Method and system for predicting service life of rolling bearing under different working conditions based on parameter decoupling personalized federation

Abstract

The invention discloses a method and a system for predicting the service life of a rolling bearing under different working conditions based on parameter decoupling personalized federation, relates to the technical field of bearing service life prediction, and aims to solve the problem that vibration data of the rolling bearing under different working conditions are always in non-independent identical distribution in the existing method. Parameter updating directions are separated from each other, and a global model is slow in convergence and even deviates from an optimal parameter. The method comprises the following steps: marking different stages of bearing life data of a training set by adopting a multi-stage degradation label representation method; the network model comprises a SEResNet block and a ConvLSTM block, the SEResNet block is used for performing feature extraction and reserving important features by connecting a maximum pooling layer, and the ConvLSTM block is used for further processing the features; and the central server decouples the network model, takes an SEResNet network as a shared presentation layer, takes ConvLSTM and a full connection layer as personalized layers, trains the shared presentation layer in a federated learning mode, aggregates parameters of each shared presentation layer, and finally obtains a rolling bearing life prediction model.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and in particular to a method and system for predicting the life of a rolling bearing under different working conditions based on parameter decoupling personalized federation. Background Art

[0002] Rolling bearings are key components of rotating machinery. Bearings that are in high-load operation for a long time are prone to damage and have poor reliability. Bearing damage may lead to failure of the entire machine or even a catastrophic accident. [1] As a key technology of predictive maintenance, the study of remaining useful life (RUL) can analyze the degradation trend of rotating machinery, thereby avoiding casualties and production losses caused by mechanical equipment failure. Therefore, the study of rolling bearing RUL is of great significance. [2] .

[0003] The prediction methods of rolling bearing RUL are mainly divided into two basic categories: model-driven methods and data-driven methods. Model-driven methods mainly predict RUL by building a mathematical or physical model of the rolling bearing performance degradation process. However, the establishment of the physical model is highly dependent on the prior knowledge of the device degradation mechanism. Therefore, it is difficult to express the complex degradation process through an accurate model, which limits the versatility and practicality of this method. [3] .

[0004] In recent years, with the rise of artificial intelligence, data-driven RUL prediction methods have gradually shown their advantages. Among them, methods represented by deep learning can mine valuable feature information from massive data, have the ability of autonomous learning, and provide an effective solution for establishing a mapping relationship between data features and equipment performance degradation trends. [4] . Reference [5] proposed a bearing RUL prediction method based on convolutional neural networks (CNN). Experimental results show that the proposed method is superior to traditional machine learning. Reference [6] proposed a bearing life prediction method based on convolutional attention long short-term memory network. This method can effectively process time series information. Experimental results show that the prediction effect is better than other traditional deep learning models. Reference [7] proposed a rolling bearing remaining service life prediction method combining CNN and LSTM. The prediction result can be close to the actual life value of the bearing. Reference [8] proposed a prediction method based on CNN-TCN-Attention. Experiments show that combining the attention mechanism with the time convolution network can effectively improve the prediction accuracy of the model. Although the above deep learning methods can extract the deep degradation features of the original vibration signal compared with the model-driven method, these methods ignore the problems of data privacy and data islands.

[0005] Federated Learning (FL) is a collaborative machine learning method that does not require centralized training data. Its purpose is to train a global model from distributed data of different users while protecting data privacy and saving a lot of network bandwidth. [9] Reference

[10] proposed a RUL prediction method based on federated learning. When the sample size is limited, it effectively solves the data island problem between different clients. Experiments show that the proposed method achieves good prediction results on both milling cutter and bearing datasets. Reference

[11] proposed a remaining life prediction method based on federated learning and cloud-edge collaboration, which can effectively reduce data transmission delay time and protect data privacy. Reference

[12] proposed a RUL prediction method based on federated learning and network pruning. The bearing experiment proves that this method effectively solves the RUL prediction problem in the data privacy protection scenario.

[0006] In the above FL algorithms, the central server uses the Federated Averaging (FedAvg) algorithm when aggregating parameters. This method aggregates multiple users into a model under the same working conditions, which can well predict the remaining service life of the bearing under the premise of protecting data privacy. However, in actual industry, the user's rolling bearing vibration data usually comes from different working conditions, and the data often presents the characteristics of non-independent and identically distributed. The performance of the global model is easily affected by the distribution of user data. Directly using the FedAvg algorithm to aggregate rolling bearing vibration data under multiple different working conditions may lead to the separation of parameter update directions, slow convergence of the global model, and even deviation from the optimal parameters. Therefore, it is not feasible to use the same global model for all clients.

[0007] In order to solve the problems existing in traditional federated learning algorithms, personalized federated learning has emerged. It can combat the adverse effects of non-IID data. Reference

[13] proposed an improved federated learning algorithm. The algorithm adds proximal terms based on FedAvg, thereby accelerating the convergence of the model and improving the overall performance of the model. Reference

[14] proposed a personalized federated learning method of adaptively learning local models to deal with the problem of non-IID data, and verified the effectiveness of the method on four public datasets. Reference

[15] proposed a personalized federated learning method of fine-tuning and head aggregation. Experiments show that data has good personalization and generalization capabilities under non-IID conditions. Reference

[16] proposed a federated representation learning algorithm to achieve personalized federated learning by learning shared representations across clients and local heads of each client. Experiments have verified that the FL method has good classification effects in non-IID environments.

[0008] The above personalized federated learning method can not only reduce the risk of privacy leakage, but also more easily capture the personalized characteristics of users when the client data is not in an independent and identically distributed condition. However, the personalized model needs to adapt to the local needs of each client and also needs to share certain model parameters, which will further bring difficulties to the personalized federated learning strategy in the field of bearing life prediction.

[0009] Therefore, to address the problem of poor model prediction performance when the bearing vibration data of different users in federated learning are not independent and identically distributed, a rolling bearing life prediction method based on parameter decoupling personalized federation (PDPF) is proposed. This method decouples the local prediction model into a shared representation layer and a local personalized layer, learns the shared representation and personalized feature information of each client, and models each client separately, which helps to improve the overall performance of the prediction model. Summary of the invention

[0010] The technical problems to be solved by the present invention are:

[0011] Existing methods treat rolling bearing vibration data from different working conditions as non-independent and identically distributed data, which may lead to separation of parameter update directions, slow convergence of the global model, and even deviation from the optimal parameters.

[0012] The present invention adopts the following technical solutions to solve the above technical problems:

[0013] The present invention provides a method for predicting the life of a rolling bearing under different working conditions based on parameter decoupling personalized federation (PDPF). The method is based on a federated learning model, and the federated learning model includes a central server and multiple clients, and includes the following steps:

[0014] Step 1: Each client collects bearing life data respectively, and randomly divides the life into a training set and a test set. By calculating the FPT point and HRDT point of the bearing life data of the training set, a multi-level degradation label representation method is used to mark the different stages of the bearing life data of the training set, and the data is processed by wavelet transform;

[0015] Step 2, the central server constructs a network model based on SeResNet and ConvLSTM, the network model includes SEResNet blocks and ConvLSTM blocks, the SEResNet blocks are used for feature extraction, and important features are retained by connecting the maximum pooling layer, and the ConvLSTM blocks are used for further processing of the features; the central server initializes the network model, decouples the network model, uses the SEResNet network as a shared representation layer, uses the ConvLSTM and the fully connected layer as a personalization layer, and sends the shared representation layer to each client;

[0016] Step 3: Each client trains the received shared representation layer and local personalized layer based on the training set, and sends the trained shared representation layer parameters to the central server;

[0017] Step 4: The central server aggregates the received shared representation layer parameters, updates the parameters, obtains the global model, and sends the global model parameters to each client;

[0018] S5, repeating S3 to S4 until the model converges or reaches the maximum number of iterations, to obtain a rolling bearing life prediction model;

[0019] S6. Predicting the bearing life based on the rolling bearing life prediction model.

[0020] Furthermore, step 1 specifically includes the following process:

[0021] First, the monotonicity indicators of the four time domain features, namely, root mean square value (RMS), mean square amplitude (SRA), kurtosis (KU), and margin factor (MF), are calculated:

[0022]

[0023] In the formula, is the difference in eigenvalue change, is the number of difference values ​​greater than 0;

[0024] Secondly, calculate the trend indicators of the above four time domain characteristics:

[0025]

[0026] In the formula, x n and n Represent characteristic value and time value respectively;

[0027] The weights of constructing comprehensive indicators are determined based on monotonicity and trend indicators:

[0028]

[0029] In the formula, ω i Represents the weight of each time domain indicator;

[0030] Construct the time domain comprehensive index CI, namely:

[0031] CI=ω1×u RMS +ω2×u SRA +ω3×u KU +ω4×u MF

[0032] Then, the failure threshold is set according to the mean and standard deviation of the comprehensive index, namely:

[0033] T h =λ+μ

[0034] In the formula, λ represents the standard deviation of CI, μ represents the mean of CI, and T h is the threshold value;

[0035] Finally, the time corresponding to when the time domain comprehensive index CI exceeds the set threshold for the first time is taken as the first prediction time FPT point of the bearing performance degradation curve;

[0036] The method for determining the high risk degradation point HRDT is:

[0037] t h =3(λ+μ)

[0038] The formula for determining multi-level degenerate labels is:

[0039]

[0040] Where, t f represents the FPT moment, t n Indicates the current time, t h represents the HRDT time, t w Represents the total life span.

[0041] Furthermore, the network model in step 2 includes 4 stacked SEResNet blocks, and a BN normalization layer is constructed before the ReLU activation function of the SEResNet module. At the same time, a maximum pooling layer is connected after the SEResNet module, followed by 3 ConvLSTM blocks, and finally the bearing degradation characteristics are obtained through a fully connected layer.

[0042] Furthermore, in step 3, the local client iteratively updates the received shared representation layer and the local personalized layer together using a stochastic gradient descent method until the number of local update iterations is reached;

[0043] The optimization objectives of the federated learning model are:

[0044]

[0045] In the formula, n is the number of samples in the data set, n k is the number of samples of the kth user, F k (ω) is the loss of the kth user.

[0046] The present invention provides a rolling bearing life prediction system under different working conditions based on parameter decoupling personalized federation. The system has a program module corresponding to the steps of the method described in any one of the above technical solutions, and executes the steps in the above-mentioned rolling bearing life prediction method under different working conditions based on parameter decoupling personalized federation during operation.

[0047] The present invention provides a computer-readable storage medium, which stores a computer program. The computer program is configured to implement the steps of the rolling bearing life prediction method under different working conditions based on parameter decoupling personalized federation described in any one of the above-mentioned technical solutions when called by a processor.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] (1) A multi-level degradation label representation method is proposed to label the degradation data of bearings at different stages.

[0050] (2) A local prediction model for rolling bearings based on the improved stacked SeResNet and ConvLSTM is proposed to fully exploit the degradation characteristics and thus improve the prediction accuracy.

[0051] (3) A parameter-decoupled personalized federated learning algorithm suitable for the prediction of the remaining useful life of rolling bearings is proposed. The algorithm can decouple the local prediction model SeResNet-ConvLSTM into a shared representation layer and a personalized layer, and establish a separate personalized prediction model for each client data, thereby realizing the prediction of the remaining useful life of rolling bearings under different working conditions while protecting data privacy.

[0052] (4) The method of the present invention is universal, and the models trained under different data sets all show good performance and can effectively adapt to the service life prediction needs of bearings under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 Schematic diagram of SE block network architecture in an embodiment of the present invention;

[0054] Figure 2 Schematic diagram of the SEResNet network architecture in an embodiment of the present invention;

[0055] Figure 3 Schematic diagram of the ConvLSTM block network architecture in an embodiment of the present invention;

[0056] Figure 4 Schematic diagram of a federated learning framework in an embodiment of the present invention;

[0057] Figure 5 Schematic diagram of a PDPF rolling bearing life prediction method in an embodiment of the present invention;

[0058] Figure 6 Schematic diagram of the local prediction model structure in an embodiment of the present invention;

[0059] Figure 7 A diagram of a bearing test bench in an embodiment of the present invention;

[0060] Figure 8 is the FPT point and HRDT point diagram of the bearing 1_1 in the embodiment of the present invention;

[0061] Fig. 9 It is a wavelet transform time spectrum diagram in an embodiment of the present invention;

[0062] Fig.10 This is a fitting result diagram of bearing 2_6 in an embodiment of the present invention;

[0063] Fig.11 This is a fitting result diagram of bearing 2_6 in scheme 1 in an embodiment of the present invention;

[0064] Fig.12 This is a fitting result diagram of bearing 2_6 of scheme 2 in an embodiment of the present invention;

[0065] Fig.13 This is a comparison diagram of prediction errors of different algorithms in an embodiment of the present invention;

[0066] Fig.14 A comparison chart of performance evaluation index values ​​of different algorithm models in an embodiment of the present invention;

[0067] Fig.15 This is a fitting result diagram of bearing 1_3 in an embodiment of the present invention;

[0068] Fig.16 This is a comparison diagram of prediction errors of different algorithms in an embodiment of the present invention;

[0069] Fig.17 1 is a comparison chart of performance evaluation index values ​​of different algorithm models in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to enable those skilled in the art to better understand the scheme of the present invention, exemplary implementations or embodiments of the present invention will be described below in conjunction with the accompanying drawings. Obviously, the described implementations or embodiments are only implementations or embodiments of a part of the present invention, not all of them. Based on the implementations or embodiments of the present invention, all other implementations or embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0071] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0072] In a typical embodiment of the present invention, a method for predicting the life of a rolling bearing under different working conditions based on parameter decoupling personalized federation (PDPF) is provided. The method is based on a federated learning model, and the federated learning model includes a central server and multiple clients, and includes the following steps:

[0073] Step 1: Each client collects bearing life data respectively, and randomly divides the life into a training set and a test set. By calculating the FPT point and HRDT point of the bearing life data of the training set, a multi-level degradation label representation method is used to mark the different stages of the bearing life data of the training set, and the data is processed by wavelet transform;

[0074] Step 2, the central server constructs a network model based on SeResNet and ConvLSTM, the network model includes SEResNet blocks and ConvLSTM blocks, the SEResNet blocks are used for feature extraction, and important features are retained by connecting the maximum pooling layer, and the ConvLSTM blocks are used for further processing of the features; the central server initializes the network model, decouples the network model, uses the SEResNet network as a shared representation layer, uses the ConvLSTM and the fully connected layer as a personalization layer, and sends the shared representation layer to each client;

[0075] Step 3: Each client trains the received shared representation layer and local personalized layer based on the training set, and sends the trained shared representation layer parameters to the central server;

[0076] Step 4: The central server aggregates the received shared representation layer parameters, updates the parameters, obtains the global model, and sends the global model parameters to each client;

[0077] S5, repeating S3 to S4 until the model converges or reaches the maximum number of iterations, to obtain a rolling bearing life prediction model;

[0078] S6. Predicting the bearing life based on the rolling bearing life prediction model.

[0079] In one embodiment of the present invention, the root mean square (RMS) and square root amplitude (SRA) gradually increase with the degradation of bearing performance, but are insensitive to early bearing failure; Kurtosis (KU) is sensitive to early bearing failure, but is easily interfered by noise; Margin Factor (MF) is not affected by the operating conditions of the equipment and only depends on the shape of the probability density function. Therefore, the above four time domain indicators that can reflect the degradation trend of the bearing are combined to construct a time domain comprehensive index (CI). The calculation method of the time domain index is shown in formulas (1)-(4):

[0080]

[0081] In the formula, N represents the number of eigenvalues, x i Represents the value corresponding to the i-th feature point, represents the mean, X σ Represents the sample standard deviation.

[0082] First, the monotonicity index of the above four time domain features is calculated, as shown in formula (5):

[0083]

[0084] In the formula, is the difference in the change of eigenvalues. In actual calculation, it is the next eigenvalue minus the previous eigenvalue. is the number of difference values ​​greater than 0.

[0085] Secondly, the trend indicators of the above four time domain characteristics are calculated, as shown in formula (6):

[0086]

[0087] In the formula, x n and n Represent the eigenvalue and time value respectively. The range of T is [-1, 1], where -1 means the eigenvalue is strictly decreasing and 1 means the eigenvalue is strictly increasing.

[0088] Therefore, the weight of constructing the comprehensive index is determined according to the calculated monotonicity and trend index, as shown in formula (7):

[0089]

[0090] In the formula, ωi Represents the weight of each time domain indicator.

[0091] The constructed time domain comprehensive index CI is calculated by formula (8), namely:

[0092] CI=ω1×u RMS +ω2×u SRA +ω3×u KU +ω4×u MF (8)

[0093] Then, the failure threshold is set according to the mean and standard deviation of the comprehensive index. That is:

[0094] T h =λ+μ (9)

[0095] In the formula, λ represents the standard deviation of CI, μ represents the mean of CI, and T h is the threshold value.

[0096] Finally, the time corresponding to when the time domain comprehensive index CI exceeds the set threshold for the first time is taken as the first prediction time FPT point of the bearing performance degradation curve.

[0097] Determining an appropriate first prediction time (FPT) is crucial for accurately predicting RUL, and a reasonable method for constructing FPT will help improve the prediction accuracy of bearing RUL.

[0098] In the late stage of bearing degradation, the vibration signal rapidly degrades and the bearing is on the verge of failure. That is, there is a high risk degradation point (High Risk Degradation Time, HRDT) in the bearing degradation stage. The determination method is shown in formula (10):

[0099] t h =3(λ+μ) (10)

[0100] Through FPT and HRDT, the entire life cycle of rolling bearings can be divided into three different degradation stages: healthy zone, low-risk degradation zone and high-risk degradation zone.

[0101] Based on the different degradation stages of the bearing, a method for constructing a multi-level segmented label is proposed. In the healthy area, since the bearing has not yet begun to degrade, the label of this area is set to 0; in the low-risk degradation area, the life percentage is used as the label of this area; after entering the high-risk degradation area, the bearing is prone to failure, so the label of this area is set to 1. The determination of the multi-level degradation label is shown in formula (11):

[0102]

[0103] Where, tf represents the FPT moment, t n Indicates the current time, t h represents the HRDT time, t w Represents the total life span.

[0104] In one embodiment of the present invention, a Squeeze-and-Excitation Network (SE) is used to adaptively recalibrate channel-level feature responses by establishing interdependence between channels; SE can enhance favorable channel information and suppress unfavorable channel information, and is a channel attention mechanism. Its network architecture is as follows Figure 1 shown. Figure 1 Medium, F tr Represents convolution calculation, X∈R H'×W'×C' , U∈R H×W×C , its output is shown in formula (12):

[0105]

[0106] In the formula, u c represents the c-th two-dimensional matrix in U, X is the input feature, v c represents the parameters of the cth convolution kernel, x s represents the input of the sth channel, Represents the action on X corresponding to v c The two-dimensional convolution kernel of a single channel, * represents the convolution operation.

[0107] The Squeeze operation performs compression on the channel, which is equivalent to the global average pooling operation, compressing the H×W×C features to 1×1×C, expressed as:

[0108]

[0109] Then, the channel dependency is captured through the Excitation operation, and the weight corresponding to each channel is shown in formula (14):

[0110] s=F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (14)

[0111] Where σ represents the sigmoid activation function, δ represents the ReLU activation function, and W1∈R C×C / r and W2∈R C / r×C Represent the fully connected operations of reducing and increasing dimensions respectively.

[0112] The output of the SE block is obtained by re-adjusting the input feature X by the weights of the dependencies between each channel:

[0113]

[0114] In the formula, · represents the product operation.

[0115] The SE block is flexible and can be used as a non-identity mapping branch of the residual module. Figure 2 Describes the network architecture of SEResNet.

[0116] The convolution operator in the convolutional long short-term memory network replaces the fully connected layer inside the long short-term memory network structure to reduce the redundancy of the network and enhance its nonlinear modeling ability; at the same time, it integrates the time-frequency features and retains the advantages of long-term data dependencies. Its basic output is shown in formulas (16)-(21):

[0117] i t =σ(W xi *X t +W hi *H t-1 +b i ) (16)

[0118] f t =σ(W xf *X t +W hf *H t-1 +b f ) (17)

[0119] o t =σ(W xo *X t +W ho *H t-1 +b o ) (18)

[0120]

[0121] H t =o t ⊙tanh(C t ) (twenty one)

[0122] Where W xi , W xf , W xo , W xs Respectively represent the weight values ​​of the input gate, forget gate, output gate and memory unit between the input layer and the hidden layer at time t; W hi , W hf , W ho , W hs Respectively represent the weight values ​​of the input gate, forget gate, output gate and memory unit of the hidden layer between time t-1 and time t; b i 、b f、b o 、b s Respectively represent the biases of the three gates and memory nodes; * represents convolution operation, ⊙ represents element-by-element multiplication operation; σ represents the sigmoid activation function. The ConvLSTM network architecture is as follows Figure 3 shown.

[0123] During the federated learning training process, the user's data does not need to be uploaded to the central server, and is only completed by uploading and sending model parameters. The optimization goal of FL is to minimize the loss function f(ω) on the entire data, that is:

[0124]

[0125] In the formula, n is the number of samples in the data set, n k is the number of samples of the kth user, F k (ω) is the loss of the kth user.

[0126] FL consists of two steps, local training and global aggregation. In local training, edge devices download models from the central server and use local data to calculate updated models. The central server aggregates these updated models mainly through the federated averaging algorithm. That is:

[0127]

[0128] Where K represents the number of users participating in federated training, ω t+1 is the global model parameter in the t+1 iteration, is the local model parameter of the kth user. The federated learning framework is as follows Figure 4 shown.

[0129] The PDPF method of the present invention realizes personalization by decoupling local private model parameters from global aggregated model parameters. The private model parameters are trained locally on the client and are not shared with the FL server, thereby establishing a separate personalized prediction model for each client.

[0130] In the personalized federated learning method proposed in the present invention, the forward propagation operation of the jth client is set as:

[0131]

[0132] In the formula, K B and K P Respectively represent the number of shared presentation layers and personalization layers on each client; represents the weight matrix of the shared representation layer, Represents the corresponding activation function; represents the weight matrix of the personalization layer, Represents the corresponding activation function.

[0133] Use P j Denote the joint distribution generated at the jth client, and l(·,·) denotes the loss function of each sample. The learning objective of the parameter-decoupled personalized federated learning method is:

[0134]

[0135] In the formula, K represents the number of users participating in the training, W B represents the set of weight matrices of the shared representation layer, A collection of weight matrices representing the personalization layer.

[0136] During model training, the shared representation layer and the personalization layer participate in local updates together:

[0137]

[0138] Where u represents the update parameter of the personalization layer, φ represents the update parameter of the shared representation layer, γ represents the learning rate, and t represents the total number of federation rounds.

[0139] The central server aggregates the shared representation layer model parameters using a federated averaging algorithm:

[0140]

[0141] In one embodiment of the present invention, the constructed training set full-life data and the test set non-full-life data are subjected to continuous wavelet transform to construct a two-dimensional image data set as an input of the prediction model.

[0142] The network model uses stacked 4 SEResNet blocks to extract features. BN normalization is performed before the ReLU activation function in the SEResNet block. At the same time, a maximum pooling layer is added after the SEResNet module to reduce the dimension, thereby retaining important features. Then a 3-layer ConvLSTM network is used to further process the features, and finally a fully connected layer is used to obtain the bearing degradation features. The specific structure of the prediction model is as follows: Figure 6 shown.

[0143] During the iterative update process of the central server and local client, the prediction model parameters are first randomly initialized, and the prediction model is decoupled into a shared representation layer and a personalized layer. The SEResNet network is used as the shared representation layer, and ConvLSTM and FC are used as the personalized layer. Secondly, the local user uses the stochastic gradient descent algorithm to iteratively update the shared representation layer and the local personalized layer aggregated by the central server. When the number of local update iterations is reached, the central server uses FedAvg to aggregate the model parameters of the shared representation layer and sends the aggregated parameters to each client. The above process is repeated until the maximum number of federated iterations is reached. Finally, a corresponding personalized prediction model is generated for each user, and the model training is completed.

[0144] The wavelet transformed test set data is input into the model of the corresponding working condition of the training set for prediction. The degradation curve is obtained by fitting the degradation state through the double exponential model. When the ordinate label reaches 1, the corresponding abscissa is the predicted full life value. The RUL of the bearing can be obtained by subtracting the current time from the full life time:

[0145] RUL=t w -t n (29).

[0146] The present invention proposes a method (algorithm) for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation, which is the underlying technical core of the present invention. Various products can be derived based on the algorithm.

[0147] Based on the method proposed in the present invention, a rolling bearing life prediction system under different working conditions based on parameter decoupling personalized federation is developed using a programming language. The system has program modules corresponding to the steps of the above-mentioned technical solution, and executes the steps in the above-mentioned rolling bearing life prediction method under different working conditions based on parameter decoupling personalized federation during operation.

[0148] The computer program of the developed system (software) is stored on a computer-readable storage medium, and the computer program is configured to implement the steps of the above-mentioned method for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation when called by a processor. That is, the present invention is materialized on a carrier to become a computer program product.

[0149] Various implementations of the systems and techniques described herein can be realized in digital electronic circuit systems, integrated circuit systems, dedicated ASICs (application specific integrated circuits), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0150] The computer programs (also referred to as programs, software, software applications, or codes) of the present invention include machine instructions for programmable processors, and these computer programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, device, and / or device (e.g., disk, optical disk, memory, programmable logic device PLD) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium that receives machine instructions as machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.

[0151] The beneficial effects of the present invention will be described below in conjunction with specific embodiments.

[0152] Example 1

[0153] The PHM2012 bearing data set is selected for verification in this embodiment. The sampling interval of the data set is 10s, the sampling time is 0.1s each time, and the sampling frequency is 25.6kHz. It contains vibration data in both horizontal and vertical directions. The data in the horizontal acceleration direction can provide more useful information, so this embodiment only uses the vibration data in the horizontal acceleration direction for the experiment. The test bench is as follows: Figure 7 shown.

[0154] This dataset contains bearing data under three different working conditions, and its working condition information is shown in Table 1. It contains 6 sets of full-life training set data and 11 sets of non-full-life test set data, and its detailed information is shown in Table 2.

[0155] Table 1 Working condition information

[0156]

[0157] Table 2 Experimental data

[0158]

[0159]

[0160] Experimental setup

[0161] The hardware environment used in the experiment of this embodiment is: the CPU model is Intel Core i5-13500H; the memory is 64GB; the GPU model is: NVIDIA GeForce RTX4050.

[0162] In order to simulate the condition that the data of each client in federated learning is not independent and identically distributed, the users are set according to Table 3.

[0163] Table 3. Federation local user data set settings

[0164]

[0165] According to formula (9), the sum of the single mean and standard deviation of the comprehensive index of bearing 1_1 is 0.1156, and the time point when the value is exceeded for the first time is 782×10s=7820s. According to formula (10), the sum of the triple mean and standard deviation is 0.3468, and the time point when the value is exceeded for the first time is 2762×10s=27620s. Therefore, the FPT and HRDT points are 7820s and 27620s. The comprehensive index curve and the corresponding degradation nodes are shown in Figure 1. Figure 8 shown.

[0166] The PHM2012 bearing data set is subjected to wavelet transform. Taking a section of data from bearing 1_1 as an example, the wavelet transform results in

[0167] The time-frequency spectrum is as follows Fig. 9 As shown in Table 4. The darker the color of the energy bar in the wavelet transform spectrum, the greater the energy. During the bearing degradation process, the change in color can show that the bearing gradually begins to fail over time. The FPT points and HRDT points of the PHM2012 bearing data set are shown in Table 4. According to Table 4 and formula (11), three-level degradation labels can be set for the six groups of full life data in the training set.

[0168] Table 4 Results of FPT and HRDT points

[0169]

[0170] In this embodiment, the SEResNet network is used as the shared representation layer, and the central server aggregates the model parameters of this part; the three-layer ConvLSTM and FC network are used as the personalization layer of the local model and are retained locally. Each client uses the stochastic gradient descent algorithm to update its shared representation layer and personalization layer parameters locally. The parameter settings of federated learning are shown in Table 5.

[0171] Table 5 Hyperparameter settings for the federated learning phase

[0172]

[0173]

[0174] The double exponential function can effectively capture the degradation trend of the bearing, and the model established based on the double exponential function can effectively predict the RUL. Therefore, the double exponential function is used to fit the degradation state obtained from the test set, as shown in formula (30):

[0175]

[0176] This embodiment uses the evaluation indicators provided by the PHM2012 Bearing Data Challenge to perform evaluation, and it can be concluded whether the prediction result is an advanced prediction or a lagging prediction.

[0177]

[0178] In the formula, A i Defined as:

[0179]

[0180] In the formula, Er i is the prediction error. A value less than 0 indicates a lagging prediction, and a value greater than 0 indicates an advanced prediction.

[0181]

[0182] Taking user 2 test data 2_6 as an example, the fitting results are as follows Fig.10As shown, since the sampling interval of the data set is 10s, when the label of the ordinate reaches 1, the predicted full life time of bearing 2_6 is 685×10s=6850s. This embodiment uses a 95% confidence interval. The full life time in this interval is [644, 723] (10s), the predicted RUL is 6850-5720=1130s, the actual full life time is 7010s, and the actual RUL is 7010-5720=1290s. According to formula (33), the prediction error of bearing 2_6 can be calculated to be 12.40%. Similarly, the prediction results and errors of the remaining test set bearings can be obtained, as shown in Table 6. Therefore, according to formula (31) and formula (32), the average prediction score can be calculated to be 0.4572.

[0183] Table 6 Prediction results and prediction errors of PDPF method for different bearings

[0184]

[0185]

[0186] In order to verify the superiority of the network model SEResNet-ConvLSTM in this embodiment, the existing CNN-LSTM network model [7] The prediction results are compared with those of the previous one, and the comparison results of the prediction errors are shown in Table 7.

[0187] Table 7 Comparison of prediction errors of different models

[0188]

[0189]

[0190] It can be seen from Table 7 that the average prediction error of the network model SEResNet-ConvLSTM proposed in the present invention is 7.08% lower than that of the existing model, and the average score is increased by 0.1916, which proves that the proposed method can effectively improve the prediction accuracy.

[0191] In order to verify the superiority of the PDPF method proposed in this embodiment, two federation schemes are set as comparative experiments. In the comparative experiments, the network prediction model structures of all RULs are the same.

[0192] Solution 1: The fully connected layer is used as the personalized layer, and other network layers are used as the shared representation layer. The personalized federated learning method is used as a federated comparison experiment 1. Taking user 2 test data 2_6 as an example, the prediction results are as follows: Fig.11 shown.

[0193] According to the prediction method of Scheme 1, the prediction results of all the test set bearings can be obtained, as shown in Table 8. Although the prediction error of some bearings in the PDPF method proposed in the present invention is higher than that in Scheme 1, from the perspective of the overall average error, the method proposed in the present invention is 1.3% lower than that in Scheme 1, and can still maintain a good prediction effect.

[0194] Table 8 Comparison of prediction error results between Scheme 1 and PDPF method

[0195]

[0196] Solution 2: Set the personalized layer number of the PDPF method to 0, and the algorithm evolves into FedAvg. In this embodiment, FedAvg is used as the federated comparison experiment 2. The prediction results obtained are shown in Table 9, where the full life prediction results of bearing 2_6 are as follows: Fig.12 shown.

[0198] Table 9 Comparison of prediction error results between Scheme 2 and PDPF method

[0199]

[0200] The prediction errors of the PDPF method proposed in this invention and the algorithms of scheme 1 and scheme 2 for the bearings of the PHM2012 test set are compared as follows: Fig.13 As shown. Among them, the vertical axis prediction error less than 0 indicates lagging prediction, and greater than 0 indicates advanced prediction. According to formula (31), the average prediction score of each method can be obtained, and the calculation results and average error are shown in Table 10.

[0201] Table 10 Comparison of average scores and errors

[0202]

[0203] It can be seen from Table 10 that the average score of the method proposed in the present invention is improved by 0.0566 and the average error is reduced by 1.3% compared with the first solution, and the average score is improved by 0.155 and the average error is reduced by 7.52% compared with the second solution, which proves the effectiveness of the proposed method.

[0204] Mean absolute error (MAE) and root mean square error (RMSE) are commonly used indicators for evaluating the performance of prediction models. In order to prove the prediction effect of the method proposed in this invention, MAE and RMSE evaluation indicators are used to evaluate the prediction results. The corresponding calculation methods are shown in formulas (34) and (35). Comparison results of MAE and RMSE of three federation methods Fig.14 shown.

[0205]

[0206]

[0207] Depend on Fig.14 It can be seen that the MAE and RMSE values ​​of the method proposed in the present invention are reduced by 0.0414 and 0.0394 respectively compared with scheme 1, and are reduced by 0.1707 and 0.2061 respectively compared with scheme 2, which further proves the superiority of the PDPF method in RUL prediction.

[0208] Example 2

[0209] In order to further verify the effectiveness of the proposed method, this embodiment also uses the bearing data set of Xi'an Jiaotong University XJTU-SY for experimental verification. The sampling frequency of this data set is 25.6kHz, the sampling interval is 1min, the single sampling time is 1.28s, and it contains 15 sets of full life data under three working conditions. The detailed information of the data set is shown in Table 11.

[0210] Table 11 Experimental data

[0211]

[0212] According to the setting method of Example 1, the data of each working condition is taken as a user, and different training models for each working condition data are trained. The prediction results and prediction errors of the XJTU-SY bearing data set using the PDPF method are shown in Table 12. The prediction results of bearing 1_3 are as follows Fig.15 shown.

[0213] Table 12 RUL prediction results and prediction errors

[0214]

[0215] According to formula (31), the average score of each method can be obtained, and the calculation results and average error are shown in Table 13.

[0216] Table 13 Comparison of average scores and errors

[0217]

[0218] It can be seen from Table 13 that the average score of the method proposed in the present invention is 0.0455 higher than that of solution 1, and the average error is reduced by 4.12%. The average score of solution 2 is 0.1135 higher than that of solution 2, and the average error is reduced by 7%, which proves that the proposed method has a good prediction effect.

[0219] The same comparative experiment was set up according to Example 1. The comparison results of the prediction errors of the three federation methods are as follows: Fig.16 As shown, the comparison results of MAE and RMSE are as follows Fig.17 shown.

[0220] Depend on Fig.16 and Fig.17 It can be seen that the method proposed in the present invention can achieve the same prediction effect on the bearing data set of XJTU-SY.

[0221] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

[0222] The documents cited in this invention include:

[0223] [1] Liu Xiaofeng, Li Junfeng, Bi Yuanliang, et al. Bearing fault diagnosis method based on optimal frequency band cyclic pulse index spectrum[J]. Chinese Journal of Scientific Instrument, 2024, 45(06): 297-306.

[0224] Liu Xiaofeng, Li Junfeng, Bi Yuanliang, et al. Bearing fault diagnosis method based on cyclic pulse index spectrum of optimal band [J]. Chinese Journal of Scientific Instrument, 2024, 45(06): 297-306.

[0225] [2]Ding W, Li J, Mao W, et al. Rolling bearing remaining useful lifeprediction based on dilated causal convolutional DenseNet and an exponential model [J]. Reliability Engineering & System Safety, 2023, 232: 109072.

[0226] [3] Zhang Fan, Yao Dechen, Yao Shengzhuo, et al. Bearing life prediction based on Transformer-LSTM network[J]. Vibration and Shock, 2024, 43(06): 320-328.

[0227] Zhang Fan,Yao Dechen,Yao Shengzhuo,et al.Lifetime prediction ofbearings based on the Transformer-LSTM network[J].Journal of Vibration andShock,2024,43(06):320-328.

[0228] [4] Shi Peng. Research on the prediction method of remaining service life of medium carbon steel based on recurrent neural network and deep belief network[D]. Northeastern University, 2019.3-4.

[0229] Shi Peng.Research on Remaing Useful Life Prediction Method of MediumCarbon Steel Based on Recurrent Neural Networkand Deep Belief Network[D].Northeastern University,2019.3-4.

[0230] [5]Ren L, Sun Y, Wang H, et al. Prediction of bearing remaining usefullife with deep convolution neural network [J]. IEEE Access, 2018, 6: 13041-13049.

[0231] [6] Zhou Jianmin, Gao Sen, Li Jiahui, et al. Bearing life prediction method based on convolutional attention long short-term memory network[J]. Control Theory and Applications, 2023, 40(06): 1140-1148.

[0232] Zhou Jianmin,Gao Sen,Li Jiahui,et al.Bearing life prediction methodbased on convolutional attention long-short term memory network[J].ControlTheory&Applications,2023,40(06):1140-1148.

[0233] [7] Wang Yujing, Li Shaopeng, Kang Shouqiang, et al. Prediction method of remaining service life of rolling bearings based on CNN and LSTM[J]. Vibration. Test and Diagnosis, 2021, 41(03): 439-446+617.

[0234] Wang Yujing,Li Shaopeng,Kang Shouqiang,et al.Method of PredictingRemaining Useful Life of Rolling Bearing Combining CNN and LSTM[J].Journal ofVibration, Measurement&Diagnosis, 2021,41(03):439-446+617.

[0235] [8] Sun Danming, Chen Changzheng, Sun Yepeng. Rolling bearing life prediction based on CNN and TCN neural network[J]. Mechanical Design and Manufacturing, 2024, (08): 160-165.

[0236] Sun Danming,Chen Changzheng,Sun Yepeng.Rolling Bearing LifePrediction Combined with CNN and TCN Neural Network[J].Machinery Design&Manufacture, 2024,(08):160-165.

[0237] [9]Bemani A, N.Aggregation strategy on federated machinelearning algorithm for collaborative predictive maintenance[J].Sensors, 2022, 22(16):6252.

[0238]

[10] Guo L, Yu Y, Qian M, et al. FedRUL: A new federated learning method for edge-cloud collaboration based on remaining useful life prediction of machines [J]. IEEE / ASME Transactions on Mechatronics, 2022, 28(1): 350-359.

[0239]

[11] Yu Zhenjun, Lei Ningbo, Mo Yu, et al. Remaining life prediction based on federated learning and cloud-edge collaboration[J / OL]. Journal of Tsinghua University (Science and Technology), 2024, 22(034): 1-11.

[0240] Yu Zhenjun,Lei Ningbo,Mo Yu,et al.Remaining useful life predictionbased on federated learning and cloud-edge collaboration[J / OL].Journal ofTsinghua University(Science and Technology),2024,22(034):1-11.

[0241]

[12] Chen

[0242]

[13] Li T, Sahu AK, Zaheer M, et al. Federated optimization inheterogeneous networks [J]. Proceedings of Machine Learning and Systems, 2020, 2: 429-450.

[0243]

[14] Chen D, Yao L, Gao D, et al. Efficient personalized federated learning via sparse model-adaptation[C] / / International Conference on Machine Learning.PMLR, 2023:5234-5256.

[0244]

[15] Wang Y,Xu H,Ali W,et al.FedFTHA:A fine-tuning and headaggregation method in federated learning[J].IEEE Internet of Things Journal,2023.12749-12762.

[0245]

[16] Collins L,Hassani H,Mokhtari A,et al.Exploiting sharedrepresentations for personalized federated learning[C] / / InternationalConference on Machine Learning.PMLR,2021:2089-2099.

Claims

1. A method for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation, characterized in that: The method is based on a federated learning model, which includes a central server and multiple clients, and includes the following steps: Step 1: Each client collects bearing life data respectively, and randomly divides the life into a training set and a test set. By calculating the FPT point and HRDT point of the bearing life data of the training set, a multi-level degradation label representation method is used to mark the different stages of the bearing life data of the training set, and the data is processed by wavelet transform; Step 2, the central server constructs a network model based on SeResNet and ConvLSTM, the network model includes SEResNet blocks and ConvLSTM blocks, the SEResNet blocks are used for feature extraction, and important features are retained by connecting the maximum pooling layer, and the ConvLSTM blocks are used for further processing of the features; the central server initializes the network model, decouples the network model, uses the SEResNet network as a shared representation layer, uses the ConvLSTM and the fully connected layer as a personalization layer, and sends the shared representation layer to each client; Step 3: Each client trains the received shared representation layer and local personalized layer based on the training set, and sends the trained shared representation layer parameters to the central server; Step 4: The central server aggregates the received shared representation layer parameters, updates the parameters, obtains the global model, and sends the global model parameters to each client; S5, repeating S3 to S4 until the model converges or reaches the maximum number of iterations, to obtain a rolling bearing life prediction model; S6. Predicting the bearing life based on the rolling bearing life prediction model.

2. The method for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation according to claim 1 is characterized in that: Step 1 specifically includes the following process: First, the monotonicity indicators of the four time domain features, namely, root mean square value (RMS), mean square amplitude (SRA), kurtosis (KU), and margin factor (MF), are calculated: In the formula, is the difference in eigenvalue change, is the number of difference values ​​greater than 0; Secondly, calculate the trend indicators of the above four time domain characteristics: In the formula, x n and n Represent characteristic value and time value respectively; The weights of constructing comprehensive indicators are determined based on monotonicity and trend indicators: In the formula, ω i Represents the weight of each time domain indicator; Construct the time domain comprehensive index CI, namely: CI=ω1×u RMS +ω2×u SRA +ω3×u KU +ω4×u MF Then, the failure threshold is set according to the mean and standard deviation of the comprehensive index, namely: T h =λ+μ In the formula, λ represents the standard deviation of CI, μ represents the mean of CI, and T h is the threshold value; Finally, the time corresponding to when the time domain comprehensive index CI exceeds the set threshold for the first time is taken as the first prediction time FPT point of the bearing performance degradation curve; The method for determining the high risk degradation point HRDT is: t h =3(λ+μ) The formula for determining multi-level degenerate labels is: Where, t f represents the FPT moment, t n Indicates the current time, t h represents the HRDT time, t w Represents the total life span.

3. The method for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation according to claim 1 is characterized in that: The network model in step 2 includes 4 stacked SEResNet blocks. A BN normalization layer is constructed before the ReLU activation function of the SEResNet module. At the same time, a maximum pooling layer is connected after the SEResNet module, followed by 3 ConvLSTM blocks, and finally the bearing degradation characteristics are obtained through a fully connected layer.

4. The method for predicting the life of rolling bearings under different working conditions based on parameter decoupling personalized federation according to claim 1 is characterized in that: In step 3, the local client uses a stochastic gradient descent method to iteratively update the received shared representation layer and the local personalized layer together until the number of local update iterations is reached; The optimization objectives of the federated learning model are: In the formula, n is the number of samples in the data set, n k is the number of samples of the kth user, F k (ω) is the loss of the kth user.

5. A rolling bearing life prediction system under different working conditions based on parameter decoupling personalized federation, characterized in that: The system has a program module corresponding to the steps of the method described in any one of claims 1 to 4 above, and executes the steps of the rolling bearing life prediction method under different working conditions based on parameter decoupling personalized federation during operation.

6. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of the method for predicting the life of a rolling bearing under different working conditions based on parameter decoupling personalized federation according to any one of claims 1 to 4 when called by a processor.

Images