Life migration prediction method based on improved compression aggregation excitation network

Abstract

The invention belongs to the technical field of mechanical life prediction, and discloses a life migration prediction method based on an improved compression aggregation excitation network, and the specific technical scheme is as follows: constructing health indexes of a rolling bearing by using a deep convolution auto-encoder, and measuring the DTW distance between the health indexes to divide the degradation type of the rolling bearing; dividing the degeneration type of the bearing into gradual change and mutability; an improved compression aggregation excitation network combined with the maximum mean value difference is built, a service life prediction strategy under different degradation types is provided, and the service life of the rolling bearing is predicted under multiple working conditions; a life migration prediction model based on degradation type classification and an improved compression aggregation excitation network is provided, the model divides degradation types of the bearing under different working conditions and realizes residual life prediction, and the model has higher prediction precision and better robustness than other methods under three working conditions of a PHM2012 data set.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical life prediction, and in particular relates to a method based on an improved transferable compression aggregation excitation network for classifying rolling bearing degradation types under different working conditions and estimating the remaining life of the bearing. Background Art

[0002] With the popularization of industrial production equipment and the development of mechanical equipment towards integration, automation and complexity, traditional maintenance strategies are difficult to meet the needs of efficient and intelligent maintenance. Therefore, the remaining service life of equipment has become the focus of many maintenance personnel. Accurately predicting the remaining service life of mechanical equipment is helpful to formulate maintenance plans, reduce production costs, improve production efficiency and ensure production safety. Bearings are one of the most important basic mechanical components, and their operating conditions directly affect the safety and stability of mechanical equipment. They are widely used in mechanical equipment in various fields, such as automobiles, aerospace, electric power, engineering machinery, etc. Bearings are the most frequently used parts in mechanical equipment. However, due to long-term rotation and friction, bearings are also the parts with the highest failure frequency. 45%-55% of rotating machinery failures are caused by rolling bearings. At the same time, bearing failures can also lead to catastrophic accidents, causing serious mechanical failures and safety risks. Therefore, predicting the life of bearings can not only improve production efficiency, but also eliminate safety hazards and prevent economic losses and casualties.

[0003] At present, the prediction of bearing life has become a hot research topic in engineering practice. In order to solve the various problems faced in the life prediction process, scholars have proposed many models and methods in many aspects and achieved fruitful results. According to the differences in existing research methods and models, the remaining service life prediction technology can be summarized as model-based methods and data-driven methods.

[0004] The model-based prediction method describes the degradation process of rolling bearings by constructing a mathematical model that describes the physical characteristics of rolling bearings. Although the model-based prediction method can only provide an approximate solution, this method can help researchers gain a deeper understanding of the degradation process of rolling bearings and lay the foundation for the subsequent development of new prediction methods. The Paris-Erdogan model is one of the most widely used models in the field of rolling bearings. In 2016, Lei et al. combined particle filtering with the Paris-Erdogan model to predict the remaining service life of rolling bearings; Sheng proposed an improved Paris-Erdogan model to predict the remaining service life of bearings after initial failure after constructing health indicators. In addition to the Paris-Erdogan model, models for predicting remaining service life include the Forman model and the Walker model. The accuracy of the model-based prediction method depends on the researchers' understanding of the failure mechanism of rolling bearings. However, due to the variability of the operating environment of rolling bearings, the differences in bearing materials, and the influence of multiple nonlinear strong coupling effects of various mechanical components, the model finally established is only an approximate solution. Therefore, there is a large error between the remaining service life predicted by this method and the actual remaining service life.

[0005] Data-driven prediction methods mainly describe the degradation process of equipment by extracting health indicators or degradation characteristics from the test data of the equipment, thereby establishing a connection between data and remaining life. This type of method does not require in-depth research on the degradation mechanism of the equipment and is widely used in current applications. Data-driven prediction methods can be divided into a variety of methods based on probability statistics and machine learning. The probability statistics-based method does not rely on any physical failure mechanism, but describes the random process of equipment degradation based on empirical knowledge. This method can effectively predict the uncertainty of the remaining useful life (RUL) of the equipment.

[0006] Wang et al. established a nonlinear model driven by the Wiener process to describe the degradation process of the degraded system under incomplete maintenance. Sun et al. applied the Wiener process and Copula function to construct a developed reliability analysis model with multiple degradation processes and random shocks. However, these statistical methods are highly dependent on historical data. As the equipment degrades, these statistical features may gradually lose the ability to describe the true degradation process of the equipment, resulting in a decrease in the prediction accuracy of the model over time. With the increasing development of machine learning, more and more machine learning techniques are being used in the field of remaining life prediction, among which shallow learning algorithms such as support vector machine (SVM) and relevance vector machine (RVM) have been widely used in the field of remaining life prediction. Khelif et al. developed a RUL prediction model based on SVM, which can directly link the features extracted by sensors with the RUL of components, further improving the prediction efficiency. Huang et al. proposed a new remaining life prediction method based on multi-support vector regression fusion, which is suitable for small sample conditions. Li et al. proposed a multi-technique fusion method using relevance vector machine (RVM) and deep separable convolutional gated recurrent network (DSCGRN) for life prediction of rotating machinery. Wang et al. proposed an adaptive RVM that can overcome the inevitable uncertainty problem of mechanical systems during performance degradation due to changes in load, operating and usage conditions.

[0007] Since in shallow machine learning, manual feature engineering, i.e., manual selection and extraction of features, is required, this becomes difficult and time-consuming when processing large-scale data. Deep learning (DL) is a development project based on artificial neural networks (ANN). It uses deep neural networks to make the model more complex, thereby enhancing the model's ability to understand the data. DL has a deeper level of abstraction and dimensionality reduction capabilities and can effectively process high-dimensional monitoring data. In the field of remaining life prediction, DL has shown a strong ability to extract nonlinear system features and process large amounts of data. Yang et al. proposed a remaining useful life (RUL) prediction method based on a dual convolutional neural network (CNN) model architecture, which realizes the identification of early failure points and the identification and prediction of dual degradation models. Wang et al. proposed a recursive convolutional neural network and combined it with variational inference to quantify the uncertainty in the prediction results of the model.

[0008] Deep learning methods require that the training set and the test set have the same distribution. However, in practical applications, the working conditions of bearings are constantly changing. Therefore, the distribution of the collected bearing data features is often inconsistent, which leads to poor results in judging the fault status of the bearing through the trained network model. Transfer learning can mine the common features of two data sets with different distributions, effectively solving the problem of low accuracy of bearing fault diagnosis under variable working conditions using deep learning methods.

[0009] At present, transfer learning has been widely used in the field of remaining life prediction. Common transfer learning methods include domain adaptation, parameter transfer and feature transfer. Mao et al. extracted domain-invariant features by transfer component analysis (TCA) to correct the distribution transfer of deep features. This is the first application of DA in RUL prediction. Paulo et al. proposed a deep domain adaptation method to solve the distribution difference problem in RUL prediction. Huang et al. constructed a transfer deep separable convolutional recurrent network for RUL prediction under different working conditions from the same public data set. In view of the problem that different failure behaviors of bearings lead to large differences in data distribution, Chen et al. proposed a transferable convolutional neural network. By adding domain adaptation operations after the fully connected layer of the convolutional neural network, the feature distribution differences between the source domain bearing and the target domain bearing were successfully reduced. This method helps to improve the generalization ability of the model. Li et al. proposed a prediction method that combines the ideas of deep autoencoder neural network (DANN) and generative adversarial network (GAN). The joint action of DANN and GAN is used to better capture the features and distribution in the data. It is used in the life prediction research of mechanical equipment and improves the accuracy and stability of life prediction. Most studies are based on the assumption that the conditional probability distributions of the two domains are similar, and only the marginal probability distributions in domain adaptation are matched. To address this problem, He et al. proposed a new domain adaptation method to achieve migration RUL prediction by minimizing the distance between the marginal probability distribution and the conditional probability distribution in different domains. However, these transfer learning methods do not take into account the impact of degradation type on the migration prediction results when predicting the life migration of bearings. Different degradation types will affect the accuracy and effectiveness of life migration prediction. Summary of the invention

[0010] In order to solve the technical problems existing in the prior art, the present invention proposes a method based on an improved transferable compressed aggregation excitation network for dividing the degradation types of rolling bearings under different working conditions and estimating the remaining life of the bearings. A deep convolutional autoencoder is used to construct a priori gradual health indicators, and the degradation types of the bearings are further divided into gradual and sudden types by measuring the dynamic time warping distance. The improved compression and excitation module is used to automatically learn the mapping relationship between features and life, and combined with the maximum mean difference in the transfer learning method, the degradation type classification of rolling bearings under different working conditions is completed to realize the remaining life prediction of the bearings.

[0011] To achieve the above purpose, the technical solution adopted by the present invention is: a life migration prediction method based on an improved compressed aggregation excitation network, the specific steps are:

[0012] 1) Install acceleration sensors on experimental equipment under different working conditions, and obtain vibration signal data by setting appropriate sampling rate, sampling interval, duration of each sampling, and shutdown conditions to obtain the full life data of the bearing.

[0013] 2) After the data collected under different working conditions are preprocessed by Hilbert envelope spectrum, the envelope spectrum data sequence is input into DACE for autonomous learning of degradation features to construct health indicators. The DTW distance is used to measure the health indicators of the bearing to be predicted and the health indicators of the prior bearing, and the degradation failure types of rolling bearings are divided into two types: sudden failure and gradual failure.

[0014] 3) Select the full-life bearing signal data set collected under one working condition as the source domain data set, and filter out the full-life vibration signal data with the same failure type as the source domain collected under other working conditions through step 2).

[0015] 4) Input the source domain and target domain data sets into the ITSAE network, automatically learn the mapping relationship between features and lifespan, obtain the MSE between the real lifespan and the predicted lifespan in the source domain, and the MMD between the source domain features and the target domain features. Through the back propagation algorithm, the network can gradually learn and update the parameters, and finally converge to the optimal solution to obtain a well-trained network model.

[0016] 5) Input the bearing test data collected under the same working conditions as the target domain into the trained improved ISAE model to complete the remaining life prediction of the bearing.

[0017] The convolutional autoencoder is an autoencoder model based on the convolutional neural network structure. Compared with the traditional fully connected autoencoder, the convolutional autoencoder has better robustness and feature extraction capabilities. It uses convolutional layers and pooling layers to extract the features of the input data, and learns the effective representation of the data through the structure of the encoder and decoder, realizing the feature extraction of unsupervised learning based on the traditional autoencoder. Its constituent units are one-dimensional convolution and one-dimensional deconvolution neural networks.

[0018] The encoding structure is to convert the input data X = [x1, x2, ..., x n ] T After encoding the convolutional layer, a feature map is generated and k convolution kernels are initialized. Each convolution kernel is composed of parameters h and b. k Composition, generate k feature maps h, the process can be described by formula (1).

[0019] h k =σ(x×w k +b k ) (1)

[0020] Where w is the weight from the input layer to the convolutional layer, b is the bias of the convolutional layer, and σ is the ReLU activation function.

[0021] The feature maps are concatenated in a cascaded manner to form an intermediate layer. The decoding process is to reconstruct this intermediate layer through a one-dimensional deconvolution layer to finally generate a reconstructed output. This process can be described by formula (2).

[0022]

[0023] In the formula, is the weight from the intermediate layer to the deconvolution layer, is the bias of the deconvolution layer.

[0024] The deep residual shrinkage network is essentially a new improvement on the residual network. The network structure adds soft thresholding as a nonlinear layer to ResNet, aiming to improve the feature learning effect of deep learning methods on noisy or complex data. The threshold required for soft thresholding is essentially set according to the attention mechanism.

[0025] The residual module is the basic component of the deep residual network. The cuboid represents a feature map with C channels, W width and 1 height. A residual module contains two batch normalizations, two rectified linear unit activation functions, two convolutional layers and an identity map. The identity map is the core contribution of the deep residual network, which greatly reduces the difficulty of deep neural network training.

[0026] Assume that the input of a certain network is x and the expected output is H(x). In the residual network structure, the input x is directly transmitted to the output as the initial result through a shortcut connection. What the network learns is the difference between the input and output, that is, the residual H(x) = x + F(x). The purpose of training is to make the residual result close to 0. Compared with training an equivalent mapping, the difficulty of training is greatly reduced.

[0027] Soft thresholding is the core step of many noise reduction methods. Its main function is to set the part of the attribute whose absolute value is less than a predetermined threshold to zero, so that other attributes are also adjusted to zero, which is called shrinkage. The choice of threshold directly affects the noise reduction effect. The operation relationship of soft thresholding is shown in formula (3):

[0028]

[0029] Soft thresholding is a nonlinear transformation where x represents an input feature, y represents an output feature, and τ represents a threshold. Similar to the ReLU activation function, the gradient of soft thresholding can be either 0 or 1, so it can be considered an activation function. In the current shrinking network, soft thresholding is combined with the popular attention mechanism and automatically learned. This means that the process of determining the [-τ,τ] range can be automatically adjusted according to the sample itself, thus avoiding the problem of manually setting the threshold and improving accuracy.

[0030] In essence, the residual shrinkage module is an integration of residual blocks, attention mechanisms, and soft threshold functions. Among them, the attention mechanism can focus on specific local information to promote faster convergence of the model. Specifically, the attention mechanism includes two main stages: first, discovering local useful information through overall scanning; second, strengthening useful information and limiting the propagation of redundant information.

[0031] The basic modules of the one-dimensional deep residual condensation network are similar to the attention mechanism modules, in which C, W, and H represent the number of channels (batch), width, and height of the input data, respectively. This module consists of a small subnetwork that learns a set of thresholds and then performs soft thresholding on them. By adding a constant path, the complexity of modeling is reduced, which can balance the time cost of the model and improve the accuracy of the model.

[0032] The SAE network is an architectural design that combines spatial, channel, and global representations. In the SAE module, a structure similar to the residual shrinkage network is used. In this network structure, the number of input channels is first reduced to 1 through a compression operation, thereby reducing the dimension of the data. Then, the reduced input is passed to the aggregation layer to learn more representations. Finally, the results are concatenated and restored to the original shape through an excitation operation to maintain the information richness and representation ability of the network. This network structure that combines the three modules of compression, aggregation, and excitation can have better performance when learning complex feature representations.

[0033] The core part of the SAE module is evolved from the compression and excitation module, which consists of a combination of compression and excitation methods. The compression module contains the input to be compressed using a fully connected layer, and the number of input channels is reduced to 1 through the compression operation, thereby reducing the dimension of the data. The reduced input is then passed to the aggregation layer to learn more representations. Finally, the results are connected and restored to the original shape through the excitation operation to maintain the information richness and representation ability of the network. The final output is resized to its original shape.

[0034] Unlike the SE module, in the SAE module, the compression and excitation module stage is better utilized by increasing the cardinality between layers. The SAE module increases the cardinality of the first fully connected layer after the compression module, and since the fully connected layer can learn from the global representation of the network, the fully connected layer is used as an aggregation module, and the aggregated FC layer results are connected in the excitation stage and their output shape is restored to pass to the next layer.

[0035] At the same time, the SAE module is combined with a deep residual module. The residual module contains a shortcut connection, which is attached after a specific layer to transfer the learning information of the network. This helps to build a deeper network by avoiding the gradient vanishing problem. The purpose of selecting the residual module is that it transfers information every time, and if the residual module learns better, the network's learning ability will be better. The compression operation after the global average pooling layer is used to obtain channel statistics, which are then reduced and input into the compression operation, followed by the excitation layer. The residual module contains repeated convolutional layers that form a module after a specific layer. This type of module is repeated regularly to transfer the learned gradients and prevent them from disappearing when going deeper. The process of compressing the aggregation excitation module can be described by formula (4):

[0036] SAEnet=x+F(x·Ex(∑sq(x))) (4)

[0037] Here, the Sq(·) function represents the compression operation on the input x, and the Ex(·) function represents the excitation operation to reshape the channel-modified input to its original shape without reduction. This is followed by a scaling operation that restores it to its original form together with the input.

[0038] In view of the problem that different failure degradation types of bearings will affect the RUL prediction accuracy, the present invention proposes a remaining life prediction method based on a migration-based compressed aggregation excitation life prediction network improved by degradation type classification. The method includes a failure type classification module and an improved migration-based compressed aggregation excitation prediction module, wherein the failure type classification module is implemented based on convolutional autoencoder and DTW distance, and the convolutional autoencoder can extract the global features of the entire life cycle of the bearing, which can well characterize the degradation process of the bearing. However, the degradation process and life cycle of the bearing are different, and the DTW distance has the advantage of being able to measure the similarity of different time series lengths. Therefore, the advantages of both are combined to finally realize the failure type classification of the bearing. There is a certain similarity in the bearing monitoring signals under the same failure behavior. Therefore, based on the improved compressed aggregation excitation network, high similarity and generalization features under the same failure behavior can be extracted, but the problem of inconsistent data distribution still exists, so the MMD distance metric is introduced to realize cross-domain life migration prediction.

[0039] The present invention uses a data-driven prediction method to predict the remaining service life of bearings. Since bearings are susceptible to failures due to a variety of failure modes, the monitoring data obtained from different units may present different failure types. If the failure type is not correctly identified and processed, on the one hand, it may become a source of noise in the data, and on the other hand, the data distribution is highly diverse, thereby reducing the accuracy of RUL prediction. Therefore, in order to improve the accuracy of life prediction, it is necessary to classify the failure types in the bearing life cycle.

[0040] As introduced in the basic theory, the deep convolutional autoencoder automatically extracts data feature expressions through a learning method that minimizes reconstruction errors, without any experience, which helps to build universal and generalizable health indicators. The DTW distance can handle the similarity comparison problem of two data of different time series lengths and is suitable for pattern recognition problems. Therefore, after the present invention uses DCAE to construct a priori health indicators, based on the distance between the online bearing health indicator and the prior health indicator measured by DTW, the bearing failure types are divided into sudden failure behaviors and gradual failure behaviors. The divided failure types can lay a foundation for further improving the accuracy of life prediction.

[0041] The compressed aggregation excitation network can automatically learn the relationship between features and mappings. The deep residual shrinkage module is suitable for situations where the noise content of each sample is different, and can effectively alleviate the problem of gradient disappearance. Therefore, an improved compressed aggregation excitation network model is established to learn generalized lifespan features.

[0042] The network structure of the ITSAE model is mainly composed of input layer, DRSM, convolution layer, batch normalization layer, activation layer, concatenation layer, fully connected layer and output layer. In the ITSAE model, the deep residual shrinkage module is cross-introduced into each feature extraction layer of the compression aggregation excitation network to achieve noise reduction, which helps to improve the robustness of the network. The prior bearing vibration signal is used as the input of ITSAE to learn the generalized life characteristics. The goal that ISAE needs to optimize is the loss between the actual life and the predicted life. The mean absolute error is used to measure the prediction loss. This is a loss function commonly used in solving regression problems. The loss is denoted by l. RUL .

[0043] When the target to be predicted and the prior bearing have the same failure behavior, the bearing degradation characteristics have certain similarities, but the problem of inconsistent data distribution still exists under different working conditions. Therefore, the MMD distance metric is introduced into the ISAE model to construct the ITSAE model to achieve the migration prediction of the remaining life by minimizing the distribution difference between the source domain and the target domain. The basic idea of ​​MMD is to measure the difference between the two distributions by comparing their means in the feature space, as shown in formula (5), which measures the source domain h S and the target domain hT The distance between.

[0044]

[0045] in, It is a mapping used to map the original variable to the reproducing kernel Hilbert space. H indicates that this distance is determined by Mapping data into the regenerating Hilbert space for measurement, and Represent the source domain h S and the target domain h T The data sample, N S and N T The source domain h S and the target domain h T By minimizing the difference in the number of samples, knowledge transfer and sharing can be achieved, thereby improving the generalization ability and performance of the model.

[0046] In order to train a model that can achieve life prediction under different working conditions, the weighted sum of the source domain remaining life fitting loss function MSE and the working condition generalization MMD loss is taken as the optimization target of ITSAE. The final loss function can be expressed as formula (6):

[0047] loss = λl RUL +μl MMD (6)

[0048] In the formula, λ and μ are compromise parameters that are trained through back propagation to achieve an optimal model parameter and ultimately achieve accurate prediction of remaining life.

[0049] The present invention proposes a life migration prediction model of an improved compression aggregation excitation network based on degradation type classification, which includes two sub-parts: degradation failure type discrimination and life migration prediction. First, the DCAE network is used to construct the prior gradual health index and the general health index of the rolling bearing, and then the failure degradation type is discriminated by measuring the DTW distance between the two. Then, the compression aggregation excitation module is used to automatically learn the mapping relationship between features and life, and combined with MMD to generalize the working conditions, so as to accurately predict the RUL of rolling bearings under different working conditions. Finally, the PHM2012 dataset verifies that the proposed method has better effectiveness and robustness than other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the convolutional autoencoder structure.

[0051] Figure 2 Schematic diagram of the residual module.

[0052] Figure 3 Schematic diagram of the basic module of the residual shrinkage network.

[0053] Figure 4 Schematic diagram of the internal working principle of the SAE module.

[0054] Figure 5 Schematic diagram of the network structure of the life prediction model.

[0055] Figure 6 Figure 1 is a schematic diagram of the ITSAE network structure.

[0056] Figure 7 It is the overall flow chart of the present invention.

[0057] Figure 8 It is the PRONOSTIA experimental platform.

[0058] Fig. 9 Health indicator diagram for 16 bearings.

[0059] Fig.10 This is the life prediction curve obtained with bearing1_2 as the source domain, bearing3_3 as the target domain training set, and bearing3_1, bearing3_2 and bearing3_3 as the target domain test set.

[0060] Fig.11 The life prediction curve is obtained by taking bearing2_1 as the source domain, bearing3_1 as the target domain training set, and bearing3_1, bearing3_2 and bearing3_3 as the target domain test set.

[0061] Fig.12 The life prediction curve is obtained with bearing3_1 as the source domain, bearing1_2 as the target domain training set, and bearing1_2, bearing1_4, bearing1_5 and bearing1_7 as the target domain test set.

[0062] Fig.13 This is the life prediction curve under gradual failure behavior. Fig.13 (a) is the life prediction curve obtained by using bearing1_1 as the target domain test set. Fig.13 (b) is the life prediction curve obtained by using bearing1_3 as the target domain test set.

[0063] Fig.14 Indicator graphs for different migration tasks. Fig.14 (a) is the RMSE indicator diagram, Fig.14 (b) is the indicator diagram of MAE. DETAILED DESCRIPTION

[0064] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0065] like Figure 7 As shown, the life migration prediction method based on the improved compressed aggregation excitation network has the following specific steps:

[0066] Step 1: Install the acceleration sensor on the experimental equipment under different working conditions, and obtain the vibration signal data by setting the appropriate sampling rate, sampling interval, duration of each sampling and shutdown conditions, so as to obtain the full life data of the bearing.

[0067] Step 2: After the data collected under different working conditions are preprocessed with the Hilbert envelope spectrum, the envelope spectrum data sequence is input into DACE for autonomous learning of degradation features to construct health indicators. The DTW distance is used to measure the health indicators of the bearing to be predicted and the health indicators of the prior bearing, and the degradation failure types of the rolling bearing are divided into two types: sudden failure and gradual failure.

[0068] Step 3: Select the full-life bearing signal data set collected under one working condition as the source domain data set, and filter out the full-life vibration signal data collected under other working conditions with the same failure type as the source domain through step 2.

[0069] Step 4: Input the source domain and target domain data sets into the ITSAE network, automatically learn the mapping relationship between features and lifespan, obtain the MSE between the real lifespan and the predicted lifespan in the source domain, and the MMD between the source domain features and the target domain features. Through the back propagation algorithm, the network can gradually learn and update the parameters, and finally converge to the optimal solution to obtain a fully trained network model.

[0070] Step 5: Input the bearing test data collected under the same working conditions as the target domain into the trained improved ISAE model to complete the remaining life prediction of the bearing.

[0071] Convolutional autoencoder is an autoencoder model based on convolutional neural network structure. Compared with traditional fully connected autoencoders, convolutional autoencoders have better robustness and feature extraction capabilities. It uses convolutional layers and pooling layers to extract the features of input data, and learns the effective representation of data through the structure of encoder and decoder, realizing the feature extraction of unsupervised learning based on traditional autoencoders. Its constituent units are one-dimensional convolution and one-dimensional deconvolution neural networks. The model structure is as follows Figure 1 shown.

[0072] The encoding structure is to convert the input data X = [x1, x2, ..., x n ] T After encoding the convolutional layer, a feature map is generated and k convolution kernels are initialized. Each convolution kernel is composed of parameters h and b. k Composition, generate k feature maps h, the process can be described by formula (1).

[0073] h k =σ(x×w k +b k ) (1)

[0074] Where: w is the weight from the input layer to the convolutional layer; b is the bias of the convolutional layer; σ is the ReLU activation function.

[0075] The feature maps are concatenated in a cascade manner to form an intermediate layer. The decoding process is to reconstruct this intermediate layer through a one-dimensional deconvolution layer to finally generate a reconstructed output. This process can be described by formula (2).

[0076]

[0077] Where: is the weight from the intermediate layer to the deconvolution layer; is the bias of the deconvolution layer.

[0078] The Deep Residual Shrinkage Network (DRSN) is essentially a new improvement on the Residual Network (ResNet). This network structure adds soft thresholding as a nonlinear layer to ResNet, aiming to improve the feature learning effect of deep learning methods on noisy or complex data. The threshold required for soft thresholding is essentially set according to the attention mechanism.

[0079] The residual building unit (RBU) is a basic component of the deep residual network. Figure 2As shown in the figure, the cuboid represents a feature map with C channels, W width and 1 height; a residual module contains two batch normalizations (BN), two rectifier linear unit activation functions (ReLU), two convolutional layers and an identity shortcut. The identity shortcut is the core contribution of the deep residual network, which greatly reduces the difficulty of training deep neural networks. In the figure, K represents the number of convolution kernels in the convolution layer, which matches the number of channels C of the input feature map, and the size of the output feature map is the same as the size of the input feature map.

[0080] Suppose the input of a certain network is x and the expected output is H(x). In the residual network structure, the input x is directly transmitted to the output as the initial result through a shortcut connection. What the network learns is the difference between the input and output, that is, the residual H(x) = x + F(x). The purpose of training is to make the residual result close to 0. Compared with training an equivalent mapping, the difficulty of training is greatly reduced.

[0081] Soft thresholding is the core step of many noise reduction methods. Its main function is to set the part of the attribute whose absolute value is less than a predetermined threshold to zero, so that other attributes are also adjusted to zero. This is called shrinkage. The choice of threshold directly affects the noise reduction effect. The operation relationship of soft thresholding is shown in formula (3):

[0082]

[0083] Soft thresholding is a nonlinear transformation where x represents an input feature, y represents an output feature, and τ represents a threshold. Similar to the ReLU activation function, the gradient of soft thresholding can be either 0 or 1, so it can be considered an activation function. In the current shrinking network, soft thresholding is combined with the popular attention mechanism and automatically learned, which means that the process of determining the [-τ,τ] range can be automatically adjusted according to the sample itself, thus avoiding the problem of manually setting the threshold and improving accuracy.

[0084] In essence, the residual shrinkage module is an integration of residual blocks, attention mechanisms, and soft threshold functions. The attention mechanism can focus on specific local information to promote faster model convergence. Specifically, the attention mechanism includes two main stages: first, discovering local useful information through overall scanning; second, strengthening useful information and limiting the spread of redundant information.

[0085] The basic module of the one-dimensional deep residual condensation network is similar to the attention mechanism module. In these modules, C, W, and H represent the number of channels (batch), width, and height of the input data, respectively. This module consists of a small sub-network that learns a set of thresholds and then performs soft threshold processing on them. By adding a constant path, the complexity of modeling is reduced, which can not only take into account the time cost of the model, but also improve the accuracy of the model. The structure of the basic module is as follows Figure 3 shown.

[0086] The SAE-Net is an architecture design that combines spatial, channel, and global representations. In the SAE module, a structure similar to the residual shrinkage network is adopted, such as Figure 4 As shown in the figure. In this network structure, the number of input channels is first reduced to 1 through a compression operation, thereby reducing the dimension of the data. Then, the reduced input is passed to the aggregation layer to learn more representations. Finally, the results are concatenated and restored to the original shape through an excitation operation to maintain the information richness and representation ability of the network. This network structure that combines the three modules of compression, aggregation, and excitation can have better performance in learning complex feature representations.

[0087] The core part of the SAE module is evolved from the compression and excitation module, which consists of a combination of compression and excitation methods. The compression module contains the input to be compressed using the fully connected (FC) layer, and the number of channels of the input is reduced to 1 through the compression operation, thereby reducing the dimension of the data. The reduced input is then passed to the aggregation layer to learn more representations. Finally, the results are concatenated and restored to the original shape through the excitation operation to maintain the information richness and representation ability of the network. The final output is resized to its original shape.

[0088] Unlike the SE (Squeeze-and-Excitation) module, in the SAE module, the compression and excitation module stage is better utilized by increasing the cardinality between layers. The SAE module increases the cardinality of the first fully connected layer (FC) after the compression module. At the same time, since the fully connected layer can learn from the global representation of the network, the fully connected layer is used as an aggregation module. The aggregated FC layer results are connected in the excitation stage and their output shape is restored to pass to the next layer.

[0089] At the same time, the SAE module is combined with a deep residual module. The residual module contains a shortcut connection, which is attached after a specific layer to transfer the learning information of the network, which helps to build a deeper network by avoiding the gradient vanishing problem. The purpose of selecting the residual module is that it transfers information every time, and if the residual module learns better, the network's learning ability will be better. The compression operation after the global average pooling layer is used to obtain channel statistics, which are then reduced and input into the compression operation. The residual module contains repeated convolutional layers that form a module after a specific layer. This type of module is repeated regularly to transfer the learned gradients and prevent them from disappearing when going deeper. The process of compressing the aggregation excitation module can be described by formula (4):

[0090] SAEnet=x+F(x·Ex(∑sq(x))) (4)

[0091] Here, the Sq(·) function represents the compression operation on the input x, and the Ex(·) function represents the excitation operation to reshape the channel-modified input to its original shape without reduction; this is followed by a scaling operation that restores it to its original form together with the input.

[0092] Aiming at the problem that different failure degradation types of bearings will affect the RUL prediction accuracy, the present invention proposes a remaining life prediction method based on a migration-compression-aggregation-excitation life prediction network improved by degradation type classification. The network structure diagram is shown in Figure 5 As shown. The method includes a failure type classification module and an improved transferable compressed aggregation excitation prediction module, wherein the failure type classification module is implemented based on convolutional autoencoder and DTW distance. The convolutional autoencoder can extract the global features of the entire life cycle of the bearing, which can well characterize the degradation process of the bearing. The degradation process and life cycle of the bearing are different. The DTW distance has the advantage of being able to measure the similarity of different time series lengths. Therefore, the advantages of both are combined to finally realize the failure type classification of the bearing. There is a certain similarity in the bearing monitoring signals under the same failure behavior. Therefore, based on the improved compressed aggregation excitation network, high similarity and generalization features under the same failure behavior can be extracted, but the problem of inconsistent data distribution still exists. Therefore, the MMD distance metric is introduced to realize cross-domain life migration prediction.

[0093] The present invention uses a data-driven prediction method to predict the remaining service life of bearings. Since bearings are susceptible to failures due to a variety of failure modes, the monitoring data obtained from different units may present different failure types. If the failure type is not correctly identified and processed, on the one hand, it may become a source of noise in the data, and on the other hand, the data distribution is highly diverse, thereby reducing the accuracy of RUL prediction. Therefore, in order to improve the accuracy of life prediction, it is necessary to classify the failure types in the bearing life cycle.

[0094] The deep convolutional autoencoder (DCAE) automatically extracts data feature expressions through a learning method that minimizes reconstruction errors, without any experience, which helps to construct universal and generalizable health indicators. The DTW distance can handle the similarity comparison problem of two data of different time series lengths and is suitable for pattern recognition problems. Therefore, after the present invention uses DCAE to construct a priori health indicators, it divides the bearing failure types into sudden failure behaviors and gradual failure behaviors based on the distance between the online bearing health indicators and the priori health indicators measured by DTW. The divided failure types can lay a foundation for further improving the accuracy of life prediction.

[0095] The compressed aggregate excitation network (SAE) can automatically learn the relationship between features and mappings. The deep residual shrinkage module (DRSM) is suitable for the situation where the noise content of each sample is different, and can effectively alleviate the problem of gradient disappearance. Therefore, the present invention establishes an improved compressed aggregate excitation network (ISAE) model to learn generalized life characteristics, such as Figure 6 shown.

[0096] The network structure of the ITSAE model is mainly composed of input layer, DRSM, convolution layer, batch normalization layer, activation layer, concatenation layer, fully connected layer and output layer. In the ITSAE model, the deep residual shrinkage module (DRSM) is cross-introduced into each feature extraction layer of the compressed aggregation excitation network (SAE) network to achieve noise reduction, which helps to improve the robustness of the network. The prior bearing vibration signal is used as the input of ITSAE to learn the generalized life characteristics. The goal that ISAE needs to optimize is the loss between the actual life and the predicted life. The mean absolute error (MSE) is used to measure the prediction loss. This is a loss function commonly used in solving regression problems. The loss is denoted by l RUL .

[0097] When the target to be predicted and the prior bearing have the same failure behavior, the bearing degradation characteristics have certain similarities, but the problem of inconsistent data distribution still exists under different working conditions. Therefore, the MMD distance metric is introduced into the ISAE model to construct the ITSAE model to achieve the migration prediction of the remaining life by minimizing the distribution difference between the source domain and the target domain. The basic idea of ​​MMD is to measure the difference between the two distributions by comparing their means in the feature space. As shown in formula (5), the source domain h is measured. S and the target domain h T Distance between:

[0098]

[0099] in, It is a mapping used to map the original variable to the reproducing kernel Hilbert space. H indicates that this distance is determined by Mapping the data into the Reproducing Hilbert Space (RKHS) for measurement, and Represent the source domain h S and the target domain h T The data sample, N S and N T The source domain h S and the target domain h T The number of samples is minimized; knowledge transfer and sharing are achieved by minimizing this difference, thereby improving the generalization ability and performance of the model.

[0100] In order to train a model that can achieve life prediction under different working conditions, the weighted sum of the source domain remaining life fitting loss function MSE and the working condition generalization MMD loss is taken as the optimization target of ITSAE. The final loss function can be expressed as formula (6):

[0101] loss = λl RUL +μl MMD (6)

[0102] In the formula, λ and μ are compromise parameters that are trained through back propagation to achieve an optimal model parameter and ultimately achieve accurate prediction of remaining life.

[0103] The present invention proposes a life migration prediction model of an improved compression aggregation excitation network based on degradation type classification, which includes two sub-parts: degradation failure type discrimination and life migration prediction. First, the DCAE network is used to construct the prior gradual health index and the general health index of the rolling bearing, and then the failure degradation type is discriminated by measuring the DTW distance between the two. Then, the compression aggregation excitation module is used to automatically learn the mapping relationship between features and life, and combined with MMD to generalize the working conditions, so as to accurately predict the RUL of rolling bearings under different working conditions. Finally, the PHM2012 dataset verifies that the proposed method has better effectiveness and robustness than other methods.

[0104] In order to further demonstrate the effectiveness of the proposed RUL prediction method, multiple remaining life migration prediction experiments were carried out on the full life cycle signals of the bearing. The superiority of the proposed method was verified through comparative analysis and ablation experiments with different methods.

[0105] The PHM Challenge dataset was provided by the FEMTO-ST Institute and was conducted on a laboratory experimental platform (PRONOSTIA) that is capable of accelerating bearing degradation under constant and / or variable operating conditions while collecting online health monitoring data (speed, load force, temperature, vibration). Figure 8PRONOSTIA is a platform dedicated to testing and validating bearing fault detection, diagnosis and prediction methods. Designed and implemented at the AS2M department of the FEMTO-ST Institute, PRONOSTIA's main purpose is to provide real experimental data describing the degradation of ball bearings throughout their service life, until complete failure.

[0106] The bearing degradation data mainly consists of two parts, namely vibration data and temperature data. The vibration sensor consists of two micro accelerometers positioned at 90° to each other, such as Fig. 9 As shown, the first is placed on the vertical axis and the second on the horizontal axis. Two accelerometers are placed radially on the outer ring of the bearing with a sampling frequency of 25.6kHz, and the temperature sensor is a resistance temperature detector placed in a hole close to the outer bearing ring with a sampling frequency of 0.1Hz.

[0107] The data set is divided into three different working conditions, and each working condition has different training sets and test sets. The load of working condition 1 is 4000N and the speed is 1800rpm; the load of working condition 2 is 4200N and the speed is 1650rpm; the load of working condition 3 is 5000N and the speed is 1500rpm. The specific characteristics of the training set and test set are shown in Table 1.

[0108] Table 1 Characteristics of PHM2012 training set and test set

[0109]

[0110] Model establishment and parameter setting: The determination of failure behavior is based on DCAE implementation, and the cross-domain life prediction is based on ITSAE. The detailed parameters of the constructed network model are shown in Tables 2 and 3. For the DCAE framework, the encoder contains 3 convolutional layers and 3 pooling layers. The first convolutional layer uses a wide convolution kernel with a step size of 4 and a size of 1×64. The second and third convolutional layers use small convolution kernels of 1×4. The pooling layer size and step size are both 4. The number of kernels in the first to third convolutional layers is 16, 32, and 64, respectively. The decoder layer includes 3 upsampling layers and 3 deconvolution layers. The parameters of each layer correspond to the encoding layer one by one. The hidden layer node is 1, which is the degradation feature extracted by the present invention. This feature is used as the basis for failure behavior determination. For the ITSAE framework, life feature extraction is based on alternately connected DRSN and convolutional layers. While enhancing the anti-noise performance of the network, more detailed life features can be extracted. Both models use the rectified linear unit (ReLU), the optimization algorithm is the adaptive moments (Adam) algorithm, and the learning rate is 0.0001. The input of the present invention is the envelope spectrum data sample, firstly the sample data is imported into the DCAE model for autonomous learning of degradation characteristics to realize failure behavior judgment, and then the sample data is imported into the ITSAE model for autonomous learning of life characteristics to realize the remaining life migration prediction.

[0111] The computer configuration used in the experiment is as follows: CPU: Core TM i7-12700H CPU@2.3GHz, 16GB memory, all experiments were conducted on the TensorFlow platform, and the average prediction accuracy was calculated after 10 independent experiments. In order to ensure fair comparison, all methods were implemented under the same network structure.

[0112] Table 2 Detailed experimental parameters of the proposed DCAE

[0113] Network Layer Activation Function Step Length size Number Output size Input Layer - - - - 1×1280 Convolution 1 + batch normalization layer Leaky relu 1×4 1×64 16 16×320 Pooling Layer - 1×4 1×4 16 16×80 Convolution 2 + batch normalization layer Leaky relu 1×1 1×4 32 32×80 Pooling Layer - 1×4 1×4 32 32×20 Convolution 3 + batch normalization layer Leaky relu 1×1 1×4 64 64×20 Pooling Layer - 1×4 1×4 64 64×5 Max Average Pooling Layer - - - - 1×64 Output Layer - - - - 1×1 Fully connected layer - - - - 1×320 Conversion Layer - - - - 64×5 Upsampling layer + batch normalization layer - 1×4 1×1 64 64×20 Deconvolution 4 Leaky relu 1×1 1×4 32 32×20 Upsampling layer + batch normalization layer - 1×4 1×1 32 32×80 Deconvolution 5 Leaky relu 1×1 1×4 16 16×80 Upsampling layer + batch normalization layer - 1×4 1×1 16 16×320 Reconstruction layer Leaky relu 1×1 1×4 1 1×1280

[0114] Table 3 Detailed experimental parameters of the proposed ITSAE

[0115]

[0116]

[0117] Experimental results and analysis: 1. Construct health indicators: First, use the bearing1_1 training set in the working condition 1 in the PHM2012 dataset table to train the DCAE network, and further construct the health indicator of bearing1_1. The health indicator graph is shown in the figure. Then, the other 16 training sets and test sets in the PHM2012 dataset are input into the DCAE network, and the health indicators are constructed respectively. The health indicator graphs are obtained in turn, as shown in the figure. Fig. 9 As shown, it can be seen that the trends of health indicators of each test set and training set, the trends of health indicators of bearing1_6, bearing2_4 and bearing2_5 are obviously different from others, so bearing1_6, bearing2_4 and bearing2_5 are not involved in the experiment.

[0118] The DTW distances between other bearings and bearing1_1 are calculated, as shown in Tables 4 and 5. It can be seen from the table that the DTW distance between bearing1_3 and bearing1_1 is the smallest, which is 0.008, while the DTW distances between other data sets and bearing1_1 are all above 0.01, among which bearing2_7 has the largest value of 0.210. It is determined that bearing1_1 and bearing1_3 are of the same degradation failure type, namely, gradual failure bearings, and other bearings are of the same degradation type, namely, sudden failure bearings.

[0119] Table 4 DTW distances between other bearings and bearing1_1

[0120]

[0121] Table 5 Division of source domain and target domain of sudden failure bearing

[0122]

[0123]

[0124]

[0125] Except for bearing1_1 and bearing1_3, the others are sudden failure bearings. Therefore, when sudden failure bearings are selected for experiments, bearing1_2, bearing2_1 and bearing3_1 are used as source domains in turn, and the remaining bearings are used as target domains for cross-validation. The divided target domain test set and training set are shown in Table 4. When predicting the life of gradually failed bearings, bearing1_1 and bearing1_3 are selected for two-fold cross-validation.

[0126] The life prediction curve obtained by taking bearing1_2 as the source domain, bearing3_3 as the target domain training set, and bearing3_1, bearing3_2 and bearing3_3 as the target domain test set is as follows: Fig.10 shown.

[0127] Fig.10 The life prediction curve is obtained by taking bearing1_2 as the source domain, bearing1_2 as the target domain training set, and bearing1_2, bearing1_4, bearing1_5, and bearing1_7 as the target domain test sets.

[0128] The life prediction curve obtained by taking bearing2_1 as the source domain, bearing3_1 as the target domain training set, and bearing3_1, bearing3_2 and bearing3_3 as the target domain test set is as follows: Fig.11 shown.

[0129] Fig.11 The life prediction curve is obtained by taking bearing2_1 as the source domain, bearing3_1 as the target domain training set, and bearing3_1, bearing3_2 and bearing3_3 as the target domain test sets.

[0130] The life prediction curve obtained by taking bearing3_1 as the source domain, bearing1_2 as the target domain training set, and bearing1_2, bearing1_4bearing1_5 and bearing1_7 as the target domain test set is as follows: Fig.12 shown.

[0131] Fig.12 The life prediction curve is obtained by taking bearing3_1 as the source domain, bearing1_2 as the target domain training set, and bearing1_2, bearing1_4, bearing1_5, and bearing1_7 as the target domain test sets.

[0132] The life prediction curve obtained by taking bearing1_3 as the source domain, bearing1_1 as the target domain training set, and bearing1_1 as the target domain test set is as follows: Fig.13 (a) shows that MAE = 0.045, RMSE = 0.063. The life prediction curve obtained by taking bearing1_1 as the source domain, bearing1_3 as the target domain training set, and bearing1_3 as the target domain test set is shown in Fig.13 As shown in (b), MAE = 0.050, RMSE = 0.065.

[0133] In this part, in order to verify the feasibility and superiority of the proposed RUL prediction method, bearing1_2, bearing2_1 and bearing3_1 in the PHM2012 dataset are used as source domains, and the remaining bearing crosses are used as target domains. The proposed method is compared with other machine learning algorithms to verify the superiority of the method. In addition, an ablation study is performed on the constructed network architecture to demonstrate its advantages.

[0134] As shown in Table 6, in the ablation method study, the proposed method is simply referred to as “ISAE” without the maximum mean difference (MMD). In the comparison method study, the proposed method is compared with the transfer convolutional neural network (TCNN) proposed by Chen, whose network parameters are carefully readjusted according to the above references. The experimental results are shown in Table 7.

[0135] Table 6 Ablation experiment and comparative experiment

[0136]

[0137]

[0138] Table 7 Ablation experiment and comparative experiment results

[0139]

[0140]

[0141]

[0142] Among the regression model evaluation indicators, the mean absolute error (MAE) and the root mean square error (RMSE) are used to measure the degree of error between the model prediction value and the true value. From the table, it can be seen that when bearing3_1 is the source domain, bearing2_2 is the target domain training set, and bearing2_3 and bearing2_6 are the test sets, compared with the other two methods, the proposed method has smaller MAE and RMSE values, which means that the proposed method can more accurately predict the RUL results and enhance the robustness of the method.

[0143] In order to reduce the impact of random errors and improve the reliability of the evaluation results, the MAE value and RMSE value obtained by each method in each group of experiments were summed and averaged, as shown in Table 6 and Fig.14As shown in the figure, in all experiments with bearing1_2, bearing2_1, and bearing3_1 as source domains, the average values ​​of MAE and RMSE of the proposed method are significantly lower than those of other methods. In particular, when bearing3_1 is used as the source domain, the RMSE value is reduced by 0.134 and the MAE value is reduced by 0.148 compared with Chen's method. This is because the proposed method performs migration after determining the failure behavior, which well extracts the common features of the source domain and the target domain, and effectively verifies the accuracy of the proposed method.

[0144] The present invention proposes a life migration prediction model of an improved compression aggregation excitation network based on degradation type classification, which includes two sub-parts: degradation failure type discrimination and life migration prediction. First, the DCAE network is used to construct the prior gradual health index and the general health index of the rolling bearing, and then the failure degradation type is discriminated by measuring the DTW distance between the two. Then, the compression aggregation excitation module is used to automatically learn the mapping relationship between features and life, and combined with MMD to generalize the working conditions, so as to accurately predict the RUL of rolling bearings under different working conditions. Finally, the PHM2012 dataset verifies that the proposed method has better effectiveness and robustness than other methods.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A life migration prediction method based on an improved compressed aggregation excitation network, characterized in that: The specific steps are as follows: Step 1: Use deep convolutional autoencoders to construct health indicators for rolling bearings, and measure the DTW distance between health indicators to classify the degradation types of rolling bearings, and classify the degradation types of bearings into gradual and sudden degradation types. Step 2: Build an improved compression aggregation excitation network combined with the maximum mean difference, propose a life prediction strategy under different degradation types, and complete the prediction of rolling bearing life under multiple working conditions; Step 3: A life migration prediction model based on degradation type classification and improved compressed aggregation excitation network is proposed. This model divides the degradation types of bearings under different working conditions and realizes remaining life prediction.

2. The method for predicting life migration based on improved compressed aggregation excitation network according to claim 1 is characterized in that: In step 1, the specific structure of the deep convolutional autoencoder is: The input data X = [x1, x2, ..., x n ] T After encoding the convolutional layer, a feature map is generated and k convolution kernels are initialized. Each convolution kernel is composed of parameters h and b. k Composition, generate k feature maps h, the process is shown in formula (1); h k =σ(x×w k +b k ) (1) Where w is the weight from the input layer to the convolutional layer, b is the bias of the convolutional layer, and σ is the ReLU activation function; The feature maps are concatenated to form an intermediate layer. The decoding process reconstructs the intermediate layer through a one-dimensional deconvolution layer to generate the reconstructed output. The process is shown in formula (2). In the formula, is the weight from the intermediate layer to the deconvolution layer, is the bias of the deconvolution layer.

3. The method for predicting life migration based on improved compressed aggregation excitation network according to claim 2 is characterized in that: In step 2, a deep residual shrinkage network is constructed, in which soft thresholding is added as a nonlinear layer to the residual network, and the threshold required for soft thresholding is set according to the attention mechanism; Assume that the input of a certain network is x and the expected output is H(x). In the residual network structure, the input x is directly transmitted to the output as the initial result through a shortcut connection. The network learns the difference between the input and output, that is, the residual H(x) = x + F(x). The operation relationship of soft thresholding is shown in formula (3): Soft thresholding is a nonlinear transformation, where x represents an input feature, y represents an output feature, and τ represents a threshold.

4. The method for predicting life migration based on improved compressed aggregation excitation network according to claim 3 is characterized in that: In step 2, in the compression aggregation excitation network, the number of input channels is reduced to 1 through compression operation; The compression aggregation excitation module uses a fully connected layer as an aggregation module. The aggregated FC layer results are connected in the excitation stage and their output shape is restored to be passed to the next layer; The process of compressing the aggregation incentive module is shown in formula (4): SAEnet=x+F(x·Ex(∑sq(x))) (4) Among them, the Sq(·) function represents the compression operation on the input x, and the Ex(·) function represents the excitation operation.

5. The method for predicting life migration based on improved compressed aggregation excitation network according to claim 4 is characterized in that: The improved compressed aggregation excitation network consists of an input layer, a deep residual shrinkage module, a convolution layer, a batch normalization layer, an activation layer, a concatenation layer, a fully connected layer and an output layer. In the improved compressed aggregation excitation network, the deep residual shrinkage module is cross-introduced into each feature extraction layer of the compressed aggregation excitation network to achieve noise reduction; The maximum mean difference distance metric is introduced into the improved compressed aggregation incentive network to construct an improved compressed aggregation incentive network model, which minimizes the distribution difference between the source domain and the target domain and predicts the remaining life of the migration. As shown in formula (5), the source domain h is measured. S and the target domain h T The distance between in, It is a mapping used to map the original variable to the reproducing kernel Hilbert space. H represents Map the data into the regenerating Hilbert space for distance measurement, Represents the source domain h S The data sample, represents the target domain h T The data sample, N S is the source domain h S The number of samples, N T The target domain h T The number of samples; The weighted sum of the source domain remaining life fitting loss function MSE and the operating condition generalization MMD loss is taken as the optimization target of the transferable compressed aggregation excitation network. The final loss function is expressed as formula (6): loss=λl RUL +μl MMD (6) In the formula, λ and μ are compromise parameters that are trained through back propagation to achieve an optimal model parameter and ultimately achieve accurate prediction of remaining life.

Images