Method for predicting residual life of multi-stage degraded bearing under cross working conditions

Abstract

The invention provides a method for predicting the residual life of a multi-stage degraded bearing under cross-working conditions, and relates to the technical field of mechanical equipment prediction and health management. According to the cross-domain health state recognition method, firstly, feature alignment among different cross-domain degradation stages is realized by using a depth feature migration model based on Transform, so that cross-domain health state recognition is carried out; secondly, in order to effectively mine key information in the performance degradation data, a TCN-SE model fused with an extrusion excitation SE module is constructed; on the basis, the time feature distribution difference is reduced in combination with MMD and domain confrontation methods, cross-domain invariant features are obtained, and RUL prediction of the multi-stage degraded bearing under the cross working condition is achieved through a staged loss function. According to the method, cross-domain health stage identification and RUL prediction are combined, the RUL of the multi-stage degraded bearing under cross-working conditions can be accurately predicted, and the problem that a traditional deep learning method is insufficient in generalization ability when facing different working condition data is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment prediction and health management, and in particular to a method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions. Background Art

[0002] In engineering practice, it is difficult to obtain sufficient supervised fault data because the equipment is usually in a normal state. At the same time, factors such as the operating conditions, application scenarios, and environmental noise of the bearing are complex and changeable, resulting in inconsistencies between offline and online working conditions, causing distribution differences between training data and test data, thereby reducing the performance and generalization of deep learning prediction methods. Remaining Useful Life (RUL) prediction methods based on deep learning usually assume that training data and test data follow the same distribution, but in actual engineering applications, this assumption is often difficult to meet. The operating conditions of mechanical equipment (such as speed, load, etc.) are complex and changeable, resulting in training set and test set data from different working conditions. The resulting data distribution differences reduce the generalization ability of the model, making it difficult for the model trained in the source domain to obtain satisfactory results in the target domain. Especially for bearings, the performance differences under different working conditions make it difficult for traditional models to effectively predict cross-domain RUL. To solve this problem, researchers introduced domain adaptation (DA) technology. As an important branch of transfer learning, DA aims to improve the generalization ability of the model in the target domain by minimizing the distribution difference between the source domain (training data) and the target domain (test data).

[0003] Existing DA methods can be mainly divided into distance metric-based methods and adversarial-based methods. For example, the maximum mean discrepancy (MMD) is used to learn cross-domain invariant features, thereby improving the accuracy of bearing RUL prediction under different working conditions. Although existing transfer learning methods have been successful in cross-working condition RUL prediction, most DA methods only focus on the consistent overall feature distribution of data under different working conditions, ignoring that the bearing degradation process is staged and different degradation stages show different characteristics. Simply aligning global time features may lead to key information loss and model performance degradation. In addition, most DA methods usually only use one of MMD or adversarial mechanisms. However, MMD does not work well in high-dimensional feature spaces, while adversarial mechanisms perform poorly when domain shifts are significant. In fact, given that MMD and adversarial mechanisms measure inter-domain differences from different perspectives, combining the advantages of the two can further narrow the inter-domain differences, which is of great research significance. In addition, traditional RUL prediction methods usually assume that each channel and feature in the network contributes equally to RUL. In fact, different channels and features contain different levels of degradation information, and their impact on RUL prediction is also different. Summary of the invention

[0004] In response to the above technical problems, the present invention comprehensively considers the global temporal characteristics between the source domain and the target domain and the distribution differences of the local degradation stages, and proposes a RUL prediction method for multi-stage degraded bearings under cross-working conditions. First, the Transformer-based deep feature migration model is used to achieve feature alignment between different degradation stages across domains, so as to perform cross-domain health status identification. Secondly, in order to effectively mine the key information in the performance degradation data, a TCN-SE (Temporal Convolutional Network with Squeeze-and-Excitation) model integrating the Squeeze-and-Excitation (SE) module is constructed. On this basis, the MMD and domain adversarial methods are combined to reduce the distribution differences of temporal features, obtain cross-domain invariant features, and realize the RUL prediction of multi-stage degraded bearings under cross-working conditions through a staged loss function.

[0005] Specifically, the present invention provides a method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions, comprising the following steps:

[0006] Step 1: Obtaining a full life cycle vibration signal of a bearing with a known remaining life and a vibration signal of a bearing with an unknown remaining life, wherein the vibration signals of the bearing with a known remaining life and the bearing with an unknown remaining life are collected under different working conditions;

[0007] Step 2: Calculate the multivariate features of the vibration signal, wherein the multivariate features include time domain, frequency domain and entropy value features to form a feature matrix;

[0008] Step 3: normalizing the feature matrix;

[0009] Step 4: Perform sliding time window processing on the normalized feature matrix so that the feature matrix is ​​encapsulated in a time window of fixed length to obtain a processed feature matrix;

[0010] Step 5: Perform stage change point detection on the health indicators of bearings with known remaining life, identify slow degradation points and fast degradation points respectively, obtain classification labels, and thus obtain a feature matrix with classification labels, which is recorded as the source domain data set;

[0011] Step 6: Use the Tansformer-based selective adversarial network to align the local degradation stages under different working conditions, so as to obtain the feature matrix with classification labels of the bearing with unknown remaining life, which is recorded as the target domain dataset;

[0012] Step 7: Construct a remaining life label set of the source domain dataset;

[0013] Step 8: Input the source domain dataset and target domain dataset containing the remaining life labels into the domain adaptation network model based on TCN-SE to learn cross-domain invariant features, then train the predictor of the model based on the cross-domain invariant features of the source domain, input the target domain data into the trained model, and predict the remaining life of the target bearing.

[0014] As a further illustration of the present invention, in step 3, when performing normalization processing, a minimum-maximum normalization method is used to scale each eigenvalue to the interval [0,1]:

[0015]

[0016] In the formula, is the normalized eigenvalue, and are the maximum and minimum values ​​of the j-th column feature, respectively.

[0017] As a further explanation of the present invention, in step 4, the sliding time window processing is performed on the normalized feature matrix, specifically:

[0018] By setting the time window size to w and the step size to s, we can get the processed data set: Where L is the number of samples after sliding window processing, calculated by the following formula:

[0019]

[0020] As a further illustration of the present invention, in step 5, the Chebyshev method is used to detect stage change points of the health index based on the root mean square, and the slow degradation points and the fast degradation points are respectively identified.

[0021] As a further illustration of the present invention, in step 6, the optimization goal of the model based on the Tansformer selective adversarial network is:

[0022]

[0023] In the formula, θ class is the model parameter set.

[0024] As a further illustration of the present invention, in step 7, the constructed remaining life label set is calculated by the following formula:

[0025]

[0026] In the formula, t i Indicates the time point corresponding to the i-th sample, y i It represents the RUL corresponding to the i-th sample, SDP is the slow degradation point, and FDP is the fast degradation point.

[0027] As a further illustration of the present invention, in the TCN-SE-based domain adaptation network model, the TCN-SE module is used to extract deep degradation feature information, wherein the introduced channel attention mechanism is used to perform weighted reorganization of features; and by combining the maximum mean difference and domain adversarial methods, the feature distribution distance between different domains is reduced, thereby learning cross-domain invariant characteristics.

[0028] As a further illustration of the present invention, the loss function of the maximum mean difference MMD is:

[0029]

[0030] In the formula, and are the deep features extracted by the feature extractor from the source domain and the target domain, respectively, n S With n T are the sample sizes of deep features in the source domain and target domain, respectively.

[0031] As a further illustration of the present invention, the domain adversarial method adopts an adversarial optimization strategy to confuse the feature distributions of different domains by minimizing the discriminative ability of the domain classifier. The domain adversarial loss function is:

[0032]

[0033] In the formula, n is the sample size, D i is the real domain label, the source domain is 0, the target domain is 1, is the predicted domain label.

[0034] As a further illustration of the present invention, in step 8, the prediction optimization target of the TCN-SE-based domain adaptive network model is:

[0035]

[0036] In the formula, θ rul is the model parameter set, β1 and β2 represent adjustable parameters.

[0037] Compared with the prior art, the present invention has the following beneficial technical effects:

[0038] Firstly, based on the ideas of transfer learning and domain adaptation, this invention proposes a transfer learning algorithm based on a selective adversarial strategy to realize the cross-domain stage recognition task. This method can effectively reduce the feature distribution differences between cross-domain data and improve the accuracy of bearing health stage recognition under cross-working conditions. Secondly, the feature distribution alignment is performed by combining the maximum mean difference and domain adversarial technology to obtain the domain invariant characteristics of the source domain and the target domain, thereby realizing cross-domain RUL prediction. By combining cross-domain health stage recognition with RUL prediction, the RUL of multi-stage degraded bearings under cross-working conditions can be accurately predicted, overcoming the problem of insufficient generalization ability of traditional deep learning methods when faced with different working condition data.

[0039] Other features and advantages of the technical solution will be described in the subsequent description, and partly become apparent from the description, or understood by implementing the technical solution. The purpose and other advantages of the technical solution can be achieved and obtained through the structures specifically pointed out in the written description and the drawings.

[0040] The technical solution of the present technical solution is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings are used to provide a further understanding of the technical solution and constitute a part of the specification. Together with the embodiments of the technical solution, they are used to explain the technical solution and do not constitute a limitation of the technical solution. In the accompanying drawings:

[0042] Figure 1 It is a selective adversarial network structure based on Tansformer in the present invention;

[0043] Figure 2 The internal structure of the extrusion and excitation block in the present invention;

[0044] Figure 3 It is the domain adaptive network model framework based on TCN-SE in the present invention;

[0045] Figure 4 (a)-(b) are t-SNE feature visualizations of bearings 2-6 in the present invention; Figure 4 (a) is the feature without transfer training, and Figure 4 (b) is the feature after transfer training;

[0046] FIG5(a)-(b) are feature distribution comparisons of different working conditions with and without transfer training in the present invention, wherein FIG5(a) is feature distribution without transfer training, and FIG5(b) is feature distribution after transfer training;

[0047] Figure 6 (a)-(d) is a comparison of the RUL prediction results with and without the migration TCN-SE in the present invention, wherein Figure 6 (a) is the prediction result of T1: bearing 2-6, and Figure 6 (b) is the prediction result of T2: bearing 3-3; Figure 6 (c) is the prediction result of T3: bearing 1-1, and Figure 6 (d) is the prediction result of T5: bearing 1-3;

[0048] Figure 7(a)-(b) is a t-SNE feature visualization of bearing 1-1 in the present invention, wherein Figure 7(a) is the feature without migration training; Figure 7(b) is the feature after migration training. DETAILED DESCRIPTION

[0049] The preferred embodiments of the present technical solution are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present technical solution, and are not used to limit the present technical solution.

[0050] The present invention provides a method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions, comprising the following steps:

[0051] Step 1: Obtain a full life cycle vibration signal of a bearing with a known remaining life and a vibration signal of a bearing with an unknown remaining life, wherein the vibration signals of the bearing with a known remaining life and the bearing with an unknown remaining life are collected under different operating conditions.

[0052] Specifically, you can use Represents the acquired vibration signal data set, where x i =(x1,x2,…,x M ), N and M represent the sample size and sample length respectively.

[0053] Step 2: Calculate the multivariate features of the vibration signal, which include time domain, frequency domain and entropy value features, to form a feature matrix k is the feature dimension.

[0054] Step 3: For the feature matrix Perform normalization.

[0055] Specifically, during normalization, the minimum-maximum normalization method is used to scale each eigenvalue to the [0,1] interval:

[0056]

[0057] In the formula, is the normalized eigenvalue, and are the maximum and minimum values ​​of the j-th column feature, respectively.

[0058] Step 4: Perform sliding time window processing on the normalized feature matrix so that the feature matrix is ​​encapsulated in a time window of fixed length to obtain a processed feature matrix.

[0059] By setting the time window size to w and the step size to s, we can get the processed data set: Where L is the number of samples after sliding window processing, calculated by the following formula:

[0060]

[0061] Step 5: Perform stage change point detection on the health index HI of the bearing with known remaining life, identify the slow degradation points and fast degradation points respectively, obtain classification labels, and thus obtain a feature matrix with classification labels, which is recorded as the source domain data set.

[0062] Specifically, the Chebyshev method is used to detect the stage change points of the health indicators based on the root mean square (RMS), and the slow degradation point (SDP) and fast degradation point (FDP) are identified respectively.

[0063] Step 6: Use the Tansformer-based selective adversarial network to align the local degradation stages under different working conditions, so as to obtain the feature matrix with classification labels for bearings with unknown remaining life, which is recorded as the target domain dataset.

[0064] Specifically, in this step, a Tansformer-based selective adversarial network is used to achieve local degradation stage alignment under different working conditions, including the following healthy stage recognition method based on Transformer deep feature migration:

[0065] In order to more effectively identify the trend changes of bearing health status under different working conditions, a selective adversarial network is introduced to realize the cross-domain stage recognition task. In industrial scenarios, the degradation trend appearing in the actual monitoring data is often unknown and may usually be a subset of the source domain data. In this case, if global domain adaptation is directly applied, state type samples in the source domain that do not exist in the target domain may cause negative transfer and reduce the performance of the model in the target domain. In addition, the model may learn features that do not exist or are irrelevant in the target domain, affecting the recognition accuracy. In order to match different health state spaces C s ≠C t In the DA process, it is necessary to filter and remove samples of outlier conditions and irrelevant conditions from the source domain. Therefore, a method based on Tansformer selective adversarial network to achieve local health stage matching is proposed.

[0066] First, the domain discriminator D is split into |C s |Domain discriminators D divided by categories k , where k represents the health status of the source domain, that is, k = 1,...,|C s |. Each D k Domain adaptation is performed on source domain samples and target domain samples with health status k. Since the health status space of the target domain |C t | is unknown, so it is difficult to determine which health state D needs to be executed k However, for the input sample x i , its predicted output Is a C s This describes the probability distribution of x i The probability of being assigned to each health state. Therefore, the probability output can be used to convert each input sample x i Assign to the corresponding D k . This process can be expressed as:

[0067]

[0068] Where D k represents the kth domain discriminator, d i represents the domain label and defines the shared feature space of the two domains, L ce is the cross entropy loss, which can be calculated as follows:

[0069]

[0070] Where y is the label of the input sample x, is the predicted label, and K is the number of health states. In order to achieve domain-invariant learning using cross-entropy loss, a gradient reversal layer (GRL) is introduced in the shared feature layer for adversarial learning.

[0071] Compared with the traditional adversarial loss, the established multi-discriminator network can effectively avoid the negative transfer of abnormal samples and promote the positive transfer of each instance through the instance-level probability weighting mechanism. In addition, in the selective adversarial network, a class-level weighting strategy is also introduced to further reduce the impact of abnormal samples and improve the connection of shared health status. In theory, the health status C of the target domain is t The domain discriminator of can promote the positive transfer of the model, while being responsible for the abnormal health status in the source domain The domain discriminator of the source domain usually introduces noise, which weakens the positive transfer effect of the model. Therefore, by reducing the weight of the abnormal health state domain discriminator of the source domain, the introduction of noise can be reduced and the positive transfer effect can be improved. The loss can be expressed as:

[0072]

[0073] In the formula, Represents the class-level weight of health state k.

[0074] Although the loss function L DIS Relevant knowledge can be selectively transferred, but it depends heavily on Therefore, the entropy minimization strategy is introduced to further enhance the separability between different health states and improve the prediction accuracy of the target sample. This process is to minimize the probability The prediction entropy L E The loss function L E It is expressed as:

[0075]

[0076] Where H(·) represents the conditional entropy loss.

[0077] There are two fully connected layers in the classifier, where the cross entropy loss is used to learn the discriminative features, which is defined as:

[0078]

[0079] The total loss of the Tansformer-based selective adversarial network includes the three terms given in formula (5), formula (6) and formula (7), which can be expressed as:

[0080] L total-class =L C (x s ,y s )+λ1L DIS (x s ,x t )+λ2L E (x t ) (8)

[0081] Where λ1 and λ2 represent trade-off parameters.

[0082] The network parameters are updated as follows:

[0083]

[0084] Where η class is the learning rate, θ F ,θ C and θ D They are respectively based on the parameters of the feature extractor, classifier and domain discriminator in the Tansformer selective adversarial network.

[0085] In summary, the optimization objective of the above Tansformer-based selective adversarial network model can be expressed as:

[0086]

[0087] In the formula, θ class is the model parameter set.

[0088] Through these improvements, we aim to build a cross-condition bearing health stage identification model that is more adaptable to actual industrial scenarios. This model can not only effectively handle the distribution differences between the source domain and the target domain, but also flexibly deal with unknown degradation patterns that may appear in actual monitoring data.

[0089] Step 7: Construct a remaining lifespan label set for the source domain dataset.

[0090] Specifically, in the RUL prediction of bearings, RUL labels are mostly in the form of linear annotation from 1 to 0, so the constructed remaining life label set can be calculated by the following formula:

[0091]

[0092] In the formula, t i Indicates the time point corresponding to the i-th sample, y i It represents the RUL corresponding to the i-th sample, SDP is the slow degradation point, and FDP is the fast degradation point.

[0093] As the equipment continues to operate, the bearings begin to show varying degrees of damage. i For samples with SDP≤t, the bearing is in a healthy stage. Since the early stable signal has no obvious degradation trend, RUL prediction is not performed. i ≤T end For samples of , the RUL value decreases linearly from 1 to 0.

[0094] Step 8: Input the source domain dataset containing the remaining life label and the target domain dataset (without RUL label) into the domain adaptation network model based on TCN-SE to learn the cross-domain invariant features, and then train the predictor of the model based on the cross-domain invariant features of the source domain. Input the target domain data into the trained model to predict the remaining life of the target bearing.

[0095] The structure of the above TCN-SE module and its calculation steps are as follows:

[0096] In order to improve the RUL prediction performance based on the Temporal Convolutional Network (TCN) model, a channel attention mechanism is introduced. Although traditional deep learning models perform well in feature representation and capturing deep relationships within input data, they lack sensitivity to information features between different channels. To overcome this limitation, this study adopts the SE block, also known as the channel attention mechanism. This mechanism can dynamically adjust the weights of feature channels, strengthen key information and suppress redundant content, thereby enhancing the model's ability to recognize important temporal patterns in the RUL prediction task. Based on this, the TCN-SE model is proposed, which can not only effectively improve the expression ability of TCN, but also enhance its sensitivity to multi-channel data, thereby achieving more accurate and reliable results in the RUL prediction task. Figure 2 The structure of the channel attention mechanism is shown, and its specific calculation steps are described as follows.

[0097] (1) Squeeze operation: The spatial information is input into the feature map U = [u1,u2,...,u C ], Compression is performed to generate a channel descriptor, thereby capturing the global spatial information of each feature channel. It is expressed as:

[0098]

[0099] In the formula, F sq (·) indicates extrusion operation, represents the cth channel in the input feature map, z c is the global information of channel c, i.e., the output of the squeeze operation, C is the number of channels of the feature map, H and W represent the number of rows and columns of channel c, respectively, and u c (i,j) is the element in the i-th row and j-th column of channel c.

[0100] (2) Excitation operation: Using two linear layers and nonlinear activation functions, the compressed features are adaptively learned for importance, and nonlinear relationships between channels are constructed to generate weight coefficients for each feature channel. Specifically, the first linear layer maps the channel descriptor to a low-dimensional space to reduce model complexity; the second linear layer remaps the features to the original dimension. ReLU and Sigmoid functions are used after the two linear layers to construct nonlinear relationships between channels. The output of the Sigmoid function is directly used as the weight coefficient of each channel. This design enables the model to adaptively recalibrate feature channels, effectively improving attention to key signals, thereby enhancing the model's ability to recognize important features. It is expressed as:

[0101] s=F ex(z,W)=σ(g(z,W))=σ(W2δ(W1z)) (13)

[0102] In the formula, F ex (·) indicates the incentive operation, is the generated channel importance weight vector, δ represents the ReLU activation function, σ represents the Sigmoid activation function, and are the weights of the two linear layers, where r is the dimensionality reduction ratio.

[0103] (3) Reweighting: Apply the learned weight coefficients to the corresponding channels of the original feature map, realize the adaptive recalibration of features, and enhance the model's ability to model the dependencies between channels. It is expressed as:

[0104]

[0105] In the formula, F scale (·) represents the channel-level multiplication operation, Represents the reweighted feature map.

[0106] In the TCN-SE architecture, multiple TCN-SE modules are stacked together in sequence, and each module consists of a TCN hidden layer and a SE layer. The TCN hidden layer is used to capture long-term degradation information, while the SE layer evaluates and reorganizes the acquired information to further highlight key performance degradation signals while weakening the interference of non-critical signals. After each TCN layer, it enters the SE layer, and the input and output shapes of the SE layer remain consistent. The TCN-SE module structure is expressed as:

[0107] X conn =SE(TCN(X real )) (15)

[0108] Where, X real is the input sequence, X conn is the output sequence, TCN is the TCN hidden layer, and SE is the SE layer.

[0109] After stacking multiple layers of the network, it is necessary to avoid the loss of information on the real input sequence due to performance degradation caused by gradient vanishing. The solution is to use the skip connection technology, as shown in Equation (16).

[0110] X key =ReLU(X real +X conn ) (16)

[0111] Specifically, the dimension-transformed X real With X connAdd together to get the key virtual signal X learned from the TCN-SE block key Since TCN-SE can effectively extract long-term degradation patterns in the input sequence while being highly sensitive to performance degradation, the output can more prominently reflect the degradation characteristics of the system, providing a reliable data basis for subsequent predictions.

[0112] The specific prediction process of the domain adaptation network of TCN-SE in step 8 is described in detail below:

[0113] The present invention uses the TCN-SE module to extract deep degradation feature information, in which the channel attention mechanism is integrated to perform weighted reorganization of features. The MMD and domain adversarial methods are combined to reduce the feature distribution distance between different domains, thereby learning cross-domain invariant characteristics. Finally, the trained model is used for RUL prediction of bearings.

[0114] The model framework of the TCN-SE domain adaptive network proposed in the present invention is as follows Figure 3 As shown in Figure 1, its optimization objective consists of three parts: the RUL regression loss term of the source domain data, the MMD loss term between the deep features of the source domain and the target domain, and the domain adversarial loss term.

[0115] (1) The present invention uses labeled source domain data to train the RUL predictor and optimizes the model parameters by minimizing the prediction loss. Finally, the trained model is applied to the test data in the target domain to predict the RUL of the bearing. The prediction performance is evaluated using the staged prediction loss function L RUL , calculated as follows:

[0116]

[0117] In the formula, y RUL is the real RUL value, To predict the RUL value, k is the importance parameter, k = 0.6, SDP is the slow degradation point that begins to enter the slow degradation stage, and FDP is the fast degradation point. end is the end of bearing life.

[0118] In the slow degradation stage, RMSE is used as the loss function to ensure the convergence of the prediction; in the fast degradation stage, the weighted sum-based Score function is used. Compared with the symmetric constraint of RMSE, the Score function imposes a more severe penalty on delayed predictions, which is particularly important because a predicted RUL greater than the actual RUL may result in missing the best preventive maintenance opportunity. By adopting a staged prediction loss function strategy, this model can effectively balance prediction accuracy and timeliness in different degradation stages of the bearing, thereby significantly improving the overall prediction performance.

[0119] (2) In order to reduce the feature distribution difference between the source domain and the target domain, the present invention obtains domain-invariant features by minimizing MMD. The loss function of the domain difference metric MMD is:

[0120]

[0121] In the formula, and are the deep features extracted by the feature extractor from the source domain and the target domain, respectively, n S With n T are the sample sizes of deep features in the source domain and target domain, respectively.

[0122] (3) The domain adversarial method adopts an adversarial optimization strategy to confuse the feature distribution of different domains by minimizing the discriminative ability of the domain classifier. In the adversarial mechanism, H-distance is used to measure the distribution difference between different domains. However, H-distance is difficult to calculate and can be approximated by learning a classifier for distinguishing source domain and target domain samples. The classifier consists of a fully connected layer and a single neuron binary classification layer, and uses binary classification cross entropy as the loss function. The model parameters are continuously updated and optimized through back propagation, so that the domain discriminator cannot distinguish the source of the features generated by the feature extractor. This process proves that the feature extractor can generate feature representations with domain invariance, thereby realizing cross-domain transfer learning. The domain adversarial loss function is:

[0123]

[0124] In the formula, n is the sample size, D i is the real domain label, the source domain is 0, the target domain is 1, is the predicted domain label.

[0125] Among them, the introduction of GRL between the feature extractor and the domain discriminator is the key to this method. GRL automatically reverses the gradient during the back-propagation process, effectively promoting the adversarial process between the feature extractor and the domain discriminator, enabling the feature extractor to generate domain-invariant feature representations. By minimizing the distribution difference between the source domain and the target domain, a latent feature space is found so that the source domain and the target domain obey similar feature distributions.

[0126] Combining the loss functions of RUL prediction and domain adaptation, the final optimization function is summarized as follows:

[0127] L total-rul =L RUL +β1L MMD (F s ,F t )+β2L D (F s ,F t ) (twenty two)

[0128] Where β1 and β2 represent adjustable parameters, which are used to adjust the influence weights of domain discrimination loss and MMD loss. s and F t Represent the deep features of data in the source domain and target domain respectively.

[0129] This model uses the Adam optimization algorithm to minimize the overall objective function to update the model parameters. In each iterative training, the network parameters are updated as follows:

[0130]

[0131] Where η rul is the learning rate, θ f ,θ r and θ d are the parameters of the feature extractor, RUL predictor and domain discriminator, respectively.

[0132] In summary, the prediction optimization goal of the domain adaptive network model based on TCN-SE in the present invention is:

[0133]

[0134] In the formula, θ rul is the model parameter set.

[0135] Example Application

[0136] The present invention uses the bearing degradation data set provided by FEMTO-ST Institute in the IEEE PHM 2012 Data Challenge to verify the effectiveness of the proposed method. The data set is collected by the PRONOSTIA experimental platform, which can accelerate the degradation of bearings under constant or variable operating conditions, while collecting online monitoring data of speed, load force, temperature and vibration signals. The PRONOSTIA platform can perform degradation experiments from the start of operation to failure, and automatically stop the test when the vibration signal amplitude exceeds 20g to prevent damage from spreading and ensure safety. The experimental platform uses NSK 6804DD rolling bearings, and an accelerometer is installed on the horizontal axis and vertical axis of its outer ring to measure and collect vibration signals in two directions. The sampling frequency of the vibration signal is 25.6kHz. The sampling time is 0.1 seconds, sampling is performed every 10 seconds, and 2560 data sampling points are generated each time.

[0137] Since the life of rolling bearings in actual work can reach tens of thousands of hours, in order to obtain full life cycle data, the experiment accelerates the failure of bearings by increasing the speed or adding external loads. Data collection involves three working conditions: working condition 1: load 4000N, speed 1800r / min; working condition 2: load 4200N, speed 1650r / min; working condition 3: load 5000N, speed 1500r / min. Studies have shown that the use of data collected by horizontal vibration sensors to analyze bearing degradation trends is better than the use of data collected by vertical vibration sensors. Horizontal vibration data is selected for analysis. The PHM 2012 dataset provides 17 rolling bearing datasets that run completely to failure. According to the requirements, the datasets under different working conditions are divided into source domains and target domains. The migration tasks constructed are listed in Table 1. This example conducts 6 groups of experimental tasks to evaluate the effectiveness of the proposed method. The labeled working conditions 1, 2 and 3 are used as source domains, and migrated to the remaining two unlabeled working condition datasets.

[0138] Table 1. Specific details of the migration prediction task

[0139]

[0140]

[0141] Step 1: Cross-domain phase identification and verification analysis

[0142] The time domain feature RMS is used in combination with Chebyshev inequality to offline divide the operating status of the bearing. According to the division results, the classification labels of the training data based on the Tansformer selective adversarial network can be obtained, and the multivariate degradation features are input for classification modeling. According to the specific migration task details in Table 1, the online health status of the rolling bearing under cross-working conditions is identified. In order to reduce the impact of model randomness, 5 experiments are conducted and the average value is taken as the final result. The experimental results are shown in Table 2.

[0143] Table 2 Multi-stage identification results under cross-operating conditions

[0144]

[0145] The experimental results show that the multi-stage recognition accuracy of online test bearings varies due to different samples. The average test results of the proposed method on 6 migration tasks are all greater than 85%, indicating that the proposed method can well divide the multiple health states of bearings even when used under different working conditions, and achieve efficient and accurate cross-domain stage recognition. At the same time, this also shows that the effective integration of Tansformer and the selective adversarial mechanism constructs a powerful cross-domain classification model, which helps to realize online condition monitoring of bearings and lays the foundation for subsequent RUL prediction.

[0146] To verify the effectiveness of transfer training, feature visualization analysis was performed. The t-Distributed Stochastic Neighbor Embedding (t-SNE) technology was used to simplify the features of the source and target domains into two dimensions. Figure 4 shows the t-SNE feature visualization of bearing 2-6, where Figure 4(a) shows the features without transfer training, and Figure 4(b) shows the features after transfer training. As can be seen from Figure 4(a), there is a significant distribution difference between the source and target domain features of bearing 2-6 without transfer training, and the sample points of different status categories have obvious overlaps in the feature space, and the separability is poor. In contrast, Figure 4(b) shows the results after transfer training: the feature distribution difference between the source and target domains is significantly reduced, the health status features and fault features show good separability, and the clustering effect of samples of the same health status is also improved. This shows that the deep features after transfer can provide effective discriminant information for classification, have high sensitivity to changes in health status, and the proposed stage recognition method based on transfer learning shows rationality and effectiveness.

[0147] Step 2: RUL prediction results and comparative analysis

[0148] In order to verify the effectiveness of the proposed method in reducing the distribution differences between different domains, in the example application, the migration task from working condition 2 to working condition 3 is selected, and its distribution is shown in Figure 5, where Figure 5(a) shows the feature distribution before migration training, and Figure 5(b) shows the feature distribution after migration training. It can be seen from Figure 5(a) that although there is some overlap in the feature distribution of the two working conditions, the overall distribution difference is significant. This intuitively illustrates the challenge of cross-working condition RUL prediction, because the prediction model trained with source domain data under different working conditions cannot directly achieve the ideal RUL prediction results in the target domain. Therefore, it is necessary to eliminate the distribution differences of the extracted features under different working conditions. As shown in Figure 5(b), after using the transfer learning method, the overlapping area of ​​the feature distribution between the source domain and the target domain increases significantly. However, there are still some differences in the feature distribution between the two domains, indicating that the proposed method retains the target domain-specific features while aligning the features. Therefore, the proposed method can not only retain the target domain-specific features, but also make the target domain features similar to the source domain features, thereby improving the accuracy of cross-domain RUL prediction.

[0149] The six groups of migration tasks in Table 1 are used to verify the effectiveness of the proposed method in RUL prediction under different working conditions. The RUL prediction results of different migration tasks are shown in Figure 6 and Table 3. Among them, Figure 6 (a) is the prediction result of T1: bearing 2-6, Figure 6 (b) is the prediction result of T2: bearing 3-3; Figure 6 (c) is the prediction result of T3: bearing 1-1, and Figure 6 (d) is the prediction result of T5: bearing 1-3. It can be seen from the figure that the RUL prediction curve of the method based on the TCN-SE domain adaptive network proposed in the present invention is significantly closer to the true value, which is better than the TCN-SE method that does not use migration knowledge. This result shows that by solving the problem of distribution differences in different working conditions, the prediction performance of the proposed method has been significantly improved. In addition, Figure 6 (d) shows that there is a large prediction error in the migration prediction experiment (T5) from working condition three to working condition one. This may be because when working condition three is used as the source domain, the bearing data set has too few samples. The model cannot fully capture the data features in the source domain during the learning process, and it is difficult to find the shared feature space from the source domain to the target domain, resulting in limited prediction effect of the model and increased error.

[0150] Table 3 shows the comparison results of the six groups of transfer tasks with and without transfer training. 2 , SCORE as evaluation indicators, the results show that the proposed method is significantly better than the TCN-SE method without transfer knowledge when predicting RUL under different working conditions. This verifies the superiority of the RUL prediction method based on the TCN-SE domain adaptive network, which can effectively improve the prediction performance of the model.

[0151] Table 3 Evaluation indicators of prediction results of each transfer learning task

[0152]

[0153]

[0154] In addition, in order to intuitively demonstrate the effectiveness of transfer training, the extracted deep degradation features are visualized and analyzed, and the multi-dimensional nonlinear features are mapped to two-dimensional space. Taking T3 as an example, the t-SNE method is used to reduce the dimension of the features and visualize them to show their distribution in space. As shown in Figure 7, Figure 7(a) shows the features without transfer training; Figure 7(b) shows the features after transfer training. The blue circles represent the source domain samples, and the red circles represent the target domain samples. As can be seen from Figure 7(a), without the application of DA, the target domain samples are scattered in the feature space, which is significantly different from the source domain feature distribution, increasing the difficulty of cross-domain prediction. However, after transfer training, as shown in Figure 7(b), the distribution difference between the source domain and target domain features is significantly reduced overall, but the two do not completely overlap. Therefore, the domain adaptation strategy proposed in this patent effectively reduces the feature distribution difference and constructs a domain-invariant feature space. This helps the target domain samples to be better distributed in the manifold of the source domain features and promotes the realization of cross-domain transfer learning tasks.

[0155] Obviously, those skilled in the art can make various changes and modifications to the technical solution without departing from the spirit and scope of the technical solution. Thus, if these modifications and variations of the technical solution fall within the scope of the claims of the technical solution and their equivalents, the technical solution is also intended to include these modifications and variations.

Claims

1. A method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions, characterized in that: The steps include: Step 1: Obtaining a full life cycle vibration signal of a bearing with a known remaining life and a vibration signal of a bearing with an unknown remaining life, wherein the vibration signals of the bearing with a known remaining life and the bearing with an unknown remaining life are collected under different working conditions; Step 2: Calculate the multivariate features of the vibration signal, wherein the multivariate features include time domain, frequency domain and entropy value features to form a feature matrix; Step 3: normalizing the feature matrix; Step 4: Perform sliding time window processing on the normalized feature matrix so that the feature matrix is ​​encapsulated in a time window of fixed length to obtain a processed feature matrix; Step 5: Perform stage change point detection on the health indicators of bearings with known remaining life, identify slow degradation points and fast degradation points respectively, obtain classification labels, and thus obtain a feature matrix with classification labels, which is recorded as the source domain data set; Step 6: Use the Tansformer-based selective adversarial network to align the local degradation stages under different working conditions, so as to obtain the feature matrix with classification labels of the bearing with unknown remaining life, which is recorded as the target domain dataset; Step 7: Construct a remaining life label set of the source domain dataset; Step 8: Input the source domain dataset and target domain dataset containing the remaining life labels into the domain adaptation network model based on TCN-SE to learn cross-domain invariant features, then train the predictor of the model based on the cross-domain invariant features of the source domain, input the target domain data into the trained model, and predict the remaining life of the target bearing.

2. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In step 3, when normalizing, the minimum-maximum normalization method is used to scale each eigenvalue to the [0,1] interval: In the formula, is the normalized eigenvalue, and are the maximum and minimum values ​​of the j-th column feature, respectively.

3. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In step 4, the normalized feature matrix is ​​subjected to sliding time window processing, specifically: Setting the time window size to w and the step size to s, we can get the processed data set: X input =(x1,x2,…,x L )∈R L×w×k , where L is the number of samples after the sliding window processing, calculated by the following formula:

4. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In step 5, the Chebyshev method is used to detect the stage change points of the health index based on the root mean square, and the slow degradation points and fast degradation points are identified respectively.

5. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In step 6, the optimization goal of the Tansformer-based selective adversarial network model is: In the formula, θ class is the model parameter set.

6. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In step 7, the remaining life label set constructed is calculated by the following formula: Where, t i Indicates the time point corresponding to the i-th sample, y i It represents the RUL corresponding to the i-th sample, SDP is the slow degradation point, and FDP is the fast degradation point.

7. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1 is characterized in that: In the TCN-SE-based domain adaptation network model, the TCN-SE module is used to extract deep degraded feature information, and the introduced channel attention mechanism is used to weighted reorganize the features; and the feature distribution distance between different domains is reduced by combining the maximum mean difference and domain adversarial methods, thereby learning cross-domain invariant characteristics.

8. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 7 is characterized in that: The loss function of the maximum mean difference MMD is: In the formula, and are the deep features extracted by the feature extractor from the source domain and the target domain, respectively, n S With n T are the sample sizes of deep features in the source domain and target domain, respectively.

9. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 7 is characterized in that: The domain adversarial method adopts an adversarial optimization strategy to confuse the feature distribution of different domains by minimizing the discriminative ability of the domain classifier. The domain adversarial loss function is: In the formula, n is the sample size, D i is the real domain label, the source domain is 0, and the target domain is 1. is the predicted domain label.

10. The method for predicting the remaining life of a multi-stage degraded bearing under cross-operating conditions as claimed in claim 1, characterized in that: In step 8, the prediction optimization target of the TCN-SE-based domain adaptive network model is: In the formula, θ rul is the model parameter set, β1 and β2 represent adjustable parameters.

Images