Prediction model training method, prediction method and related device for train surface wind pressure

Abstract

The present application discloses a prediction model training method, prediction method and related devices for train surface wind pressure, wherein the method includes: using a training set to train a preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a target prediction model. The preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measurement point based on the training sample, and the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measurement point based on the graph structure information and the first predicted wind pressure data. In this way, information from different models can be integrated to improve the overall prediction performance, and the second preset model uses the graph structure information to help predict or interpolate missing values ​​through the graph structure. Therefore, the target prediction model can be applied to predict wind pressure data in any scenario, thereby improving the breadth of application.

Description

Technical Field

[0001] The present application belongs to the field of model application technology, and in particular relates to a prediction model training method, a prediction method and related devices for train surface wind pressure. Background Art

[0002] As a high-speed transportation system, the wind pressure data of the maglev train's pressure measurement points are mainly measured by sensors. However, in actual tests, these sensors are easily affected by vibration, impact, pressure changes, electrical interference, and chemical corrosion, causing failure and loss of critical data.

[0003] Therefore, under the influence of high-speed operation and environmental factors (such as crosswind), it is very likely that failure problems of pressure measuring points will occur frequently.

[0004] At present, traditional data interpolation technology is generally used to predict and interpolate failed pressure measurement points. However, traditional data interpolation technology is effective in some cases, but performs poorly when dealing with data missing problems caused by sensor failure. Therefore, traditional data interpolation technology has certain limitations. Summary of the Invention

[0005] The embodiments of the present application provide a prediction model training method, prediction method, device, equipment and storage medium for train surface wind pressure, which can improve the wide application of predicted wind pressure data.

[0006] In a first aspect, an embodiment of the present application provides a method for training a prediction model of wind pressure on a train surface, the method comprising:

[0007] Obtaining a training set; the training set includes a plurality of training samples, wherein each training sample includes an influencing parameter of wind pressure and wind pressure data labels of a plurality of pressure measuring points on the surface of the train, wherein the wind pressure data labels of the plurality of pressure measuring points include wind pressure data labels of failed pressure measuring points;

[0008] The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a target prediction model; wherein, the preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, and the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data, the first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label, the graph structure information is used to characterize the spatial position relationship of the multiple pressure measuring points, and the target prediction model is used to predict the wind pressure data of the failure pressure measuring point.

[0009] In some embodiments, the above-mentioned first loss function value is calculated by a first target loss function; wherein, the first target loss function includes a first loss function and a second loss function; the first loss function is used to determine the mean square error between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the second loss function is the residual term of the Reynolds-Navier-Stokes equation.

[0010] In some embodiments, the above-mentioned use of the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain the target prediction model includes: using the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain the initial prediction model; obtaining a test set; the test set includes a plurality of the training samples; using the test set to test and train the initial prediction model until the second loss function value of the initial prediction model meets the training stop condition to obtain the target prediction model; wherein the second loss function value is calculated by the second target loss function; the second target loss function includes a third loss function and a fourth loss function, the third loss function is used to determine the mean loss between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the fourth loss function is obtained by the sum of the squares of the weights, the regularization parameter, and the product of the preset coefficient.

[0011] In some embodiments, the above-mentioned training samples include experimental training samples and simulation training samples, and the training set is used to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain the target prediction model, including: using the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain an initial prediction model; obtaining a validation set; the validation set includes multiple training samples; based on a Gaussian process regression model, the validation set is used to evaluate the model performance of the initial prediction model to adjust the network parameters of the initial prediction model to obtain the target prediction model; wherein the Gaussian process regression model represents that the high-fidelity data is a linear transformation of the low-fidelity data plus a preset deviation function, and the low-fidelity data and the deviation function obey a multivariate Gaussian process, the mean of the multivariate Gaussian process is 0, and the covariance matrix is ​​composed of the covariance function of the low-fidelity data and the covariance function of the deviation function, the high-fidelity data is the second predicted wind pressure data corresponding to the experimental training sample, and the low-fidelity data is the second predicted wind pressure data corresponding to the simulation training sample.

[0012] In some embodiments, the first preset model is a back-propagation neural network model, and the second preset model is a graph neural network model.

[0013] In a second aspect, an embodiment of the present application provides a method for predicting wind pressure on a train surface, the method comprising:

[0014] Obtaining target influencing parameters of wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points;

[0015] The target influencing parameters, the target wind pressure data and the target graph structure information are input into the target prediction model to obtain the wind pressure data of the failure pressure measuring point output by the target prediction model; wherein, the target prediction model is obtained by the prediction model training method for train surface wind pressure described in any one of the first aspects.

[0016] In a third aspect, an embodiment of the present application provides a train surface wind pressure prediction model training device, the device comprising:

[0017] A first acquisition module is configured to acquire a training set; the training set includes a plurality of training samples, wherein each training sample includes an influencing parameter of wind pressure and wind pressure data labels of a plurality of pressure measuring points on the surface of the train, wherein the wind pressure data labels of the plurality of pressure measuring points include wind pressure data labels of failed pressure measuring points;

[0018] A training module is used to train the preset hybrid model using the training set until the first loss function value of the preset hybrid model meets the training stop condition to obtain a target prediction model; wherein, the preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, and the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data, the first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label, the graph structure information is used to characterize the spatial position relationship of the multiple pressure measuring points, and the target prediction model is used to predict the wind pressure data of the failure pressure measuring point.

[0019] In a third aspect, an embodiment of the present application provides a train surface wind pressure prediction model training device, the device comprising:

[0020] A second acquisition module is used to obtain target influencing parameters of the wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points;

[0021] A prediction module is used to input the target influencing parameters, the target wind pressure data and the target graph structure information into a target prediction model to obtain the wind pressure data of the failure pressure measuring point output by the target prediction model; wherein, the target prediction model is obtained by the prediction model training method for train surface wind pressure described in any one of the first aspects.

[0022] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising: a processor and a memory storing computer program instructions; when the processor executes the computer program instructions, the prediction model training method for train surface wind pressure as described in the first aspect is implemented;

[0023] Or implement the method for predicting the train surface wind pressure as described in the second aspect.

[0024] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, implements the train surface wind pressure prediction model training method as described in the first aspect;

[0025] Alternatively, the method for predicting train surface wind pressure as described in the second aspect is implemented.

[0026] In a seventh aspect, an embodiment of the present application provides a computer program product. When instructions in the computer program product are executed by a processor of an electronic device, the electronic device executes the train surface wind pressure prediction model training method as described in the first aspect.

[0027] Alternatively, the method for predicting train surface wind pressure as described in the second aspect is executed.

[0028] The train surface wind pressure prediction model training method, prediction method, device, equipment, and storage medium of the embodiments of the present application train surface wind pressure prediction model trains a preset hybrid model using a training set until the first loss function value of the preset hybrid model satisfies the training stop condition, thereby obtaining a target prediction model. In this way, the first preset model captures some data features that the second preset model fails to capture. By combining the prediction results of the two, information from different models can be integrated, thereby improving the overall prediction performance. The second preset model utilizes graph structure information to learn the spatial relationship between wind pressure measurement points, so that each node (wind pressure measurement point) can receive information from its neighboring nodes. In interpolation tasks, the wind pressure data of known nodes can be propagated to unknown nodes through the graph structure, helping to predict or interpolate missing values. Therefore, the target prediction model obtained in the present application can be applied to predict wind pressure data of failed wind pressure measurement points in any scenario of the train, thereby improving the breadth of application. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0030] Figure 1 1 is a flow chart of a method for training a prediction model of train surface wind pressure provided in an embodiment of the present application;

[0031] Figure 2 This is a schematic diagram of a BPNN and GNN hybrid neural network model provided in an embodiment of the present application;

[0032] Figure 3 A schematic diagram of a preset train provided in an embodiment of the present application;

[0033] Figure 4 A three-dimensional view of the numerical simulation calculation area and boundary condition strategy provided in the embodiment of the present application;

[0034] Figure 5 1 is a flow chart of a method for predicting train surface wind pressure provided in an embodiment of the present application;

[0035] Figure 6 A schematic diagram of the structure of a train surface wind pressure prediction model training device provided in an embodiment of the present application;

[0036] Figure 7 A schematic diagram of the structure of a device for predicting train surface wind pressure provided in an embodiment of the present application;

[0037] Figure 8 It is a structural diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0038] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain the present application, rather than to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present application by illustrating the examples of the present application.

[0039] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, the elements defined by the phrase "comprising..." do not exclude the presence of other identical elements in the process, method, article, or device comprising the elements.

[0040] To solve the related technical problems, the present invention provides a method, apparatus, device, storage medium, and computer program product for training a prediction model of train surface wind pressure. The following first introduces the method for training a prediction model of train surface wind pressure provided by the present invention.

[0041] Figure 1 The flowchart of the train surface wind pressure prediction model training method provided by the embodiment of the present application is shown. Figure 1 As shown, the train surface wind pressure prediction model training method specifically includes the following steps:

[0042] Step S101: Obtain a training set.

[0043] Step S102: Use the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a target prediction model.

[0044] The above training set includes multiple training samples.

[0045] Each training sample includes the influencing parameters of wind pressure and wind pressure data labels of multiple pressure measurement points on the train surface.

[0046] The above-mentioned influencing parameters may be train operation status information and information related to the wind acting on the train surface, such as speed, acceleration, ambient wind speed (wind speed acting on the train surface), wind direction, and yaw angle.

[0047] Among the wind pressure data tags of the multiple pressure measuring points, there is a wind pressure data tag of a failed pressure measuring point.

[0048] The above-mentioned failed pressure measuring point may be a pressure measuring point at which wind pressure data cannot be measured or a pressure measuring point at which wind pressure data measured has a large error.

[0049] In one embodiment, the wind pressure data tag of the above-mentioned failed pressure measuring point has a failure mark, that is, the wind pressure data tag of the failed pressure measuring point is marked as a failure mark to indicate that the wind pressure data tag is a failed wind pressure data tag, and the wind pressure data tag of the failed pressure measuring point can be marked.

[0050] In one embodiment, the wind pressure data tag may have a corresponding pressure measurement point identifier to indicate the pressure measurement point corresponding to the wind pressure data tag.

[0051] For example, suppose the train has pressure measuring points, and the wind pressure data of each pressure measuring point is , the status is marked as , The value range is 1 to , the influencing parameters are , The value ranges from 1 to G, where G is the number of parameters included in the influencing parameters. The training samples can form the measurement point feature matrix Y and the influencing parameter feature information s as follows:

[0052] Measurement point characteristic matrix Y:

[0053] ;

[0054] Each row represents the wind pressure data and status mark of a pressure measuring point.

[0055] ;

[0056] in, They represent different influencing parameters respectively.

[0057] The preset hybrid model includes a first preset model and a second preset model.

[0058] The first preset model is used to predict and obtain first predicted wind pressure data corresponding to the failure pressure measuring point based on the training samples.

[0059] The second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data.

[0060] The first predicted wind pressure data corresponding to the failure pressure measurement point may be the first predicted wind pressure data corresponding to each failure wind pressure data label in the training sample. The second predicted wind pressure data corresponding to the failure pressure measurement point may be the second predicted wind pressure data corresponding to each failure wind pressure data label in the training sample.

[0061] The above-mentioned first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label.

[0062] The above graph structure information is used to represent the spatial position relationship of multiple pressure measurement points.

[0063] The above target prediction model is used to predict the wind pressure data of the failed pressure measuring point.

[0064] In one embodiment, the above-mentioned training set is used to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain the target prediction model. The training sample can be input into the preset hybrid model, and the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, and the graph structure information and the first predicted wind pressure data are input into the second preset model, and the second preset model is used to output the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data. Based on the second predicted wind pressure data corresponding to the failure pressure measuring point and the wind pressure data label corresponding to the failure pressure measuring point, the first loss function value is determined, and the network parameters of the preset hybrid model are adjusted based on the first loss function value, and the training iteration is continued until the first loss function value meets the training stop condition to obtain the target prediction model.

[0065] In some implementations, the training samples may further include graph structure information.

[0066] The above graph structure information is used to describe the relationship between nodes (vertices), and the graph structure information can be defined by an adjacency matrix.

[0067] The adjacency matrix is ​​a two-dimensional matrix used to represent the connection relationship between nodes.

[0068] Matrix dimensions: Assume that the graph has nodes, the adjacency matrix A is a × The matrix of .

[0069] Matrix elements: =1 means there is an edge between node c and node d, =0 means there is no edge between node c and node d.

[0070] In this embodiment, the above nodes can be pressure measuring points. When there are pressure measuring points, the adjacency matrix A can be expressed as:

[0071] .

[0072] Therefore, the input process of the model can be as follows:

[0073] First, data preparation: obtain the wind pressure data of all pressure measurement points as , Status mark of each measuring point And the influencing parameters .

[0074] Data organization: Wind pressure data and status marks are combined into the measuring point feature matrix Y, the influencing parameter feature vector s is constructed, and the adjacency matrix A is constructed. The measuring point feature matrix Y, the influencing parameter feature vector s, and the adjacency matrix A are used as training samples.

[0075] Model processing: The training samples are input into the first preset model. The first preset model extracts global information and outputs intermediate features (the first predicted wind pressure data corresponding to the failed pressure measuring point). The intermediate features and the adjacency matrix A output by the first preset model are input into the second preset model. The second preset model uses the spatial relationship between the pressure measuring points to further process the data and output the second predicted wind pressure data corresponding to the failed pressure measuring point.

[0076] In one embodiment, the above-mentioned first preset model can be used to predict the first predicted wind pressure data corresponding to each pressure measuring point based on the training samples, and the first predicted wind pressure data corresponding to each pressure measuring point includes the first predicted wind pressure data corresponding to the failed pressure measuring point. The above-mentioned second preset model is used to predict the second predicted wind pressure data corresponding to each pressure measuring point based on the graph structure information and the first predicted wind pressure data, and the second predicted wind pressure data corresponding to each pressure measuring point includes the second predicted wind pressure data corresponding to the failed pressure measuring point.

[0077] Taking the first preset model as the back propagation neural network (BPNN) and the second preset model as the graph neural network (GNN) as an example, refer to Figure 2 , Figure 2 The schematic diagram of the BPNN and GNN hybrid neural network model provided in the embodiment of the present application, the input process of the model can be as follows:

[0078] First, data preparation: obtain the wind pressure data of all pressure measuring points, the status mark of each measuring point and the influencing parameters.

[0079] Data organization: Wind pressure data and status marks are combined into the measuring point feature matrix Y, the influencing parameter feature vector s is constructed, and the adjacency matrix A is constructed. The measuring point feature matrix Y, the influencing parameter feature vector s, and the adjacency matrix A are used as training samples.

[0080] Model processing: training sample x iThe data is input into the back propagation neural network (BPNN) through the input layer. The BPNN extracts global information and outputs intermediate features (the first predicted wind pressure data corresponding to the failed pressure measurement point). The intermediate features and the adjacency matrix A output by the BPNN are input into the graph neural network (GNN). The GNN uses the spatial relationship between the pressure measurement points to further process the data and outputs the training sample x through the output layer. i The second predicted wind pressure data corresponding to the corresponding failure pressure measuring point (Prediction value). Where i ranges from 1 to m, and m is the number of training samples. i Perform backpropagation and update iteration with the corresponding wind pressure data label (true value).

[0081] In one embodiment, the above-mentioned training set may include training samples corresponding to multiple trains of different models. The training samples may also include model information characterizing the train model. The model information may include at least one of the height, width, and shape of the train.

[0082] In some embodiments, the first loss function value is calculated by using a first target loss function.

[0083] The first objective loss function includes a first loss function and a second loss function.

[0084] The first loss function is used to determine the mean square error between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the second loss function is the residual term of the Reynolds-Navier-Stokes equation.

[0085] In one embodiment, the first objective loss function can be expressed as follows: (1);

[0086] (2);

[0087] (3);

[0088] Among them, Loss represents the first objective loss function, which is used to back-propagate and update parameters. represents the first loss function, It is a pure data fitting loss, which generally measures the difference between the predicted value and the true value. represents the second loss function, is the residual loss of the equation, used to "softly constrain" the network to satisfy the fluid dynamics equations. ω is a hyperparameter that controls the weighting of the two losses and is typically selected through cross-validation or empirical tuning. N is the total number of samples used to calculate the data loss, meaning that N represents the number of second-order predicted wind pressure data output for each training sample. represents the true label of the i-th sample, i represents the sequence number in the second predicted wind pressure data output for the training sample, Indicates the wind pressure data label corresponding to the i-th second predicted wind pressure data. Represents the network's predicted value for the i-th input sample, so that Same dimension, that is Represents the i-th second predicted wind pressure data. Represents the squared error between the prediction and the truth. Factor represents the mean squared error averaged over all samples. M is the total number of spatial / temporal points used to sample residuals in the constrained loss. The velocity vector function representing the flow field depends on the spatial coordinate x, that is, the three-dimensional velocity field (u x ,u y ,u z )). ▽ represents the spatial gradient operator, which acts on a scalar or vector field. Indicates the convection term for velocity (i.e. ), which describes the transmission of the fluid's own velocity. is the pressure gradient. ν: kinematic viscosity coefficient, usually , where μ is the dynamic viscosity. : The Laplace operator acts on the velocity vector, which is equivalent to . : At the jth spatial point x j , square the norm of the vector residual in the brackets.

[0089] In some embodiments, the above step S102 may also include but is not limited to the following:

[0090] The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining an initial prediction model.

[0091] Get the test set.

[0092] The above test set includes multiple training samples.

[0093] In one embodiment, the training samples in the test set are different from the training samples in the training set.

[0094] The initial prediction model is tested and trained using the test set until the second loss function value of the initial prediction model meets the training stop condition, and the target prediction model is obtained.

[0095] The second loss function value is calculated by the second objective loss function.

[0096] The above-mentioned second objective loss function includes a third loss function and a fourth loss function.

[0097] The third loss function is used to determine the mean loss between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample.

[0098] The fourth loss function is obtained by multiplying the square of the weight, the regularization parameter, and the preset coefficient.

[0099] In one embodiment, the second objective loss function can be expressed as follows:

[0100] (4);

[0101] in, is the network parameter vector of the preset hybrid model, For a single point loss, represents the loss between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, which can be square error or cross entropy, etc., and N represents the number of second predicted wind pressure data output for the training sample.

[0102] The third loss function is ,Right now It can represent the mean value between the losses of each second predicted wind pressure data corresponding to the training sample and the corresponding wind pressure data label in the training sample.

[0103] The fourth loss function is ,in, Indicates the preset coefficient, represents the regularization parameter, is the constraint strength, represents the sum of squares of weights, The value can be .

[0104] In some embodiments, the above-mentioned training samples include experimental training samples and simulation training samples. The above-mentioned experimental training samples are derived from actual experimental operations, and the above-mentioned simulation training samples are generated through computer simulation. The above-mentioned simulation training samples can correspond to the same working conditions as the experimental training samples, that is, the simulation training samples are obtained through computer simulation under the same working conditions as the experimental training samples.

[0105] The above-mentioned train can be a maglev train, a high-speed train, an airplane, etc.

[0106] In one implementation, the Reynolds-averaged Navier-Stokes equations can be used to simulate the aerodynamic performance of a train at different yaw angles and wind speeds to generate simulation training samples. For example, the aerodynamic performance of a train at different yaw angles (15°, 30°, 45°, and 60°) and a uniform wind speed (60 m / s) can be simulated. Wind tunnel experiments are conducted under the same operating conditions as the experimental training samples to collect aerodynamic data of the train model under the same conditions, generating the experimental training samples. This aerodynamic data can include influencing parameters and corresponding wind pressure data.

[0107] In one embodiment, the above simulation training samples can be obtained in the following manner:

[0108] A preset train of preset dimensions is constructed, and a computational domain is defined around the preset train.

[0109] The above-mentioned preset dimensions may include the characteristic height H of the train 高度 , characteristic length L of the train 长度 , characteristic width W of the train 宽度 And the cross-sectional area of ​​the train, etc.

[0110] The above calculation domain is the numerical simulation calculation area.

[0111] Reference Figure 3 , Figure 3 A schematic diagram of a preset train provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, the characteristic height H of the preset train 高度 4.2, characteristic length L 长度 19.04H 高度 , characteristic width W 宽度 0.88H 高度 , the unit is meter (m), where the characteristic length L 长度 Including the characteristic length of the train head, the characteristic length of the middle of the train and the characteristic length of the tail of the train. The characteristic length of the train head, the characteristic length of the middle of the train and the characteristic length of the tail of the train are 6.42H respectively. 高度 、6.20H 高度 and 6.42H 高度 .

[0112] Based on the preset train traveling in a target direction of a calculation domain defined by the preset train, a running state of the preset train in a crosswind is simulated, and aerodynamic force data is obtained in the simulated running state of the preset train in the crosswind.

[0113] The main incoming flow is introduced through the first velocity inlet of the computational domain, and the crosswind is simulated through the second velocity inlet of the computational domain. The first boundary surface of the computational domain and the second boundary surface of the computational domain are configured as zero-pressure outlets to simulate the actual atmospheric environment at the far end of the preset train.

[0114] Construct simulation training samples based on aerodynamic data.

[0115] Reference Figure 4 , Figure 4 A three-dimensional view of the numerical simulation calculation area and boundary condition strategy provided in the embodiment of the present application, such as Figure 4 As shown, the three-dimensional view includes a first velocity inlet 1 , a second velocity inlet 2 , a first pressure outlet 1 (first boundary surface), and a first pressure outlet 2 (second boundary surface).

[0116] The ratio of the cross-sectional area of ​​the train to the cross-sectional area of ​​the computational domain is less than a preset value, which may be less than 1, such as 0.01. The height of the computational domain is not less than a first preset multiple of the train's characteristic height, which may be 8 times. The upstream distance of the computational domain is at least the first preset multiple of the train's characteristic height, and the downstream distance of the computational domain is at least the second preset multiple of the train's characteristic height, which may be twice the first preset multiple. For example, the first preset multiple may be 8 times, and the second preset multiple may be 16 times. The lateral width of the flow field on the windward side should be at least the first preset multiple of the train's characteristic height, and the lateral width of the flow field on the leeward side should be at least the second preset multiple of the train's characteristic height.

[0117] For example, the simulation train runs along the x-direction in an open field environment, and the computational domain defined around the train is a rectangular area and follows strict specifications. The ratio of the cross-sectional area of ​​the train to the cross-sectional area of ​​the computational domain is less than 0.01, the computational domain height is not less than 8 times the characteristic height, the upstream distance is at least 8 times the characteristic height, and the downstream distance is at least 16 times the characteristic height. In addition, the lateral width of the flow field is at least 8 times the characteristic height on the windward side and at least 16 times the characteristic height on the leeward side. The simulation is based on the principle of relative motion. The train remains stationary and the forward movement of the train is simulated by applying the incoming flow in front. Figure 4, boundary ABB1A1 is set as velocity inlet 1 to introduce the main incoming flow, and boundary ADD1A1 is set as velocity inlet 2 to simulate the crosswind effect. The velocity components at these two inlets are defined as v1=vcosθ, v1 is the speed of the train, v2=vsinθ, v2 is the speed of the crosswind effect, and v is the speed of the crosswind. The two construct the velocity incoming flow in the synthetic CFD calculation. The boundary surfaces CDD1C1 (pressure outlet 1) and CBB1C1 (pressure outlet 2) are configured as zero-pressure outlets to simulate the actual atmospheric environment at the far end of the train. The ground and track surfaces (ABCD) are treated as moving no-slip walls whose movement speed is opposite to the train speed to ensure accurate simulation of relative motion. The top surface A1B1C1D1 is designated as a symmetry surface, prohibiting any convective flux and ensuring that the vertical velocity component is zero. The train surface is defined as a no-slip wall to accurately simulate the interaction between the fluid and the train surface. In Figure 4 In , the characteristic length of the train is L, the length of the computational domain along the x-direction is 2L + L + L, the width of the computational domain along the y-direction is L + 2L, and the height of the computational domain along the z-direction is L.

[0118] In one embodiment, the above step S102 may also include but is not limited to the following:

[0119] The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining an initial prediction model.

[0120] Get the validation set.

[0121] The above validation set includes multiple training samples.

[0122] In one embodiment, the training samples in the validation set may be different from the training samples in the training set and the test set.

[0123] Based on the Gaussian process regression model, the validation set is used to evaluate the model performance of the initial prediction model to adjust the network parameters of the initial prediction model and obtain the target prediction model.

[0124] Among them, the Gaussian process regression model represents that the high-fidelity data is a linear transformation of the low-fidelity data plus a preset deviation function, and the low-fidelity data and the deviation function obey a multivariate Gaussian process, the mean of the multivariate Gaussian process is 0, and the covariance matrix is ​​composed of the covariance function of the low-fidelity data and the covariance function of the deviation function. The high-fidelity data is the second predicted wind pressure data corresponding to the experimental training sample, and the low-fidelity data is the second predicted wind pressure data corresponding to the simulation training sample.

[0125] In one embodiment, the Gaussian process regression model can be expressed as follows:

[0126] (5);

[0127] in, 、 are low-fidelity data and high-fidelity data, respectively. Obtained by maximum likelihood estimation. Represents the deviation function, characterizing y H (x) and The nonlinear difference between is assumed to be a Gaussian process. represents the Gaussian process prior, with a mean of 0 and a covariance matrix of ,Right now . Represents the covariance kernel function of the low-fidelity data itself, corresponding to . Represents the covariance kernel function of the deviation function δ, corresponding to . In the covariance matrix represents y L and δ are independent of each other, and each satisfies the corresponding Gaussian process prior.

[0128] In some embodiments, the above step S102 may further include:

[0129] The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a first prediction model.

[0130] Get the test set.

[0131] The first prediction model is tested and trained using the test set until the second loss function value of the first prediction model meets the training stop condition, thereby obtaining a second prediction model.

[0132] Get the validation set.

[0133] Based on the Gaussian process regression model, the validation set is used to evaluate the model performance of the second prediction model to adjust the network parameters of the second prediction model and obtain the target prediction model.

[0134] The above-mentioned method of using the test set to test and train the first prediction model until the second loss function value of the first prediction model meets the training stop condition to obtain the second prediction model can refer to the above-mentioned method of using the test set to test and train the initial prediction model until the second loss function value of the initial prediction model meets the training stop condition to obtain the target prediction model.

[0135] The above-mentioned method of using a validation set to evaluate the model performance of the second prediction model based on the Gaussian process regression model to adjust the network parameters of the second prediction model to obtain the target prediction model can refer to the above-mentioned method of using a validation set to evaluate the model performance of the initial prediction model based on the Gaussian process regression model to adjust the network parameters of the initial prediction model to obtain the target prediction model.

[0136] In some embodiments, the above step S102 may further include:

[0137] The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a first prediction model.

[0138] Get the validation set.

[0139] Based on the Gaussian process regression model, the validation set is used to evaluate the model performance of the first prediction model to adjust the network parameters of the first prediction model and obtain the second prediction model.

[0140] Get the test set.

[0141] The second prediction model is tested and trained using the test set until the second loss function value of the second prediction model meets the training stop condition, thereby obtaining the target prediction model.

[0142] The above-mentioned method of using a validation set to evaluate the model performance of the first prediction model based on the Gaussian process regression model to adjust the network parameters of the first prediction model to obtain the second prediction model can refer to the above-mentioned method of using a validation set to evaluate the model performance of the initial prediction model based on the Gaussian process regression model to adjust the network parameters of the initial prediction model to obtain the target prediction model.

[0143] The above-mentioned method of using the test set to test and train the second prediction model until the second loss function value of the second prediction model meets the training stop condition to obtain the target prediction model can refer to the above-mentioned method of using the test set to test and train the initial prediction model until the second loss function value of the initial prediction model meets the training stop condition to obtain the target prediction model.

[0144] In some embodiments, the test set is used to test and train the second prediction model until the second loss function value of the second prediction model meets the training stop condition to obtain the target prediction model, and the method may further include:

[0145] A data set is obtained, wherein the data set includes a plurality of training samples. The training samples in the data set may be different from the training samples in the validation set, the test set, or the training set.

[0146] Based on the data set, the second prediction model is trained using a cross-validation method to obtain a third prediction model.

[0147] Specifically, the data set is divided into K modules. Fold in half Training with data other than The final selection:

[0148] ;

[0149] By verifying on unseen data, we avoid excessive memorization of hyperparameters to the training set. Where K is the total number of folds, that is, the dataset is divided into K parts, resulting in K subsets. k represents the kth subset, where k ranges from 1 to K, i.e., k = 1,…,K. Indicates the use of hyperparameters The parameter vector obtained by training the data after removing the k-th fold data (k-th subset) in the data set is The parameter vector obtained by training. Represents the hyperparameters of the model, such as regularization coefficient, learning rate, etc., where the optimal value needs to be selected through cross-validation. Indicates that the trained parameter w will be used in the k-fold data D k The validation loss obtained above (can be mean square error, cross entropy, etc.). The validation loss of all K-fold data is averaged as The overall evaluation index of the value.

[0150] The third prediction model is tested and trained using the test set until the second loss function value of the third prediction model meets the training stop condition, thereby obtaining the target prediction model.

[0151] In some embodiments, the first preset model is a back propagation neural network model (BPNN), and the second preset model is a graph neural network model (GNN).

[0152] In order to better implement the above-mentioned train surface wind pressure prediction model training method, the present application embodiment provides a complete embodiment process as follows:

[0153] Step 1:

[0154] The Reynolds-averaged Navier-Stokes equations were used to simulate the aerodynamic performance of the train at different yaw angles and wind speeds, generating a large number of preliminary data sets (aerodynamic data).

[0155] Step 2:

[0156] Wind tunnel experiments are conducted in a controlled environment to collect aerodynamic data on the train model under the same conditions. The aerodynamic data includes influencing parameters and wind pressure data corresponding to each pressure measurement point.

[0157] Step 3:

[0158] The data obtained from numerical simulation and wind tunnel experiments are cleaned, formatted, and feature engineered to prepare data for neural network training. The feature engineering formula is as follows:

[0159] ;

[0160] in, is the i-th eigenvalue before feature engineering, or the i-th data point before feature engineering, is the i-th eigenvalue after feature engineering, which can be the i-th data point after feature engineering. Q is the total number of all samples of this feature in the original data set (i.e., the number of samples). It represents the mean (average value), which is the sum of all sample values ​​(eigenvalues) divided by the number of samples. The formula is , the above data points can be training samples.

[0161] Step 4:

[0162] A hybrid neural network of back-propagation neural network (BPNN) and graph neural network (GNN) was used to preliminarily configure the two neural unit modules. Unlike traditional data subset clustering or classification division, the present invention specifically constructs a spatial correlation graph matrix of pressure measurement points, and captures the nonlinear spatial correlation relationship between measurement points through the self-attention weight matrix. An appropriate number of hidden layers and neurons are designed for BPNN, and an appropriate graph structure is set for GNN. Among them, the BPNN network independently processes single-point pressure prediction, and the collaborative relationship between the independent network training process and the GNN structure is different from traditional alternating training. The self-attention mechanism is introduced into the original GNN framework to allow the model to automatically learn the correlation weights between train measurement points. The implementation method is as follows:

[0163] ;

[0164] ;

[0165] ;

[0166] in, is the feature vector (wind pressure data) of the pressure reading of node c before it is input into the GNN. is a trainable linear transformation matrix (with dimensionality transformation), which is used to map the original node features to the new representation space. 、 Node c and node d are mapped to dimensional space after the eigenvector. is a trainable attention weight vector used to Performs a linear mapping and outputs a scalar. : The scoring function in the attention mechanism corresponds to an "inactive" attention score.

[0167] is the set of neighbor nodes of node c in the graph (usually including the indexes of all nodes connected to c by edges). : For Neighbors' Day , the calculated unnormalized attention score. is the number of neighbor nodes q Sum and normalize. : The final attention weight of node c to node d (after normalization), satisfying .

[0168] : The updated feature vector output by node c in the current graph convolutional layer (GAT layer). : According to the attention weight α cd Mapping characteristics of neighbor node d Weighted summation is performed to obtain aggregated information; the sum is then passed into the activation function σ to obtain the final new feature .

[0169] Step 5:

[0170] Select an activation function and optimizer for the neural network, and set the learning rate and other hyperparameters. Split the preprocessed data into training, validation, and test sets. Train the neural network using the training data, monitor the loss and accuracy during training, and perform model tuning. The loss function uses a multi-weight approach. During training, in addition to fitting the data loss, the residual term of the Reynolds-Navier-Stokes equations is added to the network loss, so that the soft constraint model satisfies the basic equations of fluid dynamics:

[0171] (1);

[0172] (2);

[0173] (3);

[0174] Among them, Loss is the total loss of the neural network (the first objective loss function), which is used to backpropagate and update parameters. It is the pure data fitting loss (the first loss function), which generally measures the difference between the predicted value and the true value. is the residual loss (the second loss function), which is used to "softly constrain" the network to satisfy the fluid dynamics equations. ω is a hyperparameter that controls the weights of the two losses and is typically selected through cross-validation or empirical tuning.

[0175] N is the total number of samples used to calculate data loss. i : The true label of the i-th sample. : The network's predicted value for the i-th input sample, making it consistent with y i Same dimension. : The squared error between the prediction and the truth. factor It represents the mean square error averaged over all samples.

[0176] M is the total number of spatial / temporal points used to sample residuals in the constrained loss. is the velocity vector function of the flow field, which depends on the spatial coordinate x, that is, the three-dimensional velocity field (u x ,u y ,u z )). ▽ is the spatial gradient operator, which acts on a scalar or vector field. For the velocity advection term (i.e. ), which describes the transmission of the fluid's own velocity. is the pressure gradient. ν: kinematic viscosity coefficient, usually , where μ is the dynamic viscosity. : The Laplace operator acts on the velocity vector, which is equivalent to . : At the jth spatial point x j , square the norm of the vector residual in the brackets.

[0177] The validation set was used to evaluate model performance, hyperparameters were adjusted to optimize model performance, and hybrid modeling was performed using combined CFD numerical simulation and wind tunnel test data. This fully utilized the advantages of both types of data, effectively improving the model's generalization ability and prediction accuracy in actual engineering applications. Low-fidelity numerical simulation and high-fidelity experimental data were jointly modeled, and the difference between the two was captured using an autoregressive Gaussian process, fully leveraging the complementary advantages of large amounts of CFD data and small amounts of experimental data.

[0178] ;

[0179] in 、 are low-fidelity numerical simulation data and high-fidelity experimental data, respectively. Obtained by maximum likelihood estimation. : Deviation function, characterizing yH (x) and b The nonlinear difference between , itself assumed to be a Gaussian process. : Gaussian process prior, mean is 0, covariance matrix is ,Right now . : The covariance kernel function of the low-fidelity model itself, corresponding to . : The covariance kernel function of the deviation function δ, corresponding to . In the covariance matrix represents y L and δ are independent of each other, and each satisfies the corresponding Gaussian process prior.

[0180] Step 6:

[0181] The cross-validation method is used to ensure the generalization ability and stability of the model. The dataset is divided into K modules. Fold in half Training with data other than Final selection

[0182] ;

[0183] By verifying on unseen data, we avoid excessive memorization of hyperparameters to the training set. Where K is the total number of folds, that is, the dataset is divided into K parts. k : The kth subset, where k = 1,…,K. : Using hyperparameters After removing the k-th fold data from the data set, The parameter vector obtained by training. : Hyperparameters of the model (such as regularization coefficient, learning rate, etc.), where the optimal value needs to be selected through cross-validation. :The trained parameter w is used in the k-th fold data D k The validation loss obtained above (can be mean square error, cross entropy, etc.). To find the average validation loss of all K folds, The overall evaluation index of the value.

[0184] Step 7:

[0185] The performance of the neural network was evaluated using an independent test set to assess its accuracy and robustness in predicting failure data at pressure measurement points. Further adjustments and optimizations were made based on the test results, adding a parameter penalty term (secondary objective loss function) to the original empirical risk minimization objective:

[0186] ;

[0187] in, is the wind pressure data model parameter vector, is the single point loss (such as square error, cross entropy, etc.), is the constraint strength. Since there is a corresponding gradient descent update that effectively suppresses the problem of excessive parameters, in order to reduce overfitting of training noise, the following formula is proposed:

[0188] ;

[0189] After that, let the validation set loss be calculated at each cycle t during the training process ,when

[0190] When , stop training and take the parameters as:

[0191] ;

[0192] Introduce a random "dropout" mechanism in each layer of the neural network. Apply a mask to the activation vector of the lth layer :

[0193] ;

[0194] The drop probability during training is 1-p, and the scaling weight during testing is This randomization integrates exponentially in the model space and can significantly reduce overfitting.

[0195] After the model is operational, it can receive new aerodynamic data in real time to predict and interpolate data from pressure measurement points that may have failed. Regularly monitor the model's predictive performance to ensure it maintains high accuracy under new operating conditions. Adjust model parameters based on new data and actual operational feedback, repeating Step 6 and subsequent steps to adapt to changes in environmental and operating conditions.

[0196] This embodiment applies a hybrid neural network model that combines a back-propagation neural network (BPNN) and a graph neural network (GNN). Through this combination, accurate prediction and data interpolation of failed pressure measurement points of a maglev train under crosswind conditions are achieved, thereby improving the safety of train operation and the accuracy of the monitoring system.

[0197] In this embodiment, the network structure can be designed to determine the number of hidden layers, the number of neurons in each layer, and the specific network structure suitable for this problem (such as the topology of the neural network shown in the figure). The original data obtained from CFD simulation and wind tunnel experiments is cleaned and formatted, which includes processing missing values, outliers, and data standardization. Feature engineering is performed, including selecting features that significantly affect the prediction results and creating new features to improve the model's learning efficiency and prediction accuracy. The optimal training strategy is determined, using loss functions and optimizers, and adjusting the learning rate and other training parameters. Regularization techniques with dropout and early stopping are implemented to avoid overfitting.

[0198] Multiple rounds of validation and model comparison can be performed to determine which model structure or parameter configuration performs best in practical applications.

[0199] The above-mentioned neural network model includes an adaptive learning mechanism that can automatically adjust network parameters such as learning rate and weight according to different environmental changes to optimize the performance of the model in the new environment, thereby meeting the data prediction needs under different wind speed and wind direction conditions.

[0200] In the data preprocessing stage, advanced data cleaning, normalization and standardization are used, and feature engineering techniques are combined to extract key features. These processes and techniques improve the quality and consistency of input data, enhance the effectiveness of model training and the accuracy of predicted data.

[0201] Regularization techniques, Dropout and Early Stopping, as well as optimization algorithm selection, such as using Adam or SGD optimizers, can also be introduced. These techniques help reduce model overfitting and improve the model's ability to generalize to unseen data.

[0202] Using graph neural networks to process the spatial data of pressure measurement points on the train surface and learning the complex relationships between pressure points through graph structures can effectively capture and predict spatially interdependent pressure measurement point data, providing more accurate analysis and prediction for complex spatial data.

[0203] The present invention uses graph neural networks to process the spatial relationships and dependencies of pressure measurement points. GNN can effectively capture the complex interactions between measurement points and is particularly suitable for processing spatial data of pressure measurement points distributed on the surface of trains. The introduction of the self-attention mechanism significantly improves the accuracy of spatial data processing and the robustness of the model.

[0204] The present invention uses backpropagation neural networks and graph neural networks to evaluate their performance to select the best model. In order to reduce the overfitting phenomenon of the model, the generalization performance of the model is ensured through techniques such as Dropout, Early Stopping and soft constraints, avoiding the model's excessive dependence on training data. This hybrid method can optimize model performance. The present invention significantly improves the accuracy and reliability of pressure measurement by using backpropagation neural networks and graph neural networks to predict pressure measurement points. In particular, under different wind speed and yaw angle conditions, the prediction results are better than traditional methods and can effectively cope with complex environments.

[0205] The present invention proposes a data interpolation method based on a deep learning model to handle data loss caused by failure of pressure measuring points. This method does not rely on traditional statistical interpolation methods, but predicts failure data by learning deep features of data. Compared with traditional methods, this method directly uses the original pressure measuring point data for prediction without performing any form of modal dimensionality reduction or decomposition, thereby retaining the true spatial correlation characteristics between measuring point data to the maximum extent. The present invention adopts a deep neural network, which can capture the complex nonlinear relationships in sensor data and achieve efficient data reconstruction. During the operation of high-speed maglev trains, sensors may fail due to factors such as vibration and external force impact, resulting in loss of pressure data. The present invention reconstructs the faulty measuring points through a deep learning model to ensure accurate prediction and supplementation of missing data.

[0206] The neural network model of the present invention can adaptively adjust its parameters according to different meteorological conditions and measurement point failure modes to optimize prediction performance. The model's adaptive learning mechanism can dynamically adjust parameters to adapt to environmental changes and is applicable to maglev trains and even broader high-speed transportation fields such as high-speed railways and aviation. This adaptive learning mechanism improves the adaptability and stability of the model in changing environments. The present invention demonstrates the stability and adaptability of the model in dynamic environments by testing the robustness of the model under different yaw angles and multi-point failure scenarios. Even in the case of multiple sensor failures, the present invention can still maintain high prediction accuracy and reliability.

[0207] In the wind environment simulation of high-speed trains, processing large amounts of sensor data has a high computational cost. Through the optimized neural network structure, the present invention significantly improves the efficiency of data processing and reduces the consumption of computing resources while ensuring accuracy.

[0208] In some embodiments, the target prediction model can also be used to predict wind pressure data on the surface of urban infrastructure such as bridges and elevated roads.

[0209] In an embodiment of the present application, the first preset model captures some data features that the second preset model fails to capture. By combining the prediction results of the two, information from different models can be integrated, which may improve the overall prediction performance. The second preset model uses graph structure information to learn the spatial relationship between pressure measurement points, so that each node (wind pressure measurement point) can receive information from its neighboring nodes. In the interpolation task, the wind pressure data of known nodes can be propagated to unknown nodes through the graph structure to help predict or interpolate missing values. Therefore, the target prediction model obtained in the present application can be applied to predict the wind pressure data of failed pressure measurement points in any scenario of the train, thereby improving the breadth of application.

[0210] In some embodiments, the present application also provides a method for predicting the wind pressure on the surface of a train, referring to Figure 5 , Figure 5 This is a flow chart of a method for predicting train surface wind pressure provided in an embodiment of the present application. The method for predicting train surface wind pressure specifically includes the following steps:

[0211] Step S501 : obtaining target influencing parameters of the wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points.

[0212] Step S502 : inputting target influencing parameters, target wind pressure data and target graph structure information into a target prediction model to obtain wind pressure data of a failure pressure measuring point output by the target prediction model.

[0213] The target prediction model is obtained by the above-mentioned train surface wind pressure prediction model training method.

[0214] In some embodiments, the target wind pressure data may be marked with identification information indicating the corresponding target pressure measurement point.

[0215] In some embodiments, target model data of the target train can also be obtained before executing step S502. The above step S502 can also be to input the target influence parameters, target wind pressure data, target graph structure information and target model data into the target prediction model to obtain the wind pressure data of the failure pressure measuring point output by the target prediction model.

[0216] In an embodiment of the present application, the target prediction model combines the first preset model and the second preset model. Through this combination, accurate prediction and data interpolation of pressure measurement points that fail on the train under crosswind conditions can be achieved.

[0217] In order to better implement the above-mentioned train surface wind pressure prediction model training method, the present application embodiment provides a train surface wind pressure prediction model training device, referring to Figure 6 , Figure 6This is a schematic diagram of the structure of a train surface wind pressure prediction model training device provided in an embodiment of the present application. The train surface wind pressure prediction model training device 600 includes the following:

[0218] The first acquisition module 601 is used to obtain a training set; the training set includes multiple training samples, wherein each training sample includes an influencing parameter of wind pressure and wind pressure data labels of multiple pressure measuring points on the surface of the train, and the wind pressure data labels of the multiple pressure measuring points include wind pressure data labels of failed pressure measuring points.

[0219] The training module 602 is used to train the preset hybrid model using the training set until the first loss function value of the preset hybrid model meets the training stop condition to obtain the target prediction model; wherein, the preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, and the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data, the first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label, the graph structure information is used to characterize the spatial position relationship of multiple pressure measuring points, and the target prediction model is used to predict the wind pressure data of the failure pressure measuring point.

[0220] In some embodiments, the first loss function value in the above-mentioned training module 602 is calculated by a first target loss function, wherein the first target loss function includes a first loss function and a second loss function; the first loss function is used to determine the mean square error between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the second loss function is the residual term of the Reynolds-Navier-Stokes equation.

[0221] In some embodiments, the above-mentioned training module 602 is specifically used to: use the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition, and an initial prediction model is obtained; obtain a test set; the test set includes multiple training samples; use the test set to test and train the initial prediction model until the second loss function value of the initial prediction model meets the training stop condition, and a target prediction model is obtained; wherein, the second loss function value is calculated by the second target loss function; the second target loss function includes a third loss function and a fourth loss function, the third loss function is used to determine the mean loss between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the fourth loss function is obtained by the sum of the squares of the weights, the regularization parameter, and the product of the preset coefficients.

[0222] In some embodiments, the above-mentioned training module 602 is specifically used to: use the training set to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition to obtain an initial prediction model; obtain a verification set; the verification set includes multiple training samples; based on the Gaussian process regression model, use the verification set to evaluate the model performance of the initial prediction model to adjust the network parameters of the initial prediction model to obtain a target prediction model; wherein, the Gaussian process regression model represents that the high-fidelity data is a linear transformation of the low-fidelity data plus a preset deviation function, and the low-fidelity data and the deviation function obey a multivariate Gaussian process, the mean of the multivariate Gaussian process is 0, and the covariance matrix is ​​composed of the covariance function of the low-fidelity data and the covariance function of the deviation function, the high-fidelity data is the second predicted wind pressure data corresponding to the experimental training sample, and the low-fidelity data is the second predicted wind pressure data corresponding to the simulation training sample.

[0223] In some embodiments, the first preset model in the above-mentioned training module 602 is a back-propagation neural network model, and the second preset model is a graph neural network model.

[0224] The train surface wind pressure prediction model training device 600 provided in the embodiment of the present application can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar and will not be repeated here.

[0225] In order to better implement the above-mentioned train surface wind pressure prediction method, the present application embodiment provides a train surface wind pressure prediction device, referring to Figure 7 , Figure 7 This is a schematic diagram of the structure of a train surface wind pressure prediction device provided in an embodiment of the present application. The train surface wind pressure prediction device 700 includes the following:

[0226] The second acquisition module 701 is used to acquire target influencing parameters of the wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points.

[0227] Prediction module 702 is used to input the target impact parameters, target wind pressure data, and target graph structure information into a target prediction model to obtain wind pressure data at the failure pressure measurement point output by the target prediction model. The target prediction model is obtained using the aforementioned train surface wind pressure prediction model training method.

[0228] The train surface wind pressure prediction device 700 provided in the embodiment of the present application can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar and will not be repeated here.

[0229] Figure 8 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present application is shown.

[0230] The electronic device may include a processor 801 and a memory 802 storing computer program instructions.

[0231] Specifically, the processor 801 may include a central processing unit (CPU), or an application specific integrated circuit (ASIC), or may be configured to implement one or more integrated circuits of the embodiments of the present application.

[0232] The memory 802 may include a large-capacity memory for data or instructions. By way of example and not limitation, the memory 802 may include a hard disk drive (HDD), a floppy disk drive, a flash memory, an optical disk, a magneto-optical disk, a magnetic tape, or a universal serial bus (USB) drive, or a combination of two or more of these. Where appropriate, the memory 802 may include removable or non-removable (or fixed) media. Where appropriate, the memory 802 may be internal or external to the integrated gateway disaster recovery device. In a specific embodiment, the memory 802 is a non-volatile solid-state memory.

[0233] In some embodiments, the memory 802 may include read-only memory (ROM), random access memory (RAM), magnetic disk storage media devices, optical storage media devices, flash memory devices, electrical, optical, or other physical / tangible memory storage devices. Thus, in general, the memory includes one or more tangible (non-transitory) computer-readable storage media (e.g., memory devices) encoded with software including computer-executable instructions, and when the software is executed (e.g., by one or more processors), it is operable to perform the operations described with reference to the method according to an aspect of the present disclosure.

[0234] The processor 801 reads and executes computer program instructions stored in the memory 802 to implement any one of the train surface wind pressure prediction model training methods or train surface wind pressure prediction methods in the above embodiments.

[0235] In one example, the electronic device may further include a communication interface 803 and a bus 810. Figure 8 As shown, the processor 801, the memory 802, and the communication interface 803 are connected via a bus 810 and communicate with each other.

[0236] The communication interface 803 is mainly used to implement communication between various modules, devices, units and / or equipment in the embodiments of the present application.

[0237] The bus 810 includes hardware, software, or both that couples components of the electronic device to each other. By way of example, and not limitation, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industrial Standard Architecture (EISA) bus, a Front Side Bus (FSB), a HyperTransport (HT) interconnect, an Industrial Standard Architecture (ISA) bus, an InfiniBand interconnect, a Low Pin Count (LPC) bus, a memory bus, a Micro Channel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or a combination of two or more of these. Where appropriate, the bus 810 may include one or more buses. Although the embodiments of the present application describe and illustrate specific buses, the present application contemplates any suitable bus or interconnect.

[0238] The electronic device can execute the train surface wind pressure prediction model training method or train surface wind pressure prediction method in the embodiment of the present application, thereby realizing the combination of Figure 1 and Figure 6 The train surface wind pressure prediction model training method and device described, or the combination of Figure 5 and Figure 7 A method and device for predicting train surface wind pressure are described.

[0239] In addition, in conjunction with the train surface wind pressure prediction model training method or train surface wind pressure prediction method in the above-mentioned embodiments, the present application embodiments may provide a computer-readable storage medium for implementation. The computer-readable storage medium stores computer program instructions; when the computer program instructions are executed by a processor, the method for training a train surface wind pressure prediction model or the method for predicting train surface wind pressure in any of the above-mentioned embodiments are implemented.

[0240] In combination with the train surface wind pressure prediction model training method or train surface wind pressure prediction method in the above-mentioned embodiments, an embodiment of the present application provides a computer program product. When the instructions in the computer program product are executed by the processor of an electronic device, the electronic device executes any one of the train surface wind pressure prediction model training methods in the above-mentioned embodiments, or the electronic device executes any one of the train surface wind pressure prediction methods in the above-mentioned embodiments.

[0241] It should be understood that the present application is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, a detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present application is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order of the steps after understanding the spirit of the present application.

[0242] The functional blocks shown in the block diagrams described above can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they may be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, and the like. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments may be stored in a machine-readable medium or transmitted via a data signal carried in a carrier wave over a transmission medium or communication link. "Machine-readable medium" may include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROMs, flash memory, erasable ROMs (EROMs), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, and the like. Code segments may be downloaded via a computer network such as the Internet or an intranet.

[0243] It should also be noted that the exemplary embodiments mentioned in this application describe some methods or systems based on a series of steps or devices. However, this application is not limited to the order of the above steps. In other words, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0244] Aspects of the present disclosure have been described above with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each block in the flowcharts and / or block diagrams, as well as combinations of blocks in the flowcharts and / or block diagrams, 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, or other programmable data processing device to produce a machine such that execution of these instructions by the processor of the computer or other programmable data processing device enables the implementation of the functions / actions specified in one or more blocks in the flowcharts and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field programmable logic circuit. It should also be understood that each block in the block diagrams and / or flowcharts, as well as combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware that performs the specified functions or actions, or by a combination of dedicated hardware and computer instructions.

[0245] The above description is only a specific embodiment of the present application. Those skilled in the art will clearly understand that for the convenience and brevity of description, the specific working processes of the systems, modules and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here. It should be understood that the scope of protection of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed in the present application, and these modifications or replacements should be included in the scope of protection of the present application.

Claims

1. A method for training a prediction model of train surface wind pressure, characterized in that: include: Obtaining a training set; the training set includes a plurality of training samples, wherein each training sample includes an influencing parameter of wind pressure and wind pressure data labels of a plurality of pressure measuring points on the surface of the train, wherein the wind pressure data labels of the plurality of pressure measuring points include wind pressure data labels of failed pressure measuring points; The preset hybrid model is trained using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a target prediction model; wherein the preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data, the first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label, the graph structure information is used to characterize the spatial position relationship between the multiple pressure measuring points, and the target prediction model is used to predict the wind pressure data of the failure pressure measuring point; The training samples include experimental training samples and simulation training samples. The preset hybrid model is trained using the training set until a first loss function value of the preset hybrid model meets a training stop condition, thereby obtaining a target prediction model, including: The training set is used to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining an initial prediction model; Obtaining a validation set; the validation set includes multiple training samples; Based on the Gaussian process regression model, the validation set is used to evaluate the model performance of the initial prediction model to adjust the network parameters of the initial prediction model to obtain the target prediction model; wherein, the Gaussian process regression model represents: the high-fidelity data is a linear transformation of the low-fidelity data plus a preset deviation function, and the low-fidelity data and the deviation function obey a multivariate Gaussian process, the mean of the multivariate Gaussian process is 0, and the covariance matrix of the multivariate Gaussian process is composed of the covariance function of the low-fidelity data and the covariance function of the deviation function, the high-fidelity data is the second predicted wind pressure data corresponding to the experimental training sample, and the low-fidelity data is the second predicted wind pressure data corresponding to the simulation training sample.

2. The method according to claim 1, characterized in that The first loss function value is calculated by a first target loss function; Among them, the first objective loss function includes a first loss function and a second loss function; the first loss function is used to determine the mean square error between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the second loss function is the residual term of the Reynolds-Navier-Stokes equation.

3. The method according to claim 1, characterized in that The method of training a preset hybrid model using the training set until a first loss function value of the preset hybrid model satisfies a training stop condition to obtain a target prediction model includes: The training set is used to train the preset hybrid model until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining an initial prediction model; Obtaining a test set; the test set includes multiple training samples; The initial prediction model is tested and trained using the test set until the second loss function value of the initial prediction model meets the training stop condition, thereby obtaining a target prediction model; wherein, the second loss function value is calculated by a second target loss function; the second target loss function includes a third loss function and a fourth loss function, the third loss function is used to determine the mean loss between the second predicted wind pressure data and the wind pressure data label in the corresponding training sample, and the fourth loss function is obtained by multiplying the sum of the squares of the weights, the regularization parameter, and the preset coefficient.

4. The method according to claim 1, wherein The first preset model is a back-propagation neural network model, and the second preset model is a graph neural network model.

5. A method for predicting wind pressure on a train surface, characterized in that: include: Obtaining target influencing parameters of wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points; The target influencing parameters, the target wind pressure data and the target graph structure information are input into a target prediction model to obtain the wind pressure data of the failure pressure measuring point output by the target prediction model; wherein, the target prediction model is obtained by the prediction model training method for train surface wind pressure according to any one of claims 1 to 4.

6. A train surface wind pressure prediction model training device, characterized in that: The device comprises: A first acquisition module is configured to acquire a training set; the training set includes a plurality of training samples, wherein each training sample includes an influencing parameter of wind pressure and wind pressure data labels of a plurality of pressure measuring points on the surface of the train, wherein the wind pressure data labels of the plurality of pressure measuring points include wind pressure data labels of failed pressure measuring points; A training module is used to train the preset hybrid model using the training set until the first loss function value of the preset hybrid model meets the training stop condition, thereby obtaining a target prediction model; wherein, the preset hybrid model includes a first preset model and a second preset model, the first preset model is used to predict the first predicted wind pressure data corresponding to the failure pressure measuring point based on the training sample, and the second preset model is used to predict the second predicted wind pressure data corresponding to the failure pressure measuring point based on the graph structure information and the first predicted wind pressure data, the first loss function value is determined by the second predicted wind pressure data and the corresponding wind pressure data label, the graph structure information is used to characterize the spatial position relationship of the multiple pressure measuring points, and the target prediction model is used to predict the wind pressure data of the failure pressure measuring point; the training samples include experimental training samples and simulation training samples, and the training set is used to train the preset hybrid model Training is performed until the first loss function value of the preset hybrid model meets the training stop condition to obtain an initial prediction model; a validation set is obtained; the validation set includes multiple training samples; based on a Gaussian process regression model, the validation set is used to evaluate the model performance of the initial prediction model to adjust the network parameters of the initial prediction model to obtain the target prediction model; wherein, the Gaussian process regression model represents: high-fidelity data is a linear transformation of low-fidelity data plus a preset deviation function, and the low-fidelity data and the deviation function obey a multivariate Gaussian process, the mean of the multivariate Gaussian process is 0, and the covariance matrix of the multivariate Gaussian process is composed of the covariance function of the low-fidelity data and the covariance function of the deviation function, the high-fidelity data is the second predicted wind pressure data corresponding to the experimental training sample, and the low-fidelity data is the second predicted wind pressure data corresponding to the simulation training sample.

7. A train surface wind pressure prediction device, characterized in that: The device comprises: A second acquisition module is used to obtain target influencing parameters of the wind pressure on the train surface of the target train, target wind pressure data of each effective pressure measuring point, and target graph structure information of multiple pressure measuring points; A prediction module is used to input the target influencing parameters, the target wind pressure data and the target graph structure information into a target prediction model to obtain the wind pressure data of the failure pressure measuring point output by the target prediction model; wherein the target prediction model is obtained by the prediction model training method for train surface wind pressure according to any one of claims 1 to 4.

8. An electronic device, characterized in that: The electronic device includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, the prediction model training method for train surface wind pressure according to any one of claims 1 to 4 is implemented; Or implement the method for predicting train surface wind pressure as described in claim 5.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer program instructions, and when the computer program instructions are executed by the processor, the method for training a prediction model of train surface wind pressure according to any one of claims 1 to 4 is implemented; Or implement the method for predicting train surface wind pressure as described in claim 5.

Images