Fault diagnosis method of electric-hydraulic hybrid steer-by-wire system based on transfer learning

Abstract

The present invention discloses a fault diagnosis method for an electric-hydraulic composite wire-controlled steering system based on transfer learning, and the steps are as follows: training a 1DCNN-LSTM neural network combined with an attention mechanism; training an equivalent hydraulic mechanism fault value estimation module based on transfer learning; and performing real-time fault diagnosis on the electric-hydraulic composite wire-controlled steering system based on a 1DCNN-LSTM neural network combined with an attention mechanism. The method of the present invention performs transfer learning based on an electric mechanism fault diagnosis neural network with sufficient test data to obtain an accurate hydraulic mechanism fault diagnosis neural network, thereby ensuring the accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the technical field of wire-controlled steering, and in particular relates to a fault diagnosis method for an electric-hydraulic composite wire-controlled steering system based on transfer learning. Background Art

[0002] With the continuous advancement of unmanned driving technology for heavy vehicles, when unmanned driving technology reaches L4 and L5 levels, the driver will no longer have control over the vehicle. Therefore, the vehicle steering behavior will no longer be controlled by the steering wheel, but by the on-board controller directly controlling the electro-hydraulic composite steering system to drive the front wheel steering. It can be seen that the electro-hydraulic composite wire-controlled steering system is a necessary condition for realizing unmanned driving of heavy vehicles. In order to ensure the driving safety of the vehicle during automatic driving, the electro-hydraulic composite wire-controlled steering system should have automatic fault diagnosis capabilities, so as to detect faults in time and take corresponding measures to ensure driving safety.

[0003] The electric-hydraulic composite wire-controlled steering system used in heavy vehicles has a complex structure and many components, which increases the difficulty of fault diagnosis. Therefore, it is of great significance to effectively detect the fault type and fault degree of the electric-hydraulic composite wire-controlled steering system during vehicle driving. In recent years, more and more researchers have applied deep learning to the fault diagnosis of complex nonlinear systems. Although deep learning can solve the fault diagnosis problem of nonlinear systems, a large amount of data is required to ensure the diagnostic accuracy of the neural network. However, the electric-hydraulic composite wire-controlled steering system in a faulty state cannot guarantee the safety of the vehicle, making it difficult to collect enough data in a faulty state. The electric-hydraulic composite wire-controlled steering system is a system with variable loads and mutual coupling between actuators. During the steering process of heavy vehicles, the load of the hydraulic mechanism is much greater than that of the electric mechanism. If the hydraulic mechanism fails, the steering performance of the vehicle will be greatly reduced or even lost, which increases the difficulty of testing real vehicles. Therefore, it is difficult to collect data under hydraulic mechanism failure, which makes the amount of electric mechanism fault data and hydraulic mechanism fault data unbalanced, resulting in the above-mentioned deep learning being difficult to effectively perform fault diagnosis on the wire-controlled electric-hydraulic composite steering system. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the purpose of the present invention is to provide a method for fault diagnosis of an electric-hydraulic composite wire-controlled steering system based on transfer learning, so as to solve the problem that data collection under hydraulic mechanism failure of heavy-duty vehicles in the prior art is difficult, resulting in an imbalance in the amount of electric mechanism fault data and the amount of hydraulic mechanism fault data, which makes it difficult to effectively perform fault diagnosis on the wire-controlled electric-hydraulic composite steering system using deep learning; the method of the present invention performs transfer learning on an electric mechanism fault diagnosis neural network with sufficient test data to obtain an accurate hydraulic mechanism fault diagnosis neural network, thereby ensuring the accuracy of fault diagnosis.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] A fault diagnosis method for an electric-hydraulic composite wire-controlled steering system based on transfer learning of the present invention comprises the following steps:

[0007] 1) Training the 1DCNN-LSTM neural network combined with the attention mechanism:

[0008] 11) Test data collection: Collect the data of the motor current sensor, nut displacement sensor, hydraulic cylinder flow sensor, steering wheel sensor and the corresponding equivalent fault values ​​when the heavy vehicle is in normal state, electric mechanism fault state with different fault values ​​and hydraulic mechanism fault state, and establish the data set in time sequence;

[0009] It should be noted that: due to the difficulty and high risk of collecting the fault status of the hydraulic mechanism, the number of fault tests of the electric mechanism needs to be more than the number of fault tests of the hydraulic mechanism;

[0010] 12) Establishing a 1DCNN-LSTM neural network combined with an attention mechanism, wherein the 1DCNN-LSTM neural network combined with an attention mechanism includes a fault type recognition module, a 1D-CNN feature filtering module, a steering feature extraction module, an equivalent electric mechanism fault value estimation module, and an equivalent hydraulic mechanism fault value estimation module;

[0011] 13) Based on the electric mechanism fault set (obtained through a large number of real vehicle tests), supervised training is performed on the fault type recognition module, 1D-CNN feature filtering module, steering feature extraction module and equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism;

[0012] 2) Training the equivalent hydraulic mechanism fault value estimation module based on transfer learning:

[0013] 21) Migrate the equivalent electric mechanism fault value estimation module, and fine-tune the hyperparameters of the equivalent electric mechanism fault value estimation module based on the hydraulic mechanism fault data set (obtained through a small number of real vehicle tests), so that the 1DCNN-LSTM neural network can estimate the fault value of the hydraulic mechanism;

[0014] 22) Based on the electric mechanism fault data set and the hydraulic mechanism fault data set, the parameters of the 1D-CNN feature filtering module and the steering feature extraction module are corrected through a dynamic sample allocation strategy based on a joint error function so that the electric mechanism fault estimation results and the hydraulic mechanism fault estimation results can achieve good results;

[0015] 3) Real-time fault diagnosis of electric-hydraulic hybrid steer-by-wire system based on 1DCNN-LSTM neural network combined with attention mechanism:

[0016] 31) The fault type identification module identifies the current fault type according to the motor current, nut displacement and hydraulic cylinder flow collected in real time;

[0017] 32) The 1D-CNN feature filtering module extracts system features based on the motor current, nut displacement and hydraulic cylinder flow collected in real time; the steering feature extraction module extracts the driver's steering trend features based on the steering wheel angle signal collected in real time; the system features are integrated with the driver's steering trend characteristics to obtain a new feature map V including the working status of each part of the steering system and the movement trend of the steering system;

[0018] 33) Based on the current fault type in step 31), the new characteristic map V in step 32) is input into the corresponding equivalent electric mechanism fault value estimation module or the equivalent hydraulic mechanism fault value estimation module to calculate the fault value of the current electric-hydraulic composite wire-controlled steering system.

[0019] Furthermore, the equivalent fault value in step 11) is the loss rate of the output power of the current electric mechanism or hydraulic mechanism under the normal state, as shown in the following formula:

[0020]

[0021] Where τ is the equivalent fault value, W t is the current output power of the mechanism, W nor It is the output power of the mechanism under normal condition.

[0022] Furthermore, the fault type identification module in the step 12) is composed of a deep LSTM, a fully connected layer and a Softmax classifier, which determines whether the current steering system has a fault and the specific fault type based on the motor current signal, nut displacement signal and hydraulic cylinder flow signal collected by the sensor; the 1D-CNN feature filtering module is composed of two layers of one-dimensional convolutional layers and a pooling layer, which filters the motor current, nut displacement and hydraulic cylinder flow signals and extracts system features; the steering feature extraction module is composed of an LSTM neural network and a fully connected layer, which extracts the driver's steering trend feature map based on the steering wheel angle timing signal within 1 second; the equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module are both composed of a scaled attention layer, a deep LSTM and a fully connected layer.

[0023] Furthermore, in step 12), the steering feature extraction module combines the driver steering trend feature map with the feature map Z output by the 1D-CNN feature filtering module through a position-by-position addition algorithm, thereby generating a new feature map V containing the dynamic features of the steering system; the position-by-position addition algorithm is as follows:

[0024] V=Z+[h 1 ,h 2 ,…,h p ] T (2)

[0025] Among them, Z is the feature map output by the 1D-CNN feature filter module, h p is the corresponding feature in the driver's steering tendency feature map that is added to the feature of the pth row in Z.

[0026] Furthermore, in the step 12), the scaling attention layer in the equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module compresses the channel features by globally averaging the feature values ​​in each channel in the input new feature map V using the squeezing function formula (3); the average value of each channel is activated using the activation function of formula (4), thereby generating the attention matrix W A ; By scaling the attention weights of formula (5), the expressiveness of key features is improved and the expressiveness of invalid features is reduced; the attention matrix W A The new feature map D is obtained by fusing it with the feature map V, as shown in formula (6):

[0027]

[0028]

[0029]

[0030] W A =α·W E (6)

[0031] Where n is the number of eigenvalues ​​in the channel, V j i is the i-th eigenvalue in the j-th channel of the feature map V, q j is the compression value of the jth channel, α is the scaling factor, δ is the ReLu function, and is the weight obtained by training, W A is the attention matrix, W E is the weight matrix, α j is the normalized weight of the jth channel, is the weight calculated by the activation function for the jth channel, and σ is the sigmoid function.

[0032] Furthermore, the step 13) specifically includes: based on the electric mechanism fault data set, using the cross entropy function as the loss function and adopting the error back propagation method to perform supervised training on the 1DCNN-LSTM neural network combined with the attention mechanism until the estimated equivalent motor fault value meets the accuracy requirement.

[0033] Furthermore, the step 21) specifically includes: fine-tuning the hyperparameters of the scaled attention layer, deep LSTM and fully connected layer in the equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism trained in step 1) based on the hydraulic mechanism fault data set, so as to improve the feature expressiveness that is beneficial to the hydraulic mechanism estimation accuracy and reduce the expressiveness of meaningless features.

[0034] Furthermore, the dynamic sample allocation strategy in step 22) is specifically as follows: based on the joint loss function of formula (7), the proportion of the number of samples of different fault types in the sample set used for training is continuously adjusted as shown in formula (8), and the 1D-CNN feature filtering module and the steering feature extraction module are trained twice; the number of samples of the fault type with large error is increased in the sample set of the next iteration to narrow the gap in estimation accuracy between different types;

[0035]

[0036]

[0037] Where E is the average error of the sample, K is the number of samples in a mini-batch, L is the number of motor fault samples in a mini-batch, and J is * is the error in estimating the fault degree of the hydraulic mechanism, and J is the error in estimating the fault degree of the electric mechanism.

[0038] Furthermore, in the step 31), the fault type identification module classifies the fault types into normal, electric mechanism fault and hydraulic mechanism fault based on the motor current signal, the nut displacement signal and the hydraulic cylinder flow signal.

[0039] Furthermore, in step 32), the steering feature extraction module updates and stores the steering wheel angle signal with a time domain length of 1 second in real time, and extracts the driver's steering trend feature based on the steering wheel angle signal within the 1 second time domain.

[0040] Furthermore, in the step 33), the new characteristic graph V is input into the corresponding equivalent fault value estimation module according to the recognition result of the fault type recognition module and the fault type and fault value are finally calculated.

[0041] Beneficial effects of the present invention:

[0042] The present invention uses a 1DCNN-LSTM neural network based on transfer learning combined with an attention mechanism to identify fault types and estimate fault values ​​for an electric-hydraulic composite steer-by-wire system, and can accurately estimate the fault value of a hydraulic mechanism when there are few hydraulic mechanism fault data sets. Specifically, the following are performed:

[0043] 1. The 1DCNN-LSTM neural network combined with the attention mechanism fuses the long-term driver operation signal with the short-term vehicle signal so that the feature map contains more steering motion trend information, thereby improving the real-time performance and accuracy of the fault diagnosis network.

[0044] 2. A scaling attention layer is added to the 1DCNN-LSTM neural network combined with the attention mechanism to amplify the features in the feature map that are beneficial to the diagnosis accuracy and suppress the interference items in the feature map, thereby improving the accuracy and robustness of the fault diagnosis network.

[0045] 3. Hydraulic mechanism fault tests are expensive and dangerous, resulting in a small number of hydraulic mechanism fault data sets. The present invention uses transfer learning and a dynamic sample allocation strategy to obtain a high estimation accuracy even when there are fewer hydraulic mechanism fault data sets. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 Schematic diagram of the 1DCNN-LSTM neural network combined with the attention mechanism in the present invention.

[0047] Figure 2 This is a schematic diagram of the training of the equivalent hydraulic mechanism fault value estimation module based on transfer learning in the present invention. DETAILED DESCRIPTION

[0048] In order to facilitate the understanding of those skilled in the art, the present invention is further described below in conjunction with embodiments and drawings. The contents mentioned in the implementation modes are not intended to limit the present invention.

[0049] Reference Figure 1 , Figure 2 As shown, a fault diagnosis method of an electric-hydraulic composite wire-controlled steering system based on transfer learning of the present invention comprises the following steps:

[0050] 1) Training the 1DCNN-LSTM neural network combined with the attention mechanism:

[0051] 11) Test data collection: Collect the data of the motor current sensor, nut displacement sensor, hydraulic cylinder flow sensor, steering wheel sensor and the corresponding equivalent fault values ​​when the heavy vehicle is in normal state, electric mechanism fault state with different fault values ​​and hydraulic mechanism fault state, and establish the data set in time sequence;

[0052] It should be noted that: due to the difficulty and high risk of collecting the fault status of the hydraulic mechanism, the number of fault tests of the electric mechanism needs to be more than the number of fault tests of the hydraulic mechanism;

[0053] 12) Establishing a 1DCNN-LSTM neural network combined with an attention mechanism, wherein the 1DCNN-LSTM neural network combined with an attention mechanism includes a fault type recognition module, a 1D-CNN feature filtering module, a steering feature extraction module, an equivalent electric mechanism fault value estimation module, and an equivalent hydraulic mechanism fault value estimation module;

[0054] 13) Based on the electric mechanism fault set (obtained through a large number of real vehicle tests), supervised training is performed on the fault type recognition module, 1D-CNN feature filtering module, steering feature extraction module and equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism.

[0055] Specifically, the equivalent fault value in step 11) is the loss rate of the output power of the current electric mechanism or hydraulic mechanism under the normal state, as shown in the following formula:

[0056]

[0057] Where τ is the equivalent fault value, W t is the current output power of the mechanism, W nor It is the output power of the mechanism under normal condition.

[0058] Specifically, the fault type identification module in the step 12) is composed of a deep LSTM, a fully connected layer and a Softmax classifier, which determines whether the current steering system has a fault and the specific fault type based on the motor current signal, nut displacement signal and hydraulic cylinder flow signal collected by the sensor; the 1D-CNN feature filtering module is composed of two layers of one-dimensional convolutional layers and a pooling layer, which filters the motor current, nut displacement and hydraulic cylinder flow signals and extracts system features; the steering feature extraction module is composed of an LSTM neural network and a fully connected layer, which extracts the driver's steering trend feature map based on the steering wheel angle timing signal within 1 second; the equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module are both composed of a scaled attention layer, a deep LSTM and a fully connected layer.

[0059] Specifically, in step 12), the steering feature extraction module combines the driver's steering trend feature map with the feature map Z output by the 1D-CNN feature filtering module through a position-by-position addition algorithm, thereby generating a new feature map V containing the dynamic features of the steering system; the position-by-position addition algorithm is as follows:

[0060] V=Z+[h 1 ,h 2,…,h p ] T (2)

[0061] Among them, Z is the feature map output by the 1D-CNN feature filter module, h p is the corresponding feature in the driver's steering tendency feature map that is added to the feature of the pth row in Z.

[0062] Specifically, the scaling attention layer in the equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module in step 12) compresses the channel features by globally averaging the feature values ​​in each channel of the input new feature map V through the squeezing function formula (3); the average value of each channel is activated by the activation function of formula (4), thereby generating the attention matrix W A ; By scaling the attention weights of formula (5), the expressiveness of key features is improved and the expressiveness of invalid features is reduced; the attention matrix W A The new feature map D is obtained by fusing it with the feature map V, as shown in formula (6):

[0063]

[0064]

[0065]

[0066] W A =α·W E (6)

[0067] Where n is the number of eigenvalues ​​in the channel, V j i is the i-th eigenvalue in the j-th channel of the feature map V, q j is the compression value of the jth channel, α is the scaling factor, δ is the ReLu function, and is the weight obtained by training, W A is the attention matrix, W E is the weight matrix, α j is the normalized weight of the jth channel, is the weight calculated by the activation function for the jth channel, and σ is the sigmoid function.

[0068] Specifically, the step 13) specifically includes: based on the electric mechanism fault data set, using the cross entropy function as the loss function and adopting the error back propagation method to perform supervised training on the 1DCNN-LSTM neural network combined with the attention mechanism until the estimated equivalent motor fault value meets the accuracy requirement.

[0069] 2) Training the equivalent hydraulic mechanism fault value estimation module based on transfer learning:

[0070] 21) Migrate the equivalent electric mechanism fault value estimation module, and fine-tune the hyperparameters of the equivalent electric mechanism fault value estimation module based on the hydraulic mechanism fault data set (obtained through a small number of real vehicle tests), so that the 1DCNN-LSTM neural network can estimate the fault value of the hydraulic mechanism;

[0071] 22) Based on the electric mechanism fault data set and the hydraulic mechanism fault data set, the parameters of the 1D-CNN feature filtering module and the steering feature extraction module are corrected through a dynamic sample allocation strategy based on the joint error function so that both the electric mechanism fault estimation results and the hydraulic mechanism fault estimation results can achieve good results.

[0072] Specifically, the step 21) specifically includes: fine-tuning the hyperparameters of the scaled attention layer, deep LSTM and fully connected layer in the equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism trained in step 1) based on the hydraulic mechanism fault data set, so as to improve the feature expressiveness that is beneficial to the hydraulic mechanism estimation accuracy and reduce the expressiveness of meaningless features.

[0073] Specifically, the dynamic sample allocation strategy in step 22) is as follows: based on the joint loss function of formula (7), the proportion of the number of samples of different fault types in the sample set used for training is continuously adjusted as shown in formula (8), and the 1D-CNN feature filtering module and the steering feature extraction module are trained twice; the number of samples of the fault type with large error is increased in the sample set of the next iteration to narrow the gap in estimation accuracy between different types;

[0074]

[0075]

[0076] Where E is the average error of the sample, K is the number of samples in a mini-batch, L is the number of motor fault samples in a mini-batch, and J is * is the error in estimating the fault degree of the hydraulic mechanism, and J is the error in estimating the fault degree of the electric mechanism.

[0077] 3) Real-time fault diagnosis of electric-hydraulic hybrid steer-by-wire system based on 1DCNN-LSTM neural network combined with attention mechanism:

[0078] 31) The fault type identification module identifies the current fault type according to the motor current, nut displacement and hydraulic cylinder flow collected in real time;

[0079] 32) The 1D-CNN feature filtering module extracts system features based on the motor current, nut displacement and hydraulic cylinder flow collected in real time; the steering feature extraction module extracts the driver's steering trend features based on the steering wheel angle signal collected in real time; the system features are integrated with the driver's steering trend characteristics to obtain a new feature map V including the working status of each part of the steering system and the movement trend of the steering system;

[0080] 33) Based on the current fault type in step 31), the new characteristic map V in step 32) is input into the corresponding equivalent electric mechanism fault value estimation module or the equivalent hydraulic mechanism fault value estimation module to calculate the fault value of the current electric-hydraulic composite wire-controlled steering system.

[0081] In the step 31), the fault type identification module classifies the fault types into normal, electric mechanism fault and hydraulic mechanism fault based on the motor current signal, the nut displacement signal and the hydraulic cylinder flow signal.

[0082] In the step 32), the steering feature extraction module updates and stores the steering wheel angle signal with a time domain length of 1 second in real time, and extracts the driver's steering trend feature based on the steering wheel angle signal within the 1 second time domain.

[0083] In the step 33), the new characteristic graph V is input into the corresponding equivalent fault value estimation module according to the recognition result of the fault type recognition module, and the fault type and fault value are finally calculated.

[0084] The present invention has many specific application paths. The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements can be made without departing from the principle of the present invention. These improvements should also be regarded as the protection scope of the present invention.

Claims

1. A fault diagnosis method for electric-hydraulic hybrid steer-by-wire system based on transfer learning, It is characterized in that Here are the steps: 1) Training the 1DCNN-LSTM neural network combined with the attention mechanism: 11) Test data collection: Collect the data of the motor current sensor, nut displacement sensor, hydraulic cylinder flow sensor, steering wheel sensor and the corresponding equivalent fault values ​​when the heavy vehicle is in normal state, electric mechanism fault state with different fault values ​​and hydraulic mechanism fault state, and establish the data set in time sequence; 12) Establishing a 1DCNN-LSTM neural network combined with an attention mechanism, wherein the 1DCNN-LSTM neural network combined with an attention mechanism includes a fault type recognition module, a 1D-CNN feature filtering module, a steering feature extraction module, an equivalent electric mechanism fault value estimation module, and an equivalent hydraulic mechanism fault value estimation module; 13) Based on the electric mechanism fault set, supervised training is performed on the fault type recognition module, 1D-CNN feature filtering module, steering feature extraction module and equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism; 2) Training the equivalent hydraulic mechanism fault value estimation module based on transfer learning: 21) Migrate the equivalent electric mechanism fault value estimation module, and use fine-tuning based on the hydraulic mechanism fault data set to correct the hyperparameters of the equivalent electric mechanism fault value estimation module, so that the 1DCNN-LSTM neural network estimates the fault value of the hydraulic mechanism; 22) Based on the electric mechanism fault data set and the hydraulic mechanism fault data set, the parameters of the 1D-CNN feature filtering module and the steering feature extraction module are corrected through a dynamic sample allocation strategy based on a joint error function so that the electric mechanism fault estimation results and the hydraulic mechanism fault estimation results can achieve good results; 3) Real-time fault diagnosis of electric-hydraulic hybrid steer-by-wire system based on 1DCNN-LSTM neural network combined with attention mechanism: 31) The fault type identification module identifies the current fault type according to the motor current, nut displacement and hydraulic cylinder flow collected in real time; 32) The 1D-CNN feature filtering module extracts system features based on the motor current, nut displacement and hydraulic cylinder flow collected in real time; Steering The feature extraction module extracts the driver's steering trend feature based on the steering wheel angle signal collected in real time; The system characteristics are integrated with the driver's steering tendency characteristics to obtain a new characteristic map V including the working status of each part of the steering system and the motion trend of the steering system; 33) Based on the current fault type in step 31), the new characteristic map V in step 32) is input into the corresponding equivalent electric mechanism fault value estimation module or the equivalent hydraulic mechanism fault value estimation module to calculate the fault value of the current electric-hydraulic composite wire-controlled steering system.

2. The fault diagnosis method of the electric-hydraulic composite wire-controlled steering system based on transfer learning according to claim 1, It is characterized in that The equivalent fault value in step 11) is the output power loss rate of the electric mechanism or hydraulic mechanism under normal conditions, as shown in the following formula: Where τ is the equivalent fault value, W t is the current output power of the mechanism, W nor It is the output power of the mechanism under normal condition.

3. The fault diagnosis method of the electric-hydraulic composite wire-controlled steering system based on transfer learning according to claim 1, It is characterized in that In the step 12), the fault type identification module is composed of a deep LSTM, a fully connected layer and a Softmax classifier, which determines whether the current steering system has a fault and the specific fault type according to the motor current signal, nut displacement signal and hydraulic cylinder flow signal collected by the sensor; the 1D-CNN feature filtering module is composed of two layers of one-dimensional convolutional layers and a pooling layer, which filters the motor current, nut displacement and hydraulic cylinder flow signals and extracts system features; Steering The feature extraction module consists of an LSTM neural network and a fully connected layer, which extracts the driver's steering trend feature map based on the steering wheel angle timing signal within 1 second; The equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module are both composed of a scaled attention layer, a deep LSTM and a fully connected layer.

4. The fault diagnosis method of the electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that In step 12), the steering feature extraction module combines the driver's steering trend feature map with the feature map Z output by the 1D-CNN feature filtering module through a position-by-position addition algorithm, thereby generating a new feature map V containing the dynamic features of the steering system; the position-by-position addition algorithm is as follows: V=Z+[h 1 ,h 2 ,…,h p ] T (2) Among them, Z is the feature map output by the 1D-CNN feature filter module, h p is the corresponding feature in the driver's steering tendency feature map that is added to the feature of the pth row in Z.

5. The fault diagnosis method for an electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 4, It is characterized in that In the step 12), the scaling attention layer in the equivalent electric mechanism fault value estimation module and the equivalent hydraulic mechanism fault value estimation module compresses the channel features by globally averaging the feature values ​​in each channel in the input new feature map V through the squeezing function (3); The activation function of formula (4) is used to activate the average value of each channel to generate the attention matrix W A ; The attention weight scaling of formula (5) improves the expressiveness of key features and reduces the expressiveness of invalid features; the attention matrix W A The new feature map D is obtained by fusing it with the feature map V, as shown in formula (6): W A =α·W E (6) Where n is the number of eigenvalues ​​in the channel, V j i is the i-th eigenvalue in the j-th channel of the feature map V, q j is the compression value of the jth channel, α is the scaling factor, δ is the ReLu function, and is the weight obtained by training, W A is the attention matrix, W E is the weight matrix, α j is the normalized weight of the jth channel, is the weight calculated by the activation function for the jth channel, and σ is the sigmoid function.

6. The fault diagnosis method of the electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that The step 13) specifically includes: based on the electric mechanism fault data set, using the cross entropy function as the loss function and adopting the error back propagation method to perform supervised training on the 1DCNN-LSTM neural network combined with the attention mechanism until the estimated equivalent motor fault value meets the accuracy requirement.

7. The fault diagnosis method of the electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that The step 21) specifically includes: fine-tuning the hyperparameters of the scaled attention layer, deep LSTM and fully connected layer in the equivalent electric mechanism fault value estimation module in the 1DCNN-LSTM neural network combined with the attention mechanism trained in step 1) based on the hydraulic mechanism fault data set, so as to improve the feature expressiveness that is beneficial to the hydraulic mechanism estimation accuracy and reduce the expressiveness of meaningless features.

8. The fault diagnosis method for an electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that The dynamic sample allocation strategy in step 22) is specifically as follows: based on the joint loss function of formula (7), the proportion of the number of samples of different fault types in the sample set used for training is continuously adjusted as shown in formula (8), and the 1D-CNN feature filtering module and the steering feature extraction module are trained twice; the number of samples of the fault type with large error is increased in the sample set of the next iteration to narrow the gap in estimation accuracy between different types; Where E is the average error of the sample, K is the number of samples in a mini-batch, L is the number of motor fault samples in a mini-batch, and J is * is the error in estimating the fault degree of the hydraulic mechanism, and J is the error in estimating the fault degree of the electric mechanism.

9. The fault diagnosis method for an electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that In the step 32), the steering feature extraction module updates and stores the steering wheel angle signal with a time domain length of 1 second in real time, and extracts the driver's steering trend feature based on the steering wheel angle signal within the 1 second time domain.

10. The fault diagnosis method of the electric-hydraulic composite steer-by-wire system based on transfer learning according to claim 1, It is characterized in that In the step 33), the new characteristic graph V is input into the corresponding equivalent fault value estimation module according to the recognition result of the fault type recognition module, and the fault type and fault value are finally calculated.

Images