CN118643735A - Soil compartment pressure prediction model transfer learning method, device, equipment and storage medium - Google Patents

Description

Soil compartment pressure prediction model transfer learning method, device, equipment and storage medium

Technical Field

The present invention relates to the technical field of shield tunneling, and in particular to a soil compartment pressure prediction model transfer learning method, device, equipment and storage medium.

Background Art

The soil compartment pressure of the shield machine is a key parameter in shield tunneling, and plays a key role in maintaining the stability of the tunnel face and preventing surface settlement. In the existing practice, soil compartment pressure prediction data models are set up to achieve accurate prediction of soil compartment pressure. However, these models have poor generalization ability in cross-interval shield machine projects, that is, the model trained in project A has poor prediction effect in project B. On the one hand, this is due to the inconsistent geometry, geological conditions and excavation parameter control protocols of the shield machines used in different shield projects. The spatial distribution of the soil compartment pressure of the shield machine shows different patterns in different projects, and there are large probability distribution differences between data sets; on the other hand, there is not enough labeled data for training in the target prediction interval. So far, there is rarely a transfer learning method to solve the generalization ability of the soil compartment pressure prediction model.

Summary of the invention

The present invention provides a soil compartment pressure prediction model transfer learning method, device, equipment and storage medium, which are used to solve the defect of poor generalization ability of the soil compartment pressure prediction model in the prior art and achieve the technical effect of improving the generalization ability of the soil compartment pressure prediction model.

The present invention provides a soil compartment pressure prediction model transfer learning method, comprising the following steps:

Acquire a plurality of first sample data of source domains and a plurality of second sample data of target domains; wherein the first sample data and the second sample data are dimensionless processed;

Inputting the first sample data and the second sample data into a basic soil compartment pressure prediction model of a source domain for iterative training, thereby obtaining a soil compartment pressure prediction model of the source domain; wherein a maximum average difference adaptation layer is provided in the basic soil compartment pressure prediction model;

The soil compartment pressure prediction model of the source domain is migrated to the target domain, and the parameters of the soil compartment pressure prediction model are adjusted based on the second sample data to obtain the soil compartment pressure prediction model of the target domain.

According to a soil compartment pressure prediction model transfer learning method provided by the present invention, the first sample data includes first training data and first label data, and the second sample data includes second training data and second label data; the first sample data and the second sample data are input into the basic soil compartment pressure prediction model of the source domain for iterative training, comprising:

Inputting the first training data and the second training data into the basic soil compartment pressure prediction model for processing, obtaining soil compartment pressure prediction data corresponding to the first training data, a first output value of the first training data at the maximum average difference adaptation layer, and a second output value of the second training data at the maximum average difference adaptation layer;

Calculate a training loss value according to the soil cabin pressure prediction data, the first label data, the first output value, and the second output value;

The basic soil compartment pressure prediction model will be iteratively trained according to the training loss value.

According to a soil compartment pressure prediction model transfer learning method provided by the present invention, the calculating of the training loss value according to the soil compartment pressure prediction data, the first label data, the first output value and the second output value includes: calculating the first loss value according to the soil compartment pressure prediction data and the first label data;

Calculate a second loss value according to the first output value and the second output value;

The training loss value is calculated according to the first loss value and the second loss value.

According to a soil compartment pressure prediction model transfer learning method provided by the present invention, the soil compartment pressure prediction model includes a parameter feature convolution module, a time series feature extraction module, a feature series prediction module and a maximum average difference adaptation layer; the output end of the parameter feature convolution module and the output end of the time series feature extraction module are respectively connected to the input end of the feature series prediction module, and the maximum average difference adaptation layer is arranged in the feature series prediction module.

According to a soil compartment pressure prediction model transfer learning method provided by the present invention, after adjusting the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain, the method further includes:

Acquire a first construction parameter, a first geological parameter, and a first soil compartment pressure parameter within a preset time step range from the current moment in the target domain, and acquire a second construction parameter and a second geological parameter of the current time step;

performing dimensionless processing on the first construction parameter, the first geological parameter, the first soil compartment pressure parameter, the second construction parameter, and the second geological parameter respectively, and inputting the dimensionless-processed first construction parameter and the first geological parameter into the parameter feature convolution module for processing to obtain a third output value; and

Inputting the dimensionless processed first soil compartment pressure parameter into the time series feature extraction module for processing to obtain a fourth output value;

The dimensionless processed second construction parameter and the second geological parameter, as well as the third output value and the fourth output value are input into the characteristic series prediction module for processing to obtain a predicted value of the target soil compartment pressure at the current time step.

According to a soil compartment pressure prediction model transfer learning method provided by the present invention, before obtaining a plurality of first sample data in source domains and a plurality of second sample data in target domains, the method further includes:

Acquire a plurality of original construction parameters and original soil compartment pressure parameters of the source domain and the target domain;

Obtaining a dimensionless processing strategy corresponding to the original construction parameters and the original soil compartment pressure parameters;

The original construction parameters and the original soil compartment pressure parameters are respectively dimensionally processed by the corresponding dimensionless processing strategy to obtain the first sample data and the second sample data.

According to a soil cabin pressure prediction model transfer learning method provided by the present invention, the dimensionless processing strategy corresponding to the original soil cabin pressure parameters is to obtain the characteristic coefficients of the mean term, the upper and lower gradient terms, the left and right gradient terms and the local heterogeneous terms through the physical characteristic function that describes the spatial distribution of the soil cabin pressure, and calculate the corresponding dimensionless value based on the characteristic coefficients.

The present invention also provides a soil compartment pressure prediction model transfer learning device, comprising the following modules:

A first acquisition module is configured to acquire first sample data of a plurality of source domains and second sample data of a plurality of target domains; wherein the first sample data and the second sample data are dimensionless processed;

A training module is configured to input the first sample data and the second sample data into a basic soil compartment pressure prediction model of a source domain for iterative training to obtain a soil compartment pressure prediction model of the source domain; wherein the basic soil compartment pressure prediction model is provided with a maximum average difference adaptation layer;

The migration module is configured to migrate the soil compartment pressure prediction model of the source domain to the target domain, and adjust the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain.

The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the soil compartment pressure prediction model transfer learning method as described in any one of the above is implemented.

The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the soil compartment pressure prediction model transfer learning method as described in any one of the above is implemented.

The present invention also provides a computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the computer program implements any of the soil compartment pressure prediction model transfer learning methods described above.

The present invention provides a soil cabin pressure prediction model transfer learning method, device, equipment and storage medium, which inputs the first sample data and the second sample data after dimensionless processing into the basic soil cabin pressure prediction model of the source domain for iterative training to obtain the soil cabin pressure prediction model of the source domain, and then migrates the soil cabin pressure prediction model of the source domain to the target domain, and adjusts the parameters of the soil cabin pressure prediction model based on the second sample data to obtain the soil cabin pressure prediction model of the target domain. The dimensionless processing can replace the unit removal of the physical quantity in the sample data with a suitable variable, realize the feature alignment of the data and eliminate the influence of certain parameter threshold ranges, provide a data basis for subsequent migration, and improve the generalization ability of the soil cabin pressure prediction model. At the same time, a maximum average difference adaptation layer is set in the basic soil cabin pressure prediction model, and the maximum average difference adaptation layer can pull in the distribution difference between the source domain and the target domain, which can improve the generalization ability of the soil cabin pressure prediction model, realize cross-regional prediction of the same model, and effectively improve the prediction accuracy of the model.

BRIEF DESCRIPTION OF THE DRAWINGS

In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

FIG1 is one of the flow charts of the soil compartment pressure prediction model transfer learning method provided by the present invention.

FIG2 is a schematic diagram of the migration architecture of the soil compartment pressure prediction model migration learning method provided by the present invention.

FIG. 3 is a schematic flow chart of step 120 in the embodiment provided by the present invention.

FIG. 4 is a second flow chart of the soil compartment pressure prediction model transfer learning method provided by the present invention.

FIG. 5 is a schematic diagram of calculating a predicted value of a target soil compartment pressure in an embodiment provided by the present invention.

FIG. 6 is a mapping diagram of parameters and dimensionless processing strategies in the embodiment provided by the present application.

FIG. 7 is a schematic diagram of the structure of the soil compartment pressure prediction model transfer learning device provided by the present invention.

FIG8 is a schematic diagram of the structure of an electronic device provided by the present invention.

DETAILED DESCRIPTION

In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

The soil compartment pressure prediction model transfer learning method of the present invention is described below in conjunction with Figures 1 to 5.

Fig. 1 is a flow chart of a soil compartment pressure prediction model transfer learning method according to an exemplary embodiment. As shown in Fig. 1, in an exemplary embodiment, the soil compartment pressure prediction model transfer learning method includes steps 110 to 130, which are described in detail as follows.

Step 110, obtaining a plurality of first sample data of source domains and a plurality of second sample data of target domains; wherein the first sample data and the second sample data are dimensionless processed.

In the embodiment of the present invention, due to the different models and/or construction areas of shield machines in various projects, the meaning and number of shield machine measurement parameters are inconsistent, which creates difficulties for the transfer learning model, that is, the original data features are inconsistent with the feature dimensions. When training the model, the present invention performs dimensionless processing on each data. Dimensionless (nondimensionalize/dimensionless) refers to removing part or all of the units of an equation involving physical quantities by replacing them with a suitable variable to achieve the purpose of simplifying experiments or calculations. Through dimensionless data, data feature alignment is achieved and the influence of certain parameter threshold ranges is eliminated, providing a data basis for subsequent migration.

The source domain has more first sample data, while the target domain is in the early stage of the project and has only a small amount of second sample data with labeled data. It is an engineering task that requires migrating the soil compartment pressure prediction model of the source domain.

Step 120: input the first sample data and the second sample data into a basic soil compartment pressure prediction model of a source domain for iterative training to obtain a soil compartment pressure prediction model of the source domain; wherein a maximum average difference adaptation layer is provided in the basic soil compartment pressure prediction model.

In an embodiment of the present invention, a large amount of first sample data and a small amount of second sample data are input into a basic soil cabin pressure prediction model in a source domain for iterative training to obtain a soil cabin pressure prediction model in the source domain. A maximum mean discrepancy adaptation layer (Maximum Mean Discrepancy, MMD) is provided in the basic soil cabin pressure prediction model, and the basic soil cabin pressure prediction model adopts a CNN-GRU (Convolutional Neural Network-Gated Recurrent Unit) model.

Step 130 : Migrate the soil compartment pressure prediction model of the source domain to the target domain, and adjust the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain.

In an embodiment of the present invention, the soil cabin pressure prediction model of the source domain obtained through training is migrated to the target domain, and the parameters of the soil cabin pressure prediction model are fine-tuned through a small amount of second sample data to obtain the soil cabin pressure prediction model of the target domain. The soil cabin pressure prediction model of the source domain and the soil cabin pressure prediction model of the target domain have the same model structure and differ only in specific parameters.

In an embodiment of the present invention, as shown in FIG2 , the first sample data and the second sample data after dimensionless processing are input into the basic soil cabin pressure prediction model of the source domain for iterative training to obtain the soil cabin pressure prediction model of the source domain, and then the soil cabin pressure prediction model of the source domain is migrated to the target domain, and the parameters of the soil cabin pressure prediction model are adjusted based on the second sample data to obtain the soil cabin pressure prediction model of the target domain. Dimensionless processing can replace the unit removal of the physical quantity in the sample data with a suitable variable, realize the feature alignment of the data and eliminate the influence of certain parameter threshold ranges, provide a data basis for subsequent migration, and improve the generalization ability of the soil cabin pressure prediction model. At the same time, a maximum average difference adaptation layer is set in the basic soil cabin pressure prediction model, and the maximum average difference adaptation layer can pull in the distribution difference between the source domain and the target domain, which can improve the generalization ability of the soil cabin pressure prediction model, realize cross-regional prediction of the same model, and effectively improve the prediction accuracy of the model.

In an exemplary embodiment of the present invention, the first sample data includes first training data and first label data, and the second sample data includes second training data and second label data; please refer to Figure 3, in step 120, the first sample data and the second sample data are input into the basic soil compartment pressure prediction model of the source domain for iterative training, including steps 310 to 330, which are described in detail as follows.

Step 310: input the first training data and the second training data into the basic soil compartment pressure prediction model for processing to obtain the soil compartment pressure prediction data corresponding to the first training data, the first output value of the first training data at the maximum average difference adaptation layer, and the second output value of the second training data at the maximum average difference adaptation layer.

In an embodiment of the present invention, as shown in FIG2 , a large amount of first training data in a source domain and a small amount of second training data in a target domain are input into a basic soil compartment pressure prediction model for training. After the first training data is processed by the basic soil compartment pressure prediction model, predicted soil compartment pressure prediction data is finally output. The first training data has corresponding first label data, and the first label data is the data of the actual soil compartment pressure corresponding to the first training data after dimensionless processing.

When the first training data and the second training data pass through the maximum mean difference adaptation layer, corresponding output values will be obtained.

Step 320: Calculate a training loss value based on the soil compartment pressure prediction data, the first label data, the first output value, and the second output value.

In the embodiment of the present invention, the training loss value is calculated based on the first label data, the soil compartment pressure prediction data obtained after being processed by the basic soil compartment pressure prediction model, the first output value, and the second output value. During the model training process, the training loss value includes the classification loss generated by the first sample data of the source domain and the difference between the source domain and the target domain generated by the maximum mean difference adaptation layer with MMD metric, thereby narrowing the distribution difference between the source domain and the target domain.

Step 330: iteratively train the basic soil compartment pressure prediction model according to the training loss value.

In the embodiment of the present invention, the foundation soil compartment pressure prediction model is iteratively trained according to the training loss value. When the training loss value is less than the set loss threshold, the training is terminated.

In an exemplary embodiment of the present invention, in step 320, the soil compartment pressure prediction model transfer learning calculates the training loss value according to the soil compartment pressure prediction data, the first label data, the first output value and the second output value, including the following steps, which are described in detail as follows.

A first loss value is calculated according to the soil compartment pressure prediction data and the first label data.

A second loss value is calculated according to the first output value and the second output value.

The training loss value is calculated according to the first loss value and the second loss value.

In the embodiment of the present invention, as shown in FIG2 , after the predicted data of the soil compartment pressure is obtained, the first loss value _LC is calculated based on the predicted data and the first label data, the second loss value L _MMD is calculated based on the first output value and the second output value, and the training loss value L (Train Loss) is calculated based on the first loss value and the second loss value.

Specifically, the calculation formula for the training loss value is as follows.

L＝ _LC + _λLMMD

Among them, λ represents a hyperparameter. Through the above formula for calculating the training loss value, the loss can be minimized as much as possible, not only minimizing the distance between domains (or maximizing domain confusion), but also obtaining a representation that is conducive to training a strong classifier. Such a representation can learn a strong classifier that is easy to transfer across domains.

In the embodiment of the present application, directly training the classifier using only the sample data of the source domain usually leads to overfitting of the distribution of the source domain, thereby resulting in performance degradation during testing when identifying the target domain. Therefore, the present invention learns a representation that can minimize the distance between the source domain distribution and the target domain distribution, and can train the classifier on the sample data of the source domain and directly apply it to the target domain with minimal accuracy loss. In order to minimize the domain difference between the distribution of the source domain and the target domain, the standard distribution distance metric, namely the maximum mean difference (MMD), is considered. Using a feature mapping function (·), X _S represents the first sample data set of the source domain, and the first sample data set has multiple first sample data x _s , and X _T represents the second sample data set of the target domain, and the second sample data set has multiple second sample data x _t . The maximum average difference between the first sample data x _s ∈ X _S in the source domain and the second sample data x _t ∈ X _T in the target domain is calculated, and the corresponding second loss value can be obtained. The calculation formula of the second loss value is as follows.

The calculation formula for the first loss value is as follows.

In the above formula, N represents the number of predicted data of soil tank pressure. Represents the label data corresponding to the i-th sample data, represents the predicted data of the soil compartment pressure of the i-th sample data, |X _S | represents the number of first sample data in the first sample data set, and |X _T | represents the number of second sample data in the second sample data set.

In an exemplary embodiment of the present invention, please refer to Figure 3, the soil compartment pressure prediction model transfer learning includes a parameter feature convolution module, a time series feature extraction module, a feature series prediction module and a maximum average difference adaptation layer; the output end of the parameter feature convolution module and the output end of the time series feature extraction module are respectively connected to the input end of the feature series prediction module, and the maximum average difference adaptation layer is arranged in the feature series prediction module.

In an embodiment of the present invention, the model structure of the soil compartment pressure prediction model is shown in FIG2, which is mainly composed of three modules: a parameter feature convolution module, a time series feature extraction module, and a feature series prediction module. The parameter feature convolution module is composed of a convolution layer, a pooling layer, and a fully connected layer. The convolution layer is set to 2 layers, and a two-dimensional convolution (Conv2D) is used to extract the associated features of the construction parameters and the geological parameters. The Dropout layer needs to be connected after the pooling layer to reduce the redundancy of the parameters and prevent the model from overfitting. The fully connected layer is set to 4 layers, which is used to learn the nonlinearity of the associated features. The time series feature extraction module is composed of a GRU layer (gated recurrent unit) and a fully connected layer. The GRU layer is set to 2 layers, and the structure of the fully connected layer in the time series feature extraction module is the same as the structure of the fully connected layer in the parameter feature convolution module. The feature series prediction module is composed of a series layer and a fully connected layer, which can fuse the input features. The maximum average difference adaptation layer is set between the series layer and the fully connected layer of the feature series prediction module, and the fully connected layer in the feature series prediction module is set to 4 layers.

When migrating the soil cabin pressure prediction model in the source domain to the soil cabin pressure prediction model in the target domain and adjusting the parameters of the model, first freeze all parameters in the model except the fully connected layer and the maximum average difference adaptation layer in the feature series prediction module, and use the second sample data with labeled data in the target domain to update the parameters of the fully connected layer and the maximum average difference adaptation layer in the feature series prediction module. Then, use the test sample data in the target domain to calculate the determination coefficient R2 of the soil cabin pressure prediction model in the target domain. In this way, the knowledge learned by the soil cabin pressure prediction model in the source domain can be migrated to the target domain, and only a small amount of second sample data in the target domain is needed for fine-tuning (Finetune).

The hyperparameters involved in the soil cabin pressure prediction model also include the hyperparameters of each module network, such as the dimension of the input data of each module and the size, number and activation function of each layer, and the hyperparameters of model training and optimization, such as learning rate and number of rounds.

In an exemplary embodiment of the present invention, please refer to Figure 4. In step 130, the soil cabin pressure prediction model transfer learning adjusts the parameters of the soil cabin pressure prediction model based on the second sample data to obtain the soil cabin pressure prediction model of the target domain. The method also includes steps 410 to 440, which are described in detail as follows.

Step 410, obtaining a first construction parameter, a first geological parameter, and a first soil compartment pressure parameter in the target domain within a preset time step range from the current moment, and obtaining a second construction parameter and a second geological parameter of the current time step.

In the embodiment of the present invention, as shown in FIG5, after the soil compartment pressure prediction model of the target domain is obtained, the first construction parameter, the first geological parameter and the first soil compartment pressure parameter in the target domain within the preset time step range from the current moment are obtained, and the preset time step range is the time step q-(Q-1) to q-1 shown in FIG5, and the second construction parameter and the second geological parameter of the current time step are obtained at the same time, such as the time step q shown in FIG5. Specifically, the preset time step range can be set to 8 time steps (min) from the current moment.

Step 420, respectively perform dimensionless processing on the first construction parameter, the first geological parameter, the first soil compartment pressure parameter, the second construction parameter and the second geological parameter, and input the dimensionless processed first construction parameter and the first geological parameter into the parameter feature convolution module for processing to obtain a third output value.

In the embodiment of the present invention, each parameter obtained is dimensionally processed. In FIG5 , the construction parameters and geological parameters after dimensionally processed are uniformly expressed as characteristic parameters. It represents the Mth characteristic parameter at the time step Q. After dimensionless processing, the soil cabin pressure parameter is the dimensionless value of the soil cabin pressure spatial characteristic coefficient (U _a , U _b , U _c , U _d ) of the corresponding time step.

The dimensionless first construction parameter and the first geological parameter are then input into the parameter feature convolution module for processing to obtain a third output value.

Step 430: Input the dimensionless processed first soil compartment pressure parameter into the time series feature extraction module for processing to obtain a fourth output value.

In the embodiment of the present invention, the dimensionless first soil compartment pressure parameter is simultaneously input into the time series feature extraction module for processing to obtain a fourth output value.

Step 440: Input the dimensionless processed second construction parameter and the second geological parameter, as well as the third output value and the fourth output value into the characteristic series prediction module for processing to obtain a predicted value of the target soil compartment pressure at the current time step.

In the embodiment of the present invention, the dimensionless second construction parameter and the second geological parameter, as well as the third output value and the fourth output value are input into the concatenation words of the feature series prediction module for feature fusion, and are processed through the maximum average difference adaptation layer and the full connection layer to obtain the target soil cabin pressure prediction value of the current time step. The result output by the feature series prediction module is the corresponding soil cabin pressure spatial characteristic function (U _a , U _b , U _c , U _d ), which is guided by the characteristic parameters to obtain the final target soil cabin pressure prediction value.

In another embodiment of the present invention, the dimensionless processed second geological parameter, the active control value in the second construction parameter, and the third output value and the fourth output value can also be input into the characteristic series prediction module for processing to obtain the target soil compartment pressure prediction value. The construction parameters and geological parameters are the same as the subsequent construction parameters and geological data, and are not described in detail here.

In an exemplary embodiment of the present invention, before the soil compartment pressure prediction model transfer learning obtains the first sample data of multiple source domains and the second sample data of multiple target domains in step 110, the method further includes the following steps, which are described in detail as follows.

A plurality of original construction parameters and original soil compartment pressure parameters of the source domain and the target domain are obtained.

A dimensionless processing strategy corresponding to the original construction parameters and the original soil compartment pressure parameters is obtained.

The original construction parameters and the original soil compartment pressure parameters are respectively dimensionally processed by the corresponding dimensionless processing strategy to obtain the first sample data and the second sample data.

In the embodiment of the present invention, the original construction parameters include bentonite flow, foam liquid flow, foam air flow, A/C group cylinder pressure eccentricity, B/D group cylinder pressure eccentricity, propulsion speed and cutter head speed, upper discharge door opening, lower discharge door opening, screw machine speed, cutter head torque, foam gun pressure, upper and lower cylinder strokes, left and right cylinder strokes, roll, and inclination. The original construction parameters can be divided into two categories: active control values and passive control values, wherein the active control values include bentonite flow, foam liquid flow, foam air flow, A/C group cylinder pressure eccentricity, B/D group cylinder pressure eccentricity, propulsion speed and cutter head speed, upper discharge door opening, lower discharge door opening, and screw machine speed. Passive control values include cutter head torque, foam gun pressure, upper and lower cylinder strokes, left and right cylinder strokes, roll, and inclination. The original soil compartment pressure parameter is the soil compartment pressure. The dimensionless processing strategy corresponding to each original construction parameter and the original soil compartment pressure parameter is shown in Figure 6.

The original geological parameters include burial depth, groundwater level, static soil pressure, tunnel face compressive strength, and tunnel face standard penetration. The original geological parameters do not need to be dimensionally processed.

In the embodiment of the present invention, when constructing the first sample data and the second sample data, in the source domain, the original geological parameters and the original construction parameters after dimensionless processing within the preset time step range, and the original geological parameters and the original construction parameters after dimensionless processing of the next time step adjacent to the preset time step range are used as the first training data of the first sample data. That is, when the preset time step range is 8, the original geological parameters and the original construction parameters after dimensionless processing of the 8th time step and the 9th time step are used as the first training data. The original soil compartment pressure data after dimensionless processing within the corresponding preset time step range, and the original soil compartment pressure data after dimensionless processing of the next time step adjacent to the preset time step range are used as the first label data.

In the target domain, the original geological parameters and the dimensionless original construction parameters within the preset time step range, as well as the original geological parameters and the dimensionless original construction parameters of the next time step adjacent to the preset time step range are used as second training data of a second sample data. The dimensionless original soil compartment pressure data within the corresponding preset time step range, as well as the dimensionless original soil compartment pressure data of the next time step adjacent to the preset time step range are used as second label data.

When the training samples are input into the basic soil compartment pressure prediction model for training, the process of inputting the data in the training samples into each module of the basic soil compartment pressure prediction model is similar to the aforementioned steps 410 to 440.

In an exemplary embodiment of the present invention, the dimensionless processing strategy corresponding to the original soil compartment pressure parameters is to obtain the characteristic coefficients of the mean term, the upper and lower gradient terms, the left and right gradient terms and the local heterogeneous terms by describing the physical characteristic function of the spatial distribution of the soil compartment pressure, and calculate the corresponding dimensionless values based on the characteristic coefficients.

In the embodiment of the present invention, a physical characteristic function suitable for describing the spatial distribution of soil cabin pressure is constructed, which can decouple the spatial distribution characteristics of soil cabin pressure, thereby realizing the dimensionless prediction of the target variable. The physical characteristic function divides the soil cabin pressure into four items, including the mean item, the upper and lower gradient item, the left and right gradient item, and the local heterogeneous item, as shown in the following formula.

P(θ,r)=[a+bcos(θ)+csin(θ)+dcos(θ)]g(r)

In the above formula: P(θ, r) represents the soil tank pressure, the coordinate system is a polar coordinate system with the geometric center of the soil tank as the pole and the vertical upward as the polar axis; θ is the polar angle; r is the polar diameter; g(r) is the radial coefficient; a, b, c, d are the characteristic coefficients of the mean term, upper and lower gradient term, left and right gradient term and local heterogeneous term of the soil tank pressure field respectively. The calculation formulas for the dimensionless values of the spatial characteristic coefficients of the soil tank pressure _Ua , _Ub , _Uc , _Ud are shown below.

U _a = a/static earth pressure.

U _b = b/static earth pressure.

U _c = c/static earth pressure.

U _d = d/static earth pressure.

Based on this process, the soil cabin pressure parameters per minute are dimensionally processed to obtain the dimensionless values of the characteristic coefficients of the physical characteristic functions of the spatial distribution of the soil cabin pressure per minute.

The soil compartment pressure prediction model transfer learning device provided by the present invention is described below. The soil compartment pressure prediction model transfer learning device described below and the soil compartment pressure prediction model transfer learning method described above can be referred to each other. It should be noted that the device provided in the following embodiment and the method provided in the above embodiment belong to the same concept, and the specific way in which each module and unit performs the operation has been described in detail in the method embodiment, which will not be repeated here.

In an exemplary embodiment of the present invention, please refer to FIG. 7 , which is a soil compartment pressure prediction model transfer learning device shown according to an exemplary embodiment, including the following modules.

The first acquisition module 710 is configured to acquire first sample data of multiple source domains and second sample data of multiple target domains; wherein the first sample data and the second sample data are dimensionless processed.

The training module 720 is configured to input the first sample data and the second sample data into the basic soil compartment pressure prediction model of the source domain for iterative training to obtain the soil compartment pressure prediction model of the source domain; wherein the basic soil compartment pressure prediction model is provided with a maximum average difference adaptation layer.

The migration module 730 is configured to migrate the soil compartment pressure prediction model of the source domain to the target domain, and adjust the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain.

In an exemplary embodiment of the present invention, the first sample data includes first training data and first label data, and the second sample data includes second training data and second label data; the training module 720 includes the following submodules.

The input submodule is configured to input the first training data and the second training data into the basic soil compartment pressure prediction model for processing, so as to obtain the soil compartment pressure prediction data corresponding to the first training data, the first output value of the first training data at the maximum average difference adaptation layer, and the second output value of the second training data at the maximum average difference adaptation layer.

A calculation submodule is configured to calculate a training loss value based on the soil compartment pressure prediction data, the first label data, the first output value and the second output value.

The training submodule is configured to iteratively train the basic soil compartment pressure prediction model according to the training loss value.

In an exemplary embodiment of the present invention, the calculation submodule includes the following submodules:

The first calculation unit is configured to calculate a first loss value according to the soil compartment pressure prediction data and the first label data.

The second calculation unit is configured to calculate a second loss value according to the first output value and the second output value.

A third calculation unit is configured to calculate the training loss value according to the first loss value and the second loss value.

In an exemplary embodiment of the present invention, the soil compartment pressure prediction model includes a parameter feature convolution module, a time series feature extraction module, a feature series prediction module and a maximum average difference adaptation layer; the output end of the parameter feature convolution module and the output end of the time series feature extraction module are respectively connected to the input end of the feature series prediction module, and the maximum average difference adaptation layer is arranged in the feature series prediction module.

In an exemplary embodiment of the present invention, the soil compartment pressure prediction model transfer learning device also includes the following modules.

The second acquisition module is configured to acquire the first construction parameter, the first geological parameter and the first soil compartment pressure parameter in the target domain within a preset time step range from the current moment, and to acquire the second construction parameter and the second geological parameter of the current time step.

The first dimensionless processing module is configured to perform dimensionless processing on the first construction parameter, the first geological parameter, the first soil compartment pressure parameter, the second construction parameter and the second geological parameter respectively, and input the first construction parameter and the first geological parameter after dimensionless processing into the parameter feature convolution module for processing to obtain a third output value.

The first input module is configured to input the dimensionless processed first soil compartment pressure parameter into the time series feature extraction module for processing to obtain a fourth output value.

The second input module is configured to input the dimensionless processed second construction parameter and the second geological parameter, as well as the third output value and the fourth output value into the characteristic series prediction module for processing to obtain the target soil compartment pressure prediction value of the current time step.

In an exemplary embodiment of the present invention, the soil compartment pressure prediction model transfer learning device also includes the following modules.

The second acquisition module is configured to acquire a plurality of original construction parameters and original soil compartment pressure parameters of the source domain and the target domain.

The third acquisition module is configured to obtain the dimensionless processing strategy corresponding to the original construction parameters and the original soil compartment pressure parameters.

The second dimensionless processing module is configured to perform dimensionless processing on the original construction parameters and the original soil compartment pressure parameters respectively through the corresponding dimensionless processing strategy to obtain the first sample data and the second sample data.

In an exemplary embodiment of the present invention, the dimensionless processing strategy corresponding to the soil compartment pressure parameters is to obtain the characteristic coefficients of the mean term, the upper and lower gradient terms, the left and right gradient terms and the local heterogeneous terms by describing the physical characteristic function of the spatial distribution of the soil compartment pressure, and calculate the corresponding dimensionless values based on the characteristic coefficients.

FIG8 illustrates a schematic diagram of the physical structure of an electronic device. As shown in FIG8 , the electronic device may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the soil cabin pressure prediction model transfer learning method, which includes: obtaining a plurality of first sample data of source domains and a plurality of second sample data of target domains; wherein the first sample data and the second sample data are dimensionless processed; inputting the first sample data and the second sample data into the basic soil cabin pressure prediction model of the source domain for iterative training to obtain the soil cabin pressure prediction model of the source domain; wherein the basic soil cabin pressure prediction model is provided with a maximum average difference adaptation layer; migrating the soil cabin pressure prediction model of the source domain to the target domain, and adjusting the parameters of the soil cabin pressure prediction model based on the second sample data to obtain the soil cabin pressure prediction model of the target domain.

In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the 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 a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

On the other hand, the present invention also provides a computer program product, which includes a computer program, and the computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the soil cabin pressure prediction model transfer learning method provided by the above methods, and the method includes: obtaining first sample data of multiple source domains and second sample data of multiple target domains; wherein the first sample data and the second sample data are dimensionless processed; the first sample data and the second sample data are input into the basic soil cabin pressure prediction model of the source domain for iterative training to obtain the soil cabin pressure prediction model of the source domain; wherein a maximum average difference adaptation layer is provided in the basic soil cabin pressure prediction model; the soil cabin pressure prediction model of the source domain is migrated to the target domain, and the parameters of the soil cabin pressure prediction model are adjusted based on the second sample data to obtain the soil cabin pressure prediction model of the target domain.

On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the soil cabin pressure prediction model transfer learning method provided by the above-mentioned methods, the method comprising: obtaining first sample data of multiple source domains and second sample data of multiple target domains; wherein the first sample data and the second sample data are dimensionlessly processed; the first sample data and the second sample data are input into a basic soil cabin pressure prediction model of the source domain for iterative training to obtain a soil cabin pressure prediction model of the source domain; wherein a maximum average difference adaptation layer is provided in the basic soil cabin pressure prediction model; the soil cabin pressure prediction model of the source domain is migrated to the target domain, and the parameters of the soil cabin pressure prediction model are adjusted based on the second sample data to obtain a soil cabin pressure prediction model of the target domain.

The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Those of ordinary skill in the art may understand and implement it without creative effort.

Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. A soil compartment pressure prediction model transfer learning method, characterized by comprising: Acquire a plurality of first sample data of source domains and a plurality of second sample data of target domains; wherein the first sample data and the second sample data are dimensionless processed; Inputting the first sample data and the second sample data into a basic soil compartment pressure prediction model of a source domain for iterative training, thereby obtaining a soil compartment pressure prediction model of the source domain; wherein a maximum average difference adaptation layer is provided in the basic soil compartment pressure prediction model; The soil compartment pressure prediction model of the source domain is migrated to the target domain, and the parameters of the soil compartment pressure prediction model are adjusted based on the second sample data to obtain the soil compartment pressure prediction model of the target domain. 2. The soil compartment pressure prediction model transfer learning method according to claim 1, characterized in that the first sample data includes first training data and first label data, and the second sample data includes second training data and second label data; the first sample data and the second sample data are input into the basic soil compartment pressure prediction model of the source domain for iterative training, comprising: Inputting the first training data and the second training data into the basic soil compartment pressure prediction model for processing, obtaining soil compartment pressure prediction data corresponding to the first training data, a first output value of the first training data at the maximum average difference adaptation layer, and a second output value of the second training data at the maximum average difference adaptation layer; Calculate a training loss value according to the soil cabin pressure prediction data, the first label data, the first output value, and the second output value; The basic soil compartment pressure prediction model will be iteratively trained according to the training loss value. 3. The soil compartment pressure prediction model transfer learning method according to claim 2, characterized in that the calculating of the training loss value according to the soil compartment pressure prediction data, the first label data, the first output value and the second output value comprises: Calculate a first loss value according to the soil compartment pressure prediction data and the first label data; Calculate a second loss value according to the first output value and the second output value; The training loss value is calculated according to the first loss value and the second loss value. 4. The soil compartment pressure prediction model transfer learning method according to claim 1 is characterized in that the soil compartment pressure prediction model includes a parameter feature convolution module, a time series feature extraction module, a feature series prediction module and a maximum average difference adaptation layer; the output end of the parameter feature convolution module and the output end of the time series feature extraction module are respectively connected to the input end of the feature series prediction module, and the maximum average difference adaptation layer is arranged in the feature series prediction module. 5. The soil compartment pressure prediction model transfer learning method according to claim 4, characterized in that after adjusting the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain, the method further comprises: Acquire a first construction parameter, a first geological parameter, and a first soil compartment pressure parameter within a preset time step range from the current moment in the target domain, and acquire a second construction parameter and a second geological parameter of the current time step; performing dimensionless processing on the first construction parameter, the first geological parameter, the first soil compartment pressure parameter, the second construction parameter, and the second geological parameter respectively, and inputting the dimensionless-processed first construction parameter and the first geological parameter into the parameter feature convolution module for processing to obtain a third output value; and Inputting the dimensionless processed first soil compartment pressure parameter into the time series feature extraction module for processing to obtain a fourth output value; The dimensionless processed second construction parameter and the second geological parameter, as well as the third output value and the fourth output value are input into the characteristic series prediction module for processing to obtain a predicted value of the target soil compartment pressure at the current time step. 6. The soil compartment pressure prediction model transfer learning method according to any one of claims 1 to 5, characterized in that before obtaining the first sample data of multiple source domains and the second sample data of multiple target domains, the method further comprises: obtaining multiple original construction parameters and original soil compartment pressure parameters of the source domain and the target domain; Obtaining a dimensionless processing strategy corresponding to the original construction parameters and the original soil compartment pressure parameters; The original construction parameters and the original soil compartment pressure parameters are respectively dimensionally processed by the corresponding dimensionless processing strategy to obtain the first sample data and the second sample data. 7. The soil compartment pressure prediction model transfer learning method according to claim 6 is characterized in that the dimensionless processing strategy corresponding to the original soil compartment pressure parameters is to obtain the characteristic coefficients of the mean term, upper and lower gradient terms, left and right gradient terms and local heterogeneous terms by describing the physical characteristic function of the spatial distribution of the soil compartment pressure, and calculate the corresponding dimensionless value based on the characteristic coefficients. 8. A soil compartment pressure prediction model transfer learning device, characterized by comprising: A first acquisition module is configured to acquire first sample data of a plurality of source domains and second sample data of a plurality of target domains; wherein the first sample data and the second sample data are dimensionless processed; A training module is configured to input the first sample data and the second sample data into a basic soil compartment pressure prediction model of a source domain for iterative training to obtain a soil compartment pressure prediction model of the source domain; wherein the basic soil compartment pressure prediction model is provided with a maximum average difference adaptation layer; The migration module is configured to migrate the soil compartment pressure prediction model of the source domain to the target domain, and adjust the parameters of the soil compartment pressure prediction model based on the second sample data to obtain the soil compartment pressure prediction model of the target domain. 9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the soil compartment pressure prediction model transfer learning method as described in any one of claims 1 to 7 is implemented. 10. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the soil compartment pressure prediction model transfer learning method according to any one of claims 1 to 7 is implemented.