selected according to the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the burst The steps of the aircraft group collaborative decision-making plan corresponding to the status data include:

When the plan comparison results are inconsistent, determine whether the authorization data corresponding to the aircraft group is authorized;

If so, determine the aircraft group collaborative decision-making plan corresponding to the emergency situation data based on the command sequence within the preset time and the aircraft group plan;

If not, perform command sequence selection on the aircraft group data set and the ground control station data set according to the preset improved multi-agent depth deterministic policy gradient algorithm to obtain the target command sequence;

Determine whether the target instruction sequence is a preset sequence;

If so, use the aircraft group plan as the aircraft group collaborative decision-making plan corresponding to the emergency situation data;

If not, the target command sequence is used as the aircraft group collaborative decision-making solution corresponding to the emergency situation data.

Optionally, the step of determining the aircraft group collaborative decision-making plan corresponding to the emergency situation data based on the instruction sequence within the preset time and the aircraft group plan includes:

Determine whether the aircraft group and the ground control station have found the optimal command sequence within the preset time;

If so, use the optimal instruction sequence as the aircraft group collaborative decision-making solution corresponding to the emergency situation data;

If not, the aircraft group plan is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

Optionally, the step of selecting an instruction sequence from the aircraft group data set and the ground control station data set according to a preset improved multi-agent deep deterministic policy gradient algorithm to obtain a target instruction sequence includes:

Using the decision criteria corresponding to the aircraft group data set and the ground control station data set to construct an initial decision criterion set;

Select the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion;

According to the priority decision-making criterion, an instruction acquisition algorithm is used to assign preset weights to instructions corresponding to the aircraft group data set and the ground control station data set, respectively, to obtain an instruction sequence set;

Using a preset improved multi-agent deep deterministic policy gradient algorithm to select multiple initial instruction sequences that meet the preset score threshold in the instruction sequence set, and construct an initial instruction queue;

Select the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence;

The target command sequence is determined based on the constraints of the aircraft group on itself and the intermediate command sequence.

Optionally, the step of determining the target instruction sequence based on the aircraft group's own constraints and the intermediate instruction sequence includes:

Determine whether the aircraft group can execute the intermediate command sequence under its own constraints;

If so, use the intermediate instruction sequence as the target instruction sequence;

If not, delete the intermediate instruction sequence from the initial instruction queue to obtain an intermediate instruction queue;

Determine whether the intermediate instruction queue is an empty set;

If so, delete the priority decision criterion from the initial decision criterion set to obtain a target decision criterion set;

Determine whether the set of target decision criteria is an empty set;

If so, use the preset sequence as the target instruction sequence;

If not, use the target decision criterion set as the initial decision criterion set, and jump to the step of selecting the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion;

If not, use the intermediate instruction queue as the initial instruction queue, and jump to the step of selecting the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence.

The invention also provides an aircraft group collaborative decision-making system, which includes:

The event pseudo-label set obtaining module is used to, when receiving emergency situation data of an aircraft group, use a preset density peak clustering algorithm to calculate the aircraft group data set corresponding to the emergency situation data and the ground control generated by the ground control station. The event pseudo-label corresponding to the station data set is obtained to obtain the event pseudo-label set;

A target event extraction model obtaining module is used to train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model;

A solution comparison result obtaining module is used to extract events from the aircraft group data set and the ground control station data set through the target event extraction model and perform event comparison to obtain a solution comparison result;

The first acquisition module of aircraft group collaborative decision-making is used to select the optimal instruction sequence based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm when the plan comparison results are inconsistent. Obtain the aircraft group collaborative decision-making plan corresponding to the emergency situation data;

The second acquisition module of aircraft group collaborative decision-making is used to use the aircraft group plan in the aircraft group data set as the aircraft group collaborative decision-making plan corresponding to the emergency situation data when the plan comparison results are consistent.

The present invention also provides an electronic device, including a memory and a processor. A computer program is stored in the memory. When the computer program is executed by the processor, the processor executes any one of the above-based implementations. Steps of the collaborative decision-making generation method for aircraft swarms.

It can be seen from the above technical solutions that the present invention has the following advantages:

The present invention obtains an event pseudo-label set by using a preset density peak clustering algorithm to calculate the event pseudo-labels corresponding to the aircraft group data set corresponding to the emergency data and the ground control station data set generated by the ground control station. The initial event extraction model is trained based on the event pseudo-label set to obtain the target event extraction model. The target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively and perform event comparison to obtain the solution comparison results. When the plan comparison results are inconsistent, the optimal instruction sequence is selected based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the aircraft group collaborative decision-making plan corresponding to the emergency situation data. When the plan comparison results are consistent, the aircraft group plan in the aircraft group data set is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data. It solves the technical problem that the existing cooperative decision-making of the aircraft group cannot be adjusted according to emergencies, and the normal completion of the operation is easily affected by the failure of a single aircraft in the aircraft group. Fine-tune the initial event extraction model based on event pseudo-labels, and then extract events to quickly grasp the overall situation of the flight environment and formulate corresponding plans to deal with emergencies. The improved multi-agent deep deterministic policy gradient algorithm is used to find the instruction sequence and do its best to enable the aircraft group to avoid threats, complete tasks, and successfully respond to emergencies.

Description of the drawings

In order to explain the embodiments of the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting any creative effort.

Figure 1 is a step flow chart of a method for generating collaborative decision-making for an aircraft group provided by Embodiment 1 of the present invention;

Figure 2 is a step flow chart of an aircraft group collaborative decision-making generation method provided in Embodiment 2 of the present invention;

Figure 3 is a CNN-based event extraction model framework provided by Embodiment 2 of the present invention;

Figure 4 is a flow chart of the event extraction model fine-tuning process provided by Embodiment 2 of the present invention;

Figure 5 is a flow chart of an execution instruction acquisition algorithm provided by Embodiment 2 of the present invention;

Figure 6 is a flow chart of an aircraft group collaborative decision-making generation method provided in Embodiment 2 of the present invention;

Figure 7 is a schematic diagram of an emergency system applying the aircraft group collaborative decision-making generation method provided in Embodiment 2 of the present invention;

Figure 8 is a structural block diagram of an aircraft group collaborative decision-making generation system provided in Embodiment 3 of the present invention.

Detailed ways

Embodiments of the present invention provide an aircraft group collaborative decision-making generation method, system and equipment, which are used to solve the problem that the existing aircraft group collaborative decision-making cannot be adjusted according to emergencies, and it is easy to affect the operation due to the failure of a single aircraft in the aircraft group. Technical issues with normal completion.

In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, what is mentioned below The described embodiments are only some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

Please refer to FIG. 1 , which is a flow chart of a method for generating collaborative decision-making for an aircraft group according to Embodiment 1 of the present invention.

Example 1 of the present invention provides an aircraft group collaborative decision-making generation method, including:

Step 101. When receiving the emergency situation data of the aircraft group, use the preset density peak clustering algorithm to calculate the event pseudo-label corresponding to the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station. , get the event pseudo-label set.

In the embodiment of the present invention, the aircraft group can be regarded as multiple intelligent agents. When these intelligent agents encounter an emergency situation, in order to cope with the emergency situation, they will generate new instruction sequences. These new command sequences are often inconsistent with the old command sequences given to the aircraft group by the ground control station before departure. In addition, the old command sequence is called the ground control station scheme and the new command sequence is called the aircraft group scheme. The aircraft group data set includes aircraft group schemes and aircraft group data. The ground control station data set includes ground control station schemes and ground control station data.

When the emergency situation data of the aircraft group is received, the aircraft group data set generated by the aircraft group based on the emergency situation data and the ground control station data set generated by the ground control station are obtained. The feature vectors of the aircraft group data set and the ground control station data set are extracted through the preset skip model, and the feature vector set is obtained, where the preset skip model refers to the Skip-gram model. Use the preset density calculation formula to calculate the density corresponding to each point in the feature vector set to obtain the density. Substitute the density into the preset relative density calculation formula to calculate the relative density of the point and obtain the relative density. Select points whose relative density is greater than the density average corresponding to the relative density to obtain multiple cluster centers. According to the cluster center, the feature vector set is divided using the preset domain calculation formula to obtain multiple sets. According to the label corresponding to the cluster center, corresponding labels are set for the points corresponding to each set to obtain the event pseudo-label set.

Step 102: Train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model.

In the embodiment of the present invention, the initial loss function value is obtained by inputting the event pseudo-label set into the initial event extraction model. According to the initial loss function value, the initial event extraction model is fine-tuned to obtain the intermediate event extraction model. Input the event pseudo-label set into the intermediate event extraction model to obtain the target loss function value. Determine whether the target loss function value is the preset function value. If so, the intermediate event extraction model corresponding to the target loss function value is used as the target event extraction model. If not, use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

Step 103: Extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model and perform event comparison to obtain the solution comparison results.

In the embodiment of the present invention, event extraction is performed on the aircraft group data set and the ground control station data set respectively through the target event extraction model to obtain aircraft group events and ground events. According to the trigger word category and parameter similarity, the aircraft group events and ground events are compared to obtain the event comparison results. Use all event comparison results to construct plan comparison results.

Step 104. When the plan comparison results are inconsistent, select the optimal instruction sequence based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the aircraft group collaboration corresponding to the emergency situation data. Decision plan.

In the embodiment of the present invention, when the solution comparison results are inconsistent, it is determined whether the authorization data corresponding to the aircraft group is authorized. If so, the aircraft group collaborative decision-making plan corresponding to the emergency data is determined based on the command sequence and aircraft group plan within the preset time. If not, then the command sequence is selected from the aircraft group data set and the ground control station data set according to the preset improved multi-agent deep deterministic policy gradient algorithm to obtain the target command sequence. Determine whether the target instruction sequence is a preset sequence. If so, the aircraft group plan is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data. If not, the target command sequence is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

Step 105: When the plan comparison results are consistent, use the aircraft group plan in the aircraft group data set as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

In the embodiment of the present invention, when more than 80% of the events in the ground control station plan and the aircraft group plan are the same, the ground control station plan and the aircraft group plan are considered to be consistent. When the two plans are consistent, the aircraft group plan is executed, that is, the aircraft group plan in the aircraft group data set is used as the aircraft group collaborative decision-making plan corresponding to the emergency data.

In the embodiment of the present invention, the event pseudo-label set is obtained by using a preset density peak clustering algorithm to calculate the event pseudo-labels corresponding to the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station. The initial event extraction model is trained based on the event pseudo-label set to obtain the target event extraction model. The target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively and perform event comparison to obtain the solution comparison results. When the plan comparison results are inconsistent, the optimal instruction sequence is selected based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the aircraft group collaborative decision-making plan corresponding to the emergency situation data. When the plan comparison results are consistent, the aircraft group plan in the aircraft group data set is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data. It solves the technical problem that the existing cooperative decision-making of the aircraft group cannot be adjusted according to emergencies, and the normal completion of the operation is easily affected by the failure of a single aircraft in the aircraft group. Fine-tune the initial event extraction model based on event pseudo-labels, and then extract events to quickly grasp the overall situation of the flight environment and formulate corresponding plans to deal with emergencies. The improved multi-agent deep deterministic policy gradient algorithm is used to find the instruction sequence and do its best to enable the aircraft group to avoid threats, complete tasks, and successfully respond to emergencies.

Please refer to Figure 2. Figure 2 is a flow chart of a method for generating collaborative decision making for an aircraft group according to Embodiment 2 of the present invention.

Another aircraft group collaborative decision-making generation method provided by Example 2 of the present invention includes:

Step 201. When receiving the emergency situation data of the aircraft group, use the preset density peak clustering algorithm to calculate the event pseudo-label corresponding to the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station. , get the event pseudo-label set.

Further, step 201 may include the following sub-steps S11-S17:

S11. When receiving the emergency situation data of the aircraft group, obtain the aircraft group data set generated by the aircraft group based on the emergency situation data and the ground control station data set generated by the ground control station.

S12. Extract the feature vectors of the aircraft group data set and the ground control station data set through the preset word skip model to obtain a feature vector set.

S13. Use the preset density calculation formula to calculate the density corresponding to each point in the feature vector set to obtain the density.

S14. Substitute the density into the preset relative density calculation formula to calculate the relative density of the point to obtain the relative density.

S15. Select points whose relative density is greater than the density average corresponding to the relative density to obtain multiple cluster centers.

S16. According to the cluster center, use the preset domain calculation formula to divide the feature vector set to obtain multiple sets.

S17. Set corresponding labels for the points corresponding to each set according to the labels corresponding to the cluster centers to obtain the event pseudo-label set.

In the embodiment of the present invention, in order to generate more accurate event pseudo-labels to improve the accuracy of event extraction, the density peak clustering algorithm needs to be improved. Density peak clustering algorithm is a simple and effective cluster analysis method. However, there are the following limitations: During the execution of the density peak clustering algorithm, the user needs to manually select the cluster center, and this operation has a certain degree of subjectivity. The allocation strategy of the density peak clustering algorithm is to assign a point to the cluster where the density is greater than it and the point closest to it is located. Such an allocation strategy often allocates most points in a low-density cluster to a high-density cluster, which is a domino effect. In view of the above limitations, the present invention proposes a method for automatically determining cluster centers, as well as a new point allocation strategy and a preset density peak clustering algorithm.

In order to automatically determine the cluster center, the density of points and the relative density of points need to be calculated. Regarding the calculation method of density, for any point x _i , the smaller the average distance from it to its K-Nearest Neighbors (KNN), the greater its density. Therefore, the density calculation formula of point x _i , that is, the preset density calculation formula, is:

Among them, ρ(xi) represents the density of point x _i ; k represents the number of neighbors of a point, that is, using _the Euclidean distance as the metric, the top k points closest to a point; Represents the Euclidean distance between point x _i and point x _j , u represents the dimension of a feature vector concentration point, that is, the number of features, g is a subscript, used to index different features; KNN(xi ₎ represents the A set of the first k points with the smallest Euclidean distance from point x _i ; i represents the i-th point x _i in the data set; j represents the j-th point x _j in the data set. Assume there are v points in the data set, Then i=1,2,...,v; j=1,2,...,v. The data sets here refer to the aircraft group data set and the ground control station data set.

The relative density calculation method of points, that is, the preset relative density calculation formula is as follows:

Among them, r _ρ ( _xi ) represents the relative density of point x _i ; ρ ( _xi ) represents the density of point _xi ; ρ (x _j ) represents the density of point x _j ; k represents the number of neighbors of a point, that is Using Euclidean distance as the metric, the top k points closest to a point; KNN(xi ₎ represents the set of the top k points with the smallest Euclidean distance from point x _i in the data set; i represents the i-th point in the data set point x _i ; j represents the j-th point x _j in the data set.

The steps of the method of automatically determining the cluster center: normalize the relative density r _ρ of the points and the high-density nearest neighbor distance δ respectively, and still use r _ρ and δ to represent the normalized results. Let v ₁ =r _ρ δ, and use v ₂ to represent the result of sorting v ₁ from large to small. For the first point in v ₂ , if its relative density is greater than the average density of the relative densities of all points, and its high-density nearest neighbor distance is greater than the average high-density nearest neighbor distance of all points, then this point is the cluster center . Continue to examine whether the next point in v ₂ is the cluster center until the number of cluster centers reaches the set value. The number of cluster centers is a parameter and needs to be set in advance.

The new point allocation strategy is based on the domain of points. The calculation formula of the domain of points, that is, the default domain calculation formula is:

D _mn (x _p )={x _q |x _q ∈MN(x _p )∨(x _m ∈MN(x _p ))∧x _q ∈MN(x _m ))};

In the formula, D _mn (x _p ) represents the domain of point x _p ; MN (x _p ) represents the set of all neighbors of point x _p ; if x _i ∈KNN(x _j ) and x _j ∈KNN( _xi ), Then x _j is a mutual neighbor of x _i , denoted as x _j ∈MN(xi ₎ . Here, x _i and x _j both represent any points in the data set. The meanings of MN(x _m ) and MN(x _p ) are similar to the meanings of MN( _xi ). x _p represents the p-th point in the data set. Assume there are v points in the data set, then p=1,2,...,v. x _m represents any point in MN(x _p ). x _q represents any point in D _mn (x _p ). D _mn (x _p ) represents the domain of x _p , and its essence is a set of points that meet some conditions. These points are either mutual neighbors of x _p , or mutual neighbors of points in the mutual neighbors of x _p . That is to say, if x _q ∈MN(x _p ), then x _q is in the set MN(x _p ); if But x _q ∈MN(x _m ) and x _m ∈MN(x _p ), then x _q is in the set D _mn (x _p ).

New point allocation strategy: for non-cluster center points, the clustering process is performed within the domain. Starting from the point x _i with the highest density, if the high-density nearest neighbor point x _j of x _i is in D _mn ( _xi ), then x _i is assigned to the cluster where x _j is located. If the high-density nearest neighbor point x _j of x _i is not in D _mn (x _i ), use x _k to represent the point in D _mn (x _i ) that is closest to x _i and has been assigned, and assign x _j to where x _k is. cluster, thereby obtaining the set corresponding to the center of the cluster.

The ground control station plan, ground control station data, aircraft group plan and aircraft group data are called relevant data. The specific process of obtaining event pseudo-labels is as follows: the feature vector of relevant data can be extracted based on the Skip-gram model, which is a preset word skip model. These feature vectors are input into the improved density peak clustering algorithm, that is, the preset density peak clustering algorithm. The improved density peak clustering algorithm calculates the similarity between these feature vectors based on Euclidean distance, and then starts the clustering process. The improved density peak clustering algorithm outputs several cluster centers, and for any cluster center, the improved density peak clustering algorithm outputs all points in the same cluster. Here a point represents an event. The number of cluster centers is the number of event types. For any cluster center, set a label for it. For points in the same cluster as this cluster center, their labels are the same as the label of the cluster center. Assume that there are currently three event types A, B, and C. Then the improved density peak clustering algorithm will output 3 cluster centers, as well as which points are in the same cluster as which cluster center. Set the label of the cluster center corresponding to type A events to "A", set the label of the cluster center corresponding to type B events to "B", and set the label of the cluster center corresponding to type C events to "C" . "A", "B", and "C" here are event tags, and the event tags here are also called event pseudo-tags. In supervised learning algorithms, most of the labels are artificially labeled, which are relatively accurate and real labels, so they are called event pseudo-labels. The improved density peak clustering algorithm in this patent is an unsupervised learning algorithm. In the unsupervised learning algorithm, there are no human-labeled labels, only labels generated by the algorithm. In most cases, these labels are not as accurate as human-labeled ones, and there will be wrong labels, such as labeling type A events as type B events. Therefore, the labels generated by unsupervised learning algorithms are called event pseudo-labels.

Step 202: Train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model.

Further, step 202 may include the following sub-steps S21-S26:

S21. Input the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

S22. Fine-tune the initial event extraction model according to the initial loss function value to obtain the intermediate event extraction model.

S23. Enter the event pseudo-label set into the intermediate event extraction model to obtain the target loss function value.

S24. Determine whether the target loss function value is the preset function value. If so, execute step S25. If not, execute step S26.

S25. Use the intermediate event extraction model corresponding to the target loss function value as the target event extraction model.

S26. Use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

In the embodiment of the present invention, as shown in Figure 3, the initial event extraction model is a CNN-based event extraction model, and the event extraction modeling is set as a multi-classification task. Set the sentence length to sl, truncate when the sentence length is greater than sl, and pad when the sentence length is less than sl. The jieba toolkit (i.e. stuttering toolkit) is used to segment the Chinese sentences in the data set into tokens (i.e. a basic unit of text). Use the Skip-gram model (a word embedding model) to obtain the word embedding of the token, assuming the dimension is d. Assume that the number of event trigger word types is n ₁ and the number of event parameter role types is n ₂ . Next, the features of semantic units of a specific length are extracted through the convolutional layer, and the sentence features are extracted through the multi-layer perceptron. When applying a convolutional neural network to a sentence, the size of the filter is set to capture the characteristics of a certain length of semantic units. If the filter size is 3×d, the features of semantic units of length 3 are captured. This patent uses filters with sizes of 2×d, 3×d, 4×d, and 5×d to extract features of corresponding length semantic units. The number of filters of different sizes is 128. Therefore, the sentence is encoded through the convolutional layer into a vector of length 128*4=512. Concatenate the word embeddings of all tokens in a sentence to obtain a vector of length sl*d. Set the number of neurons in the first layer of the multi-layer perceptron to sl*d, and the number of neurons in subsequent layers is halved in turn. The number of neurons in the output layer is greater than or equal to 2.5 (n ₁ + n ₂ ). Assume that the number of neurons in the output layer is n. ₀ . Connect the semantic unit features obtained from the convolutional layer and the sentence features obtained from the multi-layer perceptron to obtain a vector with a length of 512+n ₀ , which is input to the convolutional layer to obtain the classification result. Based on the classification results, the loss function is calculated and the back propagation error is obtained to determine the model parameters.

As shown in Figure 4, the improved density peak clustering algorithm is applied to the relevant data to obtain event pseudo-labels and event parameter role pseudo-labels. Input the pseudo-label into the initial event extraction model to obtain the loss function value, and then back-propagate the error to fine-tune the initial event extraction model to obtain the intermediate event extraction model. The event pseudo-label set is input into the CNN-based event extraction model, that is, the intermediate event extraction model, to obtain the target loss function value. The preset function value means that the target loss function value no longer decreases. When the target loss function value no longer decreases and the model converges, the fine-tuning process is stopped, and the intermediate event extraction model corresponding to the target loss function value is used as the target event extraction model. Otherwise, use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

Step 203: Extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model and perform event comparison to obtain solution comparison results.

Further, step 203 may include the following sub-steps S31-S33:

S31. Extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model to obtain the aircraft group events and ground events.

S32. Compare the aircraft group event and the ground event according to the trigger word category and parameter similarity, and obtain the event comparison result.

S33. Use all event comparison results to construct plan comparison results.

In the embodiment of the present invention, there are two purposes for extracting events from relevant data: one is to determine whether the aircraft group scheme is consistent with the ground control station scheme based on the event extraction results; the other is to input the event extraction results into the improved multi-agent depth Deterministic policy gradient algorithm to obtain optimal instruction sequences. The fine-tuned CNN-based event extraction model, that is, the target event extraction model, is used to extract events from relevant data. Both the ground control station plan and the aircraft group plan consist of a series of instructions. Instructions correspond to events one-to-one. Let e _i represent an event in the ground control station plan, and let e _j represent an event in the aircraft group plan. If the trigger words of e _i and e _j belong to the same category, and e _i and e _j have more than 50% of the parameters. are the same, then e _i and e _j belong to the same event. When more than 80% of the events in the ground control station plan and the aircraft group plan are the same, the ground control station plan and the aircraft group plan are considered to be consistent. When the two plans are consistent, the aircraft group plan is executed and the algorithm ends, that is, step 210 is executed. Otherwise, step 204 is executed.

Step 204: When the plan comparison results are inconsistent, determine whether the authorization data corresponding to the aircraft group is authorized. If so, perform step 205. If not, perform step 206.

In the embodiment of the present invention, in order to make the collaborative decision-making algorithm practical, an authorization module is added to determine whether to authorize the aircraft group to act on its own. When authorized, perform step 205; when not authorized, perform step 206.

Step 205: Determine the aircraft group collaborative decision-making plan corresponding to the emergency situation data based on the command sequence and the aircraft group plan within the preset time.

Further, step 205 may include the following sub-steps S41-S43:

S41. Determine whether the aircraft group and the ground control station have found the optimal command sequence within the preset time. If yes, step S42 is executed. If not, step S43 is executed.

S42. Use the optimal command sequence as a collaborative decision-making scheme for the aircraft group corresponding to the emergency situation data.

S43. Use the aircraft group plan as the aircraft group collaborative decision-making plan corresponding to the emergency data.

In the embodiment of the present invention, the preset time refers to a time interval set based on actual needs. Within the preset time, if the aircraft group or ground control station has output the optimal command sequence, this command sequence will be executed. If the optimal command sequence is not output, the aircraft group plan will be regarded as the optimal command sequence and executed.

Step 206: Select the command sequence from the aircraft group data set and the ground control station data set according to the preset improved multi-agent deep deterministic policy gradient algorithm to obtain the target command sequence.

Further, step 206 may include the following sub-steps S51-S56:

S51. Use the decision criteria corresponding to the aircraft group data set and the ground control station data set to construct an initial decision criterion set.

S52. Select the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion.

S53. According to the priority decision-making criterion, use the instruction acquisition algorithm to assign preset weights to the instructions corresponding to the aircraft group data set and the ground control station data set respectively, and obtain the instruction sequence set.

S54. Use the preset improved multi-agent deep deterministic policy gradient algorithm to select multiple initial instruction sequences that meet the preset score threshold in the instruction sequence set, and construct an initial instruction queue.

S55. Select the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence.

S56. Determine the target command sequence based on the aircraft group's own constraints and the intermediate command sequence.

Further, step S56 may include the following sub-steps S561-S569:

S561. Determine whether the aircraft group can execute the intermediate command sequence under its own constraints. If so, step S562 is executed. If not, step S563 is executed.

S562. Use the intermediate instruction sequence as the target instruction sequence.

S563. Delete the intermediate instruction sequence from the initial instruction queue to obtain an intermediate instruction queue.

S564. Determine whether the intermediate instruction queue is an empty set. If so, execute step S565. If not, execute step S569.

S565. Delete the priority decision criteria from the initial decision criteria set to obtain the target decision criteria set.

S566. Determine whether the target decision criterion set is an empty set. If so, execute step S567. If not, execute step S568.

S567. Use the preset sequence as the target instruction sequence.

S568: Use the target decision criterion set as the initial decision criterion set, and jump to the step of selecting the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion.

S569: Use the intermediate instruction queue as the initial instruction queue, and jump to the step of selecting the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence.

In the embodiment of the present invention, the aircraft group and the ground control station respectively determine the instruction weight according to different decision-making criteria, and use the improved multi-agent deep deterministic strategy gradient algorithm, that is, the preset improved multi-agent deep deterministic strategy gradient algorithm to determine several candidate command sequence, and select the target command sequence based on the constraints of the aircraft group itself. Step 206 involves three aspects: the principle of the instruction acquisition algorithm; the improved multi-agent deep deterministic policy gradient algorithm; and obtaining the candidate instruction sequence based on the improved multi-agent deep deterministic policy gradient algorithm.

(1) Principle of instruction acquisition algorithm. In order to complete collaborative decision-making tasks more quickly and allocate resources more rationally, the decision-making criteria are sorted from high to low priority. On the aircraft group, high-priority decision criteria are input into the instruction acquisition algorithm to find the optimal instruction sequence. On the ground control station, low-priority decision criteria are input into the command acquisition algorithm to find the optimal command sequence. When the aircraft group can find the optimal instruction sequence based on the decision criteria with high priority, it executes this sequence; otherwise, it communicates with the ground control station to obtain the optimal instruction sequence it finds. Use the above-mentioned optimal command sequence to guide the next phase of actions of the aircraft group.

Instruction acquisition algorithm process. The decision criterion with the highest priority is taken from the set of decision criteria, and the instruction is weighted according to the decision criterion. Instructions related to the ground control station plan and the aircraft group plan are given a higher weight, and instructions related to the ground control station data and the aircraft group data are given a lower weight. That is, if the agent completes the instructions related to the ground control station plan and the aircraft group plan, the agent will be given high rewards. If the agent completes the instructions related to the ground control station data and aircraft group data, the agent will be given a low reward, which is a positive value. The improved multi-agent deep deterministic policy gradient algorithm is used to select several candidate instruction sequences with the highest scores from the above instructions and store them in the queue. Take out the command sequence with the highest score in the queue. If the remaining energy of the aircraft group can support the completion of this command sequence, then it is the optimal command sequence; otherwise, delete this command sequence from the queue and continue to examine other command sequences in the queue. . If the remaining energy of the aircraft group cannot support it to complete any instruction sequence in the queue, the decision-making criteria will be changed until the optimal instruction sequence is found. Please refer to Figure 5 for the above process.

The reason for assigning higher weights to instructions related to ground control station plans and aircraft group plans, and assigning lower weights to instructions related to ground control station data and aircraft group data: Ground control stations are formulated after careful consideration Plan, the aircraft group plan is a plan formulated after considering emergencies. Therefore, the instructions involved in the ground control station scheme and the aircraft group scheme have a high degree of trust. On the other hand, the information reflected by the ground control station data is mostly global information of the flight environment, and the information reflected by the aircraft group data is mostly local information of the flight environment. The combination of global information and local information can more accurately reflect the flight environment of the aircraft group. Therefore, not only the instructions related to the ground control station plan and the aircraft group plan are referable, but the instructions related to the ground control station data and the aircraft group data are also referable to a certain extent.

(2) The improved multi-agent deep deterministic policy gradient algorithm is the preset improved multi-agent deep deterministic policy gradient algorithm. The decision-making scenario of the present invention is one's own aircraft group and the equipment group at the target. Each of your own aircraft groups has the same functions. In order to improve the generalization ability of the collaborative decision-making method, the present invention sets the scenario as a game process between two equipment groups. Similar to a football match, there is competition between teams, cooperation within the team, and different roles within the team playing different roles. Different equipment corresponds to different roles within the team. The multi-agent deep deterministic policy gradient algorithm is a multi-agent deep reinforcement learning algorithm that solves inter-team confrontation and intra-team cooperation. However, in the multi-agent deep deterministic policy gradient algorithm, each agent has a global evaluator, and the input space of the Critic network (i.e., value network) increases as the number of agents increases, resulting in a long learning cycle for the agents. Only suitable for scenarios with a small number of agents. In response to the above problems, the present invention first improves the multi-agent deep deterministic policy gradient algorithm, and then uses the improved multi-agent deep deterministic policy gradient algorithm to find several candidate instruction sequences based on a certain decision criterion.

Improved multi-agent deep deterministic policy gradient algorithm. The central controller refers to the N value networks, and the strategy area refers to the N strategy networks equipped for N agents. In order to make the logical relationship between multiple agents clearer, a fusion layer is added between the central controller and the strategy. The fusion layer consists of two pseudo-agents, and the central controller only needs to be responsible for two pseudo-agents. The first pseudo-agent integrates the actions, observations and rewards of all its own agents into a set of actions, observations and rewards. The second pseudo-agent integrates the actions, observations and rewards of all agents in the target into a set of actions, observations and rewards. The above two sets of actions, observations and rewards are then uploaded to the central controller. The central controller is only oriented to both parties in the game and only needs to consider the competitive relationship between the two pseudo-agents. This solves to a certain extent the problem that the multi-agent deep deterministic policy gradient algorithm is difficult to handle a large number of agents. A single pseudo-agent is oriented to all real agents on one side of the game, and it only needs to consider the cooperative relationship. Overall, the multi-agent deep deterministic policy gradient algorithm after inserting the fusion layer not only alleviates the current situation that the algorithm complexity increases with the increase of the input space, but can also handle complex multi-agent relationships where cooperation and competition coexist. In this multi-agent game, what we ultimately focus on is whether our own side can win, not which agent wins the most rewards. Therefore, the entire multi-agent game scene is divided into two views. One view faces oneself, which is the first pseudo-agent. Another view faces the equipment group at the target. The observations captured by all our own agents are not completely consistent, so these observations are summed and input to the first pseudo-agent. The same goes for the integration of actions and rewards for all agents on one's side. The central controller evaluates the two pseudo-agents. The two pseudo-agents respectively help the real agents they represent to improve their own strategies. For the multiple real agents corresponding to one pseudo-agent, they have the same goal, that is, winning, and they have a complete cooperative relationship. They only need to cooperate to complete the pseudo-agent. The agent obtains improvement suggestions from the central controller.

(3) Obtain candidate instruction sequences based on the improved multi-agent deep deterministic policy gradient algorithm. The present invention sets the decision-making scenario as a confrontation scenario between equipment groups of both parties. The equipment is temporarily set to have three functions: reconnaissance, penetration, and attack, and functions can continue to be added. Let A={a ₁ ,a ₂ ,...,a _i ,...,a _n1 }, B={b ₁ ,b ₂ ,...,b _i ,...,b _n2 }, C ={c ₁ , c ₂ ,..., c _i ,..., c _n3 } respectively represent a collection of one's own equipment with reconnaissance, penetration, or strike functions. Let a _i , b _i , and c _i respectively represent the i-th type of equipment with reconnaissance, penetration, or strike functions. n ₁ , n ₂ , and n ₃ respectively represent the number of types of equipment with reconnaissance, penetration, and strike functions. There can be intersections between A, B, and C, because a piece of equipment can have one or more functions. _{Let _ _ _ _ _ _ _} ={z ₁ , z ₂ ,...,z _i ,...,z _n6 } respectively represent a collection of equipment with reconnaissance, penetration, or strike functions in the target. Let x _i , y _i , and z _i respectively represent the i-th type of equipment with reconnaissance, penetration, or strike functions. n ₄ , n ₅ , and n ₆ respectively represent the number of types of equipment with reconnaissance, penetration, and strike functions. There can be intersections between X, Y, and Z, because a piece of equipment can have one or more functions. For any set among A, B, C, X, Y, and Z, there are no duplicate elements. Let a _ij represent the number of our own i-th equipment with reconnaissance function, and b _ij and c _ij have similar meanings. Let x _ij represent the number of the i-th equipment with reconnaissance function in the target, and y _ij and z _ij have similar meanings. Let c' _i represent the number of times that one's own i-th equipment with strike function can strike, and z' _i has a similar meaning. For various types of equipment, there is no limit to the number of times their reconnaissance and penetration functions can be implemented. However, after being defeated by other intelligent entities, their service will be terminated and they can no longer perform reconnaissance, strike or penetration functions.

Set up the action space. For any intelligent agent, the actions it can take come from the set A = {strike, reconnaissance, penetration, stationary, motion}, and it is not required to have the ability to perform all actions in A.

Set up the state space. Set the state space S = {type of detected target, number of detected targets, type of own equipment, number of own equipment, and actions being performed by each equipment}.

Set reward function. When the _i- th equipment of one's side detects the target, the reward of the agent corresponding to _{the i} -th equipment of one's side is increased by r ₁ , and the reward of the agent corresponding to the detected target is reduced by r ₁ . When the b _i- th equipment of one's side successfully penetrates the defense, the reward of the agent corresponding to the b _i- th equipment of one's side is increased by r _2. The reward of the agent corresponding to the target b _i that fails to prevent the penetration of the defense is reduced by r ₂ . When one's c _i- th equipment hits the target, the reward of the agent corresponding to one's c _i-th equipment is increased by r _3. The reward of the agent corresponding to the hit target is reduced by r _3. The next target that is hit is terminated. service. As of a certain time, if the sum of rewards obtained by one's own equipment is greater than the sum of rewards obtained by the equipment at the target, one's side wins, otherwise one's side does not win.

The process of the game between the two agents is a process in which different agents perform a series of actions in a time period and a space. There are currently two decision-making plans, one is the aircraft group plan and the other is the ground control station plan. For the game process of both agents, these two decision-making solutions are two pieces of prior information. Input the two prior information into the improved multi-agent deep deterministic policy gradient algorithm respectively, and observe the victory of your own side. Then the two prior information are simultaneously input into the improved multi-agent deep deterministic policy gradient algorithm. When the two prior information conflicts, cut off the conflicting part and observe my victory. Finally, choose the first few instruction sequences that your side wins and gets the most rewards.

Step 207: Determine whether the target instruction sequence is a preset sequence. If yes, execute step 208. If not, execute step 209.

In the embodiment of the present invention, the preset sequence means that the target instruction sequence is not an optimal instruction sequence, and the preset instruction sequence may also be an empty set. Determine whether the target command sequence is the optimal command sequence by selecting command sequences from the aircraft group data set and the ground control station data set according to the preset improved multi-agent deep deterministic policy gradient algorithm.

Step 208: Use the aircraft group plan as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

In the embodiment of the present invention, in extreme cases, neither the aircraft group nor the ground control station can find the optimal instruction sequence, and the aircraft group solution is used as the optimal instruction sequence. A series of events can be extracted from the aircraft group plan, and these events correspond to instructions one-to-one. Therefore, the aircraft group plan is essentially an instruction sequence.

Step 209: Use the target command sequence as the aircraft group collaborative decision-making solution corresponding to the emergency situation data.

In the embodiment of the present invention, because the aircraft group searches for the optimal instruction sequence based on a high-priority decision criterion, when the aircraft group can find the optimal instruction sequence, this instruction sequence is executed. When the aircraft group cannot find the optimal command sequence, the optimal command sequence found by the ground control station will be executed.

Step 210: When the plan comparison results are consistent, use the aircraft group plan in the aircraft group data set as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

In the embodiment of the present invention, the specific implementation process of step 210 is similar to step 105, and will not be described again here.

In the embodiment of the present invention, as shown in Figure 6, the ground control station plan, ground control station data, aircraft group plan, and aircraft group data are input, and the optimal instruction sequence is output. Specifically, the density peak clustering algorithm is improved, event pseudo-labels are generated, the existing dynamic multi-pooling convolutional neural network model is fine-tuned, events are extracted from relevant data, and events are mapped into instructions. Based on the event extraction results, it is judged whether the ground control station plan and the aircraft group plan are consistent. If the two plans are inconsistent, determine whether the aircraft group is authorized to act on its own. If the unauthorized aircraft group can act on its own, the aircraft group and the ground control station will respectively call the instruction acquisition algorithm to obtain the optimal instruction sequence. If the aircraft group finds the optimal command sequence I ₁ based on the high-priority decision criterion, I ₁ will be used to guide the action of the aircraft group. If the aircraft group does not find the optimal command sequence, but the ground control station finds the optimal command sequence I ₂ , it will use I ₂ to guide the actions of the aircraft group. If the ground control station does not find the optimal command sequence, the aircraft group plan will be used as the optimal command sequence. If the aircraft group has been authorized to act on its own, if I ₁ or I ₂ is found within the set time, I ₁ or I ₂ will be executed; otherwise, the aircraft group plan will be used as the optimal command sequence. When the aircraft group plan is consistent with the ground control station plan, the aircraft group plan is used as the optimal command sequence.

In view of the fact that when an aircraft group encounters an emergency situation, the actual scenario is characterized by complexity, uncertainty, and few data samples. The types of emergency situations, data processing, collaborative decision-making methods, and the allocation and utilization of computing resources were studied. Detailed research. The preset density peak clustering algorithm is used to generate event pseudo-labels, which solves the problem of no event labels. Based on the event pseudo-labels, the CNN-based event extraction model is fine-tuned to extract events and quickly grasp the overall situation of the flight environment. The ground control station determines whether the aircraft group can act on its own, and regardless of whether it can act on its own, corresponding plans are formulated to deal with emergencies. The aircraft group processes high-priority decision criteria and the ground control station processes low-priority decision criteria, making reasonable use of the computing resources of the aircraft group and the ground control station. The instruction sequence is closely related to the flight path and mission execution of the aircraft group. An improved multi-agent deep deterministic strategy gradient algorithm is used to find the instruction sequence, and do its best to enable the aircraft group to avoid threats, complete tasks, and successfully respond to emergencies. Responding to emergencies faced by the aircraft group according to the proposed collaborative decision-making method can improve the accuracy and efficiency of grasping the flight scenario, and can provide a scientific decision-making basis for the aircraft's actions in the next stage.

As shown in Figure 7, the embodiment of the present invention also includes an emergency system applying the aircraft group collaborative decision generation method provided by the embodiment of the present invention. The emergency system includes a decision processing module, a data processing module, a flight control module, a sensor module, a database module and Communication module. The hardware facilities of the emergency system include three major components: aircraft group, wireless communication link system, and ground control station. The aircraft groups generally carry payloads, which are equipment used to complete specific tasks. A wireless communications link system used to transfer data between a fleet of aircraft and a ground control station. The ground control station controls the aircraft group and distributes and coordinates tasks.

The decision processor module is used to execute the improved density peak clustering algorithm. The data processing module provides computing resources to the decision processor, which is used to analyze the flight data of the aircraft group to detect flight quality or generate decision solutions. The flight control module is used in the flight management and control system. It is responsible for starting the collaborative decision-making algorithm, executing the calculation results of the decision processor module, and transmitting the results to the ground control station through the communication module. The sensor module is used to collect data such as speed, altitude, attitude, acceleration and angular rate during the flight of the aircraft group. The database module is used to store data collected by the sensor module, execution results of the decision processor module, command commands and capture data from the ground control station, store various data transmitted by the aircraft group, and data required to maintain the normal flight of the aircraft group.

The communication module is used to transmit various data of the aircraft group to the ground control station, and is used to receive various data transmitted by the ground control station to the aircraft group and to collect various data transmitted by the aircraft group to the ground control station. The control module is used to convey control instructions, remotely control the aircraft group, interact with the data processing module, analyze flight data, and generate decision-making instructions.

Furthermore, the emergency system also includes a human-computer interaction interface module, which is used to generate a visual interface for the operator to convey instructions or analyze flight data through the ground control station. The communication module is a wireless communication link and is also used to transfer data between the aircraft group and the ground control station. The sensor module stores the data into the database module. When the flight control module detects data anomalies, it indicates an emergency. The flight control module starts the decision processor module, and the decision processor module executes the collaborative decision-making algorithm. When the computing resources of the aircraft group are insufficient, the aircraft group communicates with the ground control station through the wireless communication link system, and jointly completes the collaborative decision-making algorithm with the help of the ground control station. The collaborative decision-making algorithm outputs the optimal instruction sequence and uses the optimal instruction sequence to guide the next phase of the aircraft group's actions.

Please refer to FIG. 8 , which is a structural block diagram of an aircraft group collaborative decision-making generation system provided in Embodiment 3 of the present invention.

Example 3 of the present invention provides an aircraft group collaborative decision-making system, including:

The event pseudo-label set obtaining module 801 is used to, when receiving emergency data of an aircraft group, use a preset density peak clustering algorithm to calculate the aircraft group data set corresponding to the emergency data and the ground control station generated by the ground control station. The event pseudo-label corresponding to the data set is used to obtain the event pseudo-label set.

The target event extraction model obtaining module 802 is used to train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model.

The solution comparison result obtaining module 803 is used to extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model and perform event comparison to obtain the solution comparison result.

The first acquisition module 804 of collaborative decision-making of the aircraft group is used to select the optimal instruction sequence based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm when the plan comparison results are inconsistent, to obtain Collaborative decision-making scheme for aircraft groups corresponding to emergency situation data.

The second acquisition module 805 of aircraft group collaborative decision-making is used to use the aircraft group plan in the aircraft group data set as the aircraft group collaborative decision-making plan corresponding to the emergency situation data when the plan comparison results are consistent.

Optionally, the event pseudo-label set obtaining module 801 includes:

The aircraft group data set and the ground control station data set acquisition module is used to obtain the aircraft group data set generated by the aircraft group based on the emergency data and the ground control station generated by the ground control station when receiving the emergency situation data of the aircraft group. data set.

The feature vector set obtaining module is used to extract the feature vectors of the aircraft group data set and the ground control station data set through the preset word skip model to obtain the feature vector set.

The density obtaining module is used to calculate the density corresponding to each point in the feature vector set using the preset density calculation formula to obtain the density.

The default density calculation formula is:

Among them, ρ(xi) represents the density of point x _i ; k represents the number of neighbors of a point, that is, using _the Euclidean distance as the metric, the top k points closest to a point; Represents the Euclidean distance between point x _i and point x _j , u represents the dimension of a feature vector concentration point, that is, the number of features, g is a subscript, used to index different features; KNN(xi ₎ represents the A set of the first k points with the smallest Euclidean distance from point x _i ; i represents the i-th point x _i in the data set; j represents the j-th point x _j in the data set.

The relative density obtaining module is used to substitute the density into the preset relative density calculation formula to calculate the relative density of the point and obtain the relative density.

The default relative density calculation formula is:

Among them, r _ρ ( _xi ) represents the relative density of point x _i ; ρ ( _xi ) represents the density of point _xi ; ρ (x _j ) represents the density of point x _j ; k represents the number of neighbors of a point, that is Using Euclidean distance as the metric, the top k points closest to a point; KNN(xi ₎ represents the set of the top k points with the smallest Euclidean distance from point x _i in the data set; i represents the i-th point in the data set point x _i ; j represents the j-th point x _j in the data set.

The cluster center obtaining module is used to select points whose relative density is greater than the density average corresponding to the relative density and obtain multiple cluster centers.

The set obtaining module is used to divide the feature vector set according to the cluster center and using the preset domain calculation formula to obtain multiple sets.

The default domain calculation formula is:

D _mn (x _p )={x _q |x _q ∈MN(x _p )∨(x _m ∈MN(x _p ))∧x _q ∈MN(x _m ))};

In the formula, D _mn (x _p ) represents the domain of point x _p ; MN (x _p ) represents the set of all neighbors of point x _p ; MN (x _m ) represents the set of all neighbors of point x _m ; x _p represents the feature The p-th point in the vector set; x _m represents the m-th point in the feature vector set; x _q represents the q-th point in the feature vector set.

The event pseudo-label set obtaining sub-module is used to set corresponding labels for the points corresponding to each set according to the labels corresponding to the cluster centers to obtain the event pseudo-label set.

Optionally, the target event extraction model obtaining module 802 includes:

The initial loss function value obtaining module is used to input the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

The intermediate event extraction model is obtained by a module for fine-tuning the initial event extraction model according to the initial loss function value to obtain an intermediate event extraction model.

The target loss function value obtaining module is used to input the event pseudo-label set into the intermediate event extraction model to obtain the target loss function value.

The target loss function value judgment module is used to judge whether the target loss function value is the preset function value.

The target event extraction model obtains a sub-module, which is used to use the intermediate event extraction model corresponding to the target loss function value as the target event extraction model.

The jump execution module is used to, if not, use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain the initial loss function value.

Optionally, the solution comparison result obtaining module 803 includes:

The aircraft group event and ground event obtaining module is used to extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model to obtain the aircraft group event and ground event.

The event comparison result obtaining module is used to compare aircraft group events and ground events according to trigger word categories and parameter similarities, and obtain event comparison results.

The plan comparison result acquisition sub-module is used to use all event comparison results to construct the plan comparison result.

Optionally, the aircraft group collaborative decision-making first obtaining module 804 includes:

The authorization data judgment module is used to judge whether the authorization data corresponding to the aircraft group is authorized when the plan comparison results are inconsistent.

The aircraft group collaborative decision-making obtains the first sub-module, which is used to determine the aircraft group collaborative decision-making plan corresponding to the emergency situation data based on the command sequence and the aircraft group plan within the preset time.

The target command sequence obtaining module is used to select the command sequence from the aircraft group data set and the ground control station data set according to the preset improved multi-agent depth deterministic strategy gradient algorithm to obtain the target command sequence.

The target instruction sequence judgment module is used to judge whether the target instruction sequence is a preset sequence.

The aircraft group collaborative decision-making obtains a second sub-module, which is used to, if so, use the aircraft group plan as the aircraft group collaborative decision-making plan corresponding to the emergency situation data.

The aircraft group collaborative decision-making obtains a third sub-module, which is used to use the target command sequence as the aircraft group collaborative decision-making solution corresponding to the emergency situation data if not.

Optionally, the first sub-module obtained through aircraft group collaborative decision-making can perform the following steps:

Determine whether the aircraft group and ground control station have found the optimal command sequence within the preset time;

If so, the optimal command sequence will be used as the aircraft group collaborative decision-making solution corresponding to the emergency situation data;

If not, the aircraft group plan is used as the aircraft group collaborative decision-making plan corresponding to the emergency data.

Optionally, the target instruction sequence obtaining module can perform the following steps:

Use the decision criteria corresponding to the aircraft group data set and the ground control station data set to construct an initial set of decision criteria;

Select the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion;

According to the priority decision-making criterion, the instruction acquisition algorithm is used to assign preset weights to the instructions corresponding to the aircraft group data set and the ground control station data set, respectively, to obtain the instruction sequence set;

The preset improved multi-agent deep deterministic policy gradient algorithm is used to select multiple initial instruction sequences that meet the preset score threshold in the instruction sequence set and build the initial instruction queue;

Select the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence;

The target command sequence is determined based on the aircraft group's own constraints and intermediate command sequences.

Optionally, the target instruction sequence obtaining module can also perform the following steps:

Determine whether the aircraft group can execute the intermediate command sequence under its own constraints;

If so, use the intermediate instruction sequence as the target instruction sequence;

If not, delete the intermediate instruction sequence from the initial instruction queue to obtain the intermediate instruction queue;

Determine whether the intermediate instruction queue is an empty set;

If so, delete the priority decision criteria from the initial decision criteria set to obtain the target decision criteria set;

Determine whether the target decision criteria set is an empty set;

If so, use the preset sequence as the target instruction sequence;

If not, use the target decision criterion set as the initial decision criterion set, and jump to the step of selecting the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion;

If not, use the intermediate instruction queue as the initial instruction queue, and jump to the step of selecting the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence.

Embodiments of the present invention also provide an electronic device. The electronic device includes: a memory and a processor. A computer program is stored in the memory; when the computer program is executed by the processor, it causes the processor to execute an aircraft group-based method as in any of the above embodiments. Collaborative decision making methods.

The memory may be electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM, hard disk or ROM. The memory has storage space for program code for performing any of the method steps described above. For example, the storage space for the program code may include individual program codes respectively used to implement various steps in the above method. These program codes can be read from or written into one or more computer program products. These computer program products include program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks. The program code may, for example, be compressed in a suitable form. These codes, when run by the computing processing device, cause the computing processing device to perform various steps in the aircraft group-based collaborative decision generation method described above.

Those skilled in the art can clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be described again here.

In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of units is only a logical function division. In actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated. to another system, or some features can be ignored, or not implemented. On the other hand, the coupling or direct coupling or communication connection between each other shown or discussed may be through some interfaces, and the indirect coupling or communication connection of the devices or units may be in electrical, mechanical or other forms.

A unit described as a separate component may or may not be physically separate. A component shown as a unit may or may not be a physical unit, that is, it may be located in one place, or it may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated units can be implemented in the form of hardware or software functional units.

Integrated units may be stored in a computer-readable storage medium if they are implemented in the form of software functional units and sold or used as independent products. Based on this understanding, the technical solution of the present invention is essentially or contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods of various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

The above embodiments are only used to illustrate the technical solutions of the present invention, but not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the foregoing embodiments. The recorded technical solutions may be modified, or some of the technical features thereof may be equivalently replaced; however, these modifications or substitutions shall not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims (10)

1. A method for generating collaborative decision-making for aircraft groups, which is characterized by including: When receiving the emergency situation data of the aircraft group, a preset density peak clustering algorithm is used to calculate the event pseudo-label corresponding to the aircraft group data set corresponding to the emergency situation data and the ground control station data set generated by the ground control station, Get the event pseudo-label set; Train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model; The target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively and perform event comparison to obtain solution comparison results; When the solution comparison results are inconsistent, the optimal instruction sequence is selected based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm to obtain the aircraft group corresponding to the emergency situation data. Collaborative decision-making solutions; When the plan comparison results are consistent, the aircraft group plan in the aircraft group data set is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data. 2. The aircraft group collaborative decision-making generation method according to claim 1, characterized in that when receiving emergency situation data of the aircraft group, a preset density peak clustering algorithm is used to calculate the emergency situation data corresponding to The event pseudo-labels corresponding to the aircraft group data set and the ground control station data set generated by the ground control station, and the steps to obtain the event pseudo-label set include: When receiving the emergency situation data of the aircraft group, obtain the aircraft group data set generated by the aircraft group based on the emergency situation data and the ground control station data set generated by the ground control station; Extract the feature vectors of the aircraft group data set and the ground control station data set through a preset word skip model to obtain a feature vector set; Calculate the density corresponding to each point in the feature vector set using a preset density calculation formula to obtain the density; The preset density calculation formula is: Among them, ρ(xi) represents the density of point x _i ; k represents the number of neighbors of a point, that is, using _the Euclidean distance as the metric, the top k points closest to a point; Represents the Euclidean distance between point x _i and point x _j , u represents the dimension of a feature vector concentration point, that is, the number of features, g is a subscript, used to index different features; KNN(xi ₎ represents the A set of the first k points with the smallest Euclidean distance from point x _i ; i represents the i-th point x _i in the data set; j represents the j-th point x _j in the data set; Substitute the density into the preset relative density calculation formula to calculate the relative density of the point to obtain the relative density; The preset relative density calculation formula is: Among them, r _ρ ( _xi ) represents the relative density of point x _i ; ρ ( _xi ) represents the density of point _xi ; ρ (x _j ) represents the density of point x _j ; k represents the number of neighbors of a point, that is Using Euclidean distance as the metric, the top k points closest to a point; KNN(xi ₎ represents the set of the top k points with the smallest Euclidean distance from point x _i in the data set; i represents the i-th point in the data set point x _i ;j represents the jth point x _j in the data set; Select points where the relative density is greater than the density average corresponding to the relative density to obtain multiple cluster centers; According to the cluster center, the feature vector set is divided using a preset domain calculation formula to obtain multiple sets; The calculation formula of the preset domain is: D _mn (x _p )={x _q |x _q ∈MN(x _p )∨(x _m ∈MN(x _p ))^x _q ∈MN(x _m ))}; In the formula, D _mn (x _p ) represents the domain of point x _p ; MN (x _p ) represents the set of all neighbors of point x _p ; MN (x _m ) represents the set of all neighbors of point x _m ; x _p represents the feature The p-th point in the vector set; x _m represents the m-th point in the feature vector set; x _q represents the q-th point in the feature vector set; According to the labels corresponding to the cluster centers, corresponding labels are set for the points corresponding to each set to obtain an event pseudo-label set. 3. The aircraft group collaborative decision-making generation method according to claim 1, characterized in that the step of training an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model includes: Input the event pseudo-label set into the initial event extraction model to obtain the initial loss function value; Fine-tune the initial event extraction model according to the initial loss function value to obtain an intermediate event extraction model; Input the event pseudo-label set into the intermediate event extraction model to obtain the target loss function value; Determine whether the target loss function value is a preset function value; If so, use the intermediate event extraction model corresponding to the target loss function value as the target event extraction model; If not, use the intermediate event extraction model as the initial event extraction model, and jump to the step of inputting the event pseudo-label set into the initial event extraction model to obtain an initial loss function value. 4. The aircraft group collaborative decision-making generation method according to claim 1, characterized in that the target event extraction model is used to extract events from the aircraft group data set and the ground control station data set respectively. Event comparison, the steps to obtain the solution comparison results, include: Extract events from the aircraft group data set and the ground control station data set respectively through the target event extraction model to obtain aircraft group events and ground events; Compare the aircraft group event and the ground event according to the trigger word category and parameter similarity to obtain an event comparison result; Using all the event comparison results, a plan comparison result is constructed. 5. The aircraft group collaborative decision-making generation method according to claim 1, characterized in that when the solution comparison result is inconsistent, the aircraft group is determined based on the authorization data corresponding to the aircraft group and the preset improved multi-agent depth. The steps of selecting the optimal instruction sequence using the sexual policy gradient algorithm and obtaining the aircraft group collaborative decision-making solution corresponding to the emergency data include: When the plan comparison results are inconsistent, determine whether the authorization data corresponding to the aircraft group is authorized; If so, determine the aircraft group collaborative decision-making plan corresponding to the emergency situation data based on the command sequence within the preset time and the aircraft group plan; If not, perform command sequence selection on the aircraft group data set and the ground control station data set according to the preset improved multi-agent depth deterministic policy gradient algorithm to obtain the target command sequence; Determine whether the target instruction sequence is a preset sequence; If so, use the aircraft group plan as the aircraft group collaborative decision-making plan corresponding to the emergency situation data; If not, the target command sequence is used as the aircraft group collaborative decision-making solution corresponding to the emergency situation data. 6. The aircraft group collaborative decision-making generation method according to claim 5, characterized in that the aircraft group collaborative decision-making corresponding to the emergency situation data is determined based on the instruction sequence within the preset time and the aircraft group plan. Program steps include: Determine whether the aircraft group and the ground control station have found the optimal command sequence within the preset time; If so, use the optimal instruction sequence as the aircraft group collaborative decision-making solution corresponding to the emergency situation data; If not, the aircraft group plan is used as the aircraft group collaborative decision-making plan corresponding to the emergency situation data. 7. The aircraft group collaborative decision-making generation method according to claim 5, characterized in that the multi-agent depth deterministic strategy gradient algorithm based on the preset improvement is used to compare the aircraft group data set and the ground control station data. Set the steps to select the instruction sequence and obtain the target instruction sequence, including: Using the decision criteria corresponding to the aircraft group data set and the ground control station data set to construct an initial decision criterion set; Select the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion; According to the priority decision-making criterion, an instruction acquisition algorithm is used to assign preset weights to instructions corresponding to the aircraft group data set and the ground control station data set, respectively, to obtain an instruction sequence set; Using a preset improved multi-agent deep deterministic policy gradient algorithm to select multiple initial instruction sequences that meet the preset score threshold in the instruction sequence set, and construct an initial instruction queue; Select the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence; The target command sequence is determined based on the constraints of the aircraft group on itself and the intermediate command sequence. 8. The aircraft group collaborative decision-making generation method according to claim 7, characterized in that the step of determining the target instruction sequence based on the aircraft group's own constraints and the intermediate instruction sequence includes: Determine whether the aircraft group can execute the intermediate command sequence under its own constraints; If so, use the intermediate instruction sequence as the target instruction sequence; If not, delete the intermediate instruction sequence from the initial instruction queue to obtain an intermediate instruction queue; Determine whether the intermediate instruction queue is an empty set; If so, delete the priority decision criterion from the initial decision criterion set to obtain a target decision criterion set; Determine whether the set of target decision criteria is an empty set; If so, use the preset sequence as the target instruction sequence; If not, use the target decision criterion set as the initial decision criterion set, and jump to the step of selecting the decision criterion with the highest priority in the initial decision criterion set to obtain the priority decision criterion; If not, use the intermediate instruction queue as the initial instruction queue, and jump to the step of selecting the initial instruction sequence with the highest score in the initial instruction queue as the intermediate instruction sequence. 9. An aircraft group collaborative decision-making generation system, characterized by including: The event pseudo-label set obtaining module is used to, when receiving emergency situation data of an aircraft group, use a preset density peak clustering algorithm to calculate the aircraft group data set corresponding to the emergency situation data and the ground control generated by the ground control station. The event pseudo-label corresponding to the station data set is obtained to obtain the event pseudo-label set; A target event extraction model obtaining module is used to train an initial event extraction model based on the event pseudo-label set to obtain a target event extraction model; A solution comparison result obtaining module is used to extract events from the aircraft group data set and the ground control station data set through the target event extraction model and perform event comparison to obtain a solution comparison result; The first acquisition module of aircraft group collaborative decision-making is used to select the optimal instruction sequence based on the authorization data corresponding to the aircraft group and the preset improved multi-agent deep deterministic strategy gradient algorithm when the plan comparison results are inconsistent. Obtain the aircraft group collaborative decision-making plan corresponding to the emergency situation data; The second acquisition module of aircraft group collaborative decision-making is used to use the aircraft group plan in the aircraft group data set as the aircraft group collaborative decision-making plan corresponding to the emergency situation data when the plan comparison results are consistent. 10. An electronic device, characterized in that it includes a memory and a processor. A computer program is stored in the memory. When the computer program is executed by the processor, the computer program causes the processor to execute claims 1 to 8. The steps of the aircraft group collaborative decision-making generation method described in any one of the above.