Method for life phase identification of space rolling bearing based on CAHTL

Abstract

The application discloses a kind of space rolling bearing life stage identification method based on CAHTL, utilize CAHTL network to carry out space rolling bearing life stage identification, CAHTL network includes pre-encoding network, heterogeneous transfer learning network and same domain general contrast learning network;Space rolling bearing life stage identification includes the following steps: S1, information compensation;S2, heterogeneous transfer learning network pretraining;S3, construct feature transfer loss function;S4, construct same domain general contrast learning loss function;S5, construct classification loss function;S6, construct the joint loss function of CAHTL, using random gradient descent and momentum asynchronous update mode will joint loss function training to convergence;S7, utilize CAHTL network to the sample to be measured of target domain classification.The application can utilize the known life stage training sample of small, non-equivalent, partial information defect of space rolling bearing historical working condition under current working condition to the sample to be measured and carry out higher accuracy life stage identification.

Description

Technical Field

[0001] This invention belongs to the field of aircraft design technology, and specifically relates to a method for identifying the life stage of a space rolling bearing based on CAHTL. Background Technology

[0002] Space rolling bearings are among the key components of many spacecraft. Due to their unique structure and operating conditions, they largely determine whether a spacecraft can reach its predetermined service life and achieve its intended mission objectives. However, because the operating environment of spacecraft is complex and contains many unknowns, real-time monitoring of bearing data in orbit is expensive and difficult to perform timely replacement or maintenance. Therefore, research on the life stage identification of space rolling bearings in a ground-simulated space environment can establish a basic database of long-life, highly reliable space rolling bearings. This would allow for the selection of the optimal lifespan space rolling bearings (i.e., those identified as being in the normal stage) from a large pool of candidate bearings for installation in spacecraft, thereby helping to avoid catastrophic accidents caused by the failure of critical components in orbital spacecraft.

[0003] Research on the life stage identification of space rolling bearings is still in its early stages, with only a few case studies. For example, Wu Haonian et al. used non-repeating uniform random sampling and a multi-classifier to improve the equilibrium distribution fit, and then used the improved equilibrium distribution fit to identify the life stage of space rolling bearings. Dong Shaojiang et al. used multi-layer noise reduction technology, empirical mode decomposition, and an improved convolutional neural network (ICNN) for fault diagnosis of space rolling bearings. Dong et al. used Tsallis entropy-KPCA (TEK) to obtain the entropy features of the product function signal, and then input the extracted signal entropy features into an optimized fuzzy c-means model (OFCM) after principal component analysis to identify the life stage of space rolling bearings. Chen Renxiang et al. combined linear local tangent space alignment (LLTSA) dimensionality reduction and a nearest neighbor classifier to identify the life stage of space rolling bearings in a simulated ground-based space environment. However, the aforementioned machine learning methods all rely on the premise of consistent data probability distribution. In contrast, the identification of life stages of space rolling bearings in simulated space environments is conducted under varying operating conditions, and machine learning methods based on distribution consistency cannot adapt well to such environments. Secondly, these methods require a large amount of labeled historical operating condition (i.e., source domain) data for training, and demand that the training data for each life stage have an equal proportion and complete sample data. However, due to experimental time constraints and costs, accelerated life tests in ground-based simulated space environments can only obtain a small number of full-life samples of space rolling bearings under some historical operating conditions for training the classification model. Furthermore, the uneven time spans of different life stages of space bearings often result in uneven sample numbers for each life stage. Many more space bearings can only undergo truncated tests (pre-failure tests where the bearing is stopped before reaching the accuracy failure threshold), thus some truncated samples may have continuous missing elements or inconsistent sample lengths with earlier truncated samples. These complex ground-based simulated space environment test conditions expose the limitations of the aforementioned machine learning methods for identifying the life stages of space rolling bearings in ground-based simulated space environments.

[0004] Transfer learning (TL) exhibits excellent domain generalization performance under conditions of inconsistent sample probability distributions, thus providing a novel solution for identifying the life stages of spatial rolling bearings under varying operating conditions. However, traditional transfer learning performs poorly in cases of few or incomplete samples. Contrastive learning (CL), a self-supervised learning paradigm, acquires higher-order information from data pairs by comparing their "similarity" or "difference" to perform self-supervised learning on unlabeled datasets. Its flexibility in defining positive and negative samples allows for flexible redistribution of samples between two domains in cases of few samples, making it a promising solution for few-sample classification problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a CAHTL-based method for identifying the life stage of a space rolling bearing. This method can identify the life stage of the bearing under current conditions using a small number of training samples from known life stages with limited, unequal, and partially incomplete information from historical operating conditions, large differences in sample data distribution, missing elements in some samples, a small number of training samples, and uneven sample quantities at different life stages.

[0006] The objective of this invention is achieved through the following technical solution: a CAHTL-based method for identifying the life stages of space rolling bearings, which utilizes a CAHTL network to identify the life stages of space rolling bearings. The CAHTL network sequentially includes a pre-encoding network, a heterogeneous transfer learning network, and a co-domain generalized contrastive learning network.

[0007] Identifying the life stages of a space rolling bearing includes the following steps:

[0008] S1. Input the original vibration signal sample of the spatial rolling bearing with incomplete information into the pre-encoding network for element filling, i.e., information compensation;

[0009] S2. Heterogeneous transfer learning network pre-training: The original vibration signal samples of the source domain and target domain of the spatial rolling bearing after information compensation are input into the heterogeneous transfer learning network to obtain the corresponding high-dimensional features of the two domains. The feature transfer loss function is constructed on the high-dimensional features of the two domains through the feature distribution difference measurement function. The optimal parameter set of the heterogeneous transfer learning network is obtained by optimizing the feature transfer loss function.

[0010] S3. Input the labeled samples from the source domain and the unlabeled samples from the target domain into the pre-trained heterogeneous transfer learning network to construct the feature transfer loss function of the new heterogeneous transfer learning network.

[0011] S4. Cluster the high-dimensional features of the source domain samples output by the heterogeneous transfer learning network to obtain cluster points containing information of different categories; using these cluster points as ancestors, use the same-domain generalized contrastive learning network as the target domain sample features to distinguish positive and negative samples, so that each target domain sample feature obtains a unique positive sample and λ negative samples corresponding to itself, and the discrimination result is stored in the feature memory; after obtaining the positive and negative samples, construct the same-domain generalized contrastive learning loss function;

[0012] S5. Calculate the high-dimensional features of the test sample in the target domain of the spatial rolling bearing and the similarity between different clusters, and select the class label corresponding to the cluster with the highest similarity as the predicted pseudo-class label of the test sample. At the same time, construct a classification loss function based on the predicted pseudo-class label.

[0013] S6. Combine the feature transfer loss function, the same-domain generalized contrastive learning loss function, and the classification loss function of the heterogeneous transfer learning network obtained in S3 to construct the joint loss function of CAHTL. Use stochastic gradient descent and asynchronous momentum update to train the joint loss function until convergence, obtain the optimal parameter set of the CAHTL network, and complete the training of the CAHTL network.

[0014] S7. Use the trained CAHTL network to classify the test samples in the target domain to complete the identification of the life stage of the space rolling bearing.

[0015] The beneficial effects of this invention are:

[0016] (1) The pre-encoding network constructed in the proposed CAHTL can fill in the missing information of the original vibration signal to solve the problem of information compensation of the missing sample.

[0017] (2) The heterogeneous transfer learning network in CAHTL can transfer a small number of labeled samples from historical working conditions and unlabeled samples from current working conditions (i.e., the test samples) to the common feature space, so as to achieve the purpose of distributing samples with different distributions in the common feature space; the positive and negative samples of the target domain sample features constructed by the source domain cluster points can realize the redistribution of the number of samples in the two domains; comparative learning of positive and negative samples in the two domains can make the test samples have better classification ability.

[0018] (3) CAHTL classifies the test sample by calculating the similarity between the test sample and the cluster point. This classification process does not require parameter learning, thus preventing the problem of large differences in the recognition accuracy of test samples at different life stages when the source domain has class labels and the network overfitting when there are few training samples.

[0019] (4) Using stochastic gradient descent and momentum update to asynchronously update CAHTL parameters can maintain the consistency of sample features and improve the convergence speed of CAHTL.

[0020] (5) The above advantages of CAHTL enable it to use a small number of training samples of known life stages with incomplete information from historical operating conditions of (space) rolling bearings to identify the life stages of the test samples of the current operating conditions of space rolling bearings with high accuracy. Attached Figure Description

[0021] Figure 1 This is a structural framework diagram of CAHTL;

[0022] Figure 2 This is the implementation process of a space rolling bearing life stage identification method based on CAHTL;

[0023] Figure 3 (a) is a space rolling bearing vibration monitoring platform; (b) is a vacuum pump; (c) is the internal experimental frame of the vacuum pump.

[0024] Figure 4 It is a rolling bearing life test bench from the University of Cincinnati.

[0025] Figure 5 This is the reliability evaluation curve for space rolling bearing 1;

[0026] Figure 6 This is the reliability evaluation curve for space rolling bearing 2;

[0027] Figure 7 This is the reliability evaluation curve of rolling bearing No. 1 in the second set of experiments from the University of Cincinnati;

[0028] Figure 8 This is a comparison of the average recognition accuracy of Experiment 1 for the four life stages of the space rolling bearing 1;

[0029] Figure 9 This refers to the lifetime stage identification accuracy when the total number of training samples with class labels in the source domain in Experiment 1 is [missing information].

[0030] Figure 10 This shows the trend of the lifetime stage identification accuracy in Experiment 1 as the defective sample replacement rate changes.

[0031] Figure 11 This is a comparison of the average recognition accuracy of Experiment 2 for the four life stages of the space rolling bearing 1;

[0032] Figure 12 This refers to the lifetime stage identification accuracy when the total number of labeled samples in the source domain in Experiment 2 is reached.

[0033] Figure 13 This shows the trend of the lifetime stage identification accuracy in Experiment 2 as the rate of replacement of incomplete samples changes. Detailed Implementation

[0034] To address the problem of low accuracy in identifying the life stages of space rolling bearings under varying operating conditions, caused by large differences in sample data distribution, missing elements in some samples, limited training samples, and uneven sample numbers at different life stages, this invention combines the advantages of transfer learning and contrastive learning to provide a novel unsupervised transfer learning method—Class-Contrast Cluster-Allocation Heterogeneous Transfer Learning (CAHTL)—for identifying the life stages of space rolling bearings in a ground-simulated space environment.

[0035] CAHTL can perform high-precision life stage identification of the test samples in the target domain (i.e., the current operating condition of the space rolling bearing) when there are few category labels, uneven number of samples of different categories, and some missing sample elements in the source domain (i.e., the historical operating conditions of the space rolling bearing).

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0037] A CAHTL-based method for identifying the life stages of spatial rolling bearings utilizes a CAHTL network for life stage identification. The CAHTL network sequentially comprises a pre-encoding network, a heterogeneous transfer learning network, and a domain-specific generalized contrastive learning network. For example... Figure 1 As shown, in CAHTL, a pre-encoding network is constructed to fill in the missing information in the original vibration signal samples to solve the problem of information compensation for incomplete samples. Through heterogeneous transfer learning, a small number of labeled samples from historical working conditions and unlabeled samples from the current working conditions (i.e., the test samples) are transferred to a common feature space. Positive and negative samples of the target domain sample features are constructed using the source domain cluster points to redistribute the number of samples in the two domains. Then, the positive and negative samples in the two domains are compared and learned to make the test samples have better classification ability. The test sample classification is completed by calculating the similarity between the test sample and the cluster points. This classification process does not require parameter learning, so it can prevent the problem of large differences in the accuracy of sample identification at different life stages under unequal sample conditions and the problem of overfitting the network when there are few labeled training samples. The CAHTL parameters are updated asynchronously using stochastic gradient descent and momentum update to maintain the consistency of sample features and improve the convergence speed of CAHTL.

[0038] like Figure 2 As shown, the identification of the life stage of a space rolling bearing according to the present invention includes the following steps:

[0039] S1. Input the original vibration signal sample of the spatial rolling bearing with incomplete information into the pre-encoding network for element filling, i.e., information compensation;

[0040] Due to limitations in testing cycle and cost, full-life testing is typically performed only on a small number of space rolling bearings, while truncated testing is conducted on a larger number of them. Therefore, some truncated samples may exhibit continuous elemental defects or inconsistent sample lengths compared to earlier truncated samples. Furthermore, the number of vibration acceleration samples for space rolling bearings is very small; discarding incomplete truncated samples would further reduce the number of usable samples. Therefore, a pre-encoding network is constructed to address the problem of compensating for incomplete sample information. The specific implementation method is as follows: Let S be the vibration acceleration sample set of the space rolling bearing source domain. D The target domain vibration acceleration sample set is T D ; respectively taken from S D M labeled samples and samples from T D L unlabeled samples are used as input to the pre-encoding network; assuming the source domain class space Y... s and the target domain class space Y t There are K categories (i.e., K lifetime stages); a single source domain sample and a single target domain sample are denoted as... and Where m∈(1,…,M) and l∈(1,…,L), i and u represent different elements in the source and target domain samples, respectively, n represents the sequence position of the different element, * represents the missing element; N represents the number of elements in each source or target domain sample.

[0041] First, the missing elements in the original vibration acceleration signal sample set of the space rolling bearing are preprocessed, meaning that the missing elements remain empty while the positions of the other normal elements are not changed. The preprocessed space rolling bearing samples are then sorted according to the time domain T... T Divide the region into R equal-length regions, denoted as T' = (T'1, T'2, ..., T''). n ,…,T' R Meanwhile, let the samples of the source and target domains after segmentation be S = {S...} r |r∈1,…,R}、T={T r |r∈1,…,R}, these are the inputs of the pre-encoding network, and the length of each small region is assumed to be d; due to different degrees of missing elements, the amount of sample information held by each small region after segmentation is not different; the sample domain segmentation process is expressed as:

[0042]

[0043] Before each segmented small region sample, a learnable and updatable embedding vector C = {C} needs to be added.r |r∈1,…,R};Since the Transformer encoder needs to classify multiple small region samples into a single discriminant class, the additional embedding vector C is the category information that multiple small region samples are jointly seeking; this embedding process is expressed as:

[0044]

[0045] Next, the incomplete samples are filled with elements to reconstruct the original source and target domain sample sets into the following matrix form:

[0046]

[0047]

[0048] In equation (3), E MN An element is missing at a certain position; in equation (4), E LN An element is missing at a certain position;

[0049] Before filling in elements, it is necessary to determine whether an element is missing at a certain position in the sample. The formula for this determination is as follows:

[0050]

[0051] Next, a method combining small-region sample mean deviation and element column mean compensation is used to fill the missing elements at the * marked positions; let E represent E. MN or E LN The number of normal elements in the column containing the missing position; denoted by U for E. MN or E LN The number of normal elements in the small region sample corresponding to the row containing the missing position, U≤d; using x Su x Tu Representing E respectively MN or E LN The missing elements in the row correspond to normal elements at different positions in the small region sample, u∈(1,…,U); then, the missing samples are filled with elements as follows:

[0052]

[0053] In equation (6),

[0054]

[0055] In equation (7), Let r ∈ (1,2,…,R) represent the average fluctuation deviation of normal and individually filled elements in different small regions of the source and target domains, respectively. These are the mean values ​​of normal and individually filled elements in different small regions of the source and target domains, respectively.

[0056] After element padding, although the small region samples are reassembled in their original positional order and input into the Transformer encoder, the multi-head attention module of the Transformer encoder cannot directly capture the positional information between the small region samples. Therefore, a relative position encoding process is added to the small region samples to supplement their relative positional information. This relative position encoding process is as follows:

[0057]

[0058] In equation (8), These represent positions T' and T' respectively. n ∈(T'1,T'2,…,T' R The encoding vector at position ) The 2i and 2i+1th components, where d is the dimension of the position encoding vector;

[0059] Finally, the location encoding vector set and Adding the sample to the sample after processing such as domain segmentation, element filling, and vector embedding, this process can be represented as:

[0060]

[0061] The result of addition and This refers to a source and target domain input sample of the Transformer encoder.

[0062] S2. Heterogeneous transfer learning network pre-training: The original vibration signal samples of the source domain and target domain of the spatial rolling bearing after information compensation are input into the heterogeneous transfer learning network to obtain the corresponding high-dimensional features of the two domains. The feature transfer loss function is constructed on the high-dimensional features of the two domains through the feature distribution difference measurement function. The optimal parameter set of the heterogeneous transfer learning network is obtained by optimizing the feature transfer loss function.

[0063] The specific implementation method is as follows: The obtained samples from the source domain and target domain are... The input is mapped into the Transformer encoder En(·) to obtain high-dimensional features of the source and target domains. Suppose that the parameter set of the heterogeneous transfer learning network after the g-th training iteration is θ. g The feature mapping process can be expressed as follows:

[0064]

[0065] The Transformer encoder effectively utilizes the correlation between sample elements through its multi-head attention mechanism, resulting in high-dimensional features with good class representation.

[0066] Next, the high-dimensional features of the source and target domains extracted by the Transformer encoder are transferred to the same common feature space through a heterogeneous transfer learning network:

[0067] First, the feature transfer loss function is constructed using the feature distribution difference metric function T(·):

[0068]

[0069] Then, the parameters of the heterogeneous transfer learning network are learned by optimizing the feature transfer loss function to achieve the transfer of high-dimensional features from the source domain samples to the target domain samples (that is, to transfer the high-dimensional features of the source and target domains to the same common feature space), minimizing the difference in their probability distributions. The process of optimizing the feature transfer loss function involves using stochastic gradient descent to adjust the parameters θ of the heterogeneous transfer learning network. g The update process involves θ g The expression for performing an update is as follows:

[0070]

[0071] In the formula, α is the learning rate of the parameters of the heterogeneous transfer learning network;

[0072] Repeat the update process of equations (10) to (12) until the parameters of the heterogeneous transfer learning network are trained to convergence, thus completing the optimization of the feature transfer loss function and obtaining the optimal parameter set of the heterogeneous transfer learning network. This completes the pre-training of the heterogeneous transfer learning network. Since the initial parameters of each subsequent update of the same-domain generalized contrastive learning network are obtained from the global optimal solution obtained by training the heterogeneous transfer learning network, the final updated global optimal solution (i.e., the pre-trained parameter set θ of the heterogeneous transfer learning network) is used. τ When using the initial value as the starting point (i.e., initial value) of a heterogeneous transfer learning network to learn the parameters of the network, only a small number of iterations are needed for the network to converge. This allows it to adapt to new contrastive learning tasks quickly and effectively. Therefore, pre-trained heterogeneous transfer learning networks have good generalization and domain adaptability.

[0073] S3. Input the labeled samples from the source domain and the unlabeled samples from the target domain into the pre-trained heterogeneous transfer learning network to construct the feature transfer loss function of the new heterogeneous transfer learning network.

[0074] The specific implementation method is as follows:

[0075] A small number of labeled samples (i.e., samples with known life stages) from the source domain of space rolling bearings. (Their corresponding class tags are) ) and unlabeled test samples in the target domain (i.e., with unknown lifespan stage) The input is fed into a pre-trained heterogeneous transfer learning network to obtain two-domain high-dimensional features f. S (θ τ ), f T (θ τ ), m∈1,2,…,M, l∈1,2,…,L; that is:

[0076]

[0077] Then, based on the distribution difference metric function of the high-dimensional eigenvalues ​​in the two domains, a new feature transfer loss function for the heterogeneous transfer learning network is constructed as follows:

[0078] J(θ τ )=T(f S (θ τ ),f T (θ τ (14)

[0079] Because strong and complex signal interference exists in both domains of the space rolling bearing samples under simulated space environment, the samples of different categories (i.e., different life stages) in the two domains of the space bearing are very close together or even overlap in the common feature space, which will lead to low classification (i.e. life stage identification) accuracy. At the same time, the limited number of available samples in the source domain makes the network prone to overfitting during training. Therefore, the following uses clustering and constructing a generalized contrastive learning method in the same domain to redistribute the number of high-dimensional features of the two domain samples in the common feature space and classify them.

[0080] S4. Cluster the high-dimensional features of the source domain samples output by the heterogeneous transfer learning network to obtain cluster points containing information of different categories; using these cluster points as ancestors, use the same-domain generalized contrastive learning network as the features of the target domain samples to be discriminated (i.e., the high-dimensional features of the target domain samples to be tested) to distinguish between positive and negative samples, so that each target domain sample feature obtains a unique positive sample and λ negative samples corresponding to itself, and the discrimination result is stored in the feature memory; after obtaining the positive and negative samples, construct the same-domain generalized contrastive learning loss function;

[0081] The specific implementation method is as follows: clustering clusters and constructing homo-domain generalized contrastive learning are used to redistribute the number of high-dimensional features of two-domain samples in the common feature space and classify them.

[0082] The high-dimensional source domain features f output by the heterogeneous transfer learning network S (θ τClustering into K classes yields K cluster centers (i.e., cluster points) that contain category information. This process is expressed as follows:

[0083] [Cn 1 ,Cn 2 ,…,Cn j ,…,Cn K ] = Cn(f S (θ τ (15)

[0084] In the formula, Cn(·) represents the clustering algorithm, and Cn j This represents the j-th cluster center;

[0085] Use the obtained cluster centers Cn j To distinguish between positive and negative samples in the target domain: For each target domain sample feature to be distinguished... In terms of positive samples The features of the sample to be discriminated and the cluster center Cn j The cluster center with the highest similarity after comparing them one by one is:

[0086]

[0087] In the formula, δ(·)∈[-1,1] represents the similarity measurement function; while the negative sample f x - These are the features of all other source and target domain samples except for positive samples; through the above positive and negative sample discrimination layer, each category of sample feature in the target domain has a unique positive sample corresponding to it. We have λ negative samples, where λ ≤ M + L - 1; and we store the discrimination result in the feature memory; thus, the number of features of different categories of samples in the target domain tends to be balanced, which solves the problem of redistribution of the number of samples in the two domains well.

[0088] After obtaining positive and negative samples, construct the following homo-domain general contrastive learning loss function L. * (θ n ):

[0089]

[0090] In the formula, t is a temperature hyperparameter used to control the shape of the sample distribution;

[0091] By optimizing this same-domain general contrastive learning loss function, similar positive sample pairs can be brought closer together, while dissimilar negative samples can be made further apart, which is beneficial for classifying the features of the target domain samples.

[0092] S5. Calculate the high-dimensional features of the test sample in the target domain of the spatial rolling bearing and the similarity between different clusters, and select the class label corresponding to the cluster with the highest similarity as the predicted pseudo-class label of the test sample. At the same time, construct a classification loss function based on the predicted pseudo-class label.

[0093] The specific implementation method is as follows: calculate the high-dimensional features of the test samples in the target domain and the different cluster points Cn. j The similarity is calculated, j∈1,…,K; and the class label corresponding to the cluster point with the highest similarity is selected as the predicted pseudo-class label k of the test sample in the target domain of the spatial rolling bearing. l (k l (representing the pseudo-class label corresponding to the l-th sample in the target domain), the process is expressed as follows:

[0094]

[0095] Subsequently, the pseudo-class label k of the test sample in the target domain is calculated. l The probabilities are as follows:

[0096]

[0097] Next, the sum of the negative logarithms of the probabilities that all test samples in the target domain belong to their corresponding pseudo-class labels is used as the classification loss function, which is derived as follows:

[0098]

[0099] S6. Combine the feature transfer loss function, the same-domain generalized contrastive learning loss function, and the classification loss function of the heterogeneous transfer learning network obtained in S3 to construct the joint loss function of CAHTL. Use stochastic gradient descent and asynchronous momentum update to train the joint loss function until convergence, obtain the optimal parameter set of the CAHTL network, and complete the training of the CAHTL network.

[0100] The specific implementation method is as follows: combining the feature transfer loss function J(θ) of the heterogeneous transfer learning network. τ ), Same-domain generalized contrastive learning loss function L * (θ τ and the classification loss function Φ(θ) τ ζ, the number of joint loss functions for CAHTL * (θ τ The joint loss function is expressed as follows:

[0101]

[0102] In the formula, ρ and ψ are the balance constraint parameters of the joint loss function, which are used to constrain the local optimization behavior of heterogeneous transfer learning networks and homogeneous generalized contrastive learning networks, respectively.

[0103] The joint loss function is trained to convergence using stochastic gradient descent and asynchronous momentum updates to fine-tune the CAHTL parameters.

[0104] Next, we need to work on the joint loss function ζ. * (θ τ ) Optimize and update the network parameters of CAHTL during different batch training processes. Because the parameter update method based on traditional stochastic gradient descent is slow and the cluster centers Cn... j The cluster centers will change to different degrees depending on the training batches, and a rapidly changing cluster center will reduce the consistency of all positive and negative sample features in the positive and negative sample storage (consistency here means that the positive and negative samples of each sample feature remain unchanged after discrimination), making it difficult for CAHTL to converge.

[0105] Therefore, this invention addresses the problem of CAHTL's difficulty in convergence caused by excessive changes in cluster centers during different training batches using momentum updates. After a stochastic gradient calculation is completed, the returned results are accumulated, and the mean of the parameter set after this training is calculated. The parameter set is not immediately updated at this point; instead, this mean is used as the base parameter set for the next training iteration for momentum updates. The momentum update process for this parameter set is as follows:

[0106] First, the CAHTL parameter set θ after the γth training iteration is obtained using stochastic gradient descent. γ Then, gradient accumulation is performed and the mean of the CAHTL parameter set after this training is calculated. Right now:

[0107]

[0108] In the formula, α' is the learning rate of the CAHTL parameter.

[0109] Next, the mean of the parameter set obtained from the last training update is used. The mean of the parameter set obtained from the current update (i.e., the γth update). The current mean of the network parameter set is updated using momentum as follows:

[0110]

[0111] In the formula, η∈[0,1) is the momentum parameter. The larger the value of η, the slower the parameter update. As can be seen from Equation (23), in each iteration of training, the parameter set generated by the previous training will be used to a large extent for parameter set update, and the parameters generated by the current batch training will be applied to a certain extent for update. This ensures that the cluster points will not produce large abrupt changes, while also ensuring that CAHTL can converge.

[0112] Repeating equations (22) and (23) a few times will bring CAHTL to converge, completing the fine-tuning of CAHTL's parameters. At this point, the optimal parameter set θ for CAHTL for this classification task is obtained. # This means completing the training of CAHTL.

[0113] S7. Using the trained CAHTL network, classify the test samples in the target domain to complete the life stage identification of the space rolling bearing. Input the test samples in the target domain of the space rolling bearing into the trained CAHTL to calculate the class label of the test samples in the target domain. (i.e., lifespan stage), that is:

[0114]

[0115] The effectiveness of the present invention will be further verified through experiments below.

[0116] 1. Experimental setup

[0117] The experimental data primarily comes from space rolling bearing life test data collected on a self-built ground-based simulated vacuum environment space rolling bearing vibration monitoring (i.e., accelerated life testing) platform. This platform, such as... Figure 3 As shown, the test setup mainly consists of a vacuum pump, a bearing vibration monitoring test bench, a piezoelectric accelerometer, a dual-integral signal conditioner, an NI data acquisition card, and a computer. Space rolling bearings 1 and 2, both model C36018, were mounted on the vibration monitoring test bench within the vacuum pump. Under vacuum conditions, bearings 1 and 2 were both loaded with a 7 kg axial preload (the axial load would continuously change as bearing wear intensified), and operated at two non-steady speeds of approximately 1000 r / min and approximately 3000 r / min respectively (the speed instability was caused by the non-closed-loop control of the motor) (i.e., non-steady-state conditions C1 and C2 in Table 1). The piezoelectric accelerometer monitored the vibration of these two space rolling bearings, acquiring their vibration acceleration signals every 2 hours at a sampling frequency of 25.6 kHz until both bearings completely stopped and failed. Every 1024 consecutive vibration acceleration data points were taken as a sample, resulting in a total of 744 samples collected for the entire lifespan of both space rolling bearings.

[0118] The experimental data marked as operating condition C3 in Table 1 are from the accelerated life test data of rolling bearings at the University of Cincinnati. Their test bench is as follows: Figure 4As shown, four ZA-2115 double-row roller bearings were mounted on a rotating shaft of a bearing test bench. A motor with a speed of 2000 r / min was used to drive the shaft via a belt, and a radial load of 6000 lbs was applied to the shaft and bearings via a spring mechanism. The sampling frequency was 20 kHz, and the vibration acceleration data of the bearings were collected every 10 minutes. The first 1024 consecutive points of each collected acceleration data were selected as a sample, resulting in a total of 984 samples for the entire lifespan.

[0119] Table 1 Experimental Operating Conditions

[0120]

[0121] Before the life stage identification experiment begins, the life stage of three (spatial) rolling bearings under three operating conditions needs to be divided into life stages. First, 27 life stage features from the time domain, frequency domain, and time-frequency domain are extracted for each sample of these three bearings. Then, the dimensionality of the extracted life stage features is reduced using the Local Linear Embedding (LLE) method to obtain a one-dimensional principal feature corresponding to each sample. The obtained one-dimensional principal feature is then input into a Weibull distribution model to obtain the reliability evaluation curves for these three bearings. The obtained reliability evaluation curves are shown below. Figures 5 to 7 As shown. Figure 5 The reliability evaluation curve for space rolling bearing 1 is shown. Figure 6 The reliability evaluation curve for space rolling bearing 2 is shown. Figure 7 This is the reliability assessment curve for rolling bearing No. 1 in the second set of experiments from the University of Cincinnati. The reliability assessment curve divides the entire lifespan data into four stages: normal state, early degradation stage, intermediate degradation stage, and complete failure stage. Considering the operational safety and reliability of the space rolling bearing, the time point corresponding to the first occurrence of a reliability of 0.9 is used as the dividing point between the normal and early degradation stages; this point is also considered the precision failure threshold point for the space rolling bearing. The time point corresponding to the first occurrence of a reliability of 0.5 is used as the dividing point between the early and intermediate degradation stages. The time point corresponding to the first occurrence of a reliability of 0.1 is used as the dividing point between the intermediate degradation stage and the complete failure stage.

[0122] 2. CAHTL parameter settings

[0123] The CAHTL parameters are set as follows: Number of small region segments in the pre-encoder network R = 8, small region length d = 128, embedding vector C is a one-dimensional random variable, and the dimension of the positional encoding vector is d = 129; Number of Transformer encoders β = 3, number of multi-head attention heads within the encoder is 4, and the input vector dimension of each head is 32; the learning rate α of the heterogeneous transfer learning network parameters is 4e -2 The number of clusters in the homo-domain generalized contrastive learning network is K = 4, and the temperature hyperparameter is t = 0.05; the balance constraint parameters of the joint loss function are ρ = 0.3 and ψ = 0.6; the learning rate of the CAHTL parameter is α' = 4e -2 Since the high-dimensional features of both the source and target domains in the common feature space are encoded by the same network structure (i.e., the pre-encoder network and the Transformer encoder), the consistency of the sample features between the two domains is relatively strong. Therefore, a relatively large momentum parameter, η = 0.92, is used here to prevent large abrupt changes in cluster points while ensuring CAHTL convergence. Subsequent experiments will continue to use the CAHTL parameters set above without any further changes.

[0124] 3. Experiment 1 and comparative analysis

[0125] In this experiment, the full life cycle samples of the space rolling bearing 2 under operating condition C2—including the normal state stage, early degradation stage, intermediate degradation stage, and complete failure stage—were used as source domain samples to identify the life stages of the space bearing 1 under operating condition C1 (i.e., target domain) (i.e., C2→C1). Before the experiment, 100 samples were randomly selected from each life stage of space bearing 2 and space bearing 1 respectively as the total samples of each life stage in the source domain and the total samples of each life stage in the target domain for the experiment. That is, the total number of source domain samples and the total number of target domain samples used for the experiment were both 400.

[0126] (1) In the source domain, samples of the normal state stage (Q / 9), early degradation stage (3Q / 9), intermediate degradation stage (4Q / 9), and complete failure stage (Q / 9) are selected in a ratio of 1:3:4:1 and used as labeled training samples for the source domain. That is, the total number of training samples in the source domain is Q (Q≤225). In the target domain, samples of the four life stages (normal state, early degradation, intermediate degradation, and complete failure) are also randomly selected in a ratio of 1:3:4:1 as test samples for the target domain, and the total number of test samples is also Q. According to the implementation process of the life stage identification method based on CAHTL described in this invention, the samples of the source domain and the target domain are input into the CAHTL with set parameters to identify the life stage of the space rolling bearing 1 (under operating condition C1). After the experiment, the lifetime stage identification accuracy and average identification accuracy obtained by the CAHTL method invented in this invention were compared with four other typical transfer learning methods: Central Moment Matching (CMD), Domain Transfer Multiple Kernel Learning (DTMKL), Transfer Joint Matching (TJM), and Distribution Matching Embedding (DME). To reduce the error caused by randomness, the average of the first 30 experiments for each method was taken as the final experimental result (the same applies below). As the total number of labeled training samples in the source domain decreases, the comparison results of the average lifetime stage identification accuracy are as follows: Figure 8 As shown; when the total number of training samples in the source domain is Q=27, the lifetime stage recognition accuracy and average recognition accuracy of the invented CAHTL and the other four compared methods are as follows: Figure 9 As shown. Figure 9 In the middle, from left to right, each square column represents the recognition accuracy rate for the normal state stage, the early degradation stage, the intermediate degradation stage, the complete failure stage, and the average recognition accuracy rate. Figure 10 , Figure 12 , Figure 13 (The same applies).

[0127] Depend on Figure 8-9The comparison results show that, as the number of labeled training samples in the unequal source domain decreases, although the recognition accuracy and average recognition accuracy of CAHTL and the other four transfer learning methods all decrease to varying degrees due to the imbalance and insufficiency of training, the decrease in the recognition accuracy and average recognition accuracy of CAHTL is the smallest compared to the other four methods. Moreover, the recognition accuracy and average recognition accuracy of CAHTL for the four life stages are always higher than those of the other four methods, avoiding the problem of excessively large differences in recognition accuracy for test samples at different life stages.

[0128] (2) With the total number of training samples in the source domain Q = 27, and keeping the ratio of the number of samples in each lifetime stage of the two domains unchanged at 1:3:4:1, some samples in each lifetime stage of the two domains are replaced with information-deficient samples. The replacement ratio of information-deficient samples in the two domains is as follows: Set to the same value; and the information incompleteness of each incomplete sample is also set to the same value, that is, all are 0. (This information incompleteness is the maximum value when the small region length d = 128). At this time, the trends of CAHTL's accuracy in identifying the life stage of the space rolling bearing 1 and its average accuracy as the replacement ratio θ increases are as follows: Figure 10 As shown.

[0129] Figure 10 This indicates that CAHTL maintains high accuracy and average accuracy in life stage identification up to θ = 6 / 27, with the average accuracy consistently above Γ = 55.6%. Only after θ = 8 / 27 do both accuracy and average accuracy drop to a lower range. In other words, even with a small and unequal number of training samples in the source domain, and with a missing information sample ratio of θ = 6 / 27 and a missing information degree of 12.21% in both domains, CAHTL maintains high accuracy and average accuracy in life stage identification due to its good information compensation ability for missing samples. In contrast, the compared methods CMD, DTMKL, TJM, and DME, limited by their own theories, lack the ability to compensate for missing information samples and therefore cannot perform life stage identification of spatial rolling bearings under conditions of missing information samples.

[0130] 4. Experiment 2 and comparative analysis

[0131] In this experiment, samples from the normal state stage, early degradation stage, intermediate degradation stage, and complete failure stage of rolling bearing No. 1 under operating condition C3 (i.e., full-life samples) were used as source domain samples to identify the life stages (i.e., C3→C1) of the full-life samples of space bearing 1 under operating condition C1 (i.e., target domain samples). Before the experiment, 80 samples were randomly selected from each life stage of rolling bearing No. 1 under operating condition C3 and space bearing 1 under operating condition C1 as the total samples of each life stage of the source domain and the total samples of each life stage of the target domain for the experiment. That is, the total number of source domain samples and the total number of target domain samples used for the experiment were 320 each.

[0132] (1) In the source domain, samples of the normal state stage (2Q / 5), early degradation stage (Q / 10), intermediate degradation stage (Q / 5), and complete failure stage (3Q / 10) are selected in a ratio of 4:1:2:3 to serve as labeled training samples for the source domain, i.e., the total number of all training samples in the source domain is Q (Q≤200). In the target domain, samples of the four lifespan stages—normal state, early degradation, intermediate degradation, and complete failure—are randomly selected in a ratio of 3:4:1:2 to serve as test samples for the target domain, with a total number of test samples also being Q. The accuracy and average accuracy of the CAHTL method invented in this invention for identifying lifespan stages of the test samples in the target domain are compared with four other transfer learning methods. As the total number of labeled training samples in the source domain decreases, the comparison results of the average accuracy of lifespan stage identification are as follows: Figure 11 As shown; when the total number of training samples in the source domain is Q=30, the lifetime stage recognition accuracy and average recognition accuracy of the CAHTL of this invention and the other four compared methods are as follows: Figure 12 As shown.

[0133] Depend on Figure 11-12 The comparison results show that, as the number of training samples with class labels in the unequal source domain decreases, although the accuracy and average accuracy of CAHTL and the other four transfer learning methods for identifying the life stages of the test samples of the four life stages of the space bearing 1 generally decrease to varying degrees, the decrease in the accuracy and average accuracy of CAHTL for identifying the life stages is still the smallest compared to the other four methods; and the accuracy and average accuracy of CAHTL for identifying the four life stages are always higher than those of the other four methods.

[0134] With a total training sample size of Q = 30 in the source domain and maintaining the sample size ratios of 4:1:2:3 and 3:4:1:2 for each lifespan stage in both domains unchanged, some samples in each lifespan stage of both domains are replaced with information-deficient samples. The replacement ratio θ for deficient samples in both domains is set to the same value; and the information deficiencies of each information-deficient sample are also set to the same value, i.e., ε = 12.21%. At this time, the trends of CAHTL's lifespan stage identification accuracy and average identification accuracy for the spatial rolling bearing 1 as the replacement ratio θ increases are as follows: Figure 13 As shown.

[0135] Depend on Figure 13 It can be seen that CAHTL maintains a high accuracy rate in lifecycle stage identification and an average accuracy rate before θ = 6 / 30, for example, the average accuracy rate can be maintained at Γ = 56.7%; only after θ = 8 / 30 does the accuracy rate in lifecycle stage identification and the average accuracy rate drop to a lower range. In other words, CAHTL still has a high accuracy rate in lifecycle stage identification and an average accuracy rate even when the total number of training samples in the source domain is small and uneven, and the proportion of incomplete samples in each lifecycle stage of the two domains reaches θ = 6 / 30 and the information incompleteness of the incomplete samples reaches 12.21%.

[0136] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for identifying the life stage of a space rolling bearing based on CAHTL, characterized in that, The CAHTL network is used to identify the life stage of a space rolling bearing. The CAHTL network is a heterogeneous transfer learning network for class contrast cluster allocation, which includes a pre-encoding network, a heterogeneous transfer learning network, and a homogeneous general contrast learning network. Identifying the life stages of a space rolling bearing includes the following steps: S1. Input the original vibration signal sample of the spatial rolling bearing with incomplete information into the pre-encoding network for element filling, i.e., information compensation; S2. Heterogeneous transfer learning network pre-training: The original vibration signal samples of the source domain and target domain of the spatial rolling bearing after information compensation are input into the heterogeneous transfer learning network to obtain the corresponding high-dimensional features of the two domains. The feature transfer loss function is constructed on the high-dimensional features of the two domains through the feature distribution difference measurement function. The optimal parameter set of the heterogeneous transfer learning network is obtained by optimizing the feature transfer loss function. S3. Input the labeled samples from the source domain and the unlabeled samples from the target domain into the pre-trained heterogeneous transfer learning network to construct the feature transfer loss function of the new heterogeneous transfer learning network. S4. Cluster the high-dimensional features of the source domain samples output by the heterogeneous transfer learning network to obtain clusters containing information about different categories; using these clusters as ancestral points, use a homogeneous generalized contrastive learning network as the target domain sample features to distinguish between positive and negative samples, so that each target domain sample feature obtains a unique positive sample corresponding to itself. The system obtains negative samples and stores the discrimination result in the feature memory. After obtaining positive and negative samples, it constructs a homo-domain generalized contrastive learning loss function. The specific implementation method is as follows: clustering clusters and constructing homo-domain generalized contrastive learning to redistribute the number of high-dimensional features of two-domain samples in the common feature space and class discrimination. High-dimensional source domain features output from heterogeneous transfer learning networks Clustering Class, get The process of identifying cluster centers that contain category information is expressed as follows: （15）； In the formula This represents a clustering algorithm. This represents the j-th cluster center; Use the obtained cluster centers To distinguish between positive and negative samples in the target domain: For each target domain sample feature to be distinguished... In terms of positive samples The features and cluster centers of the sample to be discriminated The cluster center with the highest similarity after comparing them one by one is: （16）； In the formula, This represents the similarity measurement function; while negative samples... This refers to the features of all other source and target domain samples except for positive samples; through the above positive and negative sample discrimination layer, each category of sample features in the target domain has a unique positive sample corresponding to it. and One negative sample, ; After obtaining positive and negative samples, the following in-domain general contrastive learning loss function is constructed. : （17）； In the formula, It is a temperature hyperparameter used to control the shape of the sample distribution; By optimizing this same-domain generalized contrastive learning loss function, similar positive sample pairs can be brought closer together, while dissimilar negative samples can be made further apart. S5. Calculate the high-dimensional features of the test sample in the target domain of the spatial rolling bearing and the similarity between different clusters, and select the class label corresponding to the cluster with the highest similarity as the predicted pseudo-class label of the test sample. At the same time, construct a classification loss function based on the predicted pseudo-class label. S6. Combine the feature transfer loss function, the same-domain generalized contrastive learning loss function, and the classification loss function of the heterogeneous transfer learning network obtained in S3 to construct the joint loss function of CAHTL. Use stochastic gradient descent and asynchronous momentum update to train the joint loss function until convergence, obtain the optimal parameter set of the CAHTL network, and complete the training of the CAHTL network. S7. Use the trained CAHTL network to classify the test samples in the target domain to complete the identification of the life stage of the space rolling bearing.

2. The method for identifying the life stage of a space rolling bearing based on CAHTL according to claim 1, characterized in that, The specific implementation method of step S1 is as follows: Let the vibration acceleration sample set of the spatial rolling bearing source domain be... The target domain vibration acceleration sample set is ; respectively taken from of Individual labeled samples and from of An unlabeled sample is used as the input to the pre-encoding network; assuming the source domain class space... and target domain class space Each exists There are 10 categories; a single source domain sample and a single target domain sample are respectively denoted as _____. and ,in, , , , These represent different elements in the source and target domain samples, respectively. Represents the sequence position of different elements. This indicates that the element is missing; Represents the number of elements in each source or target domain sample; First, the missing elements in the original vibration acceleration signal sample set of the space rolling bearing are preprocessed, meaning that the missing elements remain empty while the positions of the other normal elements are not changed. The preprocessed space rolling bearing samples are then sorted according to the time domain... Cut into A region of equal length, denoted as Meanwhile, let the samples of the source and target domains after segmentation be denoted as follows: , These are the inputs to the pre-coding network, and the length of each small region is assumed to be... Due to varying degrees of element loss, the amount of sample information held in each small region after segmentation differs; this sample domain segmentation process is represented as: （1）； Before each segmented small region sample, a learnable and updatable embedding vector needs to be added. Since the Transformer encoder needs to classify multiple small region samples into a single discriminant class, this additional embedding vector... This refers to the category information jointly sought by multiple small region samples; the embedding process is represented as: （2）； Next, the incomplete samples are filled with elements to reconstruct the original source and target domain sample sets into the following matrix form: （3）； （4）； In equation (3), represent An element is missing at a certain position; in equation (4), represent An element is missing at a certain position; Before filling in elements, it is necessary to determine whether an element is missing at a certain position in the sample. The formula for this determination is as follows: （5）； Next, a method combining small-region sample mean deviation and element column mean compensation is used to... Fill the missing elements in the marked empty positions; use express or The number of normal elements in the column containing the missing positions; using express or The number of normal elements in the small region sample corresponding to the row containing the missing position. ;use , Represent or The empty spaces in the row correspond to normal elements at different positions in a small region of the sample. Then, the incomplete samples are filled with elements as follows: （6）； In equation (6), （7）； In equation (7), , These represent the average fluctuation deviations of normal and individually filled elements in different small regions of the source and target domains, respectively. ; , These are the mean values ​​of normal and individually filled elements in different small regions of the source and target domains, respectively. After element padding, a relative position encoding process is added to the small region samples to supplement the relative position information of the small region samples. This relative position encoding process is as follows: （8）； In equation (8), , Representing positions respectively Location encoding vector The , One portion, It is the dimension of the positional encoding vector; Finally, the location encoding vector set and Adding the sample to the sample after processing such as domain segmentation, element filling, and vector embedding, this process can be represented as: （9）； The result of addition and This refers to a source and target domain input sample of the Transformer encoder.

3. The method for identifying the life stage of a space rolling bearing based on CAHTL according to claim 2, characterized in that, The specific implementation method of step S2 is as follows: The obtained samples from the source domain and target domain are... , Input to Transformer encoder The mapping yields high-dimensional features of the source and target domains. , Assume the first The parameter set of the heterogeneous transfer learning network after training is: The feature mapping process can be expressed as follows: （10）； Next, the high-dimensional features of the source and target domains extracted by the Transformer encoder are transferred to the same common feature space through a heterogeneous transfer learning network: First, we use the feature distribution difference measurement function. Construct the following feature transfer loss function: （11）； Then, the parameters of the heterogeneous transfer learning network are learned by optimizing the feature transfer loss function to achieve the transfer of high-dimensional features from the source domain samples to the target domain samples, minimizing the difference in probability distribution between them. The process of optimizing the feature transfer loss function involves using stochastic gradient descent to adjust the parameters of the heterogeneous transfer learning network. The process of updating, for The expression for performing an update is as follows: （12）； In the formula, The learning rate is used for the parameters of the heterogeneous transfer learning network. Repeat the update process of equations (10) to (12) until the parameters of the heterogeneous transfer learning network are trained to convergence, complete the optimization of the feature transfer loss function, obtain the optimal parameter set of the heterogeneous transfer learning network, and complete the pre-training of the heterogeneous transfer learning network.

4. The method for identifying the life stage of a space rolling bearing based on CAHTL according to claim 3, characterized in that, The specific implementation method of step S3 is as follows: A small number of labeled samples from the space rolling bearing source region Unlabeled test samples in the target domain The input is fed into a pre-trained heterogeneous transfer learning network to obtain high-dimensional features in two domains. , , , ;Right now: （13）； Then, based on the distribution difference metric function of the high-dimensional eigenvalues ​​in the two domains, a new feature transfer loss function for the heterogeneous transfer learning network is constructed as follows: （14）。 5. The method for identifying the life stage of a space rolling bearing based on CAHTL according to claim 1, characterized in that, The specific implementation method of step S5 is as follows: calculate the high-dimensional features and different cluster points of the test samples in the target domain. similarity, The class label corresponding to the cluster point with the highest similarity is selected as the predicted pseudo-class label for the test sample in the target domain of the spatial rolling bearing. The process is expressed as follows: （18）； Subsequently, it was calculated that the target domain sample belonged to the pseudo-class label. The probabilities are as follows: （19）； Next, the sum of the negative logarithms of the probabilities that all test samples in the target domain belong to their corresponding pseudo-class labels is used as the classification loss function, which is derived as follows: （20）。 6. The method for identifying the life stage of a space rolling bearing based on CAHTL according to claim 5, characterized in that, The specific implementation method of step S6 is as follows: combining the feature transfer loss function of the heterogeneous transfer learning network. Same-domain generalized contrastive learning loss function and classification loss function The number of joint loss functions of CAHTL The joint loss function is expressed as follows: （21）； In the formula, , These are the balance constraint parameters for the joint loss function, used to constrain the local optimization behavior of heterogeneous transfer learning networks and homogeneous generalized contrastive learning networks, respectively. The joint loss function is trained to convergence using stochastic gradient descent and asynchronous momentum updates to fine-tune the CAHTL parameters.

Images