A federated model transfer learning method and system for fault diagnosis of different specifications of rolling bearings

Abstract

The application provides a federal model transfer learning method and system for fault diagnosis of different specifications of rolling bearings, relates to the technical field of rolling bearing fault diagnosis, and aims to solve the problem that the existing federal learning method has poor fault diagnosis effect on different specifications of rolling bearings. The technical points of the method include: performing short-time Fourier transform on a plurality of user vibration data to construct a time-frequency graph data set; training a local model by each user and uploading model parameters to a server, while introducing a difference update and parameter sparsification algorithm to improve the local model parameter transmission strategy in federal learning; the server aggregates model parameters by using a federal average algorithm and updates the local model, and after iteration, a shared model for transfer learning is established; a layer-by-layer thawing strategy is proposed to retain part of the parameters of the shared model and send them to each user, the shared model is fine-tuned by using local data, and a personalized model suitable for each user is obtained. The application has high accuracy and good generalization performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of rolling bearing fault diagnosis, and particularly relates to a federated model transfer learning method and system for fault diagnosis of rolling bearings of different specifications. BACKGROUND

[0002] As one of the key components of rotating machinery, rolling bearings are widely used in industrial production, and timely fault diagnosis can effectively prevent major accidents [1] In practical engineering, rolling bearings usually work in complex and harsh environments, resulting in a lack of vibration data for some specifications of rolling bearings [2] Therefore, it is of great significance to establish a fault diagnosis model for other specifications of rolling bearings based on the prior knowledge of a certain specification of rolling bearings.

[0003] In recent years, deep learning has been widely applied in the field of bearing fault diagnosis due to its strong adaptive feature extraction capability [3] Literature [4] proposes an improved residual network, which introduces SENet into the deep residual network to realize rolling bearing fault diagnosis. Literature [5] uses one-dimensional convolutional neural network to classify the faults of aero-engine bearings, with high diagnostic accuracy. Literature [6] proposes an improved SqueezeNet model, which replaces one 7x7 convolution kernel with three 3x3 convolution kernels, and uses helicopter bearing vibration data to prove the effectiveness of the method. Literature [7] uses one-dimensional convolution kernels of different sizes to extract multi-scale features of the original vibration signal, and realizes rolling bearing fault diagnosis through multi-layer network learning. Literature [8] uses one-dimensional and two-dimensional convolutional neural networks to extract features and construct new time-frequency features, and successfully applies them to locomotive bearing fault diagnosis.

[0004] The deep learning-based fault diagnosis method usually needs a large number of known label sample sets in the model training process. Due to the difficulty in disassembly and vibration signal collection, the number of known label samples of some specifications of bearings is scarce in actual work, and the established fault diagnosis model is difficult to achieve ideal effect.

[0005] To solve the problems of deep learning-based fault diagnosis methods, many scholars have shifted their focus to the field of transfer learning, which aims to apply known knowledge from one task to a similar task. To solve the problems of complex and variable working conditions of seawater pumps and difficult fault data collection, a fault diagnosis method based on parameter transfer is proposed in [9], and the effectiveness of the method is verified on three seawater pump datasets. An improved transfer learning method is proposed in

[10] , which uses a global mean pooling layer to classify bearing faults and achieves fault diagnosis of rolling bearings with less data. In

[11] , the Gram angle difference field is used to convert one-dimensional vibration signals into two-dimensional fault maps, and then model transfer learning is used for bearing fault diagnosis, which has a high diagnosis accuracy. In

[12] , an improved convolutional neural network is trained using source domain data, the model parameters are transferred to the target domain network and fine-tuned, and the effectiveness is verified using a public bearing dataset. In

[13] , four methods are proposed to convert original vibration signals into image samples, and transfer learning is used to achieve bearing fault diagnosis under noisy conditions.

[0006] However, as data security issues have gradually received attention, governments around the world have enacted a number of relevant laws to strengthen data privacy management. The General Data Protection Regulation introduced by the European Union in 2016 and the California Consumer Privacy Act enacted by the United States in 2018 both emphasize the protection of personal privacy data. China's Personal Information Protection Law, enacted in 2021, also clearly states that personal information should be handled with the consent of the individual. At the same time, users do not want to share personal data due to data privacy issues, and data often exists in the form of an island in multiple users, making it difficult to aggregate into big data. Therefore, the transfer learning-based fault diagnosis method has the problem of a lack of prior knowledge in the source domain, and it is difficult to build a pre-trained model in the source domain.

[0007] Federal learning is a distributed machine learning method based on the above background, as a major breakthrough in the field of artificial intelligence, according to the idea of "data different model dynamic, data available invisible", aims to jointly model multiple parties' data, break down the "data barrier", and achieve "common prosperity" of multiple parties

[14] A weighted strategy in the federated learning model aggregation process is proposed in document

[15] to improve the quality of the global model, and the effectiveness of the proposed method is verified using a bearing dataset. Document

[16] designs a bearing fault diagnosis method based on federated learning, which realizes the fault diagnosis of rolling bearings in the island data situation. Document

[17] proposes a fault diagnosis method based on federated learning and convolutional neural network, which has high diagnosis accuracy on rolling bearing dataset. Document

[18] proposes a multi-condition fault diagnosis federated learning system based on dynamic routing technology, and verifies the superiority of the proposed method based on rolling bearing dataset. Document

[19] proposes a dynamic verification scheme in the federated learning framework, which aims to improve the robustness of the global model to low-quality local data, and achieves good diagnosis effect on bearing and bogie datasets.

[0008] Due to the large difference between the vibration data of different specifications of rolling bearings, directly using federated learning method to aggregate multiple user different specifications of rolling bearing vibration data is easy to lead to poor model diagnosis effect; and the user directly models using local island data will not be able to establish an effective fault diagnosis model due to the small amount of known label data. SUMMARY

[0009] In view of the above problems, the present application proposes a federated model transfer learning method and system for fault diagnosis of different specifications of rolling bearings, to solve the problem that the existing federated learning method has poor fault diagnosis effect for different specifications of rolling bearings.

[0010] According to an aspect of the present application, a federated model transfer learning method for fault diagnosis of different specifications of rolling bearings is provided, which comprises the following steps:

[0011] (1) Local dataset and local model construction

[0012] Select different types of rolling bearing vibration data as the local dataset of each user, and obtain the user local time-frequency pattern sample set through short-time Fourier transform; train the convolutional neural network using the local time-frequency pattern sample set to obtain a plurality of user local models;

[0013] (2) Shared model construction

[0014] Using an improved parameter transfer strategy, the user uploads the trained local model parameters to the server, the server aggregates the uploaded parameters using federated averaging algorithm and updates the local model, and the process is repeated until the maximum federated iteration number is reached, to obtain a shared model for transfer learning;

[0015] (3) Personalized model construction

[0016] The server sends the shared model for transfer learning to each user, decides which layer parameters of the shared model to keep according to the layer-by-layer unfreezing strategy, and fine-tunes the shared model using the local data set to obtain a personalized model applicable to the local data set of the user.

[0017] Further, for the improved parameter transmission strategy in (2) shared model construction, a parameter sparsification algorithm is introduced to improve the parameter transmission strategy in the federal learning process to solve the privacy leakage when the user uploads the parameters.

[0018] Further, the specific steps of the parameter sparsification algorithm used in the federal learning include:

[0019] (21) After the user trains the local model using the local data, the difference between the global model parameters initially issued in this round and the local model parameters trained in this round is calculated layer by layer, and the difference of each layer is a matrix;

[0020] (22) The user locally generates a matrix of the same size as the model parameters, which is composed of 0 and 1, and the number of 1s is determined by the parameter sparsification degree specified in advance;

[0021] (23) Multiply the difference matrix obtained in step (21) with the parameter sparsification matrix generated in step (22) at the corresponding positions to obtain a new difference matrix;

[0022] (24) Send the new difference matrix to the server, and the server collects the difference matrices of all users and takes their average to add to the global model parameter matrix of the last round to form the global model of the next round.

[0023] Further, for the federal averaging algorithm in (2) shared model construction, it is a model parameter aggregation method in federal learning, and its expression is as shown in the following formula:

[0024]

[0025] wherein, Gk j represents the local model parameter of the kth user in the jth federal, G j+1 G represents the global model parameter of the j+1th round.

[0026] Further, for the layer-by-layer unfreezing strategy in (3) personalized model construction, the specific method is: except for the parameters of the fully connected layer, the other model parameters of the shared model for transfer learning are migrated to the target domain network as the initialization parameters of the target domain network, and the network is unfreezed layer by layer from the last convolutional layer to the front, and the network is fine-tuned using the known labeled samples of the target domain.

[0027] According to another aspect of the present application, a federal model transfer learning system for fault diagnosis of different specifications of rolling bearings is provided, which comprises:

[0028] a local data set and a local model construction module, configured to select vibration data of different types of rolling bearings as a local data set of each user, and obtain a local time-frequency pattern sample set of the user by short-time Fourier transform; and train a convolutional neural network by using the local time-frequency pattern sample set to obtain a plurality of local models of the user;

[0029] a shared model construction module, configured to use an improved parameter transmission strategy, upload the trained local model parameters of the user to a server, aggregate the uploaded parameters of the users by using a federated averaging algorithm, and update the local model, repeatedly until a maximum number of federated iterations is reached, to obtain a shared model for transfer learning;

[0030] a personalized model construction module, configured to send the shared model for transfer learning to each user by the server, determine which layer parameters of the shared model to keep according to a layer-by-layer unfreezing strategy, and fine-tune the shared model by using the local data set to obtain a personalized model suitable for the local data set of the user.

[0031] Further, the improved parameter transmission strategy in the shared model construction module is to introduce a parameter sparsification algorithm to improve the parameter transmission strategy in the federated learning process to solve the privacy leakage when the user uploads the parameters.

[0032] Further, the specific steps of the parameter sparsification algorithm used in the shared model construction module in the federated learning include:

[0033] (21) After the user trains the local model by using the local data, the difference between the global model parameters initially issued in the current round and the local model parameters trained in the current round is calculated layer by layer, and the difference of each layer is a matrix;

[0034] (22) The user locally generates a matrix with the same size as the model parameters, and the matrix is composed of 0 and 1, and the number of 1 is determined by the parameter sparsification degree specified in advance;

[0035] (23) The difference matrix obtained in step (21) is multiplied with the parameter sparsification matrix generated in step (22) at the corresponding positions to obtain a new difference matrix;

[0036] (24) The new difference matrix is sent to the server, and the server collects the difference matrices of all users and takes their average to add to the global model parameter matrix in the last round to form the global model in the next round.

[0037] Further, the federated averaging algorithm in the shared model construction module is a model parameter aggregation method in the federated learning, and its expression is as shown in the following formula:

[0038]

[0039] wherein, denotes the local model parameters of the kth user in the jth round of federation, G j+1 denotes the global model parameters of the j+1th round.

[0040] Further, the layer-by-layer thawing strategy in the personalized model construction module is used for network fine-tuning, and the specific method is as follows: except for the parameters of the full connection layer, other model parameters of the shared model used for transfer learning are migrated to the target domain network as the initialization parameters of the target domain network, the network is thawed layer by layer from the last convolutional layer, and the network is fine-tuned using the known label samples of the target domain.

[0041] The beneficial technical effects of the present application are:

[0042] The present application is based on a federal model migration framework, uses a convolutional neural network to extract vibration signal features, and constructs a user local model; an improved parameter strategy is used to reduce the number of local model parameters and improve the security of the parameter transmission process; finally, a plurality of users cooperatively construct a shared model for transfer learning and send it to the user, the user fine-tunes the shared model using local data, and realizes fault diagnosis of different specifications of rolling bearings. Compared with the user using local data to independently model, the diagnostic accuracy can be improved by 1.2% to 7.1%, and compared with the user directly performing federation, it can be improved by at least 9.5%, proving the effectiveness of the proposed method. The generalization experiment of the NICE bearing dataset proves that the proposed method has strong generalization performance. BRIEF DESCRIPTION OF DRAWINGS

[0043] The present application can be better understood by reference to the description given in conjunction with the following drawings, which are incorporated in and constitute a part of this specification, and are intended to further illustrate preferred embodiments of the present application and to explain the principles and advantages thereof.

[0044] Figure 1 is a convolutional neural network structure diagram;

[0045] Figure 2 is a pooling method schematic diagram;

[0046] Figure 3 is a federal learning flowchart;

[0047] Figure 4 is a model migration strategy schematic diagram;

[0048] Figure 5 is a flowchart of a federal model transfer learning method for fault diagnosis of rolling bearings of different specifications according to an embodiment of the present application;

[0049] Figure 6 is a bearing test bench schematic diagram;

[0050] Figure 7 is a schematic diagram of a time domain vibration signal interception method;

[0051] Figure 8 is a radar chart of the accuracy of different reserved layer numbers;

[0052] Figure 9 is a line chart of different parameter sparsification degrees;

[0053] Figure 10 is a comparison experiment chart of the method and the traditional method;

[0054] Figure 11 is a setting experiment chart of different sample numbers;wherein (a) corresponds to 20 samples of each category of local data;(b) corresponds to 10 samples of each category of local data;(c) corresponds to 5 samples of each category of local data;

[0055] Figure 12 is a schematic diagram of a rolling bearing fault diagnosis confusion matrix;wherein (a) corresponds to target domain labeled data independent modeling;(b) corresponds to the application. DETAILED DESCRIPTION

[0056] In order for those skilled in the art to better understand the application scheme, the exemplary embodiments or examples of the application will be described in conjunction with the accompanying drawings in the following. Obviously, the described embodiments or examples are only a part of the embodiments or examples of the application, not all. Based on the embodiments or examples in the application, all other embodiments or examples obtained by those skilled in the art without creative labor should belong to the scope of protection of the application.

[0057] The embodiment of the application provides a federal model transfer learning method for fault diagnosis of rolling bearings of different specifications, which will be described in detail below.

[0058] 1. Convolutional neural network

[0059] In recent years, the convolutional neural network (CNN) has attracted widespread attention in the fields of image classification, target recognition and the like due to its good ability to extract picture features. Compared with the traditional neural network, the convolutional neural network can quickly extract picture features and reduce the number of model parameters through parameter sharing, sparse connection and the like, greatly improving the training speed

[20] . The common convolutional neural network is composed of an input layer, a convolutional layer, a pooling layer, a full connection layer and an output layer, as shown in Figure 1 .

[0060] 1.1 Convolutional layer

[0061] As the core component of a convolutional neural network, the convolutional layer is primarily responsible for extracting features from the input image. It extracts local features by using multiple convolutional kernels of pre-determined size and weighted sums with corresponding regions of the input image. The formula for feature extraction by the convolutional layer is shown in equation (1).

[0062]

[0063] in, x is the m-th convolutional kernel in the n-th convolutional layer; n This is the input for this layer; * represents a convolution operation; The bias of the m-th convolutional kernel in layer n; For input x n Features extracted by this convolution kernel.

[0064] 1.2 Pooling Layer

[0065] In CNNs, pooling layers are primarily responsible for downsampling the features extracted by convolutional layers, preserving important information in the image while reducing the number of network parameters. Common pooling methods include max pooling and average pooling. These two methods are described below. Figure 2 As shown.

[0066] 1.3 Fully Connected Layer

[0067] Fully connected layers are typically placed at the end of the entire CNN. They can fit the local features extracted by the convolutional layers into global features and play a classification role. The formula for the output of the fully connected layer is shown in Equation (2).

[0068] y(x)=f(ωx+b) (2)

[0069] Where x represents the input data of the fully connected layer, ω is the weight, b is the bias, f(·) is the activation function, and y(x) is the output of the fully connected layer.

[0070] 2. Federated Transfer Learning

[0071] 2.1 Federated Learning

[0072] Federated learning frameworks typically consist of multiple users and an aggregation server. The aggregation server does not disclose user information, and user local data remains local and is not shared with other users. A global model is collaboratively built only by sharing local model parameters. Assume there are K users participating in the federation, and each user's local dataset is D. k If k = 1, 2, ..., K, then the federated learning modeling process is as follows.

[0073] First, all modeling tasks can be viewed as an optimization process of the loss function, that is:

[0074] min(f(w)) (3)

[0075] Where f(w) is the loss function for the current task, and w is the current model parameter.

[0076] Traditional centralized model training gathers all datasets together, resulting in:

[0077]

[0078] Among them, f i (w) represents the loss for the i-th sample, and n represents the number of samples in the dataset.

[0079] Unlike traditional centralized training processes, federated learning does not require centralized data. Therefore, the objective function for federated learning can be defined as:

[0080]

[0081]

[0082] Among them, F k (w) represents the loss of the k-th user in federated learning, and n k Let be the number of samples for the k-th user.

[0083] Federated learning process flowchart as follows Figure 3 As shown, the server sends the global model to each user as their local model; users train their local model using their local data and upload the model parameters to the server; the server aggregates the local model parameters using an aggregation algorithm, updates the global model, and then sends it back to the users. This process is repeated until the maximum number of federated iterations is reached, resulting in the final federated global model.

[0084] The federated averaging algorithm, as the most commonly used method for aggregating model parameters in federated learning, is expressed as shown in equation (7):

[0085]

[0086] in, G represents the local model parameters of the k-th user in the j-th round of federation. j+1 This represents the global model parameters in the (j+1)th round.

[0087] 2.2 Model Transfer Learning

[0088] Model transfer learning aims to transfer the parameters of a model trained in a source domain to a target domain with similar tasks, and fine-tune the model parameters using a known labeled sample set for the current task to adapt it to the new task. The reason model transfer learning can utilize prior knowledge from the source domain to assist in modeling the target domain and achieve ideal results is fundamentally due to the hierarchical nature of feature extraction in convolutional neural networks. Shallow networks extract common features, while deep networks extract more abstract features. Fine-tuning, a common parameter tuning method in model transfer, leverages the universality of common features, sharing shallow network parameters while adjusting deep network parameters, allowing the source domain network to quickly adapt to the new task.

[0089] The key to fine-tuning is identifying invariants across different datasets, that is, common knowledge between the source and target domains. Therefore, this invention proposes a layer-by-layer unfreezing approach for network fine-tuning. Specifically, the parameters of the shared model (excluding fully connected layers) used for transfer learning are transferred to the target domain network as initialization parameters. This unfreezing process proceeds layer by layer from the last convolutional layer forward, and the network is fine-tuned using known labeled samples from the target domain. The parameter transfer process is as follows: Figure 4 As shown.

[0090] 2.3 Improved parameter transfer strategy

[0091] Federated learning does not require the aggregation of user data; instead, it improves model quality through collaborative modeling by aggregating model parameters. However, some issues remain. Users may not want their private data and local model parameters to be accessed by others, as information leaks could lead to significant financial losses. Therefore, the security of the parameter transfer process in federated learning is a key area worthy of research.

[0092] Parameter sparsity algorithms reduce the number of parameters by generating a matrix of the same size as the parameter matrix, consisting of 0s and 1s, and then multiplying it by the corresponding positions of the original parameter matrix. The sparsity of parameters is the ratio of the number of 1s in the matrix to the sum of the number of 0s and 1s; that is, 90% parameter sparsity means that 90% of the matrix consists of 1s. Based on this idea, this invention introduces a parameter sparsity algorithm to improve the parameter passing strategy in federated learning, addressing the privacy leakage problem when users upload parameters.

[0093] The specific steps of the parameter sparsity algorithm used in this invention within the federated learning framework are as follows:

[0094] (21) After the user trains the local model using local data, the difference between the global model parameters initially issued in this round and the local model parameters trained in this round is calculated layer by layer. The difference for each layer is a matrix.

[0095] (22) The user generates a matrix of the same size as the model parameters locally. The matrix consists of 0s and 1s, and the number of 1s is determined by the predefined sparsity of the parameters.

[0096] (23) Multiply the difference matrix obtained in step (21) with the corresponding positions of the parameter sparsification matrix generated in (22) to obtain a new difference matrix.

[0097] (24) Send the new difference matrix to the server. The server collects the difference matrices of all users, takes the average of them, and adds them to the global model parameter matrix of the previous round to form the global model of the next round.

[0098] 3. Fault diagnosis methods for rolling bearings of different specifications

[0099] The flowchart of the fault diagnosis method for rolling bearings of different specifications based on federated model transfer is as follows: Figure 5 As shown. The specific steps are as follows:

[0100] (1) Construction of local dataset and local model

[0101] Vibration data from different types of rolling bearings were selected as local datasets for each user, and short-time Fourier transform was used to obtain local time-frequency maps. Convolutional neural networks were trained using these local time-frequency map sample sets to obtain multiple local models for each user.

[0102] (2) Sharing Model Construction

[0103] Using the proposed improved parameter transfer strategy, users upload the parameters of their trained local models to the server. The server uses a federated averaging algorithm to aggregate the uploaded parameters and update the local model. This process is repeated until the maximum number of federated iterations is reached, resulting in a shared model for transfer learning.

[0104] (3) Personalized model construction

[0105] The server sends the shared model for transfer learning to each user, decides which layer parameters of the shared model to retain based on the proposed layer-by-layer unfreezing strategy, and fine-tunes the shared model using the local dataset to obtain a personalized model suitable for the user's local dataset.

[0106] 4. Application and Analysis

[0107] 4.1 Experimental Dataset and Experimental Setup

[0108] The experimental data comes from the test bench bearing dataset.

[21] Bearing data experimental platform, such as Figure 6As shown, the experimental platform consists of a motor, torque sensor, power meter, and electronic controller. The drive end uses an SKF6205 bearing, and the fan end uses an SKF6203 bearing. The sampling frequency for both is 12kHz, and the platform operates under four load conditions: 0hp, 1hp, 2hp, and 3hp. The dataset includes four states: normal, inner ring fault, outer ring fault, and rolling element fault. Bearing faults are pitting defects caused by electrical discharge machining (EDM) at each part of the bearing. Each fault type has three defect diameters: 0.1778mm, 0.3556mm, and 0.5334mm, totaling ten categories.

[0109] This invention randomly selects 1024 consecutive points as a sample. For ease of representing various state categories, N represents the normal state, and the other nine fault states are specifically represented as shown in Table 1. Each sample is 1024 points long, and the selection method is to randomly select a segment of 1024 consecutive points, as shown in Table 1. Figure 7 As shown.

[0110] Table 1. Nine Fault States

[0111]

[0112] To more clearly represent the data used in the experiment, Table 2 shows the correspondence between the dataset used in this invention and the bearing specifications, working load, and rotational speed.

[0113] Table 2 shows the correspondence between the dataset and bearing specifications, load, and speed.

[0114]

[0115] This invention sets up 8 different federated tasks, and the user's local data in each task is shown in Table 3. Taking Task 1 as an example, User 1 holds SKF6205 (0hp, 1797rpm) data, User 2 holds SKF6205 (1hp, 1772rpm) data, and User 3 holds SKF6203 (0hp, 1797rpm) data. Each user has 30 training samples per class, totaling 300 samples across 10 classes, and 100 test samples per class, totaling 1000 samples.

[0116] Table 3 User Dataset Settings for Each Task

[0117]

[0118]

[0119] The proposed method uses the ReLU function as the activation function, with 20 iterations of local training, a maximum of 50 iterations of federated training, and 50 local fine-tuning iterations during model transfer. The learning rate (lr) is set to 0.001. The experimental hardware environment consisted of an Intel Xeon W-2123 CPU, 32GB of RAM, and an NVIDIA GeForce GTX1080Ti GPU.

[0120] This experiment utilizes a stack of convolutional layers, pooling layers, and fully connected layers to construct a CNN-based rolling bearing fault diagnosis model as a local model in the federated learning process. The feature extraction layer consists of four convolutional layers and four max-pooling layers, and the classifier consists of fully connected layers. The specific structure is shown in Table 4.

[0121] Table 4 Convolutional Neural Network Parameters

[0122]

[0123] 4.2 Experiment on the Layer-by-Layer Thawing Strategy

[0124] To retain an appropriate number of layers in the shared model and verify the impact of different retained layers on classification performance, a layer-by-layer unfreezing strategy was adopted. That is, after receiving the shared model, the network parameters of some layers out of the 5 layers were retained and frozen, while the unfrozen network parameters were fine-tuned. In this section, experiments were conducted on 8 tasks with 4, 3, and 2 layers retained respectively. Specific experimental results are shown in Table 5. To more clearly illustrate the impact of different retained layers on classification performance, the experimental results are presented in radar chart form, as shown below. Figure 8 As shown.

[0125] Table 5 Comparison of accuracy rates for different retention layers / %

[0126]

[0127] Through observation Figure 8 It can be seen that, for some tasks, the classification performance is basically the same when retaining the first 4 layers of model parameters and when retaining the first 3 layers of model parameters, but for some tasks, the performance is lower when retaining the first 3 layers of model parameters than when retaining the first 4 layers of model parameters; and the performance is lower when retaining the first 2 layers of model parameters than both strategies. In summary, this invention will retain the first 4 layers of shared model parameters for subsequent experiments.

[0128] 4.3 Comparison Experiment of Sparsity with Different Parameters

[0129] To verify the impact of different parameter sparsity on classification performance, appropriate parameter sparsity was selected, and five parameter sparsity levels of 60%, 70%, 80%, 90%, and 100% were used for experimental research.

[0130] Five different parameter sparsity comparison experiments were conducted for tasks 1 to 8 respectively. The experimental results for each task are the average accuracy of three user tests, as shown in Table 6.

[0131] Table 6 Comparison of sparsity under different parameters / %

[0132]

[0133] To more clearly illustrate the impact of different sparsity parameters on classification performance, the experimental results are presented in the form of a line graph, such as... Figure 9 As shown.

[0134] Through analysis Figure 9 It can be seen that when the parameter sparsity is 90%, the classification accuracy hardly decreases compared to 100%, while the accuracy decreases to varying degrees when the parameter sparsity is 80%, 70%, and 60%. Therefore, subsequent experiments will use a parameter sparsity of 90%.

[0135] 4.4 Federal Comparison Experiment

[0136] To further verify the effectiveness and superiority of the proposed method, the federated transfer learning method proposed in this invention is compared with Method 1 and Method 2. Method 1 involves users collaboratively building a fault diagnosis model using federated learning, but without using local datasets to fine-tune the shared model, and the user dataset settings for Method 1 are the same as those for the proposed method; Method 2 involves users building a fault diagnosis model using only local data. The experiments in this section test the model classification performance on 8 different tasks. The experimental results are the average test accuracy for users 1, 2, and 3, as shown below. Figure 10 As shown. By Figure 10 It can be seen that the diagnostic accuracy of the proposed method can be improved by about 9.5% compared with method 1, and by 1.2% to 2.3% compared with method 2.

[0137] In summary, the analysis shows that Method 2 suffers from insufficient effective information to build a high-quality fault diagnosis model due to the small number of locally labeled samples. In contrast, Method 1, which directly adopts federated learning, has a lower diagnostic effect than Method 2 because of the large differences in the distribution of vibration data between different specifications of rolling bearings.

[0138] 4.5 Extended Experiment

[0139] 4.5.1 Experiments with different sample sizes on local datasets

[0140] To verify the effectiveness of the proposed method under other sample size settings, the scenario in Section 4.4 where each user holds 300 training samples was expanded to include users holding 200, 100, and 50 training samples, with all other settings remaining the same as in Section 4.4. The experimental results are the average test accuracy for users 1, 2, and 3. Figure 11 As shown. Figure 11 (a) Experimental results with 20 samples per class for each user show that the accuracy of the proposed method is about 10% higher on average than that of method 1, and up to 2.7% to 3.4% higher than that of method 2. Figure 11 (b) Experimental results with 10 samples per class for each user show that the accuracy of the proposed method is about 10% higher on average than that of method 1, and 2.2% to 4.5% higher than that of method 2. Figure 11 (c) Experimental results with 5 samples per class for each user. It can be observed that the accuracy of the proposed method is about 15% higher than that of method 1 and 4.1 to 7.1% higher than that of method 2.

[0141] In summary, the proposed method shows improvements over methods 1 and 2 to varying degrees under different sample sizes, especially when the local data sample size is small, the improvement in classification accuracy is more significant. 4.5.2 Generalization Experiment

[0142] To verify the strong generalization ability of the proposed method, the datasets set in Tasks 1 to 4 were used as the source domain, and the NICE bearing dataset was used as the target domain. The NICE bearing vibration data used in this experiment included three states: inner race fault (IR), outer race fault (OR), and normal (N), with a sampling frequency of 48.828 kHz and a rotational speed of 25 rpm. This experiment used the shared model obtained from the collaborative training of the first four tasks in Section 4.4 as the source domain pre-trained model, the NICE bearing dataset with 10 samples per class (30 samples total) as the target domain training data, and the NICE bearing dataset with 100 samples per class (300 samples total) as the target domain test data. Specific task settings are shown in Table 7. This section's experiment uses NICE data to fine-tune the shared model as Method 1, and uses only NICE bearing data to build a fault diagnosis model as Method 2. Experimental results are shown in Table 8.

[0143] Table 7 Generalization Experiment Task Settings

[0144]

[0145] Table 8. Accuracy of Generalization Experiment / %

[0146]

[0147] As shown in Table 8, the experimental results reveal that independent modeling using 10 samples of NICE bearing data per class yielded an accuracy of only 89.2%, while the proposed method demonstrated varying degrees of improvement across the four different tasks. To more intuitively illustrate the effectiveness of the proposed method in diagnosing rolling bearing faults of different specifications, a multi-class confusion matrix was introduced to analyze the diagnostic results. Taking the experimental results of Task 9 as an example, the confusion matrix is ​​plotted as follows: Figure 12 As shown.

[0148] This invention proposes a layer-by-layer unfreezing and retention strategy, preserving parameters of some layers of the shared model while fine-tuning the remaining layers using user-local data. Comparative experiments were conducted with parameters of the first few layers of the shared model retained, ultimately determining that retaining the parameters of the first four layers and fine-tuning the parameters of the last layer yielded the best classification results. The invention also proposes improving the parameter transfer strategy of federated learning using difference updates and parameter sparsity algorithms. Through comparative experiments with five different parameter sparsity degrees, 90% was selected as the parameter sparsity degree for improving the parameter transfer strategy, ensuring that the model's classification performance is not affected while further enhancing the reliability of the parameter transfer process. Furthermore, a novel fault diagnosis framework for rolling bearings of different specifications based on federated model transfer is proposed. This framework can effectively establish fault diagnosis models even when multi-user bearing data is isolated and privacy-preserving, and when labeled vibration data for some bearing specifications is scarce. Compared to users independently modeling using local data, the proposed method improves diagnostic accuracy by 1.2%–7.1%, and compared to users directly performing federated learning, it improves accuracy by at least 9.5%, demonstrating the effectiveness of the proposed method. Generalization experiments on the NICE bearing dataset demonstrate the strong generalization performance of the proposed method.

[0149] Another embodiment of the present invention provides a federated model transfer learning system for fault diagnosis of rolling bearings of different specifications, the system comprising:

[0150] The local dataset and local model building module is used to select vibration data of different types of rolling bearings as local datasets for each user, and obtain a local time-frequency map sample set for each user through short-time Fourier transform; and use the local time-frequency map sample set to train a convolutional neural network to obtain multiple local models for each user.

[0151] The shared model building module is used to utilize an improved parameter passing strategy. Users upload the parameters of their trained local models to the server. The server uses a federated averaging algorithm to aggregate the parameters uploaded by users and update the local model. This process is repeated until the maximum number of federated iterations is reached to obtain a shared model for transfer learning.

[0152] The personalized model building module is used by the server to send the shared model for transfer learning to each user, decide which layer parameters of the shared model to retain based on the layer-by-layer unfreezing strategy, and fine-tune the shared model using the local dataset to obtain a personalized model suitable for the user's local dataset.

[0153] In this embodiment, preferably, the improved parameter passing strategy in the shared model construction module is: to introduce a parameter sparsity algorithm to improve the parameter passing strategy in the federated learning process, so as to solve the privacy leakage when users upload parameters.

[0154] In this embodiment, preferably, the specific steps of the parameter sparsity algorithm used in the shared model construction module in federated learning include:

[0155] (21) After the user trains the local model using local data, the difference between the global model parameters initially issued in this round and the local model parameters trained in this round is calculated layer by layer. The difference for each layer is a matrix.

[0156] (22) The user generates a matrix of the same size as the model parameters locally. The matrix consists of 0s and 1s, and the number of 1s is determined by the predefined sparsity of the parameters.

[0157] (23) Multiply the difference matrix obtained in step (21) with the corresponding positions of the parameter sparsification matrix generated in step (22) to obtain a new difference matrix;

[0158] (24) Send the new difference matrix to the server. The server collects the difference matrices of all users, takes the average of them, and adds them to the global model parameter matrix of the previous round to form the global model of the next round.

[0159] In this embodiment, preferably, the federated averaging algorithm in the shared model building module is a model parameter aggregation method in federated learning, and its expression is shown in the following formula:

[0160]

[0161] in, G represents the local model parameters of the k-th user in the j-th round of federation. j+1 This represents the global model parameters in the (j+1)th round.

[0162] In this embodiment, preferably, the layer-by-layer unfreezing strategy in the personalized model construction module is used to fine-tune the network. The specific method is as follows: except for the parameters of the fully connected layer, the other model parameters of the shared model used for transfer learning are transferred to the target domain network as the initialization parameters of the target domain network. The network is unfrozen layer by layer from the last convolutional layer forward, and the network is fine-tuned using known label samples of the target domain.

[0163] The functionality of the federated model transfer learning system for fault diagnosis of rolling bearings of different specifications described in this embodiment can be explained by the aforementioned federated model transfer learning method for fault diagnosis of rolling bearings of different specifications. Therefore, for the parts not detailed in this embodiment, please refer to the above method embodiments, and they will not be repeated here.

[0164] Although the invention has been described with respect to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

[0165] The references cited in this invention are as follows:

[0166] [1] Kang Shouqiang, Zou Jiayue, Wang Yujing, et al. Fault diagnosis method for rolling bearings under variable load based on unsupervised feature alignment [J]. Proceedings of the CSEE, 2020, 40(01):274-281+393.

[0167] [2] Wang Yujing, Lü Haiyan, Kang Shouqiang, et al. Fault diagnosis methods for different types of rolling bearings [J]. Proceedings of the CSEE, 2021, 41(01): 267-276+416.

[0168] [3]Zhang S, Zhang S, Wang B, et al.Deep learning algorithms for bearingfault diagnostics—A comprehensive review[J]. IEEE Access, 2020, 8: 29857-29881.

[0169] [4] Tian Kewei, Dong Shaojiang, Jiang Baojun, et al. Bearing fault diagnosis method based on improved deep residual network [J]. Vibration and Shock, 2021, 40(20): 247-254.

[0170] [5] Ding Xue, Deng Aidong, Li Jing, et al. Fault diagnosis of rolling bearings based on multi-scale and attention mechanism [J]. Journal of Southeast University (Natural Science Edition), 2021, 41(01):267-276+416.

[0171] [6] Yu Zhifeng, Xiong Bangshu, Li Xinmin, et al. Fault diagnosis of helicopter rolling bearings based on improved SqueezeNet [J / OL]. Journal of Aerospace Power: 1-10 [2022-04-06].

[0172] [7] Fu L, Zhang L, Tao J. An improved deep convolutional neural network with multiscale convolution kernels for fault diagnosis of bearing rolling [C]. / / IOP Conference Series: Materials Science and Engineering. IOP Publishing, 2021, 1043(5):052021.

[0173] [8] Zhang Long, Zhen Canzhuang, Xiong Guoliang, et al. Fault diagnosis of locomotive bearings based on deep time-frequency features [J]. Journal of Transportation Engineering, 2021, 21(06): 247-258.

[0174] [9] Cui Shiyu, Zhu Zhiyu. Fault diagnosis of seawater pump based on parameter transfer and one-dimensional convolutional neural network [J]. Vibration and Shock, 2021, 40(24): 180-189.

[0175]

[10] Zhang Xining, Yu Di, Liu Shuyu. Research on small sample bearing fault diagnosis method based on transfer learning [J]. Journal of Xi'an Jiaotong University, 2021, 55(10):30-37.

[0176]

[11] Hou Dongxiao, Mu Jintao, Fang Cheng, et al. Fault diagnosis of variable speed bearings based on GADF and ResNet34 with transfer learning [J]. Journal of Northeastern University (Natural Science Edition), 2022, 43(03):383-389.

[0177]

[12] Jiang F, Shen

[0178]

[13] Fan H,Xue C,Zhang X,et al.Vibration images-driven fault diagnosisbased on CNN and transfer learning of rolling bearing under strong noise[J].Shock and Vibration,2021,2021.

[0179]

[14] Yang Q,Liu Y,Cheng Y,et al.Federated learning[J].SynthesisLectures on Artificial Intelligence and Machine Learning,2019,13(3):1-207.

[0180]

[15] Geng D Q,He H W,Lan X C,et al.Bearing fault diagnosis based onimproved federated learning algorithm[J].Computing,2021:1-19.

[0181]

[16] Li Z,Li Z,Li Y,et al.An Intelligent Diagnosis Method for MachineFault Based on Federated Learning[J]. Applied Sciences,2021,11(24):12117.

[0182]

[17] Zhang Z,Xu X,Gong W,et al.Efficient federated convolutionalneural network with information fusion for rolling bearing fault diagnosis[J].Control Engineering Practice,2021,116:104913.

[0183]

[18] Zhou F,Zhang Z,Li S.Research on federated learning method forfault diagnosis in multiple working conditions[J].Complex EngineeringSystems,2021,1(2):7.

[0184]

[19] Zhang W,Li X,Ma H.Federated learning for machinery faultdiagnosis with dynamic validation and self-supervision[J].Knowledge-BaseSystems,2021,213:106679.

[0185]

[20] Alzubaidi L,Zhang J,Humaidi A J,et al.Review of deep learning:Concepts,CNN architectures,challenges,applications,future directions[J].Journal of big Data,2021,8(1):1-74.

[0186]

[21] KENNETH A Loparo.Bearing data center[M / OL].Cleveland:Case WesternReserve University,2013.http: / / csegroups.case.edu / bearingdatacenter / home。

Claims

1. A federated model transfer learning method for fault diagnosis of rolling bearings of different specifications, characterized in that, Includes the following steps: Local dataset and local model construction: Vibration data of different types of rolling bearings are selected as local datasets for each user, and user local time-frequency graph sample sets are obtained through short-time Fourier transform; convolutional neural networks are trained using local time-frequency graph sample sets to obtain multiple user local models; Shared Model Construction: Utilizing an improved parameter transfer strategy, users upload their trained local model parameters to the server. The server uses a federated averaging algorithm to aggregate the uploaded parameters and update the local model, repeating this process until the maximum number of federated iterations is reached, resulting in a shared model for transfer learning. The improved parameter transfer strategy involves introducing a parameter sparsity algorithm to improve the parameter transfer strategy during federated learning, thereby addressing privacy leaks when users upload parameters. This includes: After the user trains the local model using local data, the difference between the initial global model parameters and the locally trained model parameters is calculated layer by layer. The difference for each layer is a matrix, resulting in a difference matrix. The user also generates a matrix of the same size as the model parameters, consisting of 0s and 1s, with the number of 1s determined by a predefined parameter sparsity, resulting in a parameter sparsity matrix. The difference matrix and the parameter sparsity matrix are multiplied at corresponding positions to obtain a new difference matrix. The new difference matrix is ​​then sent to the server, which collects the difference matrices from all users, takes the average, and adds it to the global model parameter matrix from the previous round to form the global model for the next round. Personalized model building: The server sends the shared model for transfer learning to each user, decides which layer parameters of the shared model to retain based on the layer-by-layer unfreezing strategy, and fine-tunes the shared model using the local dataset to obtain a personalized model suitable for the user's local dataset.

2. The federated model transfer learning method for fault diagnosis of rolling bearings of different specifications according to claim 1, characterized in that, The federated averaging algorithm mentioned in the shared model construction is a method for aggregating model parameters in federated learning, and its expression is shown in the following formula: ； in, This represents the local model parameters of the k-th user in the j-th round of federation. This represents the global model parameters in the (j+1)th round.

3. A federated model transfer learning method for fault diagnosis of rolling bearings of different specifications according to claim 1 or 2, characterized in that, The layer-by-layer unfreezing strategy described in the personalized model construction is used to fine-tune the network. The specific method is as follows: except for the parameters of the fully connected layer, the other model parameters of the shared model used for transfer learning are transferred to the target domain network as the initialization parameters of the target domain network. The network is unfrozen layer by layer from the last convolutional layer forward, and the network is fine-tuned using known labeled samples from the target domain.

4. A federated model transfer learning system for fault diagnosis of rolling bearings of different specifications, characterized in that, include: The local dataset and local model building module is used to select vibration data of different types of rolling bearings as local datasets for each user, and obtain a local time-frequency map sample set for each user through short-time Fourier transform; and use the local time-frequency map sample set to train a convolutional neural network to obtain multiple local models for each user. A shared model building module is used to utilize an improved parameter transfer strategy. Users upload their trained local model parameters to the server. The server uses a federated averaging algorithm to aggregate the uploaded parameters and update the local model. This process is repeated until the maximum number of federated iterations is reached, resulting in a shared model for transfer learning. The improved parameter transfer strategy introduces a parameter sparsity algorithm to improve the parameter transfer strategy during federated learning, thereby addressing privacy leaks when users upload parameters. This includes: after users train their local models using local data, they calculate the difference between the initial global model parameters and the trained local model parameters for each layer, with the difference for each layer being a matrix, resulting in a difference matrix; users generate a matrix of the same size as the model parameters, consisting of 0s and 1s, with the number of 1s determined by a predefined parameter sparsity, resulting in a parameter sparsity matrix; the difference matrix is ​​multiplied by the corresponding positions of the parameter sparsity matrix to obtain a new difference matrix; the new difference matrix is ​​sent to the server, which collects the difference matrices from all users, takes their average, and adds it to the global model parameter matrix from the previous round to form the global model for the next round. The personalized model building module is used by the server to send the shared model for transfer learning to each user, decide which layer parameters of the shared model to retain based on the layer-by-layer unfreezing strategy, and fine-tune the shared model using the local dataset to obtain a personalized model suitable for the user's local dataset.

5. A federated model transfer learning system for fault diagnosis of rolling bearings of different specifications according to claim 4, characterized in that, The federated averaging algorithm in the shared model building module is a model parameter aggregation method in federated learning, and its expression is shown in the following formula: ； in, This represents the local model parameters of the k-th user in the j-th round of federation. This represents the global model parameters in the (j+1)th round.

6. A federated model transfer learning system for fault diagnosis of rolling bearings of different specifications according to claim 4 or 5, characterized in that, The layer-by-layer unfreezing strategy in the personalized model construction module is used to fine-tune the network. The specific method is as follows: except for the parameters of the fully connected layer, the other model parameters of the shared model used for transfer learning are transferred to the target domain network as the initialization parameters of the target domain network. The network is unfrozen layer by layer from the last convolutional layer forward, and the network is fine-tuned using known labeled samples from the target domain.

Images