Transfer learning and domain adaptation method for fault diagnosis of automobile steer-by-wire system

Abstract

The invention discloses a transfer learning and domain adaptation method for fault diagnosis of an automobile steer-by-wire system, relates to the technical field of fault diagnosis and transfer learning, and aims to realize efficient transfer learning between a simulation domain and a physical domain. The method comprises the following steps: constructing a transfer learning model, and realizing knowledge transfer between a simulation domain and a real object domain; an improved transfer learning algorithm is adopted to solve the performance problem caused by the distribution difference between a simulation domain and a real object domain; defining a new domain generalization index, and quantifying the adaptability difference between the simulation domain and the physical domain; the method is applied to the steer-by-wire system, efficient migration of simulation data and physical data is achieved, and fault diagnosis is carried out. By means of the method, efficient migration between simulation data and physical data can be achieved, and a more reliable and flexible solution is provided for fault diagnosis of the steer-by-wire system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis and transfer learning, and in particular to a transfer learning and domain adaptation method for fault diagnosis of a vehicle steer-by-wire system. BACKGROUND

[0002] With the rapid development of intelligent vehicles and steer-by-wire chassis technology, steer-by-wire systems, as a new technology with broad application prospects, have received more and more attention in recent years. The system cancels the mechanical connection between the steering wheel and the front wheels in the traditional mechanical steering system, and instead uses information exchange between motors, sensors and electronic controllers to achieve steering control. However, due to the high electronicization and complexity of the structure of the steer-by-wire system, it is extremely susceptible to factors such as component aging, external interference and control strategy mismatch during actual operation, thereby causing various types of system faults. In order to ensure the safety and reliability of the vehicle, efficient and accurate fault diagnosis of the steer-by-wire system has become an important direction of current research.

[0003] However, in engineering practice, due to factors such as equipment status, cost and time, it is often difficult to obtain enough labeled fault samples on actual vehicle platforms, while transfer learning technology has shown good adaptability and practicality in the field of industrial intelligent diagnosis, and can realize knowledge transfer between high-fidelity simulation platforms and actual vehicle data collection. Therefore, there is an urgent need for a transfer learning algorithm with high robustness and generalization ability, which can model and align the feature distribution differences between the simulation domain and the real domain, and improve the adaptation performance of the model in the target domain. At the same time, a quantitative generalization ability evaluation index is established to measure and feedback the cross-domain performance of the model, so as to guide the deployment and optimization of the fault diagnosis model on the steer-by-wire platform. However, in the existing literature, there is still a lack of research on transfer learning and domain adaptation methods for fault diagnosis of steer-by-wire platforms. SUMMARY

[0004] In order to overcome the defects in the prior art, the present application provides a transfer learning and domain adaptation method for fault diagnosis of a vehicle steer-by-wire system, which realizes knowledge transfer between the simulation domain and the real domain, reduces the dependence on actual fault data (real data), and reduces the cost.

[0005] To achieve the above purpose, the present application adopts the following technical solution, comprising: A transfer learning and domain adaptation method for fault diagnosis of a vehicle steer-by-wire system, comprising the following steps: S1, constructing a transfer learning model to realize knowledge transfer between the simulation domain (source domain) and the real domain (target domain); S2, for the transfer learning model, using an improved transfer learning algorithm to solve the performance problems caused by the distribution difference between the simulation domain and the real domain; S3, for the transfer learning model, define a new domain generalization index, combine the KL divergence with the joint distribution adaptation, to further optimize the distribution alignment between the simulation domain and the real domain, and quantify the adaptive difference between the simulation domain and the real domain; S4, based on the improved transfer learning algorithm and the new domain generalization index, train the transfer learning model to realize the efficient transfer of simulation data and real data.

[0006] Preferably, in step S1, the transfer learning model selects a convolutional neural network as the model architecture, which is used to extract the features of the input data and map the features of the simulation data to the real data through multi-level feature mapping, thereby ensuring the generalization ability of the transfer learning model; the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer.

[0007] Preferably, in step S2, a new I-Softmax loss The transfer learning algorithm is improved to learn more separable features and improve the score in the transfer task, which is defined as follows:

[0008] In the formula, is the feature vector of the i-th sample is the probability distribution of and respectively represent the i-th element and other elements in , where is the score of the i-th sample for class c, is the score of the i-th sample for class j; is the feature vector of the i-th sample, i.e. the output of the input data through the convolutional neural network; n represents the number of feature vectors, i.e. the number of samples; m and k are hyperparameters that control the decision boundary.

[0009] Preferably, in step S2, the improved transfer learning algorithm uses soft labels to retain the predicted probability distribution.

[0010] Preferably, in step S3, the new domain generalization index is calculated as follows: The maximum mean difference MMD is combined with the covariance alignment CORAL to obtain the loss of the improved edge distribution alignment IMDA :

[0011] where the hyperparameter is the covariance weight coefficient; ​The improved joint distribution adaptive loss (IJDA) is obtained by comparing the magnitudes of the improved marginal distribution alignment (IMDA) and conditional distribution alignment (CDA) using KL divergence comparison, adjusting their weights, and then combining them with a weighted average. As a new domain generalization metric to address the problem of data distribution differences between the source and target domains, it is as follows:

[0012]

[0013]

[0014]

[0015] In the formula, The KL divergence of CDA is aligned with the conditional distribution; C is the number of classes; This represents the prior probability of class c in the source domain; This indicates the source domain data under category c. Distribution; This represents the expectation of the source domain data, i.e., the average feature of category c in the source domain samples; This represents the prior probability of class c in the target domain; This indicates the target domain data under category c. Distribution; This represents the expected value of the target domain data, i.e., the average feature of category c in the target domain samples; Align the KL divergence of IMDA with the improved edge distribution; and Let be the covariance matrices of the source and target domains, respectively. For each dimension of the sample data, Represents the kernel function. The number of samples in the source domain. Let ||·|| represent the number of samples in the target domain, and ||·|| denotes the norm to be calculated. To improve the loss of IMDA for edge distribution alignment, The loss of CDA is aligned with the conditional distribution.

[0016] Preferably, step S4 is as follows: Adapting the improved joint distribution to loss With I-Softmax loss The sum is used as the total loss of the model. The backpropagation algorithm is used to update the network weights and bias parameters of each layer of the model. The model is updated until the maximum number of iterations is reached, and the trained transfer learning model is obtained.

[0017] Preferably, the method is applied to a steer-by-wire system, and the trained transfer learning model is used as a fault diagnosis model for the steer-by-wire system.

[0018] The application further provides a readable storage medium, which stores a computer program, and the computer program is executed to implement the transfer learning and domain adaptation method for fault diagnosis of a steer-by-wire system of an automobile.

[0019] The application further provides an electronic device, which comprises a processor, a memory, and a computer program stored on the memory and executable on the processor, and the processor executes the computer program to implement the transfer learning and domain adaptation method for fault diagnosis of a steer-by-wire system of an automobile.

[0020] The application further provides a computer program product, which comprises computer programs / instructions, and the computer programs / instructions are executed by a processor to implement the transfer learning and domain adaptation method for fault diagnosis of a steer-by-wire system of an automobile.

[0021] The application has the following advantages: (1) The application realizes knowledge transfer between a simulation domain and a real domain by constructing a transfer learning model, reduces dependence on actual fault data (real data), reduces costs, and is used for fault diagnosis problems in the case of insufficient actual fault data.

[0022] (2) The application adopts an improved transfer learning algorithm, solves performance problems caused by distribution differences between source domains and target domains, and provides theoretical feasibility for fault diagnosis of a steer-by-wire system of an automobile.

[0023] (3) The application defines an improved joint distribution adaptation, combines KL divergence and a JDA method to further optimize distribution alignment between source domains and target domains, quantifies adaptability differences between source domains and target domains, and provides convenience for fault diagnosis of a steer-by-wire platform. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The figure is a flowchart of the method of the application.

[0025] Figure 2 The figure is a framework diagram of the transfer learning and domain adaptation method of the application.

[0026] Figure 3 The figure is a method diagram of the transfer learning of the steer-by-wire system of the application.

[0027] Figure 4 The figure is four kinds of disturbances added to the simulation data of the steer-by-wire system of the embodiment of the application.

[0028] Figure 5Perturbed simulation data was added to the steer-by-wire system of this embodiment. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Example 1

[0031] Depend on Figure 1 and Figure 2 As shown, a transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems includes the following steps: S1, Construct a transfer learning model to realize knowledge transfer between the source domain (simulation domain) and the target domain (physical domain); S2, for the transfer learning model, adopts an improved transfer learning algorithm to solve the performance problem caused by the distribution difference between the source domain and the target domain; S3 defines a new domain generalization metric for transfer learning models, namely Improved Joint Distribution Adaptation (IJDA), which combines Kullback-Leibler Divergence (KL) with the JDA method to further optimize the distribution alignment between the source and target domains and quantify the adaptive differences between the source and target domains. S4, based on an improved transfer learning algorithm and a new domain generalization metric, trains the transfer learning model to achieve efficient transfer between simulation data (source domain data) and physical data (target domain data); S5. This method is applied to the steer-by-wire system, where the trained transfer learning model is used as a fault diagnosis model and applied to the steer-by-wire system for fault diagnosis.

[0032] In this embodiment, the steer-by-wire system comprises two subsystems: a steering wheel system and a front wheel subsystem. The steering wheel system consists of a steering wheel, a feedback motor, a reducer, a feedback motor driver, a steering wheel angle sensor, and a feedback motor angle sensor. The front wheel subsystem consists of a steering motor driver, a steering motor, a reducer, a rack and pinion mechanism, front wheels, a steering motor angle sensor, a gear angle sensor, and a front wheel angle sensor. In step S1, the core objective of constructing the transfer learning model is to improve the fault diagnosis capability of the target domain through knowledge transfer between the source and target domains. The essence of transfer learning lies in reducing dependence on target domain data, especially when target domain data is scarce, by transferring knowledge from the source domain to improve the generalization ability of the target domain model. In the fault diagnosis task, the transfer learning model helps the target domain learn how to identify fault patterns using knowledge from the source domain, thereby effectively improving diagnostic accuracy.

[0033] The data output from the simulation model of the steer-by-wire platform is used as the source domain (simulation) data, and the actual fault data is used as the target domain (physical) data. The different data distributions in the source and target domains affect the model's performance. Various methods can be used to reduce the distribution differences between the source and target domains, enabling the knowledge learned in the source domain to be effectively applied in the target domain. In this invention, a Convolutional Neural Network (CNN) is selected as the model architecture. CNNs, through their multiple convolutional layers, can automatically learn spatial hierarchical features in data, exhibiting excellent feature extraction capabilities for complex data such as images and time series. CNNs can not only capture important features in the source domain but also map these features to the target domain through multi-level feature mapping, thereby ensuring the model's generalization ability.

[0034] Step S1 is as follows: S11, preprocess the dataset of the steer-by-wire platform.

[0035] S111, the dataset consists of one-dimensional data collected by an angle sensor. The data is preprocessed, firstly by min-max normalization, normalizing the data to the range [0,1] or [-1,1].

[0036] In the formula, These are the original data values. It is the minimum value in the dataset. It is the maximum value in the dataset. These are normalized data values ​​to ensure that the input data have a similar scale.

[0037] S112, Sliding sampling is a common technique used to convert 1D time series data into a data format suitable for 2D or 3D input so that CNNs can process it. Sliding sampling requires setting the window size. and step length Use a sliding window to traverse the entire data sequence to generate For each sample, refer to the following formula:

[0038] In the formula, Dataset represents the length of the original one-dimensional dataset.

[0039] S113, after sliding sampling, each sample is a 1D array. To adapt to the input requirements of CNN, it is converted to 2D, and each sample is regarded as 1 channel with a height of 1 and a width of window size w, that is, the shape is (1, w).

[0040] S12. Based on the type of the steer-by-wire platform dataset, a CNN architecture needs to be built. To extract the feature values ​​from the data, a CNN architecture suitable for processing data after sliding sampling needs to be designed. The complete CNN architecture includes an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. S121, the input layer receives the dataset and transforms it into a tensor format that the model can process. The dataset structure for the steer-by-wire platform is relatively simple, with input tensors... ,in For data height, The data width is 1, and the number of channels is 1.

[0041] S122, the convolutional layer needs to select an appropriate convolutional kernel and perform local sliding multiplication with the input feature map (tensor) to extract spatial features. The kernel size needs to be set accordingly. The step size, padding, and the number of input channels C and the number of output channels M are as follows:

[0042] Where u is the u-th dimension of the convolution kernel, c represents the channel number of the input data, m represents the channel number of the convolution output, i and j are the spatial coordinates (row i, column j) of the convolution output feature, and i+u is the position of the convolution window on the input data. These are the weight parameters of the convolution kernel. This is the region of the input data where convolution operations will be performed. It is a bias term. It is the output after the convolution operation.

[0043] Choosing a suitable activation function introduces nonlinearity; commonly used activation functions include ReLU and Sigmoid.

[0044] S123, Adding a pooling layer after the convolutional layer can directly reduce the data dimensionality, thereby reducing the computational cost. The pooling layer generally uses max pooling, selecting the maximum value within the pooling window to reduce the size of the feature map. The formula for max pooling is as follows:

[0045] In the formula, S is the pooling step size. This represents the value of the i-th row, j-th column, and c-th channel of the output feature map of the convolutional layer. Represents the first... line, number The column represents the value of the c-th channel. Max pooling, by selecting the maximum value within a window, can highlight salient features while reducing the spatial dimension of the feature map, and it also exhibits some degree of translation invariance.

[0046] S124: After the data passes through the pooling layer, it is still a high-dimensional feature map. Then, the high-dimensional feature map is flattened into a one-dimensional vector, i.e., a feature vector, through a fully connected layer.

[0047] S125, then the output layer maps the feature vectors to the final output for classification or regression tasks. Classification tasks often use the Softmax function to convert the feature vectors into class probabilities, as shown in the following formula:

[0048] Where y represents the final classification output, W is the model's weight matrix, x is the input feature vector, and b is the bias. Using this formula, the model maps the feature vector x to y and performs the classification task. It is an activation function; the Softmax function is often used in classification tasks.

[0049] Finally, the transfer learning model was constructed based on the data set of the steering-by-wire platform.

[0050] Step S2 is detailed below: Although a transfer learning model was constructed, the feature distributions between the simulation domain and the real domain are not entirely the same, thus placing high demands on the model. In multi-class classification tasks for fault diagnosis, the Softmax function is widely used in neural networks due to its probabilistic interpretation and simplicity. However, in some cases, it still cannot meet the requirements of intra-class compactness and inter-class separability.

[0051] Therefore, this invention designs a novel I-Softmax loss to learn more separable features and improve scores in transfer tasks, defined as follows:

[0052] In the formula, The feature vector of the i-th sample The probability distribution; and They represent The first in One element and other elements, among which Let be the score of the i-th sample for category c. Let be the score of the i-th sample for category j; Let be the feature vector of the i-th sample, i.e., the output of the input data after passing through the CNN; n represents the number of feature vectors, i.e., the number of samples; m and k are hyperparameters controlling the decision boundary, where m is the parameter that amplifies the decision boundary. The larger m is, the larger the margin between categories, and the more separated the features; k is the offset, which helps the model adjust how to handle the magnitude of feature values. By using k, the influence of feature values ​​can be reduced, avoiding excessively extreme feature values ​​of certain categories when calculating the loss. If m=1 and k=0, then the I-SoftMax loss is... This will equal the original Softmax loss.

[0053] Unlike the original SoftMax loss, the I-Softmax loss can separate and compress learned features. This is more helpful for achieving higher accuracy than the original SoftMax loss on multi-class classification tasks.

[0054] Furthermore, in traditional semi-supervised learning, pseudo-labels are used to forcibly assign one-hot encoded values ​​as supervision signals to the model's predictions of unlabeled data, such as [0,0,0,1][0,0,0,1]. However, this discretization method introduces noise and leads to abrupt alignment of conditional distributions. Soft labels, on the other hand, preserve the model's predicted probability distribution, such as [0.1,0.1,0.1,0.7] and [0.1,0.1,0.1,0.7], achieving smoother alignment of conditional distributions and thus improving the model's generalization ability and robustness.

[0055] Step S3 is as follows: In cross-domain scenarios, feature distribution shifts can lead to a decrease in generalization performance. Therefore, a new domain generalization metric is defined. By adding a dynamic weight mechanism to the traditional Joint Distribution Adaptation (JDA), an improved Joint Distribution Adaptation (IJDA) is obtained. As a new domain generalization metric to solve the problem of data distribution differences between the source and target domains, it can effectively guide the direction of domain generalization model training.

[0056] Maximum Mean Discrepancy (MMD) is a metric used to compare the distributions of two domains mapped onto a high-dimensional space. It is primarily used to align marginal distributions between domains. By minimizing MMD, the model's performance across different domains can be made more similar. In the calculation of MMD, the mapping function is called the kernel function, and a fixed bandwidth is used. This may lead to poor results in kernel matrix calculations; therefore, a dynamic adjustment using a Gaussian mixture kernel function can be employed. This ensures that the data is smoothly mapped to the regenerating kernel Hilbert space, improving the model's generalization ability. See the following formula for reference: ,

[0057] In the formula, It is the kernel function, x and y are two different data points representing the features of the input data, and n is the number of samples. It is the bandwidth of the kernel function, which controls the smoothness of the kernel function.

[0058] Covariance Alignment (CORAL) aligns the feature distributions of a source domain dataset with those of a target domain dataset. Its core idea is to adjust the covariance matrix of the source domain dataset to match that of the target domain dataset, thereby reducing the size and range of feature values ​​learned from the source and target domains during network training. CORAL and MMD together constitute Marginal Distribution Alignment (MDA) to ensure that the overall data feature distributions of the source and target domains are as similar as possible. Further, when calculating MDA, MMD and CORAL are weighted and combined to obtain the loss function of the improved marginal distribution alignment (IMDA). :

[0059] Among them, hyperparameters The covariance weighting coefficient balances the contributions of MMD and CORAL, optimizes and fine-tunes the model's adaptation to the overall distribution, and further reduces distributional discrepancies. Its expression is:

[0060] Where p represents the current training progress, and p∈[0,1].

[0061] Enhanced joint distribution adaptation (IJDA) is used to improve domain adaptation between the source and target domains.

[0062] Conditional Distribution Alignment (CDA) ensures that the data feature distributions of the source and target domains are as similar as possible, given the same class label. To measure the importance of Improved Marginal Distribution Alignment (IMDA) and Conditional Distribution Alignment (CDA) in transfer learning, KL divergence is used to compare the magnitudes of IMDA and CDA. Their weights are adjusted, and they are combined with a weighted average to obtain Improved Joint Distribution Adaptation (IJDA). IJDA serves as a new domain generalization metric for addressing the data distribution differences between the source and target domains. IJDA enhances domain adaptation between the source and target domains, as detailed below:

[0063]

[0064]

[0065]

[0066] In the formula, The KL divergence of CDA is aligned with the conditional distribution; C is the number of classes; This represents the prior probability of class c in the source domain; This indicates the source domain data under category c. Distribution; This represents the expectation of the source domain data, i.e., the average feature of category c in the source domain samples; This represents the prior probability of class c in the target domain; This indicates the target domain data under category c. Distribution; This represents the expected value of the target domain data, i.e., the average feature of category c in the target domain samples; Align the KL divergence of IMDA with the improved edge distribution; and Let be the covariance matrices of the source and target domains, respectively. For each dimension of the sample data, Represents the kernel function. The number of samples in the source domain. Let ||·|| represent the number of samples in the target domain, and ||·|| denotes the norm to be calculated. To improve the loss of IMDA for edge distribution alignment, The loss of CDA is aligned with the conditional distribution.

[0067] Step S4 is as follows: Adapting the improved joint distribution to loss With I-Softmax loss The sum is used as the total loss of the model. The backpropagation algorithm is used to update the weights and bias parameters of each layer of the network, and the model is updated until the maximum number of iterations is reached, resulting in the trained transfer learning model.

[0068] In this embodiment, the Case Western Reserve University (CWRU) dataset and the Southeast University Second Dataset (SEU2) dataset are used as the source domain dataset and the target domain dataset, respectively.

[0069] The sliding sampling method is used to increase the number of fault samples, and then the data is expanded to three dimensions to adapt to the input of the convolutional neural network (CNN) to extract deep features of the data. The CNN uses five 1D convolutional modules, one global average pooling layer and two fully connected layers, where each 1D convolutional module includes a set of convolutional layers, batch normalization layers and max pooling layers.

[0070] The processed dataset is input into the constructed CNN, and I-Softmax loss is applied. Adaptation loss with improved joint distribution Training is conducted to obtain the results of transfer learning.

[0071] Example 2

[0072] To verify the effectiveness of the method of the present invention, the following comparative experiments were conducted: like Figure 3 As shown, a fault scenario was simulated and data was collected on a vehicle steer-by-wire platform (SBW platform, an experimental platform used for collecting actual fault data). Simultaneously, a simulation model was built on Simulink, and transfer learning was performed between the physical and simulation domains. Data was collected from the SBW platform and the simulation model for 40 seconds at a sampling period of 0.001s. The data was processed using Matlab, and fault samples were increased through sliding window sampling. The dataset has a dimension of 3601. The 4000 labels were organized using one-hot encoding. Then, the two datasets were imported into the model as the source and target domains in Python for training and optimization. Finally, the results were output and analyzed.

[0073] Using the same sinusoidal input, the front wheel steering angle output data of the SBW system platform and simulation model were collected respectively. The simulation data included three types of periodic disturbances to simulate mechanical faults, with Gaussian noise added. The platform data included four types of periodic disturbances to simulate faults. The four disturbances were as follows: Figure 4 As shown, the simulation data of the steer-by-wire system before and after adding a disturbance are as follows:Figure 5 As shown.

[0074] The processed dataset is input into the constructed CNN, and I-Softmax loss and an improved joint distribution adaptation loss are used. Training is performed to obtain the transfer learning results. The output results are visualized by plotting the training set accuracy, test set accuracy, training set loss, test set loss, total training loss, target classification loss, source classification loss, conditional distribution alignment loss, and maximum mean difference loss, and then presented in a table.

[0075] In this embodiment, to verify the proposed transfer learning and domain adaptation method (IJDA+I-Softmax) for fault diagnosis of automotive steer-by-wire systems, a comparison is made with the traditional method (JDA+Softmax). The comparison results are shown in Table 1 below: Table 1

[0076] As shown in Table 1, the proposed transfer learning and domain adaptation method (IJDA+I-Softmax) outperforms the traditional method (JDA+Softmax) on multiple metrics. The I-Softmax loss function exhibits better inter-class separability and intra-class compactness in source and target domain classification, resulting in a significant improvement in the accuracy of both the training and test sets. Specifically, the total training loss is significantly reduced, particularly in the loss terms for source and target classification, demonstrating the model's effective extraction and utilization of cross-domain features.

[0077] Furthermore, the CDA and MDA losses are significantly reduced, indicating that the IJDA method optimized by KL divergence effectively alleviates the joint distribution difference between the source and target domains. This algorithm not only improves the model's generalization ability but also enhances its practicality in real-world fault diagnosis scenarios for steer-by-wire systems. Experiments further verify that the proposed method maintains stable and accurate recognition performance even under complex conditions such as high noise and periodic perturbations.

[0078] Therefore, this invention achieves efficient knowledge transfer between the simulation domain and the physical domain, exhibits good robustness and practicality in fault diagnosis tasks, and has significant engineering promotion value and industrial application prospects.

[0079] Example 3

[0080] An electronic device includes a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method of Embodiment 1 described above.

[0081] The electronic device in this application embodiment may be the mobile device itself, or a standalone device independent of it. The standalone device may communicate with the mobile device to receive the collected input signals from it and send the selected target decision behavior to it.

[0082] An electronic device includes one or more processors and memory. The processor may be a central processing unit (CPU) or other processing unit with data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device to perform desired functions. The memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and / or cache memory. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor may execute the program instructions to implement the decision-making behavior and decision-making methods of the various embodiments of this application described above, and / or other desired functions.

[0083] Electronic devices may also include input devices and output devices.

[0084] Example 4

[0085] In addition to the methods and devices described above, embodiments of this application may also be computer program products, which include computer program instructions that, when executed by a processor, cause the processor to perform the steps in the decision-making behavior method according to various embodiments of this application as described in Embodiment 1 above.

[0086] The computer program product can be written in any combination of one or more programming languages ​​to perform the operations of the embodiments of this application. The programming languages ​​include object-oriented programming languages ​​such as Java and C++, as well as conventional procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0087] Example 5

[0088] Embodiments of this application may also be computer-readable storage media storing computer program instructions thereon, which, when executed by a processor, cause the processor to perform the steps in the decision-making behavior decision-making method according to various embodiments of this application described in Embodiment 1 above.

[0089] The computer-readable storage medium may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may, for example, include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0090] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems, characterized in that, Includes the following steps: S1, Construct a transfer learning model to realize knowledge transfer between the simulation domain (source domain) and the physical domain (target domain); S2, for the transfer learning model, adopts an improved transfer learning algorithm to solve the performance problem caused by the distribution difference between the simulation domain and the physical domain; S3 defines a new domain generalization index for transfer learning models, which combines KL divergence with joint distribution adaptation to further optimize the distribution alignment between the simulation domain and the physical domain and quantify the adaptive differences between the simulation domain and the physical domain. S4, based on an improved transfer learning algorithm and a new domain generalization metric, trains the transfer learning model to achieve efficient transfer between simulation data and physical data.

2. The transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to claim 1, characterized in that, In step S1, the transfer learning model selects a convolutional neural network as the model architecture to extract features from the input data and maps the features of the simulation data to the real data through multi-level feature mapping, thereby ensuring the generalization ability of the transfer learning model; the convolutional neural network includes an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

3. The transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to claim 2, characterized in that, In step S2, a new I-Softmax loss was designed. An improvement to the transfer learning algorithm, aimed at learning more separable features and increasing scores in transfer tasks, is defined as follows: In the formula, The feature vector of the i-th sample The probability distribution; and They represent The first in One element and other elements, among which Let be the score of the i-th sample for category c. Let be the score of the i-th sample for category j; Let be the feature vector of the i-th sample, which is the output of the input data after passing through the convolutional neural network; n represents the number of feature vectors, i.e., the number of samples; m and k are hyperparameters that control the decision boundary.

4. The transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to claim 1, characterized in that, In step S2, the improved transfer learning algorithm uses soft labels to preserve the predicted probability distribution.

5. The transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to claim 3, characterized in that, In step S3, the new domain generalization metric is calculated as follows: By combining the maximum mean difference (MMD) with the covariance alignment (CORAL) weighted loss, an improved marginal distribution alignment (IMDA) loss is obtained. : Among them, hyperparameters These are the covariance weighting coefficients; The improved joint distribution adaptive loss (IJDA) is obtained by comparing the magnitudes of the improved marginal distribution alignment (IMDA) and conditional distribution alignment (CDA) using KL divergence comparison, adjusting their weights, and then combining them with a weighted average. As a new domain generalization metric to address the problem of data distribution differences between the source and target domains, it is as follows: In the formula, The KL divergence of CDA is aligned with the conditional distribution; C is the number of classes; This represents the prior probability of class c in the source domain; This indicates the source domain data under category c. Distribution; This represents the expectation of the source domain data, i.e., the average feature of category c in the source domain samples; This represents the prior probability of class c in the target domain; This indicates the target domain data under category c. Distribution; This represents the expected value of the target domain data, i.e., the average feature of category c in the target domain samples; Align the KL divergence of IMDA with the improved edge distribution; and Let be the covariance matrices of the source and target domains, respectively. For each dimension of the sample data, Represents the kernel function. The number of samples in the source domain. Let ||·|| represent the number of samples in the target domain, and ||·|| denotes the norm to be calculated. To improve the loss of IMDA for edge distribution alignment, The loss of CDA is aligned with the conditional distribution.

6. The transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to claim 5, characterized in that, Step S4 is as follows: Adapting the improved joint distribution to loss With I-Softmax loss The sum is used as the total loss of the model. The backpropagation algorithm is used to update the network weights and bias parameters of each layer of the model. The model is updated until the maximum number of iterations is reached, and the trained transfer learning model is obtained.

7. A transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems according to any one of claims 1 to 6, characterized in that, The method is applied to steer-by-wire systems, using the trained transfer learning model as a fault diagnosis model to perform fault diagnosis.

8. A readable storage medium, characterized in that, It stores a computer program, which, when executed, implements a transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems as described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, It includes a computer program / instruction that, when executed by a processor, implements the transfer learning and domain adaptation method for fault diagnosis of automotive steer-by-wire systems as described in any one of claims 1 to 7.

Images