A method for predicting the remaining life of rolling bearings based on federated learning and model pruning

Abstract

A method for predicting the remaining life of rolling bearings based on federated learning and model pruning is disclosed. The steps are as follows: Vibration signals throughout the entire life cycle of the rolling bearing are collected; the acquired samples are divided, stored, and preprocessed to obtain sequence data that can be used as model input; a prediction model is constructed, including a multi-scale feature expansion module, a deep feature extraction module, and a prediction module; in a federated learning framework, a central server and multiple clients collaboratively train the prediction model, and a pruning strategy is used to make the model structure lightweight; the final lightweight model is used to predict the remaining life of the rolling bearing. This invention uses federated learning to build a prediction model for distributed data that is not shared between clients, greatly improving the security of data from each client. Combined with pruning operations, it effectively improves the prediction performance and inference speed of the model.

Description

Technical Field

[0001] This disclosure belongs to the field of bearing remaining life prediction technology, and in particular to a method for predicting the remaining life of rolling bearings based on federated learning and model pruning. Background Technology

[0002] Rolling bearings are core components of rotating machinery. During use, they degrade due to wear, making them one of the most prone to failure. Monitoring the operating status of rolling bearings and conducting effective research on remaining life prediction are crucial for improving the stability of mechanical systems.

[0003] With the rapid development of deep learning and sensor technology, methods for predicting the remaining life of rolling bearings based on deep neural networks have achieved significant results. Due to their powerful nonlinear mapping capabilities, deep networks can fit highly complex bearing degradation processes with the support of massive monitoring data to achieve life prediction. However, existing models and training data have the following problems: 1) To extract deep degradation features, existing models generally have a large number of layers, parameters, and slow inference speed; 2) Collecting full-life data of rolling bearings is difficult, with few samples, resulting in data silos. Traditional methods that fuse data before training the model do not consider privacy protection for multiple data owners.

[0004] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art in this country. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this disclosure is to propose a method for predicting the remaining life of rolling bearings based on federated learning and model pruning. This method builds a life prediction model for distributed rolling bearing data that is not shared between different systems using federated learning, and incorporates pruning strategies to make the model lightweight.

[0006] To achieve the above objectives, this disclosure provides the following technical solution:

[0007] A method for predicting the remaining life of rolling bearings based on federated learning and model pruning includes the following steps:

[0008] Step 1: Collect vibration signals from the entire life cycle of the rolling bearing as samples;

[0009] Step 2: A portion of the collected samples is used as the training set. Using a single sample as the basic unit, the samples in the training set are stored in multiple clients. The samples in the training set of each client are preprocessed to obtain the sequence data as input to the model.

[0010] Step 3: Construct a global model for predicting the remaining life of rolling bearings. This model includes a multi-scale feature expansion module, a deep feature extraction module, and a prediction module. The multi-scale feature expansion module includes a fully connected layer FC1, a multi-scale convolutional layer MC1, a multi-scale convolutional layer MC2, and a multi-scale convolutional layer MC3. The deep feature extraction module includes convolutional layers C1, C2, C3, and C4. The prediction module includes a fully connected layer FC2 and a regression layer.

[0011] Step 4: In the federated learning framework, the central server and multiple clients collaborate to train the global model, and the global model is trained into a lightweight prediction model based on the pruning strategy.

[0012] Step 5: Based on the lightweight global model, predict the remaining life of rolling bearings of the same type.

[0013] In the aforementioned method for predicting the remaining life of rolling bearings based on federated learning and model pruning, the preprocessing includes performing a fast Fourier transform on the vibration signal to obtain its frequency domain amplitude signal, then normalizing the frequency domain amplitude, and finally dividing the obtained sequence into multiple short sequences as sequence data based on 1024 data points.

[0014] In the method for predicting the remaining life of rolling bearings based on federated learning and model pruning, step 4 includes:

[0015] Step 4-1) Set the maximum number of communication rounds N in federated learning. round The number of pruned features ε, and the initial prediction model β. 0 ;

[0016] Step 4-2) The central server will use the initial prediction model β 0 Distribute to all clients;

[0017] Step 4-3) In each round of communication iteration, each client updates the parameters of the received model and uploads the updated model to the central server. The central server aggregates the collected models to form a global model.

[0018] Step 4-4) The central server distributes the global model to each client. Each client evaluates the global model and marks the client k with the best prediction performance at this time.

[0019] Steps 4-5) On client k, ε output features in the deep feature extraction module are pruned, and then the pruned model is uploaded to the central server, which then distributes it to all clients.

[0020] Steps 4-6) Repeat steps 4-3) to 4-5). When the predictive performance of the global model in step 4-4) begins to decline, or the number of communication iteration rounds reaches N. round At this point, the model training process ends.

[0021] In the method for predicting the remaining life of rolling bearings based on federated learning and model pruning, the method for updating the parameters of the received model by each client in step 4-3) is as follows:

[0022] Client i (i = 1, 2, ..., M) updates the received model parameters using the stochastic gradient descent algorithm, i.e.:

[0023]

[0024] In the formula, β on the left side i (t) represents the updated model parameters for client i in round t, and δ represents the learning rate for each client. This is the loss function.

[0025] In the method for predicting the remaining life of rolling bearings based on federated learning and model pruning, step 4-3) involves the central server aggregating the collected models as follows:

[0026] The central server uses a federated averaging algorithm to aggregate all collected client model parameters, namely:

[0027] In the formula, β g (t) represents the global model generated by aggregating parameters from multiple client models in round t, where M is the number of clients.

[0028] In the method for predicting the remaining life of rolling bearings based on federated learning and model pruning, the method for each client to evaluate the global model in step 4-4) is as follows:

[0029] The root mean square error (RMSE) is used to quantitatively evaluate the predictive performance of the global model. The calculation formula is as follows:

[0030]

[0031] In the formula, RUL j and Let Z be the actual lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z be the total number of input sequence data on a client.

[0032] In the method for predicting the remaining life of rolling bearings based on federated learning and model pruning, step 4-5) involves pruning ε output features from the deep feature extraction module, which includes:

[0033] Step 4-5-1) On client k, use local data to update the parameters of the global model at this time, record all output features of the deep feature extraction module, and the gradient of each network activation function with respect to each output feature;

[0034] Step 4-5-2) For the q-th output feature h q In other words, its importance is the gradient of the activation function C of its layer with respect to that feature. and the output feature h q The absolute value of the product, i.e. Calculate the importance of all output features;

[0035] Step 4-5-3) Sort the importance of all output features in the deep feature extraction module and prune the ε output features with the lowest importance.

[0036] Compared with the prior art, the beneficial effects of this disclosure are as follows:

[0037] This invention utilizes federated learning technology to build a prediction model for distributed data that is not shared among clients, greatly improving the security of data from each client and solving the data silo problem in lifetime prediction research. The multi-scale feature expansion module constructed in this invention can effectively expand the features of the input signal, providing sufficient basic features for subsequent deep feature extraction and feature regression, thus improving the prediction accuracy of the model. This invention removes some unimportant output features from the deep feature extraction module through pruning operations, making the module lightweight and improving the inference speed of the prediction model. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0039] In the attached diagram:

[0040] Figure 1 This is a flowchart of a method for predicting the remaining life of rolling bearings based on federated learning and model pruning, provided in one embodiment of this disclosure;

[0041] Figure 2 This is a structural diagram of a prediction model provided in one embodiment of this disclosure;

[0042] Figure 3 This is an overall architecture diagram of model training under federated learning provided in one embodiment of this disclosure;

[0043] Figure 4 This is a prediction result graph on test data provided in one embodiment of the present disclosure.

[0044] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0045] The following will refer to the appendix. Figures 1 to 4 Specific embodiments of this disclosure will be described in more detail below. While specific embodiments of this disclosure are shown in the accompanying drawings, it should be understood that this disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0046] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions of preferred embodiments of this disclosure are for the purpose of implementing the general principles of the specification and are not intended to limit the scope of this disclosure. The scope of protection of this disclosure is determined by the appended claims.

[0047] To facilitate understanding of the embodiments of this disclosure, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific examples. The accompanying drawings do not constitute a limitation on the embodiments of this invention.

[0048] To better understand, such as Figures 1 to 4 As shown, a method for predicting the remaining life of rolling bearings based on federated learning and model pruning includes the following steps:

[0049] 1) Data acquisition: Collect vibration signals from multiple rolling bearings throughout their entire life cycle;

[0050] 2) Data partitioning, storage and preprocessing: The collected samples are divided into training set and test set. Then, using a single sample as the basic unit, the sample data of the training set is stored in multiple clients. The raw data of each client and the test part are preprocessed to obtain sequence data that can be used as input to the model.

[0051] 3) Model construction: Construct a rolling bearing life prediction model that includes a multi-scale feature expansion module, a deep feature extraction module, and a prediction module;

[0052] 4) Model training: In the federated learning framework, the central server and multiple clients work together to train the global model. With the help of pruning strategies, the global model becomes a lightweight prediction model without losing accuracy.

[0053] 5) Model Prediction: The obtained lightweight global model was used to predict the remaining life of the test set data. Tests showed that the lightweight prediction model performed well in predicting the remaining life of similar rolling bearings. See the data below for details.

[0054] In one embodiment, the multi-scale feature expansion module of the prediction model includes a fully connected layer FC1, a multi-scale convolutional layer MC1, a multi-scale convolutional layer MC2, and a multi-scale convolutional layer MC3; the deep feature extraction module includes a convolutional layer C1, a convolutional layer C2, a convolutional layer C3, and a convolutional layer C4; and the prediction module includes a fully connected layer FC2 and a regression layer.

[0055] In one embodiment, the steps for building the lightweight prediction model in step 4) are as follows:

[0056] 4-1) Set the maximum number of communication rounds N in federated learning. round The number of pruned features ε, and the initial model β 0 ;

[0057] 4-2) The central server will use the initial prediction model β 0 Distribute to all clients;

[0058] 4-3) In each round of communication iteration, each client updates the parameters of the received model and uploads the updated model to the central server. The central server aggregates the collected models to form a global model.

[0059] 4-4) The central server distributes the global model to each client, and each client evaluates the global model and marks the client k that performs best at this time;

[0060] 4-5) On client k, ε output features in the deep feature extraction module of the prediction model are pruned, and then the pruned model is uploaded to the central server, which then distributes it to all clients.

[0061] 4-6) Repeat steps 4-3) to 4-5). When the predictive performance of the global model starts to decline in step 4-4), or the number of communication iterations reaches N. round At this point, the model training process ends;

[0062] Furthermore, the method for each client to update the parameters of the received model in step 4-3) is as follows:

[0063] Client i (i = 1, 2, ..., M) updates the parameters of the received model using local data and gradient descent, i.e.:

[0064]

[0065] In the formula, β on the left side i (t) represents the updated model parameters for client i in round t, and δ represents the learning rate for each client. This is the loss function.

[0066] Furthermore, in step 4-3), the central server aggregates all model parameters using the following method:

[0067] The central server uses a federated averaging algorithm to aggregate all received client model parameters, namely:

[0068]

[0069] In the formula, β g (t) represents the global model generated by aggregating parameters from multiple client models in round t, where M is the number of clients.

[0070] Furthermore, the method for each client to evaluate the global model in step 4-4) is as follows:

[0071] The root mean square error (RMSE) is used as an evaluation metric to quantitatively characterize the predictive performance of the model. Its calculation formula is as follows:

[0072]

[0073] In the formula, RUL j and , where are the true lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z is the total number of sequences on a client. The smaller the RMSE value, the better the model's predictive performance.

[0074] Furthermore, the steps for pruning the ε output features in the deep feature extraction module of the prediction model in steps 4-5) are as follows:

[0075] 4-5-1) On client k, use local data to update the prediction model once, record all output features of the four convolutional layers in the deep feature extraction module of the model, as well as the gradient of each network activation function with respect to each output feature;

[0076] 4-5-2) For the q-th output feature h q In other words, its importance is the gradient of the activation function C of its layer with respect to that feature. and the output feature h q The absolute value of the product, i.e. Calculate the importance of all output features;

[0077] 4-5-3) Sort the importance of all output features in the deep feature extraction module and prune the ε output features with the lowest importance.

[0078] In one embodiment, a method for predicting the remaining life of rolling bearings based on federated learning and model pruning mainly includes the following steps:

[0079] 1) Data acquisition: Collect vibration signals throughout the entire life cycle of the rolling bearing.

[0080] In this embodiment, an accelerometer can be used to collect vibration signals of the rolling bearing during operation. The vibration signals are oriented horizontally and vertically, with a sampling frequency of 25.6 kHz, a sampling duration of 1.28 seconds, and a sampling interval of 60 seconds.

[0081] 2) Data partitioning, storage and preprocessing: The collected samples are divided into training set and test set. Then, using a single sample as the basic unit, the sample data of the training set is stored in multiple clients. The raw data of each client and the test part are preprocessed to obtain sequence data that can be used as input to the model.

[0082] In this embodiment, the preprocessing operation includes performing a fast Fourier transform on the original signal to obtain its frequency domain amplitude signal, then normalizing the frequency domain amplitude, and finally dividing the resulting sequence into multiple short sequences based on 1024 data points as input to the prediction model.

[0083] 3) Model construction: Construct a rolling bearing life prediction model that includes a multi-scale feature expansion module, a deep feature extraction module, and a prediction module.

[0084] In this embodiment, the structure of the constructed prediction model is as follows: Figure 2 As shown, the multi-scale feature expansion module includes a fully connected layer FC1, a multi-scale convolutional layer MC1, a multi-scale convolutional layer MC2, and a multi-scale convolutional layer MC3. The "transformation" operation converts the output of the FC1 layer into a feature with 64 channels and a shape of 32×1. The "connect by channel" operation stacks the outputs of the three multi-scale convolutions along the channel direction to form a feature with 96 channels. The deep feature extraction module includes four one-dimensional convolutional layers: convolutional layers C1, C2, C3, and C4. The "flatten" operation flattens the output of the C4 layer into a one-dimensional array. The prediction module includes a fully connected layer FC2 and a regression layer. The regression layer outputs the lifetime prediction result. All layers use the tanh activation function, and the dropout of the fully connected layer is 0.5. The structure and parameters of each module of the prediction model are shown in Table 1.

[0085] Table 1. Structure and parameters of each module in the global model

[0086]

[0087] 4) Model training: In the federated learning framework, the central server, together with multiple clients, builds a global model for the local data it stores. With the help of pruning strategies, the global model becomes a lightweight prediction model without losing accuracy.

[0088] In this embodiment, the overall architecture for model training under federated learning is as follows: Figure 3 As shown, the total number of clients is set to M = 3.

[0089] Furthermore, the steps for building the lightweight prediction model in step 4) are as follows:

[0090] 4-1) Set the maximum number of communication rounds N in federated learning. round =10, number of pruned features ε=5, initial model β with randomization parameters 0 ;

[0091] 4-2) The central server will use the initial model β 0 Distribute to all clients;

[0092] 4-3) In each round of communication iteration, each client updates the received model and uploads the updated model to the central server. The central server aggregates all models to form a global model.

[0093] Furthermore, the method for each client to update the received model in step 4-3) is as follows:

[0094] Client i (i = 1, 2, ..., M) uses stochastic gradient descent and its local preprocessed data to update the received model parameters using the Adam optimization algorithm, i.e.:

[0095]

[0096] In the formula, β on the left side i (t) represents the updated model parameters for client i in round t, with a learning rate δ = 0.001 for each client and a loss function. The mean squared error (MSE) is calculated using the following formula:

[0097]

[0098] RUL j and Let Z be the actual lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z be the total number of input sequences on a client.

[0099] Furthermore, in step 4-3), the central server aggregates all models using the following method:

[0100] The central server uses a federated averaging algorithm to aggregate the parameters of all client models, i.e.

[0101]

[0102] In the formula, β g (t) represents the global model generated by aggregating parameters from multiple client models in round t, where M = 3 represents the number of clients.

[0103] 4-4) The central server distributes the global model to each client, and each client evaluates the global model and marks the client k that performs best at this time;

[0104] Furthermore, in step 4-4), each client uses the following method to evaluate the global model:

[0105] The root mean square error (RMSE) is used as an evaluation metric to quantitatively characterize the predictive performance of the model. Its calculation formula is as follows:

[0106]

[0107] In the formula, RUL j and , where are the true lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z is the total number of sequences on a client. The smaller the RMSE value, the better the model's predictive performance.

[0108] 4-5) On client k, ε = 5 output features in the deep feature extraction module of the pruned prediction model are extracted. The pruned model is then uploaded to the central server, which then distributes it to all clients.

[0109] Furthermore, the steps for pruning the ε output features in the deep feature extraction module of the prediction model in steps 4-5) are as follows:

[0110] 4-5-1) On client k, use local data to update the prediction model once, record all output features of the four convolutional layers in the deep feature extraction module of the model, as well as the gradient of each network activation function with respect to each output feature;

[0111] 4-5-2) For the q-th output feature h q In other words, its importance is the gradient of the activation function C of its layer with respect to that feature. and the output feature h q The absolute value of the product, i.e. Calculate the importance of all output features;

[0112] 4-5-3) Sort the importance of all output features in the deep feature extraction module and prune the ε output features with the lowest importance.

[0113] 4-6) Repeat steps 4-3) to 4-5). When the predictive performance of the global model starts to decline in step 4-4), or the number of communication iterations reaches N. round When the value is 10, the model training process ends.

[0114] 5) Model Prediction: The remaining useful life of the test data was predicted using the final lightweight global model. The results are as follows:

[0115] In this embodiment, the method of the present invention was tested on the working condition 2 dataset of the XJTU-SY rolling bearing accelerated life test dataset. This working condition contains the full life data of five rolling bearings. After preprocessing, bearings 2_1, 2_2, 2_3, and 2_5 yielded input sequences of 720, 2400, 4080, and 4000 data points of length 1024, respectively. The first three sets of sequence data were placed into three clients as training sets for model training. Bearing 2_5 was used for model prediction. The RMSE and MAE metrics were used to evaluate the model prediction performance. The formula for calculating MAE is as follows:

[0116]

[0117] In the formula, RUL j and Let Z be the true lifetime and predicted lifetime of the j-th input sequence, respectively, and Z be the total number of sequences on a client. The smaller the value of the metric, the better the predictive performance of the model.

[0118] The model training method of this invention was compared with the traditional Federated Averaging (FedAvg) training method, and the results are shown in Table 2. It can be seen that the RMSE of the model trained using the method of this invention is improved by approximately 22.32% and the MAE is improved by approximately 16.87% compared to FedAvg. This is because the pruning operation reduces the complexity of the model, increases its saturation, and prevents overfitting. Figure 4 This is a graph showing the prediction results of the method proposed in this invention on test data. It can be seen that the prediction model can accurately track the degradation trend of rolling bearings.

[0119] Table 2 Comparison of results between the training method in this invention and the FedAvg method.

[0120] Training methods RMSE MAE This invention 0.087 0.069 FedAvg 0.112 0.083

[0121] The model structure constructed by the method of this invention is compared and analyzed with the model structure without the multi-scale feature expansion module. The results are shown in Table 3. It can be seen that the use of the multi-scale feature expansion module can significantly improve the prediction performance of the model, with an RMSE improvement of approximately 37.41% and a MAE improvement of approximately 45.67%. This indicates that the multi-scale feature expansion module is crucial to the entire model structure. The basic features it expands play a positive role in the subsequent deep feature extraction module and prediction module, which is beneficial to the accurate prediction of the model.

[0122] Table 3 Comparison of results between the proposed model structure and the model structure without multi-scale feature extension modules.

[0123] Model Structure RMSE MAE This invention 0.087 0.069 No multi-scale feature extension module 0.139 0.127

[0124] The lightweight global model obtained by the method of this invention was compared with the model before pruning, and the results are shown in Table 4. The pruning operation removed approximately 57.80% of the parameters of the deep feature extraction module, and the inference time of the resulting lightweight model was shortened by 13.53%. After multiple pruning and updates, the evaluation parameters of the prediction model were significantly improved, further proving the effectiveness of this invention.

[0125] Table 4 Comparison of the model before and after pruning in the method proposed in this invention.

[0126] Number of parameters in the deep feature extraction module Model inference time RMSE MAE Before pruning 173kB 0.776s 0.454 0.461 After pruning 73kB 0.671s 0.087 0.069

[0127] In summary, by utilizing the technical solution described above, data distributed across multiple clients is used for model training via a federated learning approach where "the data remains stationary while the model moves," coupled with pruning strategies, resulting in a lightweight life prediction model suitable for rolling bearings. The obtained model exhibits excellent predictive performance and inference speed, and the modeling process ensures the security of distributed data, enabling its application in predicting the remaining life of similar rolling bearings.

[0128] Although the embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this disclosure is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of this disclosure, and all of these are within the scope of protection of this invention.

Claims

1. A method for predicting the remaining life of rolling bearings based on federated learning and model pruning, characterized in that, Includes the following steps: Step 1: Collect vibration signals from the entire life cycle of the rolling bearing as samples; Step 2: A portion of the collected samples is used as the training set. Using a single sample as the basic unit, the samples in the training set are stored in multiple clients. The samples in the training set of each client are preprocessed to obtain the sequence data as input to the model. Step 3: Construct a global model for predicting the remaining life of rolling bearings. This model includes a multi-scale feature expansion module, a deep feature extraction module, and a prediction module. The multi-scale feature expansion module includes a fully connected layer FC1, a multi-scale convolutional layer MC1, a multi-scale convolutional layer MC2, and a multi-scale convolutional layer MC3. The deep feature extraction module includes convolutional layers C1, C2, C3, and C4. The prediction module includes a fully connected layer FC2 and a regression layer. Step 4: In the federated learning framework, the central server and multiple clients collaboratively train the global model. Based on a pruning strategy, the global model is trained into a lightweight prediction model. Step 4 includes: Step 4-1) Set the maximum number of communication rounds N in federated learning. round Number of pruned features ɛ, initial prediction model β 0 ; Step 4-2) The central server will use the initial prediction model β 0 Distribute to all clients; Step 4-3) In each round of communication iteration, each client updates the parameters of the received model and uploads the updated model to the central server. The central server aggregates the collected models to form a global model. Step 4-4) The central server distributes the global model to each client. Each client evaluates the global model and marks the client k with the best prediction performance at this time. Steps 4-5) On client k, prune ɛ output features from the deep feature extraction module, then upload the pruned model to the central server, which then distributes it to all clients. Steps 4-6) Repeat steps 4-3) to 4-5). When the predictive performance of the global model in step 4-4) begins to decline, or the number of communication iteration rounds reaches N. round At this point, the model training process ends. In step 4-3), the method for each client to update the parameters of the received model is as follows: Client i (i = 1, 2, …, M) updates the received model parameters using the stochastic gradient descent algorithm, that is: ， In the formula, the left side Let δ be the model parameters updated for client i in round t, and δ be the learning rate for each client. For loss function, loss function The mean squared error (MSE) is calculated using the following formula: ， and Let Z be the actual lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z be the total number of input sequences on a client. Step 5: Based on the lightweight global model, predict the remaining life of rolling bearings of the same type.

2. The method for predicting the remaining life of rolling bearings based on federated learning and model pruning according to claim 1, characterized in that, Preprocessing includes performing a fast Fourier transform on the vibration signal to obtain its frequency domain amplitude signal, then normalizing the frequency domain amplitude, and finally dividing the resulting sequence into multiple short sequences as sequence data based on 1024 data points.

3. The method for predicting the remaining life of rolling bearings based on federated learning and model pruning according to claim 1, characterized in that, Step 4-3) The central server aggregates the collected models using the following method: The central server uses a federated averaging algorithm to aggregate all collected client model parameters, namely: ， In the formula, M represents the global model generated from the aggregation of parameters from multiple client models at round t, where M is the number of clients.

4. The method for predicting the remaining life of rolling bearings based on federated learning and model pruning according to claim 1, characterized in that, The method for each client to evaluate the global model in step 4-4) is as follows: The root mean square error (RMSE) is used to quantitatively evaluate the predictive performance of the global model. The calculation formula is as follows: ， In the formula, and Let Z be the actual lifetime and predicted lifetime of the j-th input sequence data, respectively, and Z be the total number of input sequence data on a client.

5. The method for predicting the remaining life of rolling bearings based on federated learning and model pruning according to claim 1, characterized in that, Step 4-5) involves pruning the ɛ output features from the depth feature extraction module, which includes: Step 4-5-1) On client k, use local data to update the parameters of the global model at this time, record all output features of the deep feature extraction module, and the gradient of each network activation function with respect to each output feature; Step 4-5-2) For the q-th output feature h q In other words, its importance is the gradient of the activation function C of its layer with respect to that feature. and the output feature h q The absolute value of the product, i.e. Calculate the importance of all output features; Step 4-5-3) Sort the importance of all output features in the deep feature extraction module and prune the ɛ output features with the lowest importance.

Images