Bearing unsupervised fault diagnosis method and system for multi-user privacy data tag spatial difference

Abstract

The invention discloses a bearing unsupervised fault diagnosis method and system based on multi-user privacy data label space difference, relates to the field of bearing fault diagnosis, and aims to solve the problems that the existing method is influenced by time-varying working conditions, the label cost is high, the bearing fault diagnosis of real industry is difficult to meet, and the fault diagnosis efficiency is high. And the problem that label space differences often exist among clients in a federal transfer learning method is solved. On the basis of a federated transfer learning method, feature extraction is carried out on bearing fault data through a client, features are transmitted to a central server, MR-MMD domain adaptation loss, domain identification loss and classification loss are obtained through a universal domain adaptation model of the central server, and an edge vector is introduced to recognize unknown class samples in a target domain; performing real-time self-adaptive adjustment on weight distribution of each loss through dynamic self-adaptive weighting; issuing the obtained total loss to each client to update a local model; and a central classifier model is obtained after training is completed, and unsupervised fault diagnosis of the bearing is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to an unsupervised bearing fault diagnosis method and system based on spatial differences in multi-user private data labels. Background Art

[0002] Rolling bearings are core components of rotating machinery. They usually operate in harsh working environments and must withstand strong axial and radial pressures, making them very prone to failure. [1] Bearing failure may not only cause downtime of mechanical equipment, but also lead to major safety accidents, resulting in economic losses and risks to personal safety. [2] Therefore, effective fault diagnosis of rolling bearings in rotating mechanical equipment is of vital importance to ensure the safe and stable operation of the equipment.

[0003] With the increase of computing power, deep learning technology has become a key tool to improve the accuracy and efficiency of fault diagnosis. Common networks such as Convolutional Neural Network (CNN) [3] Generative adversarial network (GAN) [4] and Deep belief network (DBN) [5] These methods automatically learn complex features and patterns from large amounts of data. By designing specific loss functions and using backpropagation algorithms, they can adaptively adjust model parameters to achieve an automated process from data to fault diagnosis. However, these methods require that the mechanical monitoring data used to train intelligent diagnostic models must contain a large amount of labeled fault data. [6] The influence of time-varying working conditions and the high cost of labeling make this requirement difficult to meet in rolling bearing fault diagnosis in real industries.

[0004] As the latest branch of transfer learning, domain adaptation technology learns features from the source domain and applies them to tasks in the target domain. It can effectively solve the problem of limited labeled data, so it has attracted a lot of attention from researchers in recent years. There are many domain adaptation technologies in the existing literature, such as deep adaptive network [7] and deep subdomain adaptation networks, etc. [8] , which can effectively align the feature distribution of the source domain and the target domain for unsupervised training. Although the above methods solve the problem that the target domain data needs to be labeled, this domain adaptation assumes that the label space of the source domain and the target domain is the same, which is called closed set domain adaptation. [9] In practical applications, prior information about the relative relationship between source and target domain label sets is often lacking.

[10] .

[0005] Universal domain adaptation only assumes that the intersection of the source and target domain label sets is non-empty, without making further constraints on the label spaces of the two domains.

[11] In order to deal with the label space difference problem in general domain adaptation, existing research proposes to use dual classifiers to

[12] Methods to distinguish unknown classes and use entropy to separate [13-14] and sample weights

[15] However, methods that rely on a single source domain ignore the rich information provided by multiple source domains, limiting the generalization ability of the model.

[16] .

[0006] Starting from data from multiple working conditions, that is, performing universal domain adaptation tasks under multiple source domains, this is a centralized training method that can integrate fault feature information from multiple source domains and extract domain-invariant common features, thereby learning more comprehensive fault knowledge and having a significant effect on fault diagnosis under unknown working conditions. In existing methods, the domain adaptation strategy that adds weights to the traditional maximum mean discrepancy (MMD) method is combined with sample confidence to separate unknown classes.

[17] , and domain adaptation in the form of pseudo-edge vector weighting and separation of unknown classes in combination with sample confidence

[18] , have achieved certain results. However, the above methods do not fully consider the issue of data privacy when processing multi-source domain data. In actual projects, data from various source domains are usually scattered in different devices or institutions, and directly sharing data may cause privacy leakage risks.

[19] .

[0007] On the other hand, federated transfer learning combines the advantages of federated learning and transfer learning, enabling multiple clients to collaboratively train a powerful model without sharing original data.

[20] This method not only effectively protects data privacy, but also makes full use of data from different clients to perform cross-domain knowledge transfer and sharing, and has strong generalization capabilities. Federated transfer learning has made significant progress. For example, the federated transfer method based on adversarial learning

[21] By introducing an adversarial mechanism between the global model and the local model, the generalization performance of the model is effectively improved; a federated migration method based on soft label information transmission

[22] By capturing the relationship between failure modes in data from different clients, the effective transfer of private data knowledge is achieved. However, these methods may have label space differences between clients, that is, there may be unknown types of failures other than known types of failures, which limits the consistency and accuracy of the model on different clients.

[0008] While various existing technologies have achieved effective results in fault diagnosis, they still face numerous challenges. Traditional federated transfer learning methods often struggle to handle the discrepancies in label spaces between different users' data, especially when data privacy restrictions prevent them from accessing other users' data. Furthermore, they lack effective strategies for situations where labels for certain user data are difficult to obtain. Summary of the Invention

[0009] The technical problems to be solved by the present invention are:

[0010] Existing methods are affected by time-varying working conditions and high labeling costs, making it difficult to meet the requirements of bearing fault diagnosis in real industries. The use of federated transfer learning methods often results in differences in label spaces between clients, limiting the consistency and accuracy of the model on different clients.

[0011] The present invention is to solve the above technical problems using the following technical solutions:

[0012] The present invention provides a bearing unsupervised fault diagnosis method based on spatial differences in multi-user private data labels, comprising the following steps:

[0013] Step 1: The client collects rolling bearing fault data under different working conditions as the local data set of the source client, and randomly selects unlabeled data of unknown fault type under another working condition as the local data set of the target client; each client receives the network model sent by the server as the initial model of the local client;

[0014] Step 2: Extract and save the features of local data, and encrypt the extracted features;

[0015] Step 3: The encrypted features are transmitted to the central server, which decrypts and aggregates the ciphertext. The central server uses the universal domain adaptation model to obtain the MR-MMD domain adaptation loss, domain discrimination loss, and classification loss, and introduces edge vectors to identify unknown class samples in the target domain. The weight distribution of each loss is adaptively adjusted in real time through dynamic adaptive weighting. The resulting total loss is sent to each client to update the local model.

[0016] Step 4: Repeat steps 2 to 3 above until the predetermined number of iterations is reached, and then send the central classifier model parameters to the target client to perform unsupervised fault diagnosis on the bearing.

[0017] Furthermore, the network model sent by the server in step 1 is a ResNet50 model. The model gradually extracts the spatial and semantic features of the data from local to global through a combination of layer-by-layer convolution kernels and pooling layers. At the same time, regularization is used to further improve the stability and convergence speed of training.

[0018] Furthermore, in step 2, random mapping is used to encrypt the extracted features locally on the client.

[0019] Furthermore, the MR-MMD domain adaptation loss, domain discrimination loss, and classification loss in step 3 are:

[0020] MR-MMD domain adaptation loss E M for:

[0021]

[0022] Among them, ψ i is the optimization parameter, J is the number of source clients, For the distribution Next, find the mathematical expectation of x, where x is the input sample, p i For source client data distribution, is the source client label space, F(·) is the L2 normalized feature, s i For the source client, is the i-th source client sample, For the probability distribution qC t Next, we find the mathematical expectation of x, q is the target client data distribution, t is the target client, C t is the target client label space, x t For the target client sample, is the 2-norm in the Hilbert space H;

[0023] Domain discrimination loss:

[0024]

[0025] in, ζ and γ are optimization parameters, E is the expectation, and D is the domain discriminator function;

[0026] Classifier loss:

[0027]

[0028] Among them, L(.,.) represents the cross entropy loss, G is the classifier function, y represents the label, ε i and φ are optimization parameters.

[0029] Furthermore, the edge vector introduced in step 3 is used to identify unknown class samples in the target domain, specifically:

[0030] Predictors for target domain samples x Defined as:

[0031]

[0032] Among them, f(x, y) represents the classification probability that sample x belongs to the yth class, and the pseudo label of sample x is expressed as The confidence of the pseudo label is determined by the boundary threshold. is the classification probability that the sample x belongs to the pseudo label;

[0033] Then, the pseudo labels are extracted from the target client, the pseudo edge vectors are calculated, and the final recognition results are output;

[0034] calculate:

[0035]

[0036] Where w0 represents the threshold used to distinguish samples belonging to the unknown class from samples belonging to the known class; if the value of the pseudo edge is greater than the threshold, the sample x is classified as belonging to the known class; otherwise, it is classified as the unknown class.

[0037] Furthermore, the weight distribution of each loss is adaptively adjusted in real time through dynamic adaptive weighting as described in step 3, specifically:

[0038] E total =α1E M +α2E D +α3E G

[0039] Among them, α1, α2 and α3 are adaptive weight coefficients;

[0040] By calculating the mean and standard deviation of the gradient, we can dynamically determine whether the gradient of the loss term deviates significantly from the overall level. When the gradient is too large, we reduce the weight to avoid over-optimization. When the gradient is too small, we increase the weight to strengthen the optimization.

[0041] The present invention provides a bearing unsupervised fault diagnosis system based on spatial differences in multi-user private data labels. The system has a program module corresponding to the steps of the method described in any one of the above technical solutions, and executes the steps in the above-mentioned bearing unsupervised fault diagnosis method based on spatial differences in multi-user private data labels during operation.

[0042] The present invention also provides a computer-readable storage medium, which stores a computer program. The computer program is configured to implement the steps of the unsupervised bearing fault diagnosis method based on the spatial difference of multi-user privacy data labels described in any one of the above technical solutions when called by a processor.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] The method of the present invention combines federated learning with universal domain adaptation to perform unsupervised diagnosis of bearing faults. First, feature extraction is performed on the local client and the features in the transmission process are further encrypted. Secondly, a hybrid radial basis kernel-maximum mean difference method is used on the central server to capture multi-scale distribution differences, reduce the distribution deviation between source domain and target domain samples in a multi-source domain scenario, and achieve unsupervised alignment between source client and target client features. Then, the problem of source-target client label space difference is solved by introducing edge vectors, and unknown class samples in the target client are effectively separated. Finally, a dynamic weighted loss fusion strategy is designed to adaptively optimize the ratio of each loss weight, further improve the learning efficiency of the model, and ultimately achieve unsupervised fault diagnosis of rolling bearings.

[0045] Through experimental verification on two data sets, the average accuracy of the method of the present invention can reach 95.6% and 87.7%, which are 6.5% and 8.1% higher than the existing methods respectively, and save at least 27% of training time, verifying the effectiveness of the proposed method. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flowchart of a method for unsupervised bearing fault diagnosis based on spatial differences in multi-user private data labels in an embodiment of the present invention;

[0047] Figure 2 : This is a diagram of the federated learning framework in an embodiment of the present invention;

[0048] Figure 3 Schematic diagram of multi-client universal domain adaptation in an embodiment of the present invention;

[0049] Figure 4 Schematic diagram of the residual structure in an embodiment of the present invention;

[0050] Figure 5 This is a flowchart of unsupervised bearing fault diagnosis based on spatial differences in multi-user private data labels in an embodiment of the present invention;

[0051] Figure 6 Schematic diagram of a bearing test bench in an embodiment of the present invention;

[0052] Figure 7 This is a graph showing the impact of different hyperparameters L on the experimental accuracy in an embodiment of the present invention;

[0053] Figure 8 This is a graph comparing the accuracy before and after adding federated learning in an embodiment of the present invention;

[0054] Figure 9 This is a comparison result diagram of the optimization algorithms in the embodiment of the present invention;

[0055] Figure 10Graph showing experimental results comparing the method of the present invention with other methods in an embodiment of the present invention;

[0056] Figure 11 t-SNE visualization diagrams of different methods in the embodiments of the present invention;

[0057] Figure 12 This is a graph comparing the accuracy of the PU dataset before and after adding federated learning in an embodiment of the present invention;

[0058] Figure 13 This is a comparison result diagram of the PU dataset optimization algorithm in an embodiment of the present invention;

[0059] Figure 14 This is a comparison chart of the accuracy results of the PU dataset under different tasks in an embodiment of the present invention. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the present invention, exemplary embodiments or examples of the present invention will be described below with reference to the accompanying drawings. Obviously, the described embodiments or examples are only some of the embodiments or examples of the present invention, and not all of them. Based on the embodiments or examples of the present invention, all other embodiments or examples obtained by those skilled in the art without creative work should fall within the scope of protection of the present invention.

[0061] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0062] 1 Basic Theory

[0063] 1.1 Federated Learning

[0064] When dealing with distributed data learning tasks, federated learning came into being in order to effectively integrate information from different data sources and overcome issues such as data privacy.

[23] The most classic one is the federated average algorithm. The federated learning framework is shown in the figure below. Figure 2 It achieves collaborative construction of the global model by aggregating the parameter updates of the local models of each participant.

[24] .

[0065] First, each device receives the global model from the central server for subsequent local training. Next, each client device updates its local data multiple times and uploads the associated gradient information to the cloud in an encrypted format to update the global model. The cloud server then averages the model parameters uploaded by the clients to generate updated global model parameters, which are then distributed to each client. Finally, these steps are repeated until the model reaches the expected performance or the maximum number of training rounds is reached.

[0066] The emergence of federated learning effectively resolves the conflict between data privacy and data sharing across decentralized devices. Because data is not exposed to third-party central servers, federated learning is particularly suitable for applications in data privacy-sensitive environments.

[0067] 1.2 Hybrid Radial Basis Kernel

[0068] Radial basis function (RBF)

[25] It is widely used in fields such as support vector machines (SVM), kernel function networks, and kernel principal component analysis. It solves nonlinear problems by mapping data into a high-dimensional feature space. However, to adapt to complex domain adaptation tasks, a hybrid radial basis kernel is introduced. This kernel combines RBFs of different bandwidths to control the scale of the kernel function and improve its adaptability to complex data distributions. The hybrid radial basis kernel can be expressed as a weighted sum of RBFs of different bandwidths. Its calculation is as shown in formula (1):

[0069]

[0070] Among them, σ i is the bandwidth of the i-th RBF, which controls the scale of the kernel function, ||xy|| 2 It represents the Euclidean distance between input samples x and y, and L is the number of RBFs.

[0071] 1.3 General Domain Adaptation

[0072] In practical applications, domain adaptation methods often face the problem of large label space differences between source domain data and target domain data. The label space of data from a single source domain is usually limited and cannot cover all categories of the target domain, resulting in a significant drop in the performance of the model on the target domain. Therefore, in order to make up for the shortcomings of a single source domain label space, it is necessary to introduce data from multiple source domains to expand the coverage of the label space. The present invention uses multiple source clients with different label spaces to more comprehensively transfer knowledge to the target client. Through continuous training, the model gradually optimizes the feature representation, so that data features of the same category are aggregated and data features of different categories are separated. Ultimately, on the target client, the ideal effect is not only to clearly distinguish known categories, but also to effectively distinguish unknown categories. The schematic diagram of the model principle is shown in the figure. Figure 3 shown.

[0073] In general domain adaptation, source client samples and target client samples are divided into shared classes (samples in the public label space) and unknown classes (samples in the private class in the target client label space). Let p i , q represent the original data of each source client and target client respectively. There are J source clients Each source client dataset is composed of i Extracted from Sample composition, and a target client n extracted from q t samples. The source client label space and the target client label space are respectively C t Indicates that the label relationship is unknown. and D t The public label set of The private tag set is represented as D t The private tag set is represented as Tag Set Contains all known labels in each source domain dataset, the label set Indicates the source client's shared tag set that does not appear on the target client. Tag Set C S and The distribution of source client data in is represented as and Label sets C, C i 、C t The distribution of target client data in is represented as q C 、 q t .

[0074] Source Client and target client D t The commonality is defined as the Jaccard distance between the two label spaces

[11] , calculated as formula (2):

[0075]

[0076] Similarly, the source client and The commonality calculation is as follows:

[0077]

[0078] Among them, ξi, ξ∈[0, 1], the smaller the values ​​of ξi and ξ, the less knowledge the source client and the target client share, and the more challenging the domain adaptation task is.

[0079] 2. Method of the present invention

[0080] 2.1 Improved MR-MMD method

[0081] MMD is a method for measuring the difference between two probability distributions by comparing their means in feature space. MMD typically uses kernel functions for feature space mapping and measurement, measuring the difference between the distributions by calculating the distance between the expected feature mappings of the two distributions in feature space. This expected feature mapping is performed using kernel functions.

[0082] The kernel function used in this paper is a mixed radial basis kernel. The mixed radial basis kernel-maximum mean discrepancy (MR-MMD) method is proposed to further extend the MMD method. By combining multiple RBFs with different bandwidths, the small-bandwidth RBF captures subtle changes in local features, while the large-bandwidth RBF focuses on global distribution differences. This method can align the distribution differences between the source and target domains at multiple levels.

[0083] The mixed radial basis kernel can be expressed as a weighted sum of these RBFs to measure the difference between distributions. Given two distributions P and Q and a set of samples drawn from these distributions and The bandwidth calculation process of the hybrid radial basis kernel is shown in Algorithm 1.

[0084]

[0085]

[0086] The calculation of the mixed radial basis kernel-maximum mean difference is as follows:

[0087]

[0088] Where H is the high-dimensional Hilbert space, f(·) is the mapping function, and m and n are the number of samples of the two distributions, respectively.

[0089] During training, the MR-MMD distance is incorporated into the model's optimization objective as part of the loss function. By minimizing the MR-MMD distance between the source and target domain feature distributions, the feature extractor gradually learns a shared feature representation, aligning the source and target domain data distributions in feature space. Simultaneously, the source domain labels guide the model in learning the classifier. Due to the aligned feature distributions, unlabeled data in the target domain can share the classifier trained in the source domain, thus achieving unsupervised transfer.

[0090] 2.2 Federal Universal Domain Adaptation Method

[0091] In order to protect client privacy without the need for data sharing, obtain label information from different source clients, reduce the distribution differences between each client, and perform unsupervised training when the target client data is unlabeled, a new distributed domain adaptation method based on federated learning and general domain adaptation is proposed; the overall framework of the model is shown in the figure below. Figure 1 shown.

[0092] 2.2.1 Random Mapping Encryption

[0093] To avoid privacy leakage during data transmission, a random mapping method is proposed to encrypt data locally on the client. Each client obtains the secret key K from the central server and then linearly maps the extracted features. The encrypted ciphertext calculation formula of each client is formula (5):

[0094]

[0095] Among them, Z i is the local feature extracted from the i-th client, The encrypted ciphertext of the local features extracted from the i-th client, where n is the number of source clients.

[0096] When the encrypted ciphertext is transmitted to the central server, the central server decrypts the ciphertext using the secret key K and obtains the features uploaded by the client. The decryption process is calculated as formula (6):

[0097]

[0098] Among them, K -1 is the inverse matrix of the secret key K.

[0099] A random mapping method is added to the federated migration method for privacy protection. First, the data is distributedly processed on each client to avoid uploading the original data to the central server. Then, the features uploaded by each client are encrypted with a key. Even if the intermediate data is intercepted, it is difficult to use it to directly train the model or restore the original data, further enhancing the privacy of the federation process.

[0100] 2.2.2 Client Local Model

[0101] During the client-side local model construction process, feature extraction and privacy protection mechanisms are primarily implemented. The client first loads data from the local dataset and performs continuous wavelet transform preprocessing to obtain the local dataset. Each client independently trains the local feature extractor F on its local device using the pretrained ResNet50 model delivered by the server to extract data features and store them in encrypted form.

[0102] ResNet50 is composed of multiple stacked residual units. Each residual unit directly adds the input and output through a jump connection, thereby retaining the original information of the input and enhancing the transmission efficiency of the features. Its residual structure is as follows: Figure 4 shown.

[0103] Furthermore, ResNet50 utilizes a large number of convolution operations, particularly layer-by-layer combinations of convolution kernels and pooling layers, to gradually extract spatial and semantic features from data, from local to global scale, enhancing the model's expressive power. Furthermore, ResNet50 utilizes regularization techniques such as Batch Normalization to further improve training stability and convergence speed, enabling it to adapt to the diverse data distribution of each client in a distributed environment and extract more robust features.

[0104] 2.2.3 Central Server General Domain Adaptation Model

[0105] The proposed MR-MMD method is used to measure and reduce the distribution difference between the source client and the target client at the feature distribution level, and to achieve domain adaptation of the source client features and the target client features. By combining different bandwidths, the MR-MMD method can better adapt to the data distribution characteristics at various scales. MR-MMD domain adaptation loss E M Defined as formula (7):

[0106]

[0107] Among them, ψ i To optimize the parameters, J represents the number of source clients, Indicates that the distribution Next, find the mathematical expectation of x, where x is the input sample, p i Indicates the source client data distribution, represents the source client label space, F(·) represents the L2 normalized feature, s i Indicates the source client, represents the i-th source client sample, In the probability distribution qC t The mathematical expectation of x is calculated below, q represents the target client data distribution, t represents the target client, C t represents the target client label space, x t Indicates the target client sample, represents the 2-norm in the Hilbert space H;

[0108] The domain discriminator D is introduced to measure the similarity between samples and domains, helping the model better understand and distinguish the data distribution of different domains and improve the quality of features extracted by the feature extractor F. This module takes the output feature z generated by the feature extractor as input, and the output of the classifier is used to measure the similarity between the input sample x and the source domain. Ideally, the prediction target of D is to output 1 for source domain samples and 0 for target domain samples. The domain discriminator D loss E D It is defined as formula (8):

[0109]

[0110] Among them, φ i , ζ and γ are optimization parameters, E is the expectation, and D is the domain discriminator function.

[0111] To solve the problem that the target client's local data is unlabeled and cannot train the classifier independently, a new training mode is proposed, that is, deploying a unified central classifier model G on the central server. This model is trained centrally by using the features extracted from the labeled data of the source client to fully learn the data characteristics and optimize the classifier performance, thereby providing accurate classification performance for the target client. The central classifier G loss E G It is defined as formula (9):

[0112]

[0113] Among them, L(.,.) represents the cross entropy loss, G is the classifier function, y represents the label, ε i and φ are optimization parameters.

[0114] In order to enable the model G to identify the unknown categories that may exist in the target client, the present invention proposes a method of introducing edge vectors into the central classifier. This vector is used to measure the reliability of the target label predicted by the classifier and belonging to the public label set C. The prediction factor for the target domain sample x is It is defined as formula (10):

[0115]

[0116] Among them, f(x, y) represents the classification probability that sample x belongs to the yth class, and the pseudo label of sample x is expressed as The confidence of the pseudo label can be determined by a boundary threshold.

[0117] Then, the pseudo-label is extracted from the target client, the pseudo-edge vector is calculated (to assess the confidence that a sample belongs to a class known to the model during training), and the final recognition result is output. When the pseudo-edge vector is equal to 1, the sample x is most likely considered to belong to a known class, and when it is equal to 0, the sample x is most likely considered to belong to an unknown class. The calculation is as shown in Formula (11).

[0118]

[0119] Where w0 represents the threshold used to distinguish samples belonging to the unknown class from samples belonging to the known class. If the value of the pseudo edge is greater than the threshold, the sample x is classified as belonging to the known class; otherwise, it is classified as the unknown class.

[0120] 2.3 Dynamic Weighted Loss Fusion Strategy

[0121] In order to effectively optimize the multiple loss terms involved in the model training process and avoid over-optimization of the loss with too large gradient while ignoring other losses, a dynamic weighted loss fusion strategy is proposed, and a comprehensive loss function E with adaptive weights is designed. total , the optimization goal is to minimize E total , the calculation formula is formula (12):

[0122] E total =α1E M +α2E D +α3E G (12)

[0123] Among them, α1, α2 and α3 are adaptive weight coefficients. The optimization process is shown in Algorithm 2.

[0124]

[0125]

[0126] By calculating the mean μg and standard deviation of the gradient Dynamically determine whether the gradient of a loss term significantly deviates from the overall level. When the gradient is large, the weight is reduced to avoid over-optimization; when the gradient is small, the weight is increased to strengthen optimization. This method does not rely on prior knowledge and adjusts the weight distribution in real time to adapt to the changing importance of different loss terms, ensuring a dynamic balance in the optimization process.

[0127] 3 Unsupervised bearing fault diagnosis method based on multi-user private data label space differences

[0128] The unsupervised fault diagnosis process of bearings with different multi-user private data label spaces under different working conditions is as follows: Figure 5 shown.

[0129] The specific steps are:

[0130] 1) Data selection and preprocessing

[0131] The system obtains rolling bearing data of a certain specification and selects labeled data from several operating conditions with varying numbers of categories as the local dataset for each source client. Unlabeled data from another operating condition with unknown fault types is randomly selected as the local dataset for the target client. Each client then constructs a time-frequency (TF) image dataset using a continuous wavelet transform on the local dataset, which serves as the input dataset for the local model. The TF images from each client are then divided into two different datasets: a training dataset and a test dataset.

[0132] 2) Client local model construction

[0133] Each client receives the ResNet50 network sent by the server as the initial model for the local client. Local features are extracted and saved, and the extracted features are encrypted through random mapping and uploaded to the central server.

[0134] 3) Construction of a universal domain adaptation model for the central server

[0135] After receiving the encrypted features from each client, the central server performs differential decryption to obtain the client's original features. The central server's universal domain adaptation model is used to determine the losses for each component. The MMD loss, domain discrimination loss, and classification loss obtained from the central server model are adaptively weighted to obtain the final total loss for this round and transmitted to each client to update the local model. This process is called a federated iteration. After a predetermined number of federated iterations, the central classifier model parameters are distributed to the target client, completing the overall model training process.

[0136] 4) Target client local test

[0137] The target client's test set data is input into the central classifier, and the model's label prediction results for the test data are compared with the true labels to obtain the accuracy of the target client's fault diagnosis.

[0138] The unsupervised bearing fault diagnosis method (algorithm) based on the spatial difference of multi-user private data labels proposed in this invention is the underlying technical core of this invention, and various products can be derived based on the algorithm.

[0139] Based on the method proposed in the present invention, a bearing unsupervised fault diagnosis system based on spatial differences in multi-user private data labels is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned bearing unsupervised fault diagnosis method based on spatial differences in multi-user private data labels during operation.

[0140] The developed system (software) computer program is stored on a computer-readable storage medium. When called by a processor, the computer program is configured to implement the steps of the aforementioned unsupervised bearing fault diagnosis method based on spatial differences in multi-user private data labels. This materializes the present invention on a carrier, becoming a computer program product.

[0141] Various implementations of the systems and techniques described herein can be realized in digital electronic circuitry, integrated circuitry, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system comprising at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0142] The computer programs (also referred to as programs, software, software applications, or code) of the present invention include machine instructions for a programmable processor and can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or apparatus (e.g., a magnetic disk, an optical disk, a memory, a programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0143] The beneficial effects of the present invention will be described below with reference to specific embodiments.

[0144] Example 1

[0145] 4 Application and Analysis

[0146] 4.1 Experimental Dataset

[0147] The dataset used in this experiment comes from Case Western Reserve University (CWRU)[ 2 '] rolling bearing experimental data set. The data acquisition system of the experimental platform mainly consists of a drive motor, a load, a torque sensor / encoder and a control circuit. The experimental data acquisition device is as follows Figure 6As shown, the driving end and fan end of the motor are respectively installed with SKF6205 and SKF6203 rolling bearings.

[0148] This dataset uses a 16-channel data recorder to collect bearing vibration signals at a sampling frequency of 12 kHz. Data were collected at four different speeds: 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm. Data collected at each speed includes normal state (Nor), inner race fault (IF), outer race fault (OF), and rolling element fault (BF). Each fault is further categorized by damage diameter, with severity levels of 0.007 inch, 0.014 inch, and 0.021 inch, for a total of 10 categories. Table 1 shows the correspondence between health status and fault severity.

[0149] Table 1 Health status of CWRU dataset

[0150]

[0151] In this experiment, the SKF6203 bearing dataset was used. Data was extracted from the original vibration signal in segments of 1024 points. These data were then transformed into time-frequency maps using the continuous wavelet transform (CWT). The complex Morlet wavelet (cmor3-3) was used as the mother wavelet. Wavelet coefficients at various scales were calculated to generate the signal's time-frequency maps. 150 time-frequency maps were generated for each health state, divided into a training set (120 images) and a test set (30 images) in an 8:2 ratio.

[0152] According to the proposed rolling bearing cross-domain fault diagnosis method, A, B, C, and D represent data at different speeds of 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm, respectively. Data sets at any three speeds are used as local source client datasets, providing transferable features, while data sets at another speed are used as target client datasets. S1, S2, S3, and T represent the three source and target clients used during the experiment, respectively. The experimental task setup is shown in Table 2. The source and target client health states for different tasks are typically randomly selected. To better represent the experimental tasks, 0-T1, 0-T2, 0-T3, and 0-T4 represent the four tasks in which the target client has zero unknown classes, and the same applies to the other two states. All experiments were conducted on a computer with an AMD Ryzen 7-5800H CPU, an NVIDIA GeForce RTX3060 GPU, and 16GB of memory. The model was built using the Pytorch deep learning framework. The SGD optimization method was used in the training process. The learning rate was set to 0.01, the batch size was set to 16, the initial values ​​of the weight parameters α1, α2, and α3 were all set to 1, and the number of iterations was 60. To reduce the randomness of the experiment, each experiment was repeated five times and the average value was taken as the final result.

[0153] Table 2. Federation migration tasks at different speeds

[0154]

[0155] 4.2 Hyperparameter Selection Experiment

[0156] Hyperparameters are user-defined parameters before model training begins and are not learned and updated during training. Task 1-T4 is used as an example for hyperparameter selection experiments. According to formula (1), the parameter L, as the number of RBFs incorporated into the hybrid radial basis kernel, has a significant impact on the performance of the improved MR-MMD method. When the number of RBFs is too small, the multi-scale distribution features are insufficiently captured; when the number of RBFs is too large, the model may capture irrelevant or redundant features, resulting in increased learning noise. Both situations may reduce the effectiveness of the transfer task. Figure 7 The effect of changing the size of L on the model accuracy. As can be seen from the figure, in the experiment, setting L to 5 achieves the best classification accuracy.

[0157] According to formula (11), parameter w0 serves as the threshold for distinguishing between known and unknown fault types. Its selection has a significant impact on model accuracy. Table 3 shows the effect of varying w0 on model accuracy. As can be seen from the table, in the experiment, setting w0 to 0.8 achieved the best classification accuracy.

[0158] Table 3 Effect of different hyperparameters w0 on experimental accuracy (%)

[0159]

[0160] 4.3 Comparative experiments before and after improvement of each part of the proposed method

[0161] To verify the impact of each part of the proposed method on performance, this section compares the performance before and after the introduction of the MR-MMD method, the performance of centralized universal domain adaptation and federated universal domain adaptation, and the effects before and after the use of dynamic weighted fusion loss.

[0162] To verify the effectiveness of the MR-MMD method, due to space limitations, we used Task 1-T2 as an example, combining it with a margin vector method after domain adaptation to distinguish unknown classes. Feature alignment was performed using an MMD-based statistical difference metric and a traditional adversarial domain adaptation method based on a domain discriminator. The diagnostic performance of the MR-MMD method employed in this paper was compared, and the results are shown in Table 4.

[0163] Table 4 Comparison of accuracy of different feature alignment methods (%)

[0164]

[0165] Comparing the accuracy of different feature alignment methods in Table 4 shows that the accuracy of cross-domain diagnosis without any alignment method is the lowest, at only 36.5%, indicating a significant difference in feature distribution between the source and target domains. The accuracy of the MR-MMD method is 11.9% and 6.8% higher than that of the statistical difference metric method and the adversarial domain adaptation method, respectively. This demonstrates that the MR-MMD method can better align features between the two domains in this experiment, improving the accuracy of cross-domain rolling bearing fault diagnosis.

[0166] Under the premise of data silos, data is usually scattered across different devices or institutions. In order to explore the classification effect of the general domain adaptation method under the federated learning framework compared with the traditional centralized general domain adaptation method, Figure 8 shown.

[0167] Depend on Figure 8As can be seen, the classification accuracy of the federated universal domain adaptation method is slightly lower than that of the centralized universal domain adaptation method, with the maximum accuracy difference being only 1.8%. By combining federated learning with universal domain adaptation to construct a new distributed domain adaptation method, it is not only possible to share valuable knowledge between different devices or institutions, but also effectively prevent data leakage. This strategy, which trades a small amount of accuracy for the protection of data privacy and knowledge sharing between data silos, is highly worthwhile in practical applications. This result demonstrates that the proposed method strikes an ideal balance between maintaining data security and improving overall system performance, and is an effective solution to the challenges of data privacy and data silos.

[0168] Taking Task 1-T2 as an example, a comparison experiment is conducted before and after using dynamic weighted fusion loss. Figure 9 As can be seen, before the loss function was improved, the loss function exhibited severe fluctuations in the early stages of iteration, with the accuracy stabilizing around 45 iterations. After the loss function was improved, the fluctuations were less pronounced in the early stages of iteration, and the accuracy remained stable around 30 iterations. The improved loss function not only significantly reduced the instability during the initial training phase but also shortened the convergence time, saving approximately 34% of the training time. This demonstrates that the improved loss function exhibits significant advantages in improving convergence efficiency, effectively accelerating model optimization and thus completing the training process more efficiently.

[0169] 4.4 Comparative experiments with other methods

[0170] The method proposed in the present invention was compared with the following four methods to verify the superiority of the proposed method:

[0171] 1) Domain aversarial training of neural networks (DANN)

[27] As a typical closed-set domain adaptation method, adversarial training is performed through the domain discriminator to learn cross-domain invariant features.

[0172] 2) Universal Domain Adaptation Network (UAN)

[28] The domain similarity and prediction uncertainty of samples are exploited to achieve shared class domain adaptation between source and target domains.

[0173] 3) Universal multi-source domain adaptation network (UMAN)

[17] The method of multiple source domains corresponding to one target domain uses pseudo-edge vector weighted form to perform domain adaptation and combines sample confidence to separate unknown classes.

[0174] 4) Federated Transfer Learning (FTL)

[29] Federated communication is achieved through deep adversarial networks combined with federated learning, completing the alignment of each source domain with the target domain.

[0175] Since methods 1) and 2) in the comparison method cannot directly complete the migration task set by this invention, three migration scenarios were constructed, using the local data of source clients 1, 2, and 3 as the source domain and the target client dataset as the target domain. The classification results obtained from these three experiments were used as the final classification result of the target domain test set data through confidence discrimination, which was used as the accuracy of this task. Each experiment was repeated five times, and the average value was taken as the final accuracy to reduce uncertainty.

[0176] 4.4.1 Experimental Results

[0177] The fault diagnosis results of different tasks are as follows: Figure 10 As shown in the figure, it can be seen that the proposed method has good cross-domain diagnosis accuracy, most of which are above 94%.

[0178] The DANN and FTL methods achieve better diagnostic performance when the target client has 0 unknown classes, that is, in the closed-set domain adaptation task. However, these two methods are unable to identify unknown class samples, and when private classes and shared classes exist in both the source client and the target client, obvious negative transfer occurs. For example, when the target client has 2 unknown classes, the diagnostic accuracy of the closed-set domain adaptation method is only about 60%, the average accuracy of the universal domain adaptation methods UAN and UMAN can reach 75.8% and 82.9%, respectively, while the average accuracy of the method proposed in the present invention can reach 93.9%, which is a significant improvement compared to the above-mentioned comparison methods.

[0179] The proposed method achieves an average accuracy of 95.6% in each case, which is at least 6.5% higher than the average accuracy of 82.3% and 89.1% for the UAN and UMAN methods, respectively. The proposed method outperforms other methods in various domain adaptation tasks, demonstrating its superiority in solving fault diagnosis problems at different speeds.

[0180] 4.4.2 Visual Analysis

[0181] In this section, in order to more intuitively evaluate the effectiveness of the proposed method, feature visualization is performed. Specifically, taking Task 1-T2 as an example, the t-distribution stochastic neighbor embedding (t-SNE) algorithm is used to reduce the dimension of the last layer of features extracted by the model. The visualization results of different methods are shown in Figure 2. Figure 11 As shown, different colors and shapes represent different health conditions.

[0182] from Figure 11 As can be seen from the five subplots, after visualizing the features extracted using the DANN and FTL methods, features from different categories are mixed together, indicating that directly applying closed-set domain adaptation methods to general domain adaptation scenarios will result in a large number of misclassified samples. Furthermore, the feature distribution maps obtained using the UAN and UMAN methods show generally clear category boundaries, but some overlapping features still exist. In contrast, after visualizing the deep features extracted by the proposed method, features of the same category are clustered together, while features of different categories are more distant from each other, and samples of the unknown class are clustered and separated from other classes. This further demonstrates that the proposed method can effectively distinguish characteristics of different health states.

[0183] 4.5 Further verification experiments on the effectiveness of the proposed method

[0184] In order to further verify the effectiveness of the proposed method, the bearing dataset of Paderborn University (PU) was used to

[30] Rolling bearing fault diagnosis experiments under different working conditions of multiple clients are carried out.

[0185] The PU dataset uses a piezoelectric accelerometer to collect bearing seat vibration signals at a sampling frequency of 64 kHz. By varying the drive system's rotational speed, the radial force on the bearing and the load torque on the drive system are measured. The PU dataset consists of four operating states, designated A, B, C, and D. The corresponding parameters for each operating state are shown in Table 5.

[0186] Table 5 PU dataset running status

[0187]

[0188] The PU dataset consists of 6203 bearings. Fault sources include artificial damage faults (ART) and real damage faults (REAL). Artificial damage faults are created using various tools, including electrical discharge machining (EDM), drilling (DRI), and pitting (PIT). In this experiment, artificial damage faults and real damage faults are classified into different fault types. Within artificial damage faults, faults generated by different methods are also classified into different fault types, resulting in a total of seven fault types. Table 6 shows the correspondence between health status and fault severity.

[0189] The number of iterations in this section is set to 150. The other experimental environments, data preprocessing methods, and different task representations are the same as in Section 4.1. The experimental task settings are shown in Table 7.

[0190] Table 6 PU dataset

[0191]

[0192] Table 7. Federation migration tasks in different running states

[0193]

[0194] 4.5.1 Comparative experiments before and after improvements of each part of the proposed method

[0195] To further verify the impact of each part of the proposed method on performance, this section further verifies the performance before and after the introduction of the MR-MMD method, the performance of centralized universal domain adaptation, the performance of federated universal domain adaptation, and the effects before and after the use of dynamic weighted fusion loss on the PU dataset. The specific experimental settings are the same as those in Section 4.3. The experimental results are shown in Table 8. Figure 12 、 Figure 13 shown.

[0196] Table 8 shows that the accuracy of cross-domain diagnosis on the PU dataset without any alignment method is the lowest, at only 21.2%, indicating a significant difference in feature distribution between the source and target domains of the PU dataset. The accuracy of the MR-MMD method is 14.7% and 8.7% higher than that of the statistical difference metric method and the adversarial domain adaptation method, respectively. This demonstrates that the MR-MMD method can effectively reduce distribution differences even when faced with more complex feature distributions, thereby improving the accuracy of cross-domain fault diagnosis for rolling bearings.

[0197] Table 8 Comparison of the accuracy of different feature alignment methods on the PU dataset (%)

[0198]

[0199] Depend on Figure 12The classification accuracy of the federated universal domain adaptation method is slightly lower than that of the centralized universal domain adaptation method, with a maximum difference of 3.5%. Although this result is slightly higher than that on the CWRU dataset, it remains within an acceptable range overall. This demonstrates that even in more complex data scenarios, the federated universal domain adaptation method can still achieve performance close to that of the centralized universal domain adaptation method. This further demonstrates that the proposed method not only achieves an ideal balance between data security and system performance, but also has strong applicability and stability, effectively addressing data privacy and data silo issues.

[0200] Depend on Figure 13 Before the loss function was improved, it exhibited significant jitter in the early stages of training, with accuracy stabilizing around the 110th iteration. After the improvement, the jitter was less pronounced in the early stages of training, and accuracy remained stable around the 80th iteration, saving approximately 27% of training time. Further verification results showed that the improved loss function not only significantly reduced jitter in the early stages of training and shortened convergence time, but also demonstrated improved stability and applicability.

[0201] 4.5.2 Comparative Experiments with Other Methods

[0202] The experiment was conducted according to the task settings in Table 7. The comparison experiment was the same as the comparison method set in Section 4.4. The results are as follows: Figure 14 shown.

[0203] Depend on Figure 14 As can be seen, in all 12 migration tasks, when classifying fault states in a PU bearing dataset, the proposed method achieved higher fault diagnosis accuracy than the other four comparison methods. The average accuracy for each scenario remained at 87.7%, an improvement of at least 8.1% compared to the average accuracies of 79.5% and 77.9% for the UAN and UMAN methods, respectively. Experimental verification demonstrates that the proposed method can effectively extract domain-invariant features of rolling bearings under different operating conditions and solve fault diagnosis problems at different speeds, further demonstrating its effectiveness.

[0204] 5 Conclusion

[0205] (1) The present invention can establish an effective fault diagnosis model when the source client and the target client do not share original data, there are differences in the label space, and the target client lacks labeled data. Experimental results show that the fault diagnosis accuracy is high even when there are significant differences in feature distribution and label space between clients. The average accuracy on the two data sets can reach 95.6% and 87.7% respectively, which is at least 6.5% higher than the other methods compared in this embodiment. At the same time, the proposed method combines the random mapping federated feature encryption and transmission method to increase the privacy of the feature transmission process at the cost of a maximum accuracy of 3.5%.

[0206] (2) A domain adaptation method based on hybrid radial basis kernel-maximum mean discrepancy (MR-MMD) is proposed to enhance the ability of traditional MMD in processing complex data distribution. Experimental verification shows that the improved MR-MMD method can improve the accuracy of cross-domain fault identification in multi-source domain scenarios. The accuracy rates on the two data sets can reach 95.1% and 89.3% respectively, which is at least 6.8% higher than other domain adaptation methods compared in this embodiment.

[0207] (3) A dynamic weighted loss fusion strategy is proposed. The experimental results show that the proposed method can significantly accelerate the model convergence speed during the training process, save at least 27% of the training time, and achieve a more balanced training effect.

[0208] Although the present invention is disclosed as above, the scope of protection disclosed by the present invention is not limited thereto. Those skilled in the art of the present invention may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the scope of protection of the present invention.

[0209] The documents cited in this invention include:

[0210] 1.Wang Yu-Jing, Xia Lin, Kang Shou-Qiang, Xie Jin-Bao, Wang Qing-Yan.Afault diagnosis method of rolling bearings based on multi-source domainheterogeneous model transfer.Journal of Vibration and Shock, 2023, 42(24): 257-266

[0211] (Wang Yujing, Xia Lin, Kang Shouqiang, Xie Jinbao, Wang Qingyan. Rolling bearing fault diagnosis method based on multi-source domain heterogeneous model migration. Vibration and Shock, 2023, 42(24): 257-266)

[0212] 2. Liu Xiao-Feng, Huang Hong-Sheng, Bo Lin, Chen Bing-Kui. A resonancesparse decomposition method based on maximizing cyclic pulse index and its application. Chinese Journal of Scientific Instrument, 2022, 43(05): 209-217

[0213] (Liu Xiaofeng, Huang Hongsheng, Bolin, Chen Bingkui. Resonance rarefaction method for maximizing cyclic pulse index and its application. Chinese Journal of Scientific Instrument, 2022, 43(05): 209-217)

[0214] 3.Niu G, Liu E, Wang

[0215] 4. Qian G, Liu J. Fault diagnosis based on conditional generative adversarial networks in nuclear power plants. Annals of Nuclear Energy, 2022, 176: 109267

[0216] 5.Zhang Su-Ying, Zhu Xing-Mei, Xu Shu-Qing. Research of motor bearingfault diagnosis based on deep belief network and symmetric dot pattern. Noiseand Vibration Control, 2022, 42(03): 98-104+143

[0217] (Zhang Suying, Zhu Xingmei, Xu Shuqing. Research on motor bearing fault diagnosis based on deep belief network and symmetric point pattern. Noise and Vibration Control, 2022, 42(03): 98-104+143)

[0218] 6. Lu Fei-Yu, Tong Qing-Bin, Jiang Xue-Dong, Xu Jian-Jun, Huo Jing-Yi. Interpretable method for mechanical fault diagnosis based on conditionmetric transfer learning. Chinese Journal of Scientific Instrument, 2024, 45(07): 250-262

[0219] (Lu Feiyu, Tong Qingbin, Jiang Xuedong, Xu Jianjun, Huo Jingyi. An interpretable method for mechanical fault diagnosis based on conditional metric transfer learning. Chinese Journal of Scientific Instrument, 2024, 45(07): 250-262)

[0220] 7.Long M, Cao Y, Wang J, Jordan M. Learning transferable features with deep adaptation networks. In: Proceedings of the 32nd International Conference on Machine Learning. New York, USA: IEEE, 2015.97-105

[0221] 8. Zhu Y, Zhuang F, Wang J, Ke G, Chen J, Bian J, et al. Deep subdomainadaptation network for image classification. IEEE Transactions on NeuralNetworks and Learning Systems, 2021, 32(4): 1713-1722

[0222] 9.Fan Cang-Ning, Liu Peng, Xiao Ting, Zhao Wei, Tang Xiang-Long. A review of deep domain adaptation: general situation and complex situation. ActaAutomatica Sinica, 2021, 47(03): 515-548

[0223] (Fan Cangning, Liu Peng, Xiao Ting, Zhao Wei, Tang Jianglong. A review of deep domain adaptation: general and complex cases. Acta Automatica Sinica, 2021, 47(03): 515-548)

[0224] 10. He Qiu-Yan, Deng Ming-Hua. Survey of universal domain adaptation. Journal of Computer Research and Development, 2024, 61(01): 120-144

[0225] (He Qiuyan, Deng Minghua. A review of general domain adaptation. Journal of Computer Research and Development, 2024, 61(01): 120-144)

[0226] 11.You K, Long M, Cao Z, Wang J, Jordan M.Universal domain adaptation.In: Proceedings of the 32nd IEEE / CVF Conf on Computer Vision and PatternRecognition(CVPR).Piscataway, USA: IEEE, 2019.2715-2724

[0227] 12.She Bo, Liang Wei-Ge, Qin Fen-Qi, Dong Hai-Di. Open set domainadaptation method based on adversarial dual classifiers for faultdiagnosis. Chinese Journal of Scientific Instrument, 2023, 44(07): 325-334

[0228] (She Bo, Liang Weige, Qin Fenqi, Dong Haidi. Open domain adaptive fault diagnosis method based on dual recognizer confrontation. Chinese Journal of Scientific Instrument, 2023, 44(07): 325-334)

[0229] 13.Kang S, Tang X, Wang Y, Wang Q, Xie J. Cross-domain fault diagnosis method for rolling bearings based on contrastive univetsal domain adaptation. ISA Transactions, 2024, 146: 195-207

[0230] 14.Cho J.Synthetic source universal domain adaptation through contrastive learning.Sensors, 2021, 21(22): 7539

[0231] 15. Yin Y, Yang Z, Wu X, Hu H. Pseudo-margin-based universal domain adaptation. Knowledge-Based Systems, 2021, 229(1 / 2): 107315

[0232] 16.Xie Gang, Han Qin, Nie Xiao-Yin, Shi Hui, Zhang Xiao-Hong, Tian Juan. A two-stage domain generalization learning framework for fault diagnosis of bearings. Acta Automatica Sinica, 2024, 50(11): 2271-2285

[0233] (Xie Gang, Han Qin, Nie Xiaoyin, Shi Hui, Zhang Xiaohong, Tian Juan. Bearing fault diagnosis method based on two-stage domain generalization learning framework. Acta Automatica Sinica, 2024, 50(11): 2271-2285)

[0234] 17. Liang P, Tian J, Wang S, Yuan X. Multi-source information jointtransfer diagnosis for rolling bearing with unknown faults via wavelettransform and an improved domain adaptation network. Reliability Engineering & System Safety, 2024, 242: 109788

[0235] 18.Yin Y, Yang Z, Hu H, Wu X.Universal multi-source domain adaptation for image classification.Pattern Recognition, 2022, 121: 108238

[0236] 19. Zhang W, Li X. Data privacy preserving federated transfer learning in machinery fault diagnostics using prior distributions. Structural Health Monitoring, 2022, 21(4): 1329-1344

[0237] 20.Kang Shou-Qiang, Yang Jia-Wei, Wang Yu-Jing, Wang Qing-Yan, Xie Jin-Bao. Fault diagnosis method of rolling bearings under different working conditions based on federated multi-representation domain adaptation. Chinese Journal of Scientific Instrument, 2023, 44(06): 165-176

[0238] (Kang Shouqiang, Yang Jiawei, Wang Yujing, Wang Qingyan, Xie Jinbao. Rolling bearing fault diagnosis method under different working conditions based on federated multi-representation domain adaptation. Chinese Journal of Scientific Instrument, 2023, 44(06): 165-176)

[0239] 21. Zhao K, Hu J, Shao H, Hu J. Federated multi-source domain adversarial adaptation framework for machinery fault diagnosis with data privacy. Reliability Engineering & System Safety, 2023, 236: 109246

[0240] 22.Li Xiang, Fu Chun-Lin, Lei Ya-Guo, Li Nai-Peng, Yang Bin. Federatedtransfer learning method for privacy-preserving collaborative intelligentmachinery fault diagnostics. Journal of Mechanical Engineering, 2023, 59(06): 1-9

[0241] (Li Xiang, Fu Chunlin, Lei Yaguo, Li Naipeng, Yang Bin. Federated transfer learning method for equipment collaborative intelligent fault diagnosis with guaranteed data privacy. Journal of Mechanical Engineering, 2023, 59(06): 1-9)

[0242] 23. Zhang Hong-Lei, Li Yi-Dong, Wu Jun, Chen Nai-Yue, Dong Hai-Rong.Asurvey on privacy-preserving federated recommender systems. Acta Automatica Sinica, 2022, 48(09): 2142-2163

[0243] (Zhang Honglei, Li Yidong, Wu Jun, Chen Naiyue, Dong Hairong. A review of privacy-preserving federated recommendation algorithms. Acta Automatica Sinica, 2022, 48(09): 2142-2163)

[0244] 24.McMahan B,Moore E,Ramage D,Hampson S,Arcas B.Communication-efficient learning of deep networks from decentralized data.In:Proceedings ofArtificial Intelligence and Statistics.New York,USA:IEEE,2017.1273-1282

[0245] 25.Hardy R.Multiquadric equations of topography and other irregularsurfaces.Journal of geophysicalresearch,1971,76(8):1905-1915

[0246] 26.LOPARO,K A.Bearing data center.Cleveland:Case Western ReserveUniversity,2013

[0247] 27.Ganin Y,Ustinova E,Ajakan H,Germain P,Larochelle H,Laviolette F,etal.Domain-adversarial training of neural networks.The Journal of MachineLearning Research,2016,17(1):2096-2030

[0248] 28.Kan S,He Z,Cen Y,Li Y,Mladenovic V,He Z.Contrastive bayesiananalysis for deep metric learning.IEEE Transactions on Pattern Analysis andMachine Intelligence,2022,45(6):7220-7238

[0249] 29.Zhang W,Li X.Federated transfer learning for intelligent faultdiagnostics using deep adversarial networks with data privacy.IEEE / ASMETransactions on Mechatronics,2021,27(1):430-439

[0250] 30.Lessmeier C,Kimotho J K,Zimmer D,Sextro W.Condition monitoring ofbearing damage in electromechanical drive systems by using motor currentsignals of electric motors:A benchmark data set for data-drivenclassification.PHM Society European Conference.Bilbao,Spain:PHMSociety.2016.3(1)。

Claims

1. A bearing unsupervised fault diagnosis method based on spatial differences in multi-user private data labels, characterized by: The steps include: Step 1: The client collects rolling bearing fault data under different working conditions as the local data set of the source client, and randomly selects unlabeled data of unknown fault type under another working condition as the local data set of the target client; each client receives the network model sent by the server as the initial model of the local client; Step 2: Extract and save the features of local data, and encrypt the extracted features; Step 3: The encrypted features are transmitted to the central server, which decrypts and aggregates the ciphertext. The central server uses the universal domain adaptation model to obtain the MR-MMD domain adaptation loss, domain discrimination loss, and classification loss, and introduces edge vectors to identify unknown class samples in the target domain. The weight distribution of each loss is adaptively adjusted in real time through dynamic adaptive weighting. The resulting total loss is sent to each client to update the local model. Step 4: Repeat steps 2 to 3 above until the predetermined number of iterations is reached, and then send the central classifier model parameters to the target client to perform unsupervised fault diagnosis on the bearing.

2. The unsupervised bearing fault diagnosis method based on multi-user private data label space difference according to claim 1 is characterized in that: The network model sent by the server in step 1 is the ResNet50 model. The model gradually extracts the spatial and semantic features of the data from local to global through a combination of convolution kernels and pooling layers. At the same time, regularization is used to further improve the stability and convergence speed of training.

3. The unsupervised bearing fault diagnosis method based on multi-user private data label space difference according to claim 2 is characterized in that: In step 2, random mapping is used to encrypt the extracted features locally on the client.

4. The unsupervised bearing fault diagnosis method based on multi-user private data label space difference according to claim 3 is characterized in that: The MR-MMD domain adaptation loss, domain discrimination loss, and classification loss described in step 3 are: MR-MMD domain adaptation loss E M for: Among them, ψ i is the optimization parameter, J is the number of source clients, For the distribution Next, find the mathematical expectation of x, where x is the input sample, p i For source client data distribution, is the source client label space, F(·) is the L2 normalized feature, s i For the source client, is the i-th source client sample, For the probability distribution qC t Next, we find the mathematical expectation of x, q is the target client data distribution, t is the target client, C t is the target client label space, x t For the target client sample, is the 2-norm in the Hilbert space H; Domain discrimination loss: in, ζ and γ are optimization parameters, E is the expectation, and D is the domain discriminator function; Classifier loss: Among them, L(.,.) represents the cross entropy loss, G is the classifier function, y represents the label, ε i and φ are optimization parameters.

5. The unsupervised bearing fault diagnosis method based on multi-user private data label space difference according to claim 4 is characterized in that: In step 3, the edge vector is introduced to identify unknown class samples in the target domain, specifically: Predictors for target domain samples x Defined as: Among them, f(x, y) represents the classification probability that sample x belongs to the yth class, and the pseudo label of sample x is expressed as The confidence of the pseudo label is determined by the boundary threshold. is the classification probability that the sample x belongs to the pseudo label; Then, the pseudo labels are extracted from the target client, the pseudo edge vectors are calculated, and the final recognition results are output; calculate: Where w0 represents the threshold used to distinguish samples belonging to the unknown class from samples belonging to the known class; if the value of the pseudo edge is greater than the threshold, the sample x is classified as belonging to the known class; otherwise, it is classified as the unknown class.

6. The unsupervised bearing fault diagnosis method based on multi-user private data label space differences according to claim 5 is characterized in that: The dynamic adaptive weighting described in step 3 is used to adaptively adjust the weight distribution of each loss in real time, specifically: E total =α1E M +α2E D +α3E G Among them, α1, α2 and α3 are adaptive weight coefficients; By calculating the mean and standard deviation of the gradient, we can dynamically determine whether the gradient of the loss term deviates significantly from the overall level. When the gradient is too large, we reduce the weight to avoid over-optimization. When the gradient is too small, we increase the weight to strengthen the optimization.

7. A bearing unsupervised fault diagnosis system based on spatial differences in multi-user private data labels, characterized by: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 6 above, and executes the steps of the above-mentioned unsupervised bearing fault diagnosis method based on multi-user privacy data label space differences during operation.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is configured to implement the steps of the bearing unsupervised fault diagnosis method based on multi-user private data label space differences according to any one of claims 1 to 6 when called by a processor.

Images