A cross-adaptive deep transfer learning method and system based on bearing life

Abstract

The application discloses a bearing residual service life multi-feature cross adaptive deep migration learning method and system considering the diversity of degradation processes. The method collects original vibration signals of a rolling bearing, and obtains multi-dimensional vibration signal data through sliding window segmentation and time-frequency feature information extraction on the collected signals. Then, the time domain features in the multi-dimensional signals are reduced through PCA to obtain comprehensive features, and MMD is used to obtain the degradation index and gradient features of the bearing. An adaptive gradient iterative division algorithm is used to adaptively divide the data of the whole life cycle of the rolling bearing to obtain different degradation stages of the bearing. According to the actual working condition of the bearing, a concept of degradation rate is proposed, and the test set and the training set are labeled according to the concept. The data is put into a multi-feature cross migration network to realize the prediction of the incomplete target domain RUL. The method uses known and easily obtained data for training, and realizes effective prediction on incomplete life cycle data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bearing life prediction, and particularly relates to a cross-adaptive deep migration learning method and system based on bearing life. BACKGROUND

[0002] As one of the key components of industrial rotating machinery, rolling bearings are widely used in various equipment. Once a fault occurs in industrial production, it will cause unplanned equipment downtime and result in serious economic losses. Therefore, it is necessary to detect and identify various potential abnormalities and faults as early as possible, and the residual useful life (RUL) prediction of bearings plays a crucial role in avoiding dangerous situations before operational risks. As a key link of prognosis and health management, fault diagnosis and RUL prediction have been studied by a large number of scholars in recent years. Among them, fault diagnosis has achieved rich results in application after years of research, while the prediction of RUL is still in the development stage, and how to realize the prediction of multiple working conditions and cross-working conditions is still a huge problem. In recent years, transfer learning as a method that can solve the difference in feature distribution of different domains has developed rapidly.

[0003] Domain adaptation (DA) can map features in different domains to the same feature space for alignment and solution. Therefore, DA provides a new and effective solution for RUL prediction in cross-working conditions. Although the above-mentioned RUL prediction has made great progress, there are still the following problems:

[0004] (1) Most of the existing DA methods for cross-working condition RUL prediction are feature alignment for the whole life cycle of bearings. However, as we all know, the bearing degradation process is mostly in a multi-stage degradation method in most cases. During long-term operation, even in similar working environments, their degradation patterns will be very different. Therefore, it is difficult to align features from the level of degradation stage with the existing methods.

[0005] (2) The existing prediction methods for bearing degradation level division often first construct a health index curve, and then set a threshold to divide the health index curve into stages. This method ignores the difference in degradation patterns under different working conditions, and cannot automatically adapt to the actual degradation process of the bearing, so it cannot obtain accurate multi-stage information of the bearing degradation.

[0006] (3) The existing modeling method for the bearing residual life label in the modeling idea of the model is to remove the health state at the first time of the failure signal (the health state at this time is 100%) to the remaining service time to determine the RUL of the bearing. This method cannot follow the use of the bearing in the actual industrial scene, ignores the influence of the degradation degree and the degradation rate in different degradation stages, and thus affects the accuracy of the residual life prediction result of the transfer learning.

[0007] (4) Most of the existing bearing RUL transfer learning models are trained based on the source domain and the target domain with the full life cycle, which ignores the fact that it is difficult to obtain the full life cycle data of a bearing in actual engineering application. Using complete source data and partial target data from the early degradation period for domain adaptation will cause the data mismatch in the similar degradation level across the domain, thereby causing the problem of decreased bearing RUL cross-condition prediction accuracy. SUMMARY

[0008] The present application provides a bearing life-based cross-adaptive deep transfer learning method, which can solve the technical problems existing in the background art, and also provides a bearing life-based cross-adaptive deep transfer learning system.

[0009] Technical scheme: The bearing life-based cross-adaptive deep transfer learning method provided by the present application comprises the following steps:

[0010] (1) Collecting the original vibration signal of the rolling bearing, and using a sliding window segmentation method to obtain multiple vibration signal data;

[0011] (2) Data preprocessing and feature extraction are performed on the original vibration signal obtained in step (1) to obtain time domain and frequency domain feature parameters;

[0012] (3) The obtained multi-dimensional time domain features are reduced using PCA dimension reduction technology to obtain time domain comprehensive features;

[0013] (4) The time domain comprehensive features in step (3) are segmented into multiple segments using a sliding window, and the maximum mean discrepancy MMD of each segment and the first segment is calculated to obtain a degradation index and a corresponding gradient curve; the maximum mean discrepancy Max meandiscrepancy is used to measure the difference between two data distributions.

[0014] (5) The obtained gradient curve is automatically identified by the AGIP algorithm to obtain the different stages of the bearing.

[0015] (6) According to the obtained division identification information and the characteristic parameters, a training set and a test set are divided, and corresponding labels are marked according to the degradation rates, wherein the training set is a labeled source domain and an unlabeled early target domain;

[0016] (7) The data of each stage in the training set is sent into the time series prediction network for training to obtain a prediction model of different stages;

[0017] (8) The test set obtained by division is sent into the time series prediction network trained in step (7) for prediction, and a final rolling bearing residual life prediction result is output.

[0018] Further, comprising:

[0019] The step (1) specifically comprises: the original vibration signal data of the rolling bearing includes horizontal vibration signal data and vertical vibration signal data, which are collected by an acceleration sensor installed on a bearing base to obtain full life cycle data of the bearing operation, and the end of collection is taken as a signal when the vibration amplitude of the bearing exceeds a certain threshold; and the specific method of the sliding window segmentation method is to segment the entire original vibration signal data at equal distances according to a certain step length l.

[0020] Further, comprising:

[0021] The calculation of the degradation index in the step (4) comprises:

[0022] The obtained bearing degradation index is a time domain comprehensive feature obtained by dimension reduction according to the multi-dimensional time domain features obtained in step (2), and is obtained by using the sliding window segmentation calculation and the MMD of the first window.

[0023] Further, comprising:

[0024] In the step (3), the AGIP algorithm comprises:

[0025] First, the average value of the gradient of the first 5% of the gradient curve is calculated as a baseline gradient threshold [g0], and then the gradient threshold of each stage is automatically updated to determine the first prediction time point FPT and the stage transition point STP, so as to realize the division of the multi-stage degradation of the full life cycle data of the rolling bearing.

[0026] Further, comprising:

[0027] The step (5) of dividing the training set and the test set comprises:

[0028] (51) The degradation rates of different stages are calculated, and corresponding degradation labels are marked according to the different degradation rates, comprising:

[0029] The decay rate of each decay stage is calculated:

[0030]

[0031] Wherein: represents the decay rate of the nth decay stage, is the decay degree of the nth decay stage, and three decay stages are defined as 100%-60%, 60%-10% and 10%-0, respectively, which are respectively recorded as: mild decay stage, slow decay stage and accelerated decay stage, is the decay time corresponding to the decay stage;

[0032] Then the label is calculated:

[0033]

[0034] Wherein, represents the label of the ith sample, , and is the cross-stage point obtained by the AGIP algorithm;

[0035] (52) The three decay stages of the source domain and the early stage data of the target domain are used as the training set, and the slow decay stage and the rapid decay stage of the target domain are used as the test set.

[0036] Further, comprising:

[0037] In step (6), the data of each stage in the training set is input into the time series prediction network model, and the time series prediction model comprises a feature extraction layer, a multi-feature cross migration layer and a network prediction layer. The feature extraction layer is a Transformer network structure, the decay data of each stage of each source domain and the decay data of the first decay stage of the target domain are input into the Transformer network structure, after inputting the input layer and N layers of encoder and decoder, the input features are linearly transformed and activated by the activation function, and the final output features are obtained. The multi-feature cross migration layer includes multiple cross attention layers, taking the source domain as the query Q and the target domain as the key K and the value V, and a new target loss function is proposed.

[0038] Further, comprising:

[0039] The target loss function comprises:

[0040] Regression loss : For minimizing the prediction error of the training data, adaptively adjusting the network training parameters, and making the predicted RUL closer to the actual result; the present application selects a radial basis function RBF as a kernel function, and calculates the mean square error MSE As follows:

[0041]

[0042] wherein, and respectively represent the real RUL label and the model predicted RUL label, n represents the batch size in the training process;

[0043] Distribution difference loss : MMD represents the distribution difference between the source domain and the target domain, so that the network can better extract domain invariant features, as a non-parametric measure MMD, it measures the first-order distribution divergence between two domains; therefore, the distribution difference loss formula after introducing kernel learning can be represented as:

[0044]

[0045] wherein, represents the i-th element of the source domain, and respectively represent the size of each batch of the source domain and the target domain, represents the kernel function;

[0046] Knowledge distillation loss L distillation : the distillation loss is used for the cross-attention module, which is mainly used to align the feature distribution of the source domain and the target domain, in the module, the source domain and the target domain are represented as the teacher and the student respectively, the training of the source domain is used to guide the target domain, so as to align the feature distribution of the source domain and the target domain through the network;

[0047] Therefore, the total loss of the network is:

[0048]

[0049] wherein, θ represents the learnable parameters in the network.

[0050] On the other hand, the application also provides a cross-adaptive deep transfer learning system based on bearing life, comprising:

[0051] The acquisition module is used for acquiring the original vibration signal of the rolling bearing, and a sliding window segmentation method is used to obtain a plurality of vibration signal data;

[0052] The curve construction module is used for obtaining the root mean square value by using the plurality of vibration signal data, calculating the gradient of the root mean square value, and constructing the gradient curve and the transformation curve of the root mean square value through normalization;

[0053] The stage calculation module is used for automatically identifying the stage crossing point by using the AGIP algorithm, and then obtaining each different stage of the bearing.

[0054] a preprocessing module, configured to perform data preprocessing on the original vibration signal and perform feature extraction to obtain feature parameters;

[0055] a division module, configured to divide a training set and a test set according to the obtained division identification information in combination with the feature parameters, and mark corresponding labels according to the degradation rates;

[0056] a training module, configured to input data of each stage in the training set into a time series prediction network model to perform training and obtain prediction models of different stages; the time series prediction network model is an RUL prediction model based on multi-feature cross adaptive transfer learning.

[0057] a prediction module, configured to input the test set obtained by division into the time series prediction network trained in the training stage to perform prediction, and output a final rolling bearing remaining life prediction result.

[0058] Further comprising:

[0059] In the stage calculation module, the AGIP algorithm is used to adaptively and automatically identify stage crossing points, comprising:

[0060] (31) Calculate the average value of the first 5% gradient , which is generated by the AGIP algorithm is the baseline gradient threshold , wherein is a function of generating a gradient threshold, and it is worth noting that ;

[0061] (32) Obtain the gradient value of each time point in time sequence and compare it with When it exceeds , record the point as a stage crossing point P1, and the gradient value at this time is , and then the AGIP algorithm is iterated and the threshold is updated to ;

[0062] (33) Find the next stage crossing point P2 from the new , and then iteratively update the threshold, until all stage crossing points P1, P2, P3 are found, and the division of the four stages is realized.

[0063] Further comprising:

[0064] In the division module, the division of the training set and the test set comprises:

[0065] (51) Calculate the degradation rate of different stages, and mark corresponding degradation labels according to the different degradation rates of different degradation stages, comprising:

[0066] The decay rate of each decay stage is calculated:

[0067]

[0068] Wherein: represents the decay rate of the nth decay stage, is the decay degree of the nth decay stage, and three decay stages are defined as 100%-60%, 60%-10% and 10%-0, respectively, and are recorded as: mild decay stage, slow decay stage and accelerated decay stage, is the decay time corresponding to the decay stage;

[0069] Then the label is calculated:

[0070]

[0071] Wherein, represents the label of the ith sample, , and is the cross-stage point obtained by the AGIP algorithm;

[0072] (52) The three decay stages of the source domain and the early stage data of the target domain are used as the training set, and the slow decay stage and the rapid decay stage of the target domain are used as the test set.

[0073] On the basis of the above, the application further provides a computer storage medium, which stores a computer program, and the computer program realizes the above-mentioned method when executed by a computer processor.

[0074] Beneficial effects: (1) The application proposes a new division algorithm that can adaptively divide the multi-stage decay stages under different working conditions. Compared with the traditional division method, the method can update and iterate the division threshold of each stage according to the dynamic data of the bearing, and can accurately divide the multi-stage decay stages of the bearing, and the calculation efficiency is improved by about 60%.

[0075] (2) The application proposes a calculation method of bearing residual life label using the decay rate of different stages of the bearing. Compared with the traditional method, the method is more in line with the use of bearings in actual industrial scenes, and the prediction accuracy is improved by about 5% compared with the traditional method.

[0076] (3) The application proposes a multi-feature cross migration domain self-adaptive method based on early stage of degradation, which does not need to rely on the whole life cycle data like traditional migration learning, and only needs the early degradation data to learn the unobtained middle and late stage degradation features, realizes the effective completion of the migration of two domains in the case of incomplete target domain data prediction.

[0077] (4) The application designs a hierarchical adaptive RUL model based on multi-feature cross adaptive migration learning, and the multi-feature cross migration module in the model effectively migrates the multi-stage degradation features from one working condition to multiple working conditions, and compared with the traditional migration model, the RMSE loss is reduced by about 0.04, and the prediction accuracy is improved by about 5%. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 The overall flowchart of the bearing residual useful life multi-feature cross adaptive deep migration learning method considering the diversity of the degradation process is described in the application.

[0079] Figure 2 The schematic diagram of the multi-stage degradation stage adaptive division and identification method based on the AGIP algorithm is described in the application.

[0080] Figure 3 The schematic diagram of the multi-feature cross migration network structure is described in the application.

[0081] Figure 4 The schematic diagram of the comparison result of the method without the multi-feature migration module is described in the application. DETAILED DESCRIPTION

[0082] In order to more clearly understand the above-mentioned purposes, features and advantages, the application will be further described in detail below in combination with the drawings and specific embodiments. It should be noted that the application can be applied in various ways, and should not be limited to the life prediction of bearings described in the examples. On the contrary, these examples are provided to more thoroughly understand the application and to fully convey the scope of the application to those skilled in the art. In order to facilitate the understanding of the embodiments of the application, the following will be further explained and described with specific examples as examples in combination with the drawings, and each drawing does not constitute a limitation to the embodiments of the application. As shown in the drawings, Figure 1 The bearing residual useful life multi-feature cross adaptive deep migration learning method considering the diversity of the degradation process includes the following steps:

[0083] (1) Collecting the original vibration signal of the rolling bearing

[0084] The original vibration signal data of the rolling bearing is adopted, and a sliding window segmentation method is used to obtain multiple vibration signal data. The original vibration signal data of the rolling bearing includes horizontal vibration signal and vertical vibration signal data of two dimensions. The data is collected by the acceleration sensor installed on the bearing base, and the data of the whole life cycle of the bearing is obtained. The threshold value of ending collection is that the vibration amplitude of the bearing is more than 20g. The sliding window segmentation method is used to equally segment the original vibration signal according to the step length l to realize data preprocessing. The value of l in the experiment is 64.

[0085] (2) Time-frequency domain feature extraction of sliding window data

[0086] The data processing of the present application firstly carries out the feature processing in the time-frequency domain, and then extracts 10 kinds of features including root mean square value, standard deviation, waveform factor and the like. Then, the feature vectors are normalized to reduce the inaccurate problems caused by different dimensions of different variables. The present application adopts Min-Max normalization method to normalize the features to [0, 1]. The specific formula is as follows:

[0087] (3)

[0088] Among them, represents the maximum value of the feature i of all samples, represents the minimum value of the feature i of all samples. represents the i-th feature value of the sample j. After normalization, the network training speed can be improved and the sample can be constrained.

[0089] (3) Obtain degradation index and corresponding gradient change curve

[0090] The time domain feature is used to measure the statistical quantity of the change degree of data, which can well represent the deviation degree from the mean value, but a single index may not be comprehensive enough, therefore, the present application uses PCA technology to reduce the dimension of multiple time domain features to obtain a comprehensive time domain feature. On this basis, multiple data are obtained by using sliding window segmentation, and the difference value between each segment and the first segment is calculated by using MMD to represent the degradation index of the bearing. The calculation formula of MMD is as follows:

[0091] (1)

[0092] In the present application is a group of length 256 sliding window data, and each group of sliding window data obtained later is calculated with the first group to obtain the degradation index. Subsequently, in order to obtain the change trend of the degradation index, since the gradient can well reflect the change trend of the data, the gradient calculation method is adopted to reflect the change trend of the RMS, and the gradient calculation formula is as follows:

[0093] (2)

[0094] Through experiments, it is found that the greater the value of a, the more stable the obtained characteristics, and vice versa. In order to more explicitly represent the obtained gradient data, the value of a in the application is set to 3.

[0095] (4) Using the AGIP algorithm to automatically identify the phase crossing point to realize the division of the multi-stage degradation phase

[0096] The full life cycle of the bearing can be divided into a healthy phase and a degradation phase, and the point of dividing the healthy phase and the degradation phase is the FPT point, at which the data distribution of the bearing will change. According to the application, the degradation phase is further subdivided, and the bearing full life cycle data is divided into four phases. Specifically as follows:

[0097] 1) Healthy phase: In this phase, the bearing runs smoothly and there is no obvious downward trend of the health index.

[0098] 2) Mild degradation phase: The bearing starts to degrade, and the health index shows a slight downward trend.

[0099] 3) Slow degradation phase: Compared with the mild degradation phase, the health index often shows a large number of irregular fluctuations, indicating that the bearing starts to significantly degrade.

[0100] 4) Accelerated degradation phase: In this phase, the health index starts to decline at a faster rate, and the bearing will eventually fail in this phase.

[0101] The AGIP algorithm proposed in the application is based on this principle to divide the multi-stage degradation phase of the bearing. Its main principle is to use the different data distribution principles of different stages to transmit this information through the gradient, thereby realizing the multi-stage division. As shown in the following formula: Figure 2 The specific process of the AGIP algorithm is as follows:

[0102] 1) Calculate the average value of the first 5% gradient The AGIP algorithm generates as the baseline gradient threshold , wherein is a function of generating the gradient threshold, and it is worth noting that ;

[0103] 2) Obtain the gradient value of each time point in time sequence and compare with [ ] When it exceeds [ ], record the point as the stage crossing point P1, and the gradient value at this time should be Then, the AGIP algorithm is started to iteratively update the threshold value to [ ];

[0104] 3) The next stage crossing point P2 is found by the new [ ] After that, the threshold value is iteratively updated in turn until all stage crossing points P1, P2, P3 are found, and the division of the four stages is realized.

[0105] The above is the whole process of the AGIP algorithm to realize adaptive automatic division of bearing multi-stage degradation stages.

[0106] (5) Time-frequency feature data training set and test set division

[0107] The problem to be solved by the application is to complete the alignment of two domains in the case of incomplete target domain, so as to realize the prediction of the RUL of the target domain. Therefore, the training set selected is the multiple degradation stage data of the source domain, and the target domain selects the light degradation data as the training set; the test set selects different degradation stages of the target domain.

[0108] (6) Calculation of degradation rate to make data labels

[0109] The training set in step 5 is expressed in the following form , wherein , and K and k respectively represent the sample quantity and the feature parameter quantity, and the test set is similar. Since the original vibration signal does not contain a label set, it is necessary to make a label set. Based on the actual industrial situation, the application proposes that the degradation degree of the same degradation stage under different working conditions should be the same, but the life cycle is different due to different degradation rates.

[0110] Based on this concept, the application uses the degradation rate to make RUL labels, and the calculation method is as follows. First, the degradation rate of each degradation stage is calculated:

[0111] (4)

[0112] Among them: represents the degradation rate of the nth degradation stage, is the degradation degree of the nth degradation stage (here we stipulate that the three degradation stages are 100%-60%, 60%-10% and 10%-0), is the degradation time corresponding to the degradation stage. Then the label is calculated:

[0113] (5)

[0114] where, represents the label of the i-th sample, 、 and are the cross-stage points obtained by the AGIP algorithm in step 3.

[0115] (7) The training set is sent into the RUL prediction model based on multi-feature cross adaptive transfer learning to train the multi-feature cross transfer network for cross-condition bearing RUL prediction. The multi-feature cross transfer network utilizes the multiple degradation stages of the source domain and the mild degradation stage of the target domain to realize the transfer of multiple incomplete target domains. And through the segmented training method, the remaining life prediction model for different degradation stages is obtained, including the mild degradation prediction model, the slow degradation prediction model and the accelerated degradation prediction model, and the hierarchical adaptive optimization method is used to reduce the model loss of each degradation stage. As shown in Figure 3 , the model mainly includes three parts: feature extraction layer, multi-feature cross transfer layer and network prediction layer.

[0116] The bearing degradation process is divided into four stages in the application, wherein 0, 1, 2 and 3 represent bearing health, mild, slow and acceleration. Therefore, T1 represents the mild stage of the target domain, and in the application, the early target domain data is used as an example to solve the problem of needing full life cycle target domain data in the traditional transfer process. The following is the core transfer idea of the method:

[0117] Firstly, the difference between different degradation stages in the source domain is calculated, and the learned physical degradation mechanism is aligned along the degradation direction on the spatial plane (within the source domain) through the space mapping principle; in addition, for the limited unlabeled data in the target domain, the method aligns the slight degradation features between the source domain and the target domain through the cross adaptive layer, so that the degradation trajectory is aligned in a spatial plane (source domain and target domain), thereby realizing the transfer of degradation information from the source domain to the target domain. On this basis, we assume that the invisible information of the degradation stage has the same physical degradation mechanism between the source domain and the target domain. At the same time, through the above two kinds of space mapping, the invisible degradation information of the incomplete target domain can also be aligned along the degradation direction, realizing the RUL prediction without enough running to failure data.

[0118] The feature extraction layer is a Transformer network structure. Transformer abandons the traditional sequential structure, thereby obtaining better parallelism. In particular, the position encoding enables it to remember the position of each information, which makes it easier to use distributed GPU for parallel training, thereby improving the efficiency of model training. Transformer mainly includes the following structures:

[0119] Input layer: including source domain data embedding layer, target domain data embedding layer and position encoder layer. The embedding layer maps the data into a vector of dimensionality according to the task requirements of network calculation. Since the Transformer does not contain recursion and convolution, the model adds position encoding to represent the absolute position or relative position of information in the sequence, so that the network can use sequence information. The position encoding selects sine and cosine functions, as shown in the following formula:

[0120]

[0121]

[0122] wherein, is the position, i is the dimension. Therefore, each dimension of the position encoding corresponds to a sine signal, and experiments have proved that this function can make the model easily learn the rules of the position.

[0123] Encoder: composed of N stacked encoder layers, where each encoder layer is connected with two sub-layers in turn, one is multi-head attention mechanism (MHA) and the other is position feed-forward network (PWFFN). Each sub-layer adopts residual jump connection and layer normalization, so that the sub-layer parameters can be fully trained and the convergence speed is accelerated. MHA solves the defect that the self-attention mechanism will pay excessive attention to its own position when encoding the current position information. By using h different attention heads to project the data differently, different queries (Q), keys (K) and values (V) are learned, and then h groups of different Q, K and V are merged in parallel. Finally, the pool outputs of h groups are connected together and changed through another learnable linear projection to produce the final output. The calculation formula is as follows:

[0124]

[0125]

[0126] wherein, the learnable parameters include , , and . Based on this design, each head can focus on different parts of the input data.

[0127] The decoder consists of N stacked decoder layers. Each encoder has three sequentially connected sub-layers: a masked multi-head attention layer (masked MHA), MHA, and PWFFN. The masked MHA is identical to the MHA in the encoder block except for the addition of a mask. A mask indicates that certain values ​​are masked so that they do not take effect during parameter updates. In time series prediction, for a time step t, the decoder's output only needs to depend on the outputs before time t. Therefore, applying a masked MHA to the decoder's input yields information about previously predicted outputs, essentially recording information between the current input and the current time step.

[0128] Output layer: After passing through N layers of encoders and decoders, the input features undergo a linear transformation and are activated by an activation function to obtain the final output. In bearing RUL prediction, sigmoid is typically chosen as the activation function for the output layer.

[0129] The multi-feature cross-transfer layer consists of multiple cross-attention layers, with the source domain as the query (Q) and the target domain as the key (K) and value (V). This cross-learning approach enables the network to learn knowledge from the target domain in the source domain, thereby better reducing the distributional differences between the two domains.

[0130] The multi-feature cross-transfer method proposed in this invention utilizes multiple decay stages in the source domain for internal alignment, and then uses the mild decay stage of the target domain to perform feature alignment with the mild decay stage of the source domain, thereby achieving good alignment even for unknown target domain data. In backpropagation, to effectively update network weights and reduce inter-domain distribution differences, a new objective loss function is proposed, which consists of three parts:

[0131] 1) Regression loss : To minimize the prediction error of the training data, the network training parameters can be adaptively adjusted to make the predicted RUL closer to the actual result. This invention patent selects the radial basis function (RBF) as the kernel function and calculates the mean squared error (MSE) using the kernel function. as follows:

[0132]

[0133] in, and represents the actual RUL label and the model-predicted RUL label, respectively, and n represents the batch size during training.

[0134] 2) Distribution difference loss : MMD represents the distribution difference of high-dimensional features between the source domain and the target domain, so that the network can better extract domain-invariant features. As a non-parametric measure, MMD can measure the first-order distribution divergence between two domains. Therefore, the distribution difference loss formula can be expressed as:

[0135]

[0136] wherein, represents the i-th element of the source domain, and respectively represent the size of each batch of the source domain and the target domain, represents the kernel function.

[0137] 3) Knowledge distillation loss L distillation : The distillation loss is used in the cross-attention module to align the feature distribution of the source domain and the target domain. In this module, the source domain and the target domain are represented as the teacher and the student, respectively. The training of the source domain is used to guide the target domain, so that the network can align the feature distribution of the source domain and the target domain.

[0138] There is a knowledge distillation network framework, which includes a student model and a teacher model. During training, the output of the student model is constrained to force the student model to make its prediction distribution close to the feature distribution of the teacher model. Following this principle, in the cross-domain adaptive layer (i.e., the transfer between the source domain and the target domain is represented by the knowledge distillation loss between the outputs of the two domains), the feature distribution of the target domain (student model) should learn the feature distribution of the source domain (teacher model). At the same time, the knowledge distillation loss is used to evaluate the difference between the source domain and the target domain to monitor the level of knowledge learned by the target domain from the source domain.

[0139] Therefore, the total loss of the network is:

[0140]

[0141] wherein, θ represents the learnable parameters in the network.

[0142] The total loss here is the sum of the three losses, the first regression loss is used to train the prediction result, the second MMD loss is within the source domain, which divides the source domain into three stages and hopes to align it in a certain direction through the network and the difference calculation of MMD (it can be imagined that the data of the source domain is like a curve distribution, which can be changed into a straight line through the internal alignment of the stages) Then the distillation loss is to align the source domain and the target domain, which is equivalent to the internal feature learning of MMD, and the distillation loss is external feature learning, and the three losses act together.

[0143] (8) Obtain the prediction result of the test machine

[0144] The test set is fed into the RUL prediction model trained in step 6 using multi-feature cross-adaptive transfer learning to make predictions, and the final rolling bearing remaining life prediction result is output.

[0145] To verify the invention, extensive experiments were conducted on the XJTU-SY dataset under two operating conditions and seven cross-domain cases to predict the RUL of rolling bearings.

[0146] Experimental results are as follows Figure 4 As shown, where, Figure 4 (a) and Figure 4 (b) shows the RUL prediction results for bearings 1_1 and 1_3. Figure 4 (c) and Figure 4 In the figure, (d) represents the RUL prediction result for bearing 2_5. Experiments were conducted using bearing data under different operating conditions to predict the mild degradation stage, slow degradation stage, and accelerated degradation stage. The red curve represents the prediction result of the proposed multi-feature cross-transfer network using the FCA method, while the blue curve represents the prediction result without the FCA method. The results show that the red curve has a better fitting effect across operating conditions. This method not only has good accuracy but also effectively solves the problem that traditional transfer methods cannot address, namely, the inability to achieve complete and good domain alignment when facing an incomplete target domain. It achieves the transfer of multi-level domain-invariant features from one operating condition to cross-operating conditions, significantly improving prediction accuracy.

[0147] On the other hand, the present invention also provides a cross-adaptive deep transfer learning system based on bearing life, comprising:

[0148] The acquisition module is used to acquire the raw vibration signal of the rolling bearing and to obtain multiple vibration signal data using the sliding window segmentation method;

[0149] The curve construction module is used to obtain the root mean square value using the multi-segment vibration signal data, calculate the gradient of the root mean square value, and construct the gradient curve and the transformation curve of the root mean square value through normalization.

[0150] The stage calculation module is used to adaptively and automatically identify the stage transition points of the acquired gradient curve using the AGIP algorithm, thereby obtaining the different stages of the bearing.

[0151] The preprocessing module is used to preprocess the raw vibration signal and extract features to obtain feature parameters.

[0152] The segmentation module is used to divide the training set and the test set by combining the obtained segmentation and recognition information with the feature parameters, and to assign corresponding labels according to the decay rate.

[0153] The training module is configured to input data of each stage in the training set into the time series prediction network to obtain prediction models of different stages through training;

[0154] The prediction module is configured to input the test set obtained through division into the time series prediction network trained in the training stage to perform prediction and output a final rolling bearing residual life prediction result.

[0155] The bearing life-based cross-adaptive deep migration learning system and the bearing life-based cross-adaptive deep migration learning method have the same other technical features, and details are not repeated here.

[0156] Based on the above embodiment, in the embodiment of the present application, a computer readable storage medium is provided, and a computer program is stored on the computer readable storage medium, and the computer program is executed by a processor to implement the bearing life-based cross-adaptive deep migration learning method in any method embodiment.

[0157] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0158] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0159] These computer program instructions can also be stored in a computer readable storage medium that can guide the computer or other programmable data processing apparatus to work in a specific way, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0160] These computer program instructions can also be loaded into computer or other programmable data processing devices, so that a series of operations steps are performed on the computer or other programmable data processing devices to generate computer-implemented processes, thus the instructions executed on the computer or other programmable data processing devices provide processes for implementing the functions specified in the flowchart Figure 1 one flowchart or multiple flowcharts and / or blocks Figure 1 one block or multiple blocks.

[0161] Although preferred embodiments of the application have been described herein, changes and modifications can be suggested to one skilled in the art and are intended to be encompassed within the scope of the application as set forth in the appended claims. It is the intent, therefore, to be limited only as

[0162] It will be apparent to those skilled in the art that various modifications and variations can be made to the embodiments of the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. A cross-adaptive deep transfer learning method based on bearing life, characterized in that, Includes the following steps: (1) Collect the original vibration signal of the rolling bearing and use the sliding window segmentation method to obtain multiple vibration signal data; (2) Perform data preprocessing on the original vibration signal obtained in step (1) and extract features to obtain time-domain and frequency-domain feature parameters; (3) The acquired multidimensional time-domain features are reduced using PCA dimensionality reduction technology to obtain comprehensive time-domain features; (4) The time-domain integrated features in step (3) are divided into multiple segments using a sliding window, and the maximum mean difference (MMD) between each segment and the first segment is calculated to obtain the bearing degradation index and the corresponding gradient curve. (5) The obtained gradient curve is automatically identified by the AGIP algorithm to obtain the different stages of the bearing. The AGIP algorithm includes: firstly, the average value of the first 5% gradient of the gradient curve is used as the benchmark gradient threshold [g0], and then the gradient threshold of each stage is automatically updated to determine the first prediction time point FPT and the stage transition point STP, so as to realize the division of the rolling bearing full life cycle data into multi-level decay stages. (6) The training set and the test set are divided according to the obtained segmentation and recognition information and the feature parameters, and the corresponding labels are applied according to the decay rate. The training set consists of a labeled source domain and an unlabeled early target domain. (7) Input the data from each stage in the training set into the time series prediction network model for training to obtain prediction models for different stages; (8) The divided test set is fed into the time series prediction network trained in step (7) for prediction, and the final rolling bearing remaining life prediction result is output.

2. The cross-adaptive deep transfer learning method based on bearing life as described in claim 1, characterized in that, The specific steps (1) include: the original vibration signal data of the rolling bearing includes two types: horizontal vibration signal data and vertical vibration signal data. These are collected by an acceleration sensor installed on the bearing base to obtain the full life cycle data of the bearing operation. The signal to end the collection is that the bearing vibration amplitude exceeds a certain threshold. The specific method of the sliding window segmentation method is to divide the entire original vibration signal data into equal intervals according to a certain step size l.

3. The cross-adaptive deep transfer learning method based on bearing life according to claim 1, characterized in that, In step (4), the obtained bearing degradation index is obtained by reducing the dimensionality of the obtained multidimensional time domain features to obtain the time domain comprehensive features, and by using sliding window segmentation to calculate the MMD of the first window segment.

4. The cross-adaptive deep transfer learning method based on bearing life according to claim 1, characterized in that, Step (5) involves dividing the training set and the test set, including: (51) Calculate the recession rate at different stages, and assign corresponding recession labels to different recession stages based on the different recession rates, including: Calculate the rate of decline for each stage of the decline: in, This represents the rate of decline in the nth stage of decline. Let n be the degree of decline in the nth stage of decline. Here, the three stages of decline are defined as 100%-60%, 60%-10%, and 10%-0, respectively denoted as: mild decline stage, slow decline stage, and accelerated decline stage. This refers to the duration of the recession phase. Then the labels are calculated: in, The label represents the i-th sample. , and These are the cross-stage points obtained from the AGIP algorithm; (52) Use the three decay stages of the source domain and the early stage data of the target domain as the training set, and the slow decay stage and rapid decay stage of the target domain as the test set.

5. The cross-adaptive deep transfer learning method based on bearing life according to claim 4, characterized in that, In step (6), the data from each stage of the training set are fed into the time series prediction network model. The time series prediction network model includes a feature extraction layer, a multi-feature cross-transfer layer, and a network prediction layer. The feature extraction layer is a Transformer network structure. The decay data of each stage of the source domain and the decay data of the first decay stage of the target domain are input into the Transformer network structure. After passing through the input layer and N layers of encoder and decoder, the input features are linearly transformed and activated by the activation function to obtain the final output features. The multi-feature cross-transfer layer includes multiple cross-attention layers. The source domain is used as the query Q, the target domain is used as the key K and the value V, and a new target loss function is proposed.

6. The cross-adaptive deep transfer learning method based on bearing life according to claim 5, characterized in that, The target loss function includes: Regression loss : To minimize the prediction error of the training data, the network training parameters are adaptively adjusted to make the predicted RUL closer to the actual result. This invention selects the radial basis function (RBF) as the kernel function and calculates the mean square error (MSE) using the mean square error. as follows: in, and These represent the actual RUL labels and the model-predicted RUL labels, respectively, and n represents the batch size during training. Distributional difference loss : The distribution difference loss, representing the difference in the distribution of features between the source and target domains, enables the network to better extract domain-invariant features. As a nonparametric metric, MMD measures the first-order distribution divergence between the two domains. Therefore, after introducing kernel learning, the distribution difference loss formula can be expressed as: in, This represents the i-th element of the source domain. and These represent the size of each batch in the source and destination domains, respectively. Represents the kernel function; Knowledge distillation loss L distillation The distillation loss used in the cross-attention module is mainly used to align the feature distributions of the source and target domains. In this module, the source and target domains are represented as the teacher and student, respectively. The training of the source domain is used to guide the target domain, thereby aligning the feature distributions of the source and target domains through the network. Therefore, the total loss of the network is: Here, θ represents the learnable parameters in the network.

7. A cross-adaptive deep transfer learning system based on bearing life, characterized in that, include: The acquisition module is used to acquire the raw vibration signal of the rolling bearing and to obtain multiple vibration signal data using the sliding window segmentation method. The curve construction module is used to obtain the root mean square value using the multi-segment vibration signal data, calculate the gradient of the root mean square value, and construct the gradient curve and the transformation curve of the root mean square value through normalization. The stage calculation module is used to adaptively and automatically identify stage transition points from the acquired gradient curve using the AGIP algorithm, thereby obtaining the various stages of the bearing; including: (31) Calculate the average value of the first 5% gradients. Then generated by the AGIP algorithm The baseline gradient threshold [ ],in, It is a function that generates the gradient threshold. It is worth noting that... ; (32) Obtain the gradient value at each time point according to the time sequence and [ Compare them, when it exceeds [ Record this point as the stage transition point P1, and the gradient value at this point is... Then the AGIP algorithm iterates and updates the threshold to [ ]; (33) By the new [ The threshold is then iteratively updated to find the next stage crossing point P2, and then the threshold is updated iteratively until all stage crossing points P1, P2, and P3 are found, thus achieving the division of the four stages. The preprocessing module is used to preprocess the raw vibration signal and extract features to obtain feature parameters. The segmentation module is used to divide the training set and the test set by combining the obtained segmentation and recognition information with the feature parameters, and to assign corresponding labels according to the decay rate. The training module is used to feed data from each stage of the training set into the time series prediction network model for training to obtain prediction models for different stages. The prediction module is used to feed the divided test set into the time series prediction network trained in the training phase for prediction, and output the final prediction result of the remaining life of the rolling bearing.

8. The cross-adaptive deep transfer learning system based on bearing life according to claim 7, characterized in that, In the partitioning module, the partitioning of the training set and the test set includes: (51) Calculate the recession rate at different stages, and assign corresponding recession labels to different recession stages based on the different recession rates, including: Calculate the rate of decline for each stage of the decline: in: This represents the rate of decline in the nth stage of decline. Let n be the degree of decline in the nth stage of decline. Here, the three stages of decline are defined as 100%-60%, 60%-10%, and 10%-0, respectively denoted as: mild decline stage, slow decline stage, and accelerated decline stage. This refers to the duration of the recession phase. Then the labels are calculated: in, The label represents the i-th sample. , and These are the cross-stage points obtained from the AGIP algorithm; (52) Use the three decay stages of the source domain and the early stage data of the target domain as the training set, and the slow decay stage and rapid decay stage of the target domain as the test set.

Images