Bearing fault diagnosis method based on coupling dynamics and domain adversarial neural network

Abstract

The invention provides a bearing fault diagnosis method based on coupling dynamics and a domain adversarial neural network, and belongs to the technical field of bearing fault diagnosis, and the method comprises the steps: carrying out the dynamic simulation based on an axle box-vehicle-track coupling dynamics model, and obtaining a source domain bearing data set and a target domain bearing data set; pre-training the constructed domain adversarial neural network model based on the source domain bearing data set until the maximum number of iterations is reached, and obtaining a pre-trained domain adversarial neural network model; inputting the target domain bearing data set into a pre-trained domain adversarial neural network model, and identifying and outputting each fault type of the target domain bearing data set; compared with the prior art, the source domain and the target domain obtained through dynamic simulation cover more working condition bearing faults, the difference between the source domain and the target domain is reduced, and the data requirement of source domain training data is reduced.

Description

Technical Field

[0001] This application belongs to the field of bearing fault diagnosis technology, and in particular relates to a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks. Background Technology

[0002] In the field of fault prediction and health management (PHM), bearing fault diagnosis is a critical task. Currently, bearing fault diagnosis in high-speed trains often employs traditional, pre-trained deep learning neural networks. Within the PHM framework, this approach requires a large number of labeled samples and demands that the training and testing data meet a uniform data distribution to effectively improve the accuracy of bearing fault diagnosis. However, current high-speed trains face a shortage of real-world bearing fault data, resulting in a diagnostic accuracy that falls short of practical requirements in fault prediction and health management.

[0003] Existing technologies attempt to address the data shortage problem by generating bearing fault data under various operating conditions using generative models on experimental benches. However, from a PHM (Prognostics and Health Management) perspective, this method has significant drawbacks. Firstly, the source domain bearing data differs considerably from the target domain bearing data, failing to adequately adapt to real-world application scenarios. Secondly, the generated data cannot cover a wide range of bearing fault conditions, and some abnormal source domain bearing data can interfere with the diagnostic process, reducing the accuracy of bearing fault diagnosis. This hinders the achievement of accurate fault prediction and health management.

[0004] In summary, existing technologies for bearing fault diagnosis suffer from several drawbacks, including insufficient variety of bearing fault data and the reduced diagnostic accuracy due to abnormal source domain bearing data. These issues make it difficult to meet the accuracy and reliability requirements of Predictive Fault and Health Management (PHM) for bearing fault diagnosis. Summary of the Invention

[0005] This application provides a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks, which can solve the problems of insufficient operating condition types in bearing fault data and abnormal source domain bearing data reducing the accuracy of bearing fault diagnosis.

[0006] In a first aspect, embodiments of this application provide a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks, including:

[0007] The source domain bearing dataset is generated based on the axle box-vehicle-track coupled dynamic model, and the target domain bearing dataset is obtained. The source domain bearing dataset consists of labeled bearing fault data under multiple working conditions, while the target domain bearing dataset consists of unlabeled bearing data.

[0008] The pre-trained domain adversarial neural network model is pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model.

[0009] The target domain bearing dataset is input into a pre-trained domain adversarial neural network model. The pre-trained domain adversarial neural network model is adjusted through weight sharing to identify and output the various fault types of the target domain bearing dataset.

[0010] Among them, the parameters for domain discrimination in the domain adversarial neural network model include the source domain sample normalization weight, which is used to reduce the interference of abnormal bearing data in the source domain bearing dataset.

[0011] In one embodiment, the domain adversarial neural network model includes an input signal, a feature extractor, a domain adaptation module, a domain discriminator, a feature classifier, and a loss decision output;

[0012] The feature extractor is used to extract bearing fault features from the input signal, which includes a source domain bearing dataset or a target domain bearing dataset.

[0013] The input of the domain adaptation module is connected to the output of the feature extractor to receive bearing fault features. It performs dynamic feature alignment on the feature spaces corresponding to the source and target domains by measuring the distance between the source and target domains. The dynamic feature alignment is to dynamically align the edge distribution and conditional distribution between the source and target domains respectively.

[0014] The input of the domain discriminator is connected to the output of the domain adaptation module. It is used to receive bearing fault features and confuse the differences between bearing fault features corresponding to the source domain and the target domain. The bearing fault features are labeled with the source domain or the target domain through feature embedding, gradient inversion layer and domain discrimination.

[0015] The input of the feature classifier is connected to the output of the domain adaptation module to receive aligned bearing fault features and perform label classification on the bearing fault features.

[0016] The loss determination output receives the label classification to determine the classification loss, receives the domain label to determine the domain discrimination loss, and determines the joint maximum average difference loss based on feature extraction and feature classification of the source and target domains. By minimizing the loss function value, the bearing fault type of the bearing fault data is identified.

[0017] In one embodiment, the domain adversarial neural network model includes 5 convolutional layers, 1 adaptive max pooling layer, and 2 fully connected layers;

[0018] The feature extractor consists of five sequentially connected convolutional layers, each consisting of a convolution operation, batch normalization, and a ReLU activation function; the domain adaptation module includes one adaptive max pooling layer; the domain discriminator includes feature embedding, gradient inversion layer, and domain discrimination; and the feature classifier includes two fully connected layers.

[0019] In one embodiment, the pre-trained domain adversarial neural network model is pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model, including:

[0020] Initialize the network structure of the domain adversarial neural network model, and determine the maximum number of iterations and the learning rate;

[0021] Based on the first parameter, second parameter, third parameter, feature classifier, feature extractor, source domain bearing dataset, number of samples in the source domain bearing dataset, number of health categories, and indicator function, the classification loss is determined by the classification loss calculation formula.

[0022] Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the Gaussian kernel function, the feature classifier, and the feature extractor, the joint maximum average difference loss is determined by the joint maximum average difference loss calculation formula.

[0023] Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the normalized weights of the source domain samples, the feature classifier, the feature extractor, and the domain discriminator, the domain discrimination loss based on the source domain sample weights is determined by the domain discrimination loss calculation formula.

[0024] The loss function value is determined by the loss function calculation formula based on classification loss, domain discrimination loss, regularization coefficient, joint maximum average difference loss and trade-off parameters.

[0025] The first, second, and third parameters are updated using an adaptive motion estimation method and a learning rate, and then iterated repeatedly until the maximum number of iterations is reached, thus obtaining a pre-trained neural network model.

[0026] In one embodiment, the loss function is calculated as follows:

[0027]

[0028] Where, L(θ) f ,θ c ,θ d ) represents the loss function value; θ f This is the first parameter of the feature extractor;

[0029] θ c θ is the second parameter of the feature classifier. d This is the third parameter of the domain discriminator;

[0030] L c The loss is the classification loss; λ is the regularization coefficient. Loss is determined by the domain;

[0031] μ is the trade-off parameter; L JMMD This represents the combined maximum average difference loss.

[0032] In one embodiment, the classification loss is calculated as follows:

[0033]

[0034] Among them, L c For classification loss; θ f This is the first parameter of the feature extractor;

[0035] θ c This is the second parameter of the feature classifier;

[0036] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0037] K s The number of health categories; I[*] is the indicator function;

[0038] i represents the source domain sample number labeled with the actual operating condition category;

[0039] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0040] for Corresponding real-world operating condition categories;

[0041] For feature classifier G c right The k-th predicted working condition category is output.

[0042] For feature extractor G f right Output bearing fault characteristics.

[0043] In one embodiment, the domain discrimination loss is calculated as follows:

[0044]

[0045] in, For domain discrimination loss; θf This is the first parameter of the feature extractor;

[0046] θ d This is the third parameter of the domain discriminator;

[0047] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0048] n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category;

[0049] i represents the source domain sample number labeled with the actual operating condition category;

[0050] j represents the sample number of the target domain that is not labeled with the actual working condition category;

[0051] The source domain sample normalization weight is the i-th sample in the source domain bearing dataset.

[0052] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0053] For the bearing data of the i-th target domain that is not labeled with the actual operating condition category;

[0054] For feature extractor G f right Output bearing fault characteristics;

[0055] For the domain discriminator G d right The output prediction domain category;

[0056] For feature classifier G c right Output predicted operating condition categories;

[0057] For the domain discriminator G d right The output prediction domain category.

[0058] In one embodiment, the formula for calculating the joint maximum average difference loss is:

[0059]

[0060] Among them, L JMMD (θ f ,θ c ) represents the joint maximum average difference loss;

[0061] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0062] n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category;

[0063] i represents the source domain sample number labeled with the actual operating condition category;

[0064] j represents the sample number of the target domain that is not labeled with the actual working condition category;

[0065] g(·) is the Gaussian kernel function;

[0066]

[0067] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0068] For the bearing data of the i-th target domain that is not labeled with the actual operating condition category;

[0069] G c For feature classifiers; G f For feature extractors.

[0070] Secondly, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a bearing failure diagnosis method as described in any one of the first aspects above.

[0071] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the bearing fault diagnosis method as described in any one of the first aspects above.

[0072] It is understood that the beneficial effects of the second and third aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here.

[0073] The beneficial effects of the embodiments in this application compared with the prior art are:

[0074] This application provides a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks. It generates a source domain bearing dataset based on a coupled dynamics model of axle box-vehicle-track, and obtains a target domain bearing dataset. The source domain bearing dataset consists of labeled bearing fault data under multiple operating conditions, while the target domain bearing dataset consists of unlabeled bearing data. The pre-trained domain adversarial neural network model is then pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained model. The target domain bearing dataset is then input into the pre-trained model, and the pre-trained model is further refined through weight sharing. The application adjusts the domain adversarial neural network model to identify and output various fault types in the target domain bearing dataset. Compared with existing technologies, the source domain bearing data and target domain bearing data obtained by dynamic simulation based on the axle box-vehicle-track coupled dynamic model cover more bearing faults under various working conditions, reducing the difference between the source domain bearing data and the target domain bearing data, and reducing the data requirements of the domain adversarial neural network model for the source domain training data. Furthermore, the parameters for domain discrimination in the pre-trained domain adversarial neural network model include the normalized weights of the source domain samples, reducing the interference of abnormal bearing data in the source domain bearing dataset, thereby improving the accuracy of bearing fault diagnosis. Attached Figure Description

[0075] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0076] Figure 1 This is a flowchart illustrating a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks provided in an embodiment of this application.

[0077] Figure 2 This is a schematic diagram of the structure of the axle box solid model of a double-row tapered roller bearing provided in an embodiment of this application;

[0078] Figure 3 This is a schematic diagram of the structure of a vehicle multi-rigid-body dynamics model provided in an embodiment of this application;

[0079] Figure 4 This is a schematic diagram of the structure of a domain adversarial neural network model provided in an embodiment of this application;

[0080] Figure 5This is a flowchart illustrating step S2 of an embodiment of the present application, which involves pre-training the constructed domain adversarial neural network model based on the source domain bearing dataset until the maximum number of iterations is reached, to obtain the pre-trained domain adversarial neural network model.

[0081] Figure 6 This is a schematic diagram of source domain bearing data provided in an embodiment of this application;

[0082] Figure 7 This is a schematic diagram of target domain bearing data provided in an embodiment of this application;

[0083] Figure 8 This is a schematic diagram illustrating the identified bearing fault types provided in an embodiment of this application;

[0084] Figure 9 This is a confusion matrix diagram of the identification results provided in an embodiment of this application;

[0085] Figure 10 This is a t-SNE clustering diagram provided in one embodiment of this application. Detailed Implementation

[0086] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0087] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0088] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0089] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0090] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0091] Currently, high-speed train bearing fault diagnosis uses traditionally trained deep learning neural networks. However, deep learning neural networks require a large number of labeled samples and training and testing data with uniform data distribution to improve the accuracy of bearing fault diagnosis. Currently, high-speed trains suffer from insufficient real-vehicle bearing fault data, resulting in unsatisfactory diagnostic accuracy.

[0092] Existing technologies use experimental bench-based generative models to generate bearing fault data for various operating conditions. However, the source domain bearing data differs significantly from the target domain bearing data and cannot cover more bearing faults under various operating conditions. Some abnormal source domain bearing data reduces the accuracy of bearing fault diagnosis.

[0093] To overcome the aforementioned technical problems, this application provides a bearing fault diagnosis method based on coupled dynamics and a domain adversarial neural network, comprising: generating a source domain bearing dataset based on a coupled dynamics model of axle box-vehicle-track, and obtaining a target domain bearing dataset; the source domain bearing dataset consists of labeled bearing fault data under multiple working conditions, and the target domain bearing dataset consists of unlabeled bearing data; pre-training a pre-constructed domain adversarial neural network model based on the source domain bearing dataset until the maximum number of iterations is reached, obtaining a pre-trained domain adversarial neural network model; inputting the target domain bearing dataset into the pre-trained domain adversarial neural network model, and performing weight sharing... This method adjusts the pre-trained domain adversarial neural network model to identify and output various fault types in the target domain bearing dataset. Compared with existing technologies, the source domain bearing data and target domain bearing data obtained by dynamic simulation based on the axle box-vehicle-track coupled dynamic model cover more bearing faults under various operating conditions, reducing the difference between the source domain bearing data and the target domain bearing data, and reducing the data requirements of the domain adversarial neural network model for the source domain training data. Furthermore, the parameters for domain discrimination in the pre-trained domain adversarial neural network model include the normalized weights of the source domain samples, reducing the interference of abnormal bearing data in the source domain bearing dataset, thereby improving the accuracy of bearing fault diagnosis.

[0094] The technical solution of this application will be described below through specific embodiments.

[0095] Firstly, such as Figure 1 As shown, this application provides a bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks, including:

[0096] S1. Generate the source domain bearing dataset for dynamic simulation based on the axle box-vehicle-track coupled dynamic model, and obtain the target domain bearing dataset.

[0097] In this embodiment, the source domain bearing dataset consists of labeled bearing fault data under multiple operating conditions. Dynamic simulations based on an axle box-vehicle-track coupled dynamics model generate the source domain bearing dataset for dynamic simulation. Amplitude scaling is used to obtain source domain bearing datasets for various operating conditions, providing data support for studying general bearing failure mechanisms, improving the effectiveness of transfer learning, and helping to alleviate diagnostic limitations caused by insufficient real-vehicle fault data. Furthermore, to avoid redundant fault information, a random seed is used as the starting point for bearing data samples, generating both the source domain bearing dataset and the target domain bearing dataset. The target domain bearing dataset is obtained from the vehicle's operational database. This target domain bearing dataset contains unlabeled bearing data, more closely resembling the actual train operating environment and improving the reliability of bearing fault diagnosis.

[0098] In one embodiment, the coupled dynamics model of the axle box-vehicle-track containing bearing defects employs a co-simulation approach. First, a solid model of the axle box with a double-row tapered roller bearing is created in 3D modeling software (e.g., SolidWorks), such as... Figure 2 As shown, the system includes an axle box 1, a cage 2, and a wheel, comprising an inner ring, an outer ring, and rolling elements. After modeling in 3D modeling software, the axle box solid model is imported into a first multibody dynamics simulation software (e.g., Simpack). The vehicle's wheel is set as a flexible wheel, and polygonal damage data of the wheel is obtained by changing the travel distance of the flexible wheel on the track. Then, the polygonal damage data of the wheel is imported into a second multibody dynamics simulation software (e.g., RecurDyn). In the second multibody dynamics simulation software, an axle box-vehicle-track coupled dynamic model is constructed to perform dynamic simulation of bearing signals, thereby generating a source domain bearing dataset for dynamic simulation.

[0099] In one embodiment, after creating a 3D solid model of the axle box in 3D modeling software, the step of constructing the axle box-vehicle-track coupled dynamics model further includes:

[0100] (1) Construct a multi-rigid-body dynamics model for the vehicle, such as Figure 3As shown, a single-car high-speed train model was created in Recurdyn software, mainly consisting of one car body, two bogies, four wheelsets, and eight axle boxes. The car body, frame, and wheelsets were considered rigid bodies. Primary springs, primary shock absorbers, secondary air springs, and secondary lateral and vertical shock absorbers were simulated using springs and damping components. Nonlinear components such as anti-hunting shock absorbers and lateral stops were fitted using spline curves. The car body, frame, and wheelsets each have six degrees of freedom: longitudinal, lateral, vertical, roll, pitch, and yaw. The axle boxes each have one pitch degree of freedom.

[0101] (2) Establishing a flexible axle box model: To more accurately study the vibration response of the axle box, the axle box was made flexible. A flexible model of the axle box was established using the finite element flexible body method provided in Recurdyn. The constructed three-dimensional model of the axle box was imported into Hypermesh and meshed using tetrahedral elements. The meshed model was then imported into Ansys software. After that, the file generated by Ansys was imported into Recurdyn to generate the flexible body model of the axle box, and finally, the vehicle dynamics model was constructed.

[0102] (3) Establish the axle box bearing dynamic model; construct the axle box dynamic model in Recurdyn. To reduce the complexity of modeling, the influence of external factors such as oil is ignored, and therefore the established bearing model does not consider the sealing structure. The established axle box bearing model considers the rolling elements, cage, inner and outer rings as rigid bodies, and fully considers the contact between the rolling elements and the inner and outer ring raceways, the rolling elements and the inner ring large and small flanges, the rolling elements and the cage, and the cage and the inner ring.

[0103] (4) Wheel-rail excitation loading; Since the vehicle dynamics model built in Recurdyn does not consider the wheel-rail coupling part, it is necessary to apply excitation to the wheelset to more realistically simulate the external disturbances during vehicle operation. When adding excitation using displacement, velocity, and acceleration, considering that when discrete data is used as excitation input, the second derivative of the displacement and the second integral of the acceleration will produce trend terms and errors, resulting in a decrease in the accuracy of the excitation input, the vehicle's wheel-rail excitation is applied by velocity.

[0104] (5) Obtain wheelset excitation in Simpack software; build a vehicle dynamics model with the same parameters in Simpack, considering flexible wheelsets but not axle box bearings, to obtain the wheelset excitation response. The wheels use LMA wear-type treads, and the rail type is CN60. Hertzian contact theory is used to calculate the wheel-rail contact normal force, and Kalker theory is used to calculate the wheel-rail creep force. Then, the wheelset excitation is loaded into the vehicle dynamics model considering axle boxes built in Recurdyn, completing the construction of the axle box-vehicle-track coupled dynamics model.

[0105] In this embodiment, the source domain bearing dataset is The target domain bearing dataset is in, This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category. for Corresponding real-world working condition category; n s n represents the number of samples in the source domain bearing dataset labeled with real-world operating condition categories. t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category; This refers to the bearing data for the i-th target domain that is not labeled with the actual operating condition category.

[0106] S2, based on the source domain bearing dataset, pre-train the constructed domain adversarial neural network model until the maximum number of iterations is reached, and obtain the pre-trained domain adversarial neural network model.

[0107] In this embodiment, the pre-trained domain adversarial neural network model is pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model. The parameters for domain discrimination in the domain adversarial neural network model include the source domain sample normalization weights, which are used to reduce the interference of abnormal bearing data in the source domain bearing dataset. They can also reduce data distribution deviation, prevent gradient explosion, and accelerate the model convergence speed, thereby improving the bearing fault diagnosis accuracy of the trained domain adversarial neural network model.

[0108] In one embodiment, such as Figure 4As shown, the domain adversarial neural network model includes an input signal, a feature extractor, a domain adaptation module, a domain discriminator, a feature classifier, and a loss decision output. The feature extractor extracts bearing fault features from the input signal, which may be a source domain bearing dataset or a target domain bearing dataset. The input of the domain adaptation module is connected to the output of the feature extractor to receive bearing fault features. It dynamically aligns the feature spaces of the source and target domains by measuring the distance between them. This dynamic alignment involves dynamically aligning the edge distributions and conditional distributions between the source and target domains to reduce cross-domain feature distribution differences. The input of the domain discriminator is connected to the output of the domain adaptation module to receive bearing fault features and confuse the source and target domain features. To differentiate between corresponding bearing fault features, the feature classifier uses feature embedding, gradient inversion layers, and neighborhood discrimination to label bearing fault features with source or target domain labels, thereby learning domain-invariant features for more robust generalization. The input of the feature classifier is connected to the output of the neighborhood adaptation module to receive aligned bearing fault features, perform label classification on the bearing fault features, and obtain the decision boundary. The loss determination output receives the label classification to determine the classification loss, receives the domain labels to determine the domain discrimination loss, and determines the joint maximum average difference loss based on feature extraction and feature classification of the source and target domains. By minimizing the loss function value, the bearing fault type of the bearing fault data is identified. Bearing fault type labels include bearing normal label, bearing outer ring fault label, bearing roller fault label, and bearing inner ring fault label.

[0109] In one embodiment, the domain adversarial neural network model includes 5 convolutional layers, 1 adaptive max-pooling layer, and 2 fully connected layers; wherein, the feature extractor includes 5 sequentially connected convolutional layers, each of which consists of a convolution operation, batch normalization, and a ReLU activation function; the domain adaptation module includes 1 adaptive max-pooling layer; the domain discriminator includes feature embedding, a gradient inversion layer, and domain discrimination; and the feature classifier includes 2 fully connected layers. The system consists of five convolutional layers connected sequentially. The first convolutional layer has 16 output channels, a kernel size of 15×15, a stride of 1×1, padding of 7×7, and an output size of (16, 1024). The second convolutional layer has 32 output channels, a kernel size of 3×3, a stride of 1×1, padding of 2×2, and an output size of (32, 1024). The third convolutional layer has 64 output channels, a kernel size of 3×3, a stride of 1×1, padding of 1×1, and an output size of (64, 512). The fourth convolutional layer has... The first convolutional layer has 128 channels, a kernel size of 3×3, a stride of 1×1, padding of 1×1, and an output size of (128, 512). The fifth convolutional layer has 256 output channels, a kernel size of 3×3, a stride of 1×1, padding of 1×1, and an output size of (256, 256). The adaptive max pooling layer is connected to the fifth convolutional layer, with 3 output channels and an output size of (256, 3). The first fully connected layer is connected to the adaptive max pooling layer, with 256 output channels. The second fully connected layer is connected to the first fully connected layer, and the number of output channels represents the number of categories.

[0110] In one embodiment, such as Figure 5 As shown, step S2 pre-trains the constructed domain adversarial neural network model based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model, including:

[0111] S21, Initialize the network structure of the domain adversarial neural network model, and determine the maximum number of iterations and the learning rate.

[0112] S22, based on the first parameter, second parameter, third parameter, feature classifier, feature extractor, source domain bearing dataset, number of samples in the source domain bearing dataset, number of health categories, and indicator function, determines the classification loss through the classification loss calculation formula.

[0113] S23. Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the Gaussian kernel function, the feature classifier, and the feature extractor, the joint maximum average difference loss is determined by the joint maximum average difference loss calculation formula.

[0114] S24. Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the normalized weights of the source domain samples, the feature classifier, the feature extractor, and the domain discriminator, the domain discrimination loss based on the source domain sample weights is determined through the domain discrimination loss calculation formula.

[0115] S25. Based on classification loss, domain discrimination loss, regularization coefficient, joint maximum average difference loss, and trade-off parameters, the loss function value is determined through the loss function calculation formula.

[0116] S26. The first, second, and third parameters are updated using adaptive motion estimation and learning rate, and the process is iterated until the maximum number of iterations is reached, thus obtaining a pre-trained neural network model.

[0117] In this embodiment, to reduce the impact of outlier samples in the source domain, a weight allocation mechanism is introduced into the loss function, assigning weights to each source domain sample based on the evaluation results of the domain discriminator. A domain alignment strategy based on the joint maximum mean difference loss term is adopted to align the edge and conditional distributions of the source and target domains, thereby enhancing the generalization ability of feature extraction across different distribution domains. Specifically, batch normalization adjusts the output of the convolutional layer to have a mean of 0 and a variance of 1, and the maximum number of iterations is greater than or equal to 500.

[0118] In one embodiment, the loss function is calculated as follows:

[0119]

[0120] Where, L(θ) f ,θ c ,θ d ) represents the loss function value; θ f This is the first parameter of the feature extractor;

[0121] θ c θ is the second parameter of the feature classifier. d This is the third parameter of the domain discriminator;

[0122] L c The loss is the classification loss; λ is the regularization coefficient. Loss is determined by the domain;

[0123] μ is the trade-off parameter; L JMMD This represents the combined maximum average difference loss.

[0124] In one embodiment, the classification loss is calculated as follows:

[0125]

[0126] Among them, L c For classification loss; θ f This is the first parameter of the feature extractor;

[0127] θ c This is the second parameter of the feature classifier;

[0128] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0129] K s The number of health categories; I[*] is the indicator function;

[0130] i represents the source domain sample number labeled with the actual operating condition category;

[0131] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0132] for Corresponding real-world operating condition categories;

[0133] For feature classifier G c right The k-th predicted working condition category is output.

[0134] For feature extractor G f right Output bearing fault characteristics.

[0135] In one embodiment, to suppress the interference of outliers on source domain samples, the weights of source domain samples are assigned according to the discrimination difficulty of the domain discriminator, in order to train a robust model. The domain discrimination loss is calculated as follows:

[0136]

[0137] in, For domain discrimination loss; θ f This is the first parameter of the feature extractor;

[0138] θ d This is the third parameter of the domain discriminator;

[0139] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0140] n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category;

[0141] i represents the source domain sample number labeled with the actual operating condition category;

[0142] j represents the sample number of the target domain that is not labeled with the actual working condition category;

[0143] The source domain sample normalization weight is the i-th sample in the source domain bearing dataset. This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0144] For the bearing data of the i-th target domain that is not labeled with the actual operating condition category;

[0145] For feature extractor G f right Output bearing fault characteristics;

[0146] For the domain discriminator G d right The output prediction domain category;

[0147] For feature classifier G c right Output predicted operating condition categories;

[0148] For the domain discriminator G d right The output prediction domain category.

[0149] The formula for calculating the normalized weights is:

[0150]

[0151] in, Let be the normalized weight of the i-th sample in the source domain bearing dataset;

[0152] Let be the weight of the i-th sample in the source domain bearing dataset;

[0153]

[0154] w s This is the set of weights for all samples in the source domain bearing dataset.

[0155] The formula for calculating the weight of the i-th sample in the source domain bearing dataset is:

[0156]

[0157] in, Let be the weight of the i-th sample in the source domain bearing dataset;

[0158] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0159] For feature extractor G f right Output bearing fault characteristics;

[0160] For the domain discriminator G d right The output prediction domain category.

[0161] In one embodiment, the joint maximum average difference loss is achieved by calculating the difference between the joint distributions of the source and target domains to align the distributions between the domain categories. The formula for calculating the joint maximum average difference loss is:

[0162]

[0163] Among them, L JMMD (θ f ,θ c ) represents the joint maximum average difference loss;

[0164] n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories;

[0165] n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category;

[0166] i represents the source domain sample number labeled with the actual operating condition category;

[0167] j represents the sample number of the target domain that is not labeled with the actual working condition category;

[0168] g(·) is the Gaussian kernel function;

[0169]

[0170] This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category.

[0171] For the bearing data of the i-th target domain that is not labeled with the actual operating condition category;

[0172] G c For feature classifiers; G f For feature extractors.

[0173] In one embodiment, an adaptive motion estimation algorithm is used to solve for and update the first, second, and third parameters. The parameter update process can be represented as follows:

[0174]

[0175] Where, θ f θ is the first parameter of the feature extractor. c This is the second parameter of the feature classifier;

[0176] θ d L is the third parameter of the domain discriminator. c For classification loss; Loss is determined by the domain;

[0177] L JMMD The maximum average difference loss is denoted by λ; λ is the regularization coefficient; μ is the tradeoff parameter.

[0178] η is the learning rate.

[0179] S3. Input the target domain bearing dataset into the pre-trained domain adversarial neural network model, adjust the pre-trained domain adversarial neural network model through weight sharing, and identify and output the bearing fault types of the target domain bearing dataset.

[0180] In one embodiment, the weight-sharing strategy is mainly reflected in the fact that the weights of the feature extractor are shared between the source domain bearing fault dataset and the target domain bearing dataset. Samples from both the source and target domains are processed through the same feature extraction network, meaning that the data from both domains share the same network parameters (e.g., weights) during forward propagation. The aim is to extract domain-invariant features from both the source and target domains.

[0181] In one embodiment, during the pre-training phase of the domain adversarial neural network (DAN) model training, a labeled bearing fault dataset from the source domain is used as input to train the constructed DANNN. The source domain data first extracts features through a shared feature extractor, then feeds it into a label classifier to predict the fault category, and calculates the classification loss based on the true labels from the source domain. Simultaneously, the source domain data and unlabeled data from the target domain are input into a domain discriminator. An adversarial mechanism is constructed through a gradient reversal layer, causing the feature extractor to generate domain-independent features that cannot be accurately distinguished by the domain discriminator during the optimization process, thereby minimizing the feature distribution differences between the source and target domains. During training, the weighted sum of the classification loss and the domain adversarial loss is continuously optimized until the maximum number of iterations is reached, ultimately obtaining a pre-trained model with good generalization ability.

[0182] After pre-training, the fine-tuning phase begins. Unlabeled bearing data from the target domain is input into the pre-trained domain adversarial neural network (DAN) model. The DAN model is then fine-tuned through continued adversarial training, further optimizing its parameters until the bearing health status output by the DAN is greater than or equal to a certain accuracy threshold (e.g., 90% or higher). Since the source and target domains use the same network structure and share weights during feature extraction, the fine-tuning process only requires further optimization of the existing DAN model parameters, without retraining the network structure. This fine-tuning phase further enhances the DAN model's adaptability to target domain features.

[0183] Finally, a fine-tuned domain adversarial neural network model is used to identify faults in the unlabeled bearing data of the target domain. After being processed by a shared feature extractor, the target domain data is input into a label classifier for category prediction. The domain adversarial neural network model outputs the fault type corresponding to each target domain sample, thus completing the fault diagnosis task for the unlabeled bearing data of the target domain.

[0184] In one embodiment, such as Figure 6 , 7 As shown, the source domain bearing dataset and the target domain bearing dataset are input into a pre-trained domain adversarial neural network model to obtain... Figure 8 Diagram of identified bearing fault types Figure 9 The confusion matrix diagram of the recognition results and Figure 10 t-SNE clustering diagram.

[0185] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0186] In a second aspect, embodiments of this application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any one of the first aspects above.

[0187] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method as described in any one of the first aspects above.

[0188] Fourthly, embodiments of this application provide a computer program product that, when run on an electronic device, causes the electronic device to perform the method described in any one of the first aspects above.

[0189] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here.

[0190] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments of this application can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms.

[0191] The computer-readable medium may include at least: any entity or device capable of carrying computer program code to a photographic device / terminal device, recording media, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media. Examples include USB flash drives, portable hard drives, magnetic disks, or optical discs. In some jurisdictions, according to legislation and patent practice, computer-readable media may not be electrical carrier signals or telecommunication signals.

[0192] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0193] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0194] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0195] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0196] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks, characterized in that, include: The source domain bearing dataset is generated based on the axle box-vehicle-track coupled dynamic model, and the target domain bearing dataset is obtained. The source domain bearing dataset consists of labeled bearing fault data under multiple working conditions, while the target domain bearing dataset consists of unlabeled bearing data. The pre-trained domain adversarial neural network model is pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model. The target domain bearing dataset is input into a pre-trained domain adversarial neural network model. The pre-trained domain adversarial neural network model is adjusted through weight sharing to identify and output the various fault types of the target domain bearing dataset. Among them, the parameters for domain discrimination in the domain adversarial neural network model include the source domain sample normalization weight, which is used to reduce the interference of abnormal bearing data in the source domain bearing dataset.

2. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 1, characterized in that, The domain adversarial neural network model includes an input signal, a feature extractor, a domain adaptation module, a domain discriminator, a feature classifier, and a loss decision output; The feature extractor is used to extract bearing fault features from the input signal, which includes a source domain bearing dataset or a target domain bearing dataset. The input of the domain adaptation module is connected to the output of the feature extractor to receive bearing fault features. It performs dynamic feature alignment on the feature spaces corresponding to the source and target domains by measuring the distance between the source and target domains. The dynamic feature alignment is to dynamically align the edge distribution and conditional distribution between the source and target domains respectively. The input of the domain discriminator is connected to the output of the domain adaptation module. It is used to receive bearing fault features and confuse the differences between bearing fault features corresponding to the source domain and the target domain. The bearing fault features are labeled with the source domain or the target domain through feature embedding, gradient inversion layer and domain discrimination. The input of the feature classifier is connected to the output of the domain adaptation module to receive aligned bearing fault features and perform label classification on the bearing fault features. The loss determination output receives the label classification to determine the classification loss, receives the domain label to determine the domain discrimination loss, and determines the joint maximum average difference loss based on feature extraction and feature classification of the source and target domains. By minimizing the loss function value, the bearing fault type of the bearing fault data is identified.

3. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 2, characterized in that, The domain adversarial neural network model consists of 5 convolutional layers, 1 adaptive max pooling layer, and 2 fully connected layers; The feature extractor consists of five sequentially connected convolutional layers, each consisting of a convolution operation, batch normalization, and a ReLU activation function; the domain adaptation module includes one adaptive max pooling layer; the domain discriminator includes feature embedding, gradient inversion layer, and domain discrimination; and the feature classifier includes two fully connected layers.

4. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 1, characterized in that, The pre-trained domain adversarial neural network model is pre-trained based on the source domain bearing dataset until the maximum number of iterations is reached, resulting in a pre-trained domain adversarial neural network model, including: Initialize the network structure of the domain adversarial neural network model, and determine the maximum number of iterations and the learning rate; Based on the first parameter, second parameter, third parameter, feature classifier, feature extractor, source domain bearing dataset, number of samples in the source domain bearing dataset, number of health categories, and indicator function, the classification loss is determined by the classification loss calculation formula. Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the Gaussian kernel function, the feature classifier, and the feature extractor, the joint maximum average difference loss is determined by the joint maximum average difference loss calculation formula. Based on the source domain bearing dataset, the target domain bearing dataset, the number of samples in the source domain bearing dataset, the number of samples in the target domain bearing dataset, the normalized weights of the source domain samples, the feature classifier, the feature extractor, and the domain discriminator, the domain discrimination loss based on the source domain sample weights is determined by the domain discrimination loss calculation formula. The loss function value is determined by the loss function calculation formula based on classification loss, domain discrimination loss, regularization coefficient, joint maximum average difference loss and trade-off parameters. The first, second, and third parameters are updated using an adaptive motion estimation method and a learning rate, and then iterated repeatedly until the maximum number of iterations is reached, thus obtaining a pre-trained neural network model.

5. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 4, characterized in that, The formula for calculating the loss function is: Where, L(θ) f ,θ c ,θ d ) represents the loss function value; θ f This is the first parameter of the feature extractor; θ c θ is the second parameter of the feature classifier. d This is the third parameter of the domain discriminator; L c The loss is the classification loss; λ is the regularization coefficient. Loss is determined by the domain; μ is the trade-off parameter; L JMMD This represents the combined maximum average difference loss.

6. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 4, characterized in that, The formula for calculating classification loss is: Among them, L c For classification loss; θ f This is the first parameter of the feature extractor; θ c This is the second parameter of the feature classifier; n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories; K s The number of health categories; I[*] is the indicator function; i represents the source domain sample number labeled with the actual operating condition category; This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category. for Corresponding real-world operating condition categories; For feature classifier G c right The k-th predicted working condition category is output. For feature extractor G f right Output bearing fault characteristics.

7. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 4, characterized in that, The formula for calculating the domain discrimination loss is: in, For domain discrimination loss; θ f This is the first parameter of the feature extractor; θ d This is the third parameter of the domain discriminator; n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories; n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category; i represents the source domain sample number labeled with the actual operating condition category; j represents the sample number of the target domain that is not labeled with the actual working condition category; The source domain sample normalization weight is the i-th sample in the source domain bearing dataset. This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category. For the bearing data of the i-th target domain that is not labeled with the actual operating condition category; For feature extractor G f right Output bearing fault characteristics; For the domain discriminator G d right The output prediction domain category; For feature classifier G c right Output predicted operating condition categories; For the domain discriminator G d right The output prediction domain category.

8. The bearing fault diagnosis method based on coupled dynamics and domain adversarial neural networks as described in claim 4, characterized in that, The formula for calculating the joint maximum average difference loss is: Among them, L JMMD (θ f ,θ c ) represents the joint maximum average difference loss; n s The number of samples in the source domain bearing dataset labeled with the actual operating condition categories; n t The number of samples in the target domain bearing dataset that are not labeled with the actual operating condition category; i represents the source domain sample number labeled with the actual operating condition category; j represents the sample number of the target domain that is not labeled with the actual working condition category; g(·) is the Gaussian kernel function; This refers to the bearing fault data of the i-th source domain, which is labeled with the actual operating condition category. For the bearing data of the i-th target domain that is not labeled with the actual operating condition category; G c For feature classifiers; G f It is a feature extractor.

Images