Method for predicting residual life of rolling bearing based on personalized federated learning

Abstract

The invention discloses a rolling bearing residual life prediction method based on personalized federated learning, and belongs to the field of rolling bearing residual life prediction.The method comprises the steps that rolling bearing full life cycle vibration data collected under different operation conditions are stored in a plurality of clients, and a server and the clients work cooperatively; and establishing a personalized prediction model for the client. A plurality of clients are aggregated into a global model at a server side by adopting a dynamic weighted federation aggregation strategy, so that the sensing ability of the aggregated global model to multi-working-condition data is improved; an adversarial training and self-adaptive fusion strategy is adopted at a client, a global model aggregated by a server and a local model are effectively fused, and the generalization ability of the local model is improved. The client local model adopts a multi-scale convolution and gating cycle unit parallel cross attention feature extraction module to capture spatio-temporal feature information in the data. According to the method, the residual life prediction precision and generalization performance of the rolling bearing under different working conditions can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and in particular relates to a method for predicting the remaining life of a rolling bearing based on personalized federated learning. Background Art

[0002] As the core supporting component of rotating machinery, the health status of rolling bearings directly affects the operating reliability and efficiency of the whole machine. Due to the harsh operating environment and complex and changeable operating conditions, rolling bearings will inevitably suffer from problems such as wear and performance degradation. Therefore, conducting research on the health status monitoring and remaining life prediction of rolling bearings is of great significance to ensure the safe and stable operation of mechanical equipment and the smooth progress of industrial production. Accurate prediction of the remaining life of rolling bearings is like installing a "smart eye" for mechanical equipment, which can help equipment maintenance personnel to timely grasp the performance degradation status of rolling bearings so as to take preventive measures, thereby reducing the risk of unexpected downtime and improving production efficiency.

[0003] With the continuous development of artificial intelligence and the improvement of hardware computing power, deep learning has been widely used in image processing, speech recognition, natural language processing and other fields. Various deep learning methods have also appeared in the field of rolling bearing remaining life prediction. However, the remaining life prediction method based on classic deep learning often needs to rely on a large amount of identically distributed data for centralized training. In the actual industrial environment, the collected rolling bearing vibration data often has significant distribution differences. In addition to the difference in data distribution, the data island effect is also a major challenge in the health management of industrial equipment. For reasons such as data security and privacy protection, enterprises are often unwilling or unable to share data in a centralized manner, resulting in the dispersion and independence of data. This heterogeneity of data distribution and data island problems make it difficult for the remaining life prediction model based on centralized data processing to effectively capture the data features in different scenarios, thereby limiting the accuracy and generalization of the prediction model. Therefore, how to effectively utilize data information in different scenarios while protecting data privacy, realize the remaining life prediction of rolling bearings under the federated learning framework, and improve the accuracy and generalization of the prediction model is still a difficult problem to be solved. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for predicting the remaining life of rolling bearings based on personalized federated learning, improve the federated learning strategy, use data from different clients to build a personalized prediction model for the remaining life of rolling bearings under different operating conditions, enhance the generalization of the prediction model, and improve the accuracy of the remaining life prediction.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: A method for predicting the remaining life of a rolling bearing based on personalized federated learning comprises the following steps: S1. Collect vibration data of rolling bearings in their entire life cycle under different operating conditions; S2. The full life cycle vibration data of rolling bearings under different operating conditions are stored in different clients. The collected full life cycle vibration data are converted into two-dimensional images by continuous wavelet transform on each client to construct training data sets, verification data sets and test data sets. S3, using the training data set and the verification data set to initialize the client local model training, and upload the trained client local model parameters to the server; S4. A dynamic weighted federation strategy is used on the server to aggregate the model parameters uploaded by multiple clients to obtain a global model, and the aggregated global model parameters are sent to each client. S5. Adopt adversarial training and parameter adaptive fusion strategy on the client side to fuse the aggregated global model parameters with the client local model parameters, and use the training data set to fine-tune the local model parameters; S6. Use the trained client local model to analyze the local test data set to obtain the remaining life prediction result of the test bearing.

[0006] A further improvement of the technical solution of the present invention is that the whole life cycle vibration data is converted into a two-dimensional image by using continuous wavelet transform in step S2, including: Continuous wavelet transform is used to analyze each data sample in the life cycle vibration data to obtain wavelet coefficients, and the normalized wavelet coefficients are converted into a two-dimensional image.

[0007] A further improvement of the technical solution of the present invention is that the client local model in step S3 includes: A multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module, which is composed of two parallel branches and a cross-attention module. The first branch of the parallel branches is composed of a multi-scale convolution module and a channel attention module connected in series, and the second branch of the parallel branches is composed of three layers of gated recurrent units connected in series in sequence. The outputs of the two parallel branches are subjected to feature fusion through the cross-attention module for extracting spatiotemporal features of the input data samples. The prediction output module, consisting of two fully connected layers, is used to output the prediction results of the remaining life of the rolling bearing.

[0008] A further improvement of the technical solution of the present invention is that: in the step S4, a dynamic weighted federation strategy is adopted on the server side to aggregate the uploaded model parameters of multiple clients to obtain a global model, including: the uploaded model parameters include the parameters of the first branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in each client local model and the validation loss of each client's local model in k (=1,2,…, K ) indicates the k Clients, K Indicates the total number of clients.

[0009] The dynamic weighted federation strategy includes the following steps: First, the server uses the verification loss of all K clients to calculate the weight of each client in, is the preset minimum weight value, The maximum value among all client validation losses; Secondly, Perform normalization to obtain the aggregate weight of the kth client ; Finally, the server side utilizes the multi-scale convolution and gated recurrent units of each client to cross-attention feature extraction module parameters in parallel and aggregation weight , generate global model parameters , and the aggregated global model parameters Distributed to each client.

[0010] A further improvement of the technical solution of the present invention is that in step S5, adversarial training and parameter adaptive fusion strategy are adopted on the client to fuse the aggregated global model parameters with the client local model parameters, and the local model parameters are fine-tuned using the training data set, including: The adversarial training is to use the global model sent by the server as the first branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model, and the second branch and the cross-attention module in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model to form a global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module for learning useful features in the client local training data set; then, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and the data features learned by the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model are input into a binary classifier composed of two fully connected layers to determine which feature extraction module the input comes from; at the same time, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module are input into the prediction output module in the client local model for prediction analysis. The optimization goal of adversarial training is in, N k Indicates k The number of samples in the training dataset of the client, x i It is i training sample data, D is a binary classifier, y i It is i The true labels of the training samples, To predict the output module i The predicted output of training samples is λ is the weight coefficient. In the adversarial training process, first fix the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters Parallel cross-attention feature extraction module parameters of multi-scale convolution and gated recurrent unit in client local model , update the parameters of the binary classifier : Then, fix the parameters of the binary classifier and update the parameters of the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction modules and prediction output module parameters The parameter adaptive fusion strategy is to parallel cross-attention feature extraction module parameters of the updated global multi-scale convolution and gated recurrent unit obtained by adversarial training. Cross-attention feature extraction module parameters in parallel with client-side local multi-scale convolution and gated recurrent units Perform weighted fusion to obtain a new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters ; The calculation formula is as follows in, is the adaptive fusion coefficient, It is k The validation loss of the client's local model.

[0011] The method of fine-tuning the local model parameters by using the training data set is to form a remaining life prediction model by combining the new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and prediction module, and optimize the parameters of the remaining life prediction model by using the local training data set and the verification data set. The optimization target is ,in N k is the number of samples in the k-th client training data set, y i For the i The true labels of the training samples, To predict the output model for i The predicted output for each training sample.

[0012] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is: The method for predicting the remaining life of rolling bearings based on personalized federated learning proposed in the present invention adopts a dynamic weighted federated aggregation strategy on the server side, adaptively adjusts the weight according to the verification loss of each client, aggregates multiple client models into a global model, and improves the global model's perception of multi-condition data; designs a multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module on the client side to capture the spatiotemporal feature information in the data; proposes an adversarial training and adaptive fusion strategy to effectively fuse the global model aggregated on the server side with the local model on the client side, realizes the construction of a personalized prediction model for the client under different operating conditions, and further improves the adaptability of the client model to the local data distribution. The present invention improves the accuracy and generalization of the remaining life measurement of rolling bearings under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the method for predicting the remaining life of a rolling bearing based on personalized federated learning proposed in the present invention; Figure 2 It is a structural diagram of the client local model proposed by the present invention; Figure 3 is the remaining life prediction result of the rolling bearing obtained by the method proposed in the present invention; Figure 3 (a) is the prediction result of Bearing1_5. Figure 3 (b) is the prediction result of Bearing2_5. Figure 3 (c) is the prediction result of Bearing3_4; Figure 4 This is the remaining life prediction result of rolling bearings obtained by three clients under the non-federated learning framework; Figure 4 (a) is the prediction result of Bearing1_5. Figure 4 (b) is the prediction result of Bearing2_5. Figure 4 (c) is the prediction result of Bearing3_4. DETAILED DESCRIPTION

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] like Figure 1 As shown, a method for predicting the remaining life of a rolling bearing based on personalized federated learning includes the following steps: S1. Collect vibration data of rolling bearings in their entire life cycle under different operating conditions; S2. The full life cycle vibration data of rolling bearings under different operating conditions are stored in different clients. The collected full life cycle vibration data are converted into two-dimensional images by continuous wavelet transform on each client to construct training data sets, verification data sets and test data sets. S3, using the training data set and the verification data set to initialize the client local model training, and upload the trained client local model parameters to the server; S4. A dynamic weighted federation strategy is used on the server to aggregate the model parameters uploaded by multiple clients to obtain a global model, and the aggregated global model parameters are sent to each client. S5. Adopt adversarial training and parameter adaptive fusion strategy on the client side to fuse the aggregated global model parameters with the client local model parameters, and use the training data set to fine-tune the local model parameters; S6. Use the trained client local model to analyze the test data set to obtain the remaining life prediction result of the test bearing.

[0016] In one embodiment, step S1, firstly, a vibration acceleration sensor is used to collect the vibration data of the rolling bearing under different operating conditions throughout its life cycle, the sampling frequency is set to 25.6 kHz, the sampling time interval is 1 minute, and the sampling time of each data sample is 1.28 seconds; the tested equipment is a rolling bearing fatigue life acceleration test bench, and its operating conditions are: operating condition 1—speed 2100 r / min and load 12 kN, operating condition 2—speed 2250 r / min and load 11 kN, and operating condition 3—speed 2400 r / min and load 10 kN. Under each operating condition, the vibration data of the entire life cycle of 5 rolling bearings are collected. The names of the rolling bearings tested under the three operating conditions are bearing1_1-bearing1_5, bearing2_1-bearing2_5, and bearing3_1-bearing3_5.

[0017] Step S2, the rolling bearing life cycle vibration data collected under working conditions 1, 2 and 3 are stored in three clients respectively; on each client, each data sample in the life cycle vibration data of each test bearing is converted into a two-dimensional image of size 128×128 by using continuous wavelet transform: in, WT represents the wavelet coefficients obtained by continuous wavelet transform, and Represent the wavelet coefficients WT The minimum and maximum values ​​in ; COEF represents a two-dimensional image, Represents a pixel in a two-dimensional image.

[0018] In the first client, use bearing1_1, bearing1_2 and bearing1_4 in working condition 1 as training data sets, bearing1_3 as validation set, and bearing1_5 as test set; In the second client, use bearing2_1, bearing2_2 and bearing2_4 in working condition 2 as training data sets, bearing2_3 as validation set, and bearing2_5 as test set; In the third client, use bearing3_1, bearing3_2 and bearing3_5 in working condition 3 as training data sets, bearing3_3 as validation set, and bearing3_4 as test set; The normalized RMS value is used to generate the label of the remaining life of the rolling bearing. The RMS value is calculated as in, x ( j ) is the sample datax The j sampling points, N is the length of the sample data, X rms Next, all the RMS values ​​calculated from the full life cycle vibration data of each rolling bearing are normalized so that each RMS value is in the range of 0 to 1. The normalization formula is as follows: in, and represent the maximum and minimum values ​​of the RMS value, respectively. is the normalized root mean square value. Then, the normalized root mean square value is further fitted using the exponential fitting method to more accurately describe the changing trend of rolling bearing performance degradation. The fitting formula is as follows: in, a and b To fit the parameters, they are learned from the data through nonlinear optimization; y t It is calculated using the root mean square value after fitting. t The normalized remaining life corresponding to the moment. y t Discretize the sample number of the full life cycle vibration data collected for each test bearing to obtain discrete values y i ( i =1,2,…, M , M represents the number of sample data), as the true label of each sample data.

[0019] Each client's local model is Figure 2 As shown, it contains a multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and a prediction output module; A multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module, which is composed of two parallel branches and a cross-attention module. The first branch of the parallel branches is composed of a multi-scale convolution module and a channel attention module connected in series, and the second branch of the parallel branches is composed of three layers of gated recurrent units connected in series in sequence. The outputs of the two parallel branches are subjected to feature fusion through the cross-attention module for extracting spatiotemporal features of the input data samples. The prediction output module, consisting of two fully connected layers, is used to output the prediction results of the remaining life of the rolling bearing.

[0020] In step S3, the local model is first trained using the training dataset of each client, and the performance of the local model is verified using the validation dataset; the purpose of training is to adapt each client model to the specific data distribution of the client. The loss function of client model training is: in, N k is the number of samples in the k-th client training data set, y i For the i The true labels of the training samples, To predict the output model for i The predicted output for each training sample.

[0021] The batch size of the training is 64, the learning rate is 0.001, the number of local training rounds is 30, and the learning rate decay parameter is set to 0.5. After the training is completed, the multi-scale convolution and gated recurrent units in each client local model are parallel cross-attention to the first branch parameters in the feature extraction module and the validation loss of each client's local model Upload to the server, where k (=1,2,…,K) represents the k clients, K=3 represents the total number of clients.

[0022] In step S4, the server first uses the verification losses of all three clients to calculate the weight of each client ,in, is the preset minimum weight value. =0.2, The maximum value among all client validation losses; Secondly, Perform normalization to obtain the aggregate weight of the kth client ; Finally, the server side utilizes the multi-scale convolution and gated recurrent units of each client to cross-attention feature extraction module parameters in parallel and aggregation weight , generate global model parameters , and the aggregated global model parameters Distributed to each client.

[0023] In step S5, each client uses the global model sent by the server as the first branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client's local model, and the second branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client's local model and the cross-attention module to form a global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module for learning useful features in the client's local training data set; Then, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and the data features learned by the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model are input into a binary classifier composed of two fully connected layers to determine which feature extraction module the input comes from; at the same time, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module are input into the prediction output module in the client local model for prediction analysis. The optimization goal of adversarial training is in, N k Indicates k The number of samples in the training dataset of the client, x i It is i training sample data, D is a binary classifier, y i It is i The true labels of the training samples, To predict the output module i The predicted output of training samples is λ is the weight coefficient.

[0024] In the adversarial training process, we first fix the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters. Parallel cross-attention feature extraction module parameters of multi-scale convolution and gated recurrent unit in client local model , update the parameters of the binary classifier ; Then, fix the parameters of the binary classifier and update the parameters of the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction modules and prediction output module parameters In the adversarial training, the batch size is 64, the learning rate is 0.001, the number of adversarial training rounds is 5, and the weight coefficient λ is set to 0.3.

[0025] After the adversarial training is completed, the updated global multi-scale convolution and gated recurrent unit obtained from the adversarial training are parallelized to cross-attention feature extraction module parameters. Cross-attention feature extraction module parameters in parallel with client-side local multi-scale convolution and gated recurrent units Perform weighted fusion to obtain a new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters ; The calculation formula is as follows in, is the adaptive fusion coefficient, It is k The validation loss of the client's local model.

[0026] Finally, on each client, the new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and prediction module are combined into a remaining life prediction model, and the parameters of the remaining life prediction model are optimized and fine-tuned using the local training dataset and validation dataset. The fine-tuning optimization objective is ,in N k is the number of samples in the k-th client training data set, y i For the i The true labels of the training samples, To predict the output model for i After the local client fine-tuning is completed, the client local model is obtained as the initialization parameter for the next round of federated learning.

[0027] In step S6, the trained client local model is used to analyze the test data set to obtain the remaining life prediction result of the test bearing.

[0028] The root mean square error is used to measure the deviation between the predicted value and the true value, and the remaining life prediction results of the three test data sets obtained by the method proposed in the present invention are given as follows: Figure 3 At the same time, as a comparison, the prediction results of three clients on three test data sets under the non-federated learning framework are given as follows Figure 4 It can be seen that the remaining life prediction errors obtained by the method proposed in the present invention in the three test data sets are all smaller than those of the comparison method, which shows that the personalized federated learning strategy proposed in the present invention can effectively improve the generalization performance of the client local model and the accuracy of the remaining life prediction.

[0029] The present invention adopts a rolling bearing remaining life prediction method based on personalized federated learning, which can improve the prediction performance of the client's local model while achieving privacy data protection, successfully extract the effective feature information contained in the client's vibration data, and make the research objectives more targeted, which is of great significance for achieving accurate prediction of the remaining life of rolling bearings.

[0030] The embodiments described above are merely descriptions of preferred implementation modes of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for predicting the remaining life of a rolling bearing based on personalized federated learning, characterized in that: The steps include: S1. Collect vibration data of rolling bearings in their entire life cycle under different operating conditions; S2. The full life cycle vibration data of rolling bearings under different operating conditions are stored in different clients. The collected full life cycle vibration data are converted into two-dimensional images by continuous wavelet transform on each client to construct training data sets, verification data sets and test data sets. S3, using the training data set and the verification data set to initialize the client local model training, and upload the trained client local model parameters to the server; S4. A dynamic weighted federation strategy is used on the server to aggregate the model parameters uploaded by multiple clients to obtain a global model, and the aggregated global model parameters are sent to each client. S5. Adopt adversarial training and parameter adaptive fusion strategy on the client side to fuse the aggregated global model parameters with the client local model parameters, and use the training data set to fine-tune the local model parameters; S6. Use the trained client local model to analyze the local test data set to obtain the remaining life prediction result of the test bearing.

2. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 1 is characterized in that: The step S2 converts the life cycle vibration data into a two-dimensional image using a continuous wavelet transform, specifically, using a continuous wavelet transform to analyze each data sample in the life cycle vibration data to obtain a wavelet coefficient, and converting the normalized wavelet coefficient into a two-dimensional image; , in, WT represents the wavelet coefficients obtained by continuous wavelet transform, and Represent the wavelet coefficients WT The minimum and maximum values ​​in ; COEF represents a two-dimensional image, Represents a pixel in a two-dimensional image.

3. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 2 is characterized in that: The normalized RMS value is used to generate the label of the remaining life of the rolling bearing. The RMS value is calculated as follows: , in, x ( j ) is the sample data x The j sampling points, N is the length of the sample data, X rms is the root mean square value; then, all the root mean square values ​​calculated from the full life cycle vibration data of each rolling bearing are normalized so that each root mean square value is in the range of 0 to 1. The normalization formula is as follows , in, and represent the maximum and minimum values ​​of the RMS value, respectively. is the normalized RMS value; The exponential fitting method is further used to fit the normalized RMS value to more accurately describe the changing trend of rolling bearing performance degradation. The fitting formula is as follows: , in, a and b To fit the parameters, they are learned from the data through nonlinear optimization; y t It is calculated using the root mean square value after fitting. t The normalized remaining life corresponding to the moment is y t Discretize the sample number of the full life cycle vibration data collected for each test bearing to obtain discrete values y i , i =1,2,…, M , M Represents the number of sample data, which serves as the true label of each sample data.

4. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 1 is characterized in that: The client local model in step S3 includes: A multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module, which is composed of two parallel branches and a cross-attention module. The first branch of the parallel branches is composed of a multi-scale convolution module and a channel attention module connected in series, and the second branch of the parallel branches is composed of three layers of gated recurrent units connected in series in sequence. The outputs of the two parallel branches are subjected to feature fusion through the cross-attention module for extracting spatiotemporal features of the input data samples. The prediction output module consists of two fully connected layers and is used to output the prediction results of the remaining life of the rolling bearing; The loss function for client model training is: in, N k is the number of samples in the k-th client training data set, y i For the i The true labels of the training samples, To predict the output model for i The predicted output for each training sample.

5. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 1, characterized in that: The uploaded model parameters in step S4 include the parameters of the first branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in each client local model and the validation loss of each client's local model in k (=1,2,…, K ) indicates the k Clients, K Indicates the total number of clients; The dynamic weighted federation strategy described in step S4 includes the following steps: First, the server uses the verification loss of all K clients to calculate the weight of each client ,in, is the preset minimum weight value, The maximum value among all client validation losses; Secondly, Perform normalization to obtain the aggregate weight of the kth client ; Finally, the server side uses the multi-scale convolution and gated recurrent units of each client to cross-attentionally analyze the parameters of the first branch in the feature extraction module in parallel. and aggregation weight , generating global model parameters , and the aggregated global model parameters Distributed to each client.

6. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 1, characterized in that: The adversarial training described in step S5 is to use the global model sent by the server as the first branch in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model, and the second branch and the cross-attention module in the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model to form a global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module for learning useful features in the client local training data set; then, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and the data features learned by the multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module in the client local model are input into a binary classifier composed of two fully connected layers to determine which feature extraction module the input comes from; at the same time, the data features learned by the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module are input into the prediction output module in the client local model for prediction analysis.

7. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 6, characterized in that: The optimization goal of adversarial training is in, N k Indicates k The number of samples in the training dataset of the client, x i It is i training sample data, D is a binary classifier, y i It is i The true labels of the training samples, For the prediction output module i The predicted output of training samples is λ is the weight coefficient; In the adversarial training process, we first fix the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters. Parallel cross-attention feature extraction module parameters of multi-scale convolution and gated recurrent unit in client local model , update the parameters of the binary classifier : Then, fix the parameters of the binary classifier and update the parameters of the global multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction modules and prediction output module parameters .

8. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 7 is characterized in that: The parameter adaptive fusion strategy in step S5 is to parallel cross-attention feature extraction module parameters of the updated global multi-scale convolution and gated recurrent unit obtained by adversarial training Cross-attention feature extraction module parameters in parallel with client-side local multi-scale convolution and gated recurrent units Perform weighted fusion to obtain a new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module parameters ; The calculation formula is as follows in, is the adaptive fusion coefficient, It is k The validation loss of the client's local model.

9. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 8, characterized in that: In step S5, the local model parameters are fine-tuned using the training data set, which is to form a remaining life prediction model with a new multi-scale convolution and gated recurrent unit parallel cross-attention feature extraction module and prediction module, and the parameters of the remaining life prediction model are optimized using the local training data set and the verification data set. The optimization target is ,in N k is the number of samples in the k-th client training data set, y i For the i The true labels of the training samples, To predict the output model for i The predicted output for each training sample.

10. The method for predicting the remaining life of a rolling bearing based on personalized federated learning according to claim 9, characterized in that: After fine-tuning on the local client is completed, the client local model is obtained as the initialization parameter for the next round of federated learning.

Images