CN118626818A - Hypersphere discriminant feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition - Google Patents

Abstract

The present invention discloses a hypersphere discriminant feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition. With open set UAV radio frequency signal recognition as the goal, a deep neural network model is first constructed, and the model training is driven by using UAV radio frequency signal samples. The optimization goal is to minimize the hypersphere cross entropy, so as to obtain a model and hypersphere discriminant feature embedding suitable for open set UAV radio frequency signal recognition, and then evaluate the intra-class and inter-class cosine similarity of the feature embedding, and use the bimodal minimum fitting and detection algorithm to adaptively obtain the intra-class and inter-class decision threshold, and finally obtain the hypersphere discriminant feature embedding and adaptive decision threshold for identifying unknown UAV radio frequency signals, and output the predicted label. The present invention solves the problem that the existing recognition method has low accuracy in open set UAV radio frequency signal recognition due to the lack of decision module and insufficient discriminability of radio frequency signal features.

Description

Hypersphere discriminant feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition

Technical Field

The present invention relates to the field of low-altitude UAV safety, and in particular to a hypersphere discrimination feature embedding and adaptive decision threshold for open-set UAV radio frequency signal recognition.

Background Art

With the continuous advancement of drone technology, drones are rapidly becoming popular in the civilian field. With their miniaturization and diverse functions, drones have been widely used in all walks of life, promoting the development of low-altitude economy, digital transportation and low-altitude perception technology. However, with the rapid increase in the types and number of drones, malicious incidents have occurred frequently, posing serious hidden dangers to public safety. Considering that the characteristics of drone radio frequency signals are difficult to tamper with and unique, drone radio frequency signal recognition is an efficient drone authentication and identification method. In addition, in addition to the accurate identification of known types of drones, drone identification methods also need to have the ability to identify drones of unknown categories to adapt to the emergence of new types of drones.

Summary of the invention

In order to solve the deficiencies in the prior art, the present invention provides a hypersphere discrimination feature embedding and adaptive decision threshold for open set UAV RF signal recognition, which solves the problem of low accuracy in open set UAV RF signal recognition due to the insufficient discrimination of UAV RF signal features obtained by deep neural network models and the need to manually set and adjust the decision threshold in existing open set UAV RF signal recognition methods.

In order to achieve the above objectives, the present invention adopts the following technical solutions:

A hypersphere discrimination feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition specifically includes the following steps:

Step 1: Build a deep neural network model;

Step 2: Use the drone RF signal samples to drive the deep neural network model training, evaluate and optimize the hypersphere cross entropy loss between the predicted labels and the true labels of the deep neural network, and obtain the deep neural network model and hypersphere discriminant feature embedding suitable for unknown drone RF signal recognition;

Step 3: Evaluate the intra-class and inter-class cosine similarity of the hypersphere discriminant feature embedding;

Step 4: Fit the minimum value of the intra-class and inter-class cosine similarity bimodal curve to obtain an adaptive decision threshold suitable for unknown UAV RF signal recognition;

Step 5: Input the unknown UAV RF signal into the deep neural network model to obtain the predicted label of the unknown UAV RF signal.

Furthermore: the initialization of the deep neural network model in step 1 includes 9 layers of complex-valued separable convolution operations, 1 layer of flattening operations, and 2 layers of fully connected operations connected in sequence.

Further: In the step 2, the UAV RF signal sample set is used to drive the deep neural network model training, wherein the UAV RF signal sample set is composed of several groups of complex-valued baseband signals of UAVs and their corresponding real category labels in the form of one-hot encoding, and the training process specifically includes the following steps:

Step 2-1: Sample the drone RF signal Input the deep neural network model and get The characteristics of ,get The predicted class label of ,in represents the number of operation layers that make up the deep neural network model, Indicates the number of individual drones that constitute the drone RF signal sample set;

Step 2-2: transform the deep neural network model into The layer weight parameter is recorded as ,So After the last layer of fully connected operations, the components of the predicted category labels can be written as

in, represents the weight parameter of the last fully connected layer operation. Furthermore, the cross entropy loss between the true category label and the predicted category label of the drone RF signal can be written as

(1)

in , for The Code elements, for About The predicted probability value of the category;

Step 2-3, convert the cross entropy loss into a hyperspherical cross entropy loss, and apply a vector normalization operation to the weight parameters of the last fully connected layer operation, which can be written as

Furthermore, applying vector normalization to the features can be written as

Then, the components of the predicted class labels can be written as

The hypersphere cross entropy loss is obtained, which can be written as

Steps 2-4 introduce the hypersphere radius and cosine boundary to obtain a more stringent hypersphere cross entropy loss. Minimizing the hypersphere cross entropy loss on the drone RF signal sample set is the optimization goal of the deep neural network model, which can be written as

in represents the hypersphere boundary, Represents the cosine bounds.

Step 2-5, use the stochastic gradient descent algorithm to update the weight parameters of each layer of the deep neural network model.

in , is the update step size. After multiple forward and backward propagations, a deep neural network model suitable for identifying unknown UAV RF signals is obtained.

Further: In the step 2, the UAV RF signal sample set is used to drive the deep neural network model training, thereby obtaining the hypersphere discrimination feature embedding suitable for the recognition of unknown UAV RF signals. Specifically:

Step 2-6, input the drone RF signal sample set into the deep neural network model to obtain the hypersphere feature set of the drone RF signal sample set, which can be written as

Step 2-7, calculate the arithmetic mean of the feature sets belonging to the same category in the hypersphere feature set to obtain the hypersphere feature prototype set of the drone RF signal sample set,

Further: the step 3 evaluates the intra-class and inter-class cosine similarity of the hypersphere discriminant feature embedding, which specifically includes the following steps:

Step 3-1, calculate the intra-class cosine similarity of the hypersphere discriminant feature embedding,

Step 3-2, calculate the inter-class cosine similarity of the hypersphere discriminant feature embedding,

Step 3-3, obtaining the maximum and minimum values of the intra-class cosine similarity, and dividing the interval consisting of the minimum and maximum values into 20 sub-intervals, counting the number of intra-class cosine similarities in each interval, and drawing a histogram of the intra-class cosine similarity;

Step 3-4, obtain the maximum and minimum values of the inter-class cosine similarity, and divide the interval consisting of the minimum and maximum values into 20 sub-intervals, count the number of inter-class cosine similarities in each interval, and draw a histogram of the inter-class cosine similarity.

Further: the step 4 fits the minimum value of the intra-class and inter-class cosine similarity bimodal curve to obtain an adaptive decision threshold suitable for identifying unknown UAV radio frequency signals, which specifically includes the following steps:

Step 4-1, obtaining the histogram envelope of the intra-class cosine similarity and the histogram envelope of the inter-class cosine similarity, and using the cubic spline interpolation function to compensate for the missing values between the two envelopes to obtain a three-peak curve;

Step 4-2, find the minimum value of the three-peak curve. The horizontal coordinate value corresponding to the minimum value is the adaptive decision threshold for drone RF signal recognition. .

Further: In the step 5, the unknown UAV radio frequency signal is input into the deep neural network model to obtain the predicted label of the unknown UAV radio frequency signal, which specifically includes the following steps:

Step 5-1: Transmit unknown drone RF signal Input the deep neural network model to obtain its characteristics With predicted labels ;

Step 5-2, calculate the maximum cosine similarity between the features of the unknown drone RF signal and the hypersphere feature prototype set, and we have

Step 5-3, the predicted label of the unknown drone RF signal is obtained according to the following inequality, namely

The beneficial effects achieved by the present invention are as follows: In view of the problem of low accuracy in the existing open set UAV RF signal recognition method due to the insufficient discriminability of the UAV RF signal characteristics obtained by the deep neural network model and the need to manually set and adjust the decision threshold, the present invention creatively introduces hypersphere discrimination feature embedding and adaptive decision threshold, effectively realizing high-precision recognition of known UAVs and unknown UAVs, with high accuracy, stability and robustness.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of the hypersphere discrimination feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition of the present invention;

FIG2 is a deep neural network model of the present invention.

DETAILED DESCRIPTION

The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

As shown in Figure 1, a hypersphere discrimination feature embedding and adaptive decision threshold for open set UAV radio frequency signal recognition specifically includes the following steps:

Step 1: Use deep learning related knowledge to build and initialize a deep neural network model;

Step 2: Use the drone RF signal samples to drive the deep neural network model training, evaluate and optimize the hypersphere cross entropy loss between the predicted labels and the true labels of the deep neural network, and obtain the deep neural network model and hypersphere discriminant feature embedding suitable for unknown drone RF signal recognition;

Step 3: Evaluate the intra-class and inter-class cosine similarity of the hypersphere discriminant feature embedding;

Step 4: Fit the minimum value of the intra-class and inter-class cosine similarity bimodal curve to obtain an adaptive decision threshold suitable for unknown UAV RF signal recognition;

Step 5: Input the unknown UAV RF signal into the deep neural network model to obtain the predicted label of the unknown UAV RF signal.

In step 1, as shown in Figure 2, the initialized deep neural network model includes 9 layers of complex-valued separable convolution operations, 1 layer of flattening operations, and 2 layers of fully connected operations connected in sequence.

In step 2, the UAV RF signal sample set is used to drive the deep neural network model training, where the UAV RF signal sample set consists of several groups of complex-valued baseband signals of UAVs and their corresponding real category labels in the form of one-hot encoding. The training process specifically includes the following steps:

Step 2-1: Sample the drone RF signal Input the deep neural network model and get The characteristics of ,get The predicted class label of ,in represents the number of operation layers that make up the deep neural network model, Indicates the number of individual drones that constitute the drone RF signal sample set;

Step 2-2: transform the deep neural network model into The layer weight parameter is recorded as ,So After the last layer of fully connected operations, the components of the predicted category labels can be written as

in, represents the weight parameter of the last fully connected layer operation. Furthermore, the cross entropy loss between the true category label and the predicted category label of the drone RF signal can be written as

(1)

in , for The Code elements, for About The predicted probability value of the category;

Step 2-3, convert the cross entropy loss into a hyperspherical cross entropy loss, and apply a vector normalization operation to the weight parameters of the last fully connected layer operation, which can be written as

Furthermore, applying vector normalization to the features can be written as

Then, the components of the predicted class labels can be written as

The hypersphere cross entropy loss is obtained, which can be written as

Steps 2-4 introduce the hypersphere radius and cosine boundary to obtain a more stringent hypersphere cross entropy loss. Minimizing the hypersphere cross entropy loss on the drone RF signal sample set is the optimization goal of the deep neural network model, which can be written as

in represents the hypersphere boundary, Represents the cosine bounds.

Step 2-5, use the stochastic gradient descent algorithm to update the weight parameters of each layer of the deep neural network model.

in , is the update step size. After multiple forward and backward propagations, a deep neural network model suitable for identifying unknown UAV RF signals is obtained.

In step 2, the UAV RF signal sample set is used to drive the deep neural network model training, and then the hypersphere discriminant feature embedding suitable for the recognition of unknown UAV RF signals is obtained. Specifically:

Step 2-6, input the drone RF signal sample set into the deep neural network model to obtain the hypersphere feature set of the drone RF signal sample set, which can be written as

Step 2-7, calculate the arithmetic mean of the feature sets belonging to the same category in the hypersphere feature set to obtain the hypersphere feature prototype set of the drone RF signal sample set,

In step 3, the intra-class and inter-class cosine similarity of the hypersphere discriminant feature embedding is evaluated, which specifically includes the following steps:

Step 3-1, calculate the intra-class cosine similarity of the hypersphere discriminant feature embedding,

Step 3-2, calculate the inter-class cosine similarity of the hypersphere discriminant feature embedding,

Step 3-3, obtaining the maximum and minimum values of the intra-class cosine similarity, and dividing the interval consisting of the minimum and maximum values into 20 sub-intervals, counting the number of intra-class cosine similarities in each interval, and drawing a histogram of the intra-class cosine similarity;

Step 3-4, obtain the maximum and minimum values of the inter-class cosine similarity, and divide the interval consisting of the minimum and maximum values into 20 sub-intervals, count the number of inter-class cosine similarities in each interval, and draw a histogram of the inter-class cosine similarity.

In step 4, the minimum value of the intra-class and inter-class cosine similarity bimodal curve is fitted to obtain an adaptive decision threshold suitable for unknown UAV radio frequency signal recognition, which specifically includes the following steps:

Step 4-1, obtaining the histogram envelope of the intra-class cosine similarity and the histogram envelope of the inter-class cosine similarity, and using the cubic spline interpolation function to compensate for the missing values between the two envelopes to obtain a three-peak curve;

Step 4-2, find the minimum value of the three-peak curve. The horizontal coordinate value corresponding to the minimum value is the adaptive decision threshold for drone RF signal recognition. .

In step 5, the unknown UAV RF signal is input into the deep neural network model to obtain the predicted label of the unknown UAV RF signal, which specifically includes the following steps:

Step 5-1: Transmit unknown drone RF signal Input the deep neural network model to obtain its characteristics With predicted labels ;

Step 5-2, calculate the maximum cosine similarity between the features of the unknown drone RF signal and the hypersphere feature prototype set, and we have

Step 5-3, the predicted label of the unknown drone RF signal is obtained according to the following inequality, namely

The beneficial effects achieved by the present invention are as follows: In view of the problem of low accuracy in the existing open set UAV RF signal recognition method due to the insufficient discriminability of the UAV RF signal characteristics obtained by the deep neural network model and the need to manually set and adjust the decision threshold, the present invention creatively introduces hypersphere discrimination feature embedding and adaptive decision threshold, effectively realizing high-precision recognition of known UAVs and unknown UAVs, with high accuracy, stability and robustness.

Based on the same technical solution, the present invention also discloses a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device performs the above-mentioned hypersphere discrimination feature embedding and adaptive decision threshold for open set drone radio frequency signal recognition.

Based on the same technical solution, the present invention also discloses a computing device, including one or more processors, one or more memories and one or more programs, wherein the one or more programs are stored in the one or more memories and are configured to be executed by the one or more processors, and the one or more programs include instructions for executing the above-mentioned hypersphere discrimination feature embedding and adaptive decision threshold for open set drone radio frequency signal recognition.

It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

The present invention is described with reference to the flowcharts and/or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and/or box in the flowchart and/or block diagram, as well as the combination of the processes and/or boxes in the flowchart and/or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device produce a device for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device that implements the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

These computer program instructions may also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one or more processes in the flowchart and/or one or more boxes in the block diagram.

The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims (9)

1. The super-sphere discrimination feature of the open set unmanned aerial vehicle radio frequency signal identification is embedded into the self-adaptive decision threshold, the method is characterized by comprising the following steps of:

Step one, constructing a deep neural network model;

Secondly, training a deep neural network model by using an unmanned aerial vehicle radio frequency signal sample, and evaluating and optimizing hypersphere cross entropy loss between a predicted label and a real label of the deep neural network to obtain a deep neural network model suitable for unknown unmanned aerial vehicle radio frequency signal identification and hypersphere distinguishing feature embedding;

Step three, evaluating cosine similarity between intra-class classes embedded by the hyper-sphere discrimination features;

fitting minimum values of inter-class cosine similarity bimodal curves in the classes to obtain a self-adaptive decision threshold suitable for the identification of the radio frequency signals of the unknown unmanned aerial vehicle;

inputting the radio frequency signal of the unknown unmanned aerial vehicle into a deep neural network model to obtain a prediction tag of the radio frequency signal of the unknown unmanned aerial vehicle.

2. The open set unmanned aerial vehicle-oriented radio frequency signal identification-oriented supersphere discrimination feature embedding and self-adaptive decision threshold according to claim 1, wherein the deep neural network model in the first step comprises 9-layer complex-valued separable convolution operation, 1-layer flattening operation and 2-layer full-connection operation which are sequentially connected.

3. The open set unmanned aerial vehicle radio frequency signal identification-oriented superball discrimination feature embedding and self-adaptive decision threshold according to claim 1, wherein in the step two, the unmanned aerial vehicle radio frequency signal sample set is used for driving the deep neural network model training, wherein the unmanned aerial vehicle radio frequency signal sample set is composed of a plurality of groups of complex value baseband signals of unmanned aerial vehicles and corresponding true class labels in a single-heat coding form, and the training process specifically comprises the following steps:

Step 2-1, sampling radio frequency signals of the unmanned aerial vehicle Inputting the deep neural network model to obtainIs characterized by (1)ObtainingPredictive category labels of (c)WhereinRepresenting the number of operational layers that make up the deep neural network model,Representing the number of unmanned aerial vehicle individuals constituting a unmanned aerial vehicle radio frequency signal sample set;

Step 2-2, modeling the deep neural network The layer weight parameter is recorded asThenAfter the last layer of full-join operation processing, each component of the obtained predictive category label can be written as

Wherein, the method comprises the steps of, wherein,Weight parameters representing the operation of the last full-connection layer; further, the cross entropy loss between the true class tag and the predicted class tag of the unmanned radio frequency signal may be written as

，

Wherein the method comprises the steps of，Is thatThe first of (3)A number of symbols of a symbol,Is thatRegarding the firstA predicted probability value for the category;

step 2-3, converting the cross entropy loss into hyperspheric cross entropy loss, and applying vector normalization operation to the weight parameters of the last layer of full-connection layer operation, which can be written as

Further, the feature is subjected to a vector normalization operation, which can be written as

Then, the components of the resulting predictive category label may be written as

Obtaining hypersphere cross entropy loss, which can be written as

; Step 2-4, introducing hypersphere radius and cosine boundary to obtain harsher hypersphere cross entropy loss, taking the hypersphere cross entropy loss on the minimized unmanned aerial vehicle radio frequency signal sample set as an optimization target of the deep neural network model, and writing as follows

WhereinIndicating the boundary of the super-sphere,Representing cosine boundaries;

Step 2-5, updating the weight parameters of each layer operation of the deep neural network model by adopting a random gradient descent algorithm, wherein the weight parameters comprise Wherein，In order to update the step length, forward propagation and reverse propagation are carried out for a plurality of times, and a deep neural network model suitable for the identification of the radio frequency signals of the unknown unmanned aerial vehicle is obtained.

4. The open set unmanned aerial vehicle radio frequency signal identification-oriented superball discrimination feature embedding and adaptive decision threshold according to claim 1, wherein in the second step, the unmanned aerial vehicle radio frequency signal sample set is utilized to drive deep neural network model training, so as to obtain the superball discrimination feature embedding suitable for unknown unmanned aerial vehicle radio frequency signal identification, in particular

Step 2-6, inputting the unmanned aerial vehicle radio frequency signal sample set into a deep neural network model to obtain an hyperspherical characteristic set of the unmanned aerial vehicle radio frequency signal sample set, which can be written as

Step 2-7, calculating an arithmetic average value of feature sets belonging to the same category in the hyperspherical feature set to obtain a hyperspherical feature prototype set of the unmanned aerial vehicle radio frequency signal sample set, wherein the hyperspherical feature prototype set comprises

。

5. The open set unmanned aerial vehicle radio frequency signal identification-oriented hypersphere discrimination feature embedding and self-adaptive decision threshold of claim 1, wherein the step three evaluates the intraclass inter-class cosine similarity of the hypersphere discrimination feature embedding, which specifically comprises the following steps:

Step 3-1, calculating the similarity of the cosine in the class embedded by the hyper-sphere discrimination characteristics, wherein the similarity is as follows

; Step 3-2, calculating cosine similarity between classes embedded by the hyper-sphere discrimination features, including

; Step 3-3, obtaining the maximum value and the minimum value of the in-class cosine similarity, dividing a section consisting of the minimum value and the maximum value into 20 subsections, counting the number of the in-class cosine similarity in each section, and drawing a histogram of the in-class cosine similarity;

and 3-4, obtaining the maximum value and the minimum value of the inter-class cosine similarity, dividing a section formed by the minimum value and the maximum value into 20 subsections, counting the number of the inter-class cosine similarity in each section, and drawing a histogram of the inter-class cosine similarity.

6. The method for embedding and self-adapting decision threshold for identifying the radio frequency signals of the unmanned aerial vehicle facing the open set according to claim 1, wherein the step four is to fit minimum values of inter-class cosine similarity bimodal curves to obtain the self-adapting decision threshold for identifying the radio frequency signals of the unknown unmanned aerial vehicle, and the method specifically comprises the following steps:

Step 4-1, obtaining a histogram envelope line of the intra-class cosine similarity and a histogram envelope line of the inter-class cosine similarity, and compensating a missing value between the two envelope lines by using a cubic spline interpolation function to obtain a trimodal curve;

Step 4-2, obtaining the minimum value of the trimodal curve, wherein the abscissa value corresponding to the minimum value is the self-adaptive decision threshold of unmanned aerial vehicle radio frequency signal identification 。

7. The method for identifying the hypersphere discrimination features for the radio frequency signals of the open-set unmanned aerial vehicle according to claim 1 is characterized by comprising the following steps of:

step 5-1, the unknown unmanned aerial vehicle radio frequency signal Inputting the deep neural network model to obtain the characteristics thereofAnd predictive labels；

Step 5-2, calculating the maximum value of cosine similarity between the features of the radio frequency signal of the unknown unmanned aerial vehicle and the hypersphere feature prototype set, wherein the maximum value is

; Step 5-3, obtaining a predictive tag of the radio frequency signal of the unknown unmanned aerial vehicle according to the following inequality, namely

。

8. A computer readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a computing device, cause the computing device to perform the method of any of claims 1-7.

9. An electronic device comprising one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs comprising instructions for performing the method of any of claims 1-7.