A mechanical fault classification and identification method for variable working conditions and partial migration conditions

Abstract

The application belongs to the field of computer fault classification and recognition, and discloses a mechanical fault classification and recognition method under variable working conditions and partial migration conditions. The method is used to solve the problem of model performance decline under the condition that the working condition changes and the label distribution is inconsistent. An improved local maximum mean difference is used as an index for measuring the distribution difference between the source domain and the target domain. While aligning the distribution of the same category related subdomains in the source domain and the target domain, the problem of reduced intelligent fault diagnosis model fault recognition accuracy caused by inconsistent labels is overcome, so as to better realize cross-domain fault diagnosis. On the basis of the existing local subdomain adaptive method, the accuracy of fault classification and recognition of the fault diagnosis model is further improved, and finally the fault classification and recognition of different fault positions of engineering equipment under different working conditions are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer classification and identification of industrial faults, and particularly relates to a mechanical fault classification and identification method under variable working conditions and partial migration conditions. BACKGROUND

[0002] Fault diagnosis is of great significance to improve the production efficiency of complex mechanical systems and reduce the accident rate. Nowadays, with the rapid rise and vigorous development of industrial internet and internet of things technology, the multi-source sensor network of mechanical equipment is densely arranged, and the interaction of monitoring data is increasing day by day, which makes mechanical fault diagnosis enter the era of "big data". Deep learning has achieved great success in many knowledge engineering fields, which greatly promotes the development of intelligent fault diagnosis. However, when the working condition of mechanical equipment changes, that is, the data used for training the model and the data collected by the model during actual diagnosis have distribution differences, the diagnostic performance of the model may be greatly reduced or even completely failed. We should study and use advanced new theories and new technologies to solve this problem, in order to promote the successful application of intelligent fault diagnosis in new scenarios with data distribution differences.

[0003] Deep learning can automatically learn fault feature information that is helpful for accurate fault identification from a large amount of historical data, so it can be applied to the field of mechanical fault diagnosis. However, many deep learning methods are subject to the same assumption: training data and test data must be obtained from the same feature space and need to have the same distribution. However, in many real application scenarios, this assumption may not hold. When the distribution of data changes, most models need to be rebuilt using newly collected data, which requires extremely expensive cost, or even is impossible. Transfer learning relaxes the assumption that test data must be independent and identically distributed with training data, and transfer learning is more and more applied to fault diagnosis of industrial equipment because it can use known knowledge to solve different but related field problems.

[0004] Transfer learning is divided into supervised transfer learning and unsupervised transfer learning, and the actual problem faced in the field of intelligent fault diagnosis is the unsupervised transfer learning problem in which the target domain label is not available. The traditional unsupervised domain adaptation method is mainly a global domain adaptation method based on distance measurement, which aligns the feature distributions of the source domain data and the target domain data by minimizing the maximum mean discrepancy (MMD), and this method only considers the global feature alignment, and some irrelevant data may become indistinguishable with the feature mapping. The deep sub-domain adaptation network aligns the feature distributions of each sub-domain (a sub-domain contains samples in the same class) by minimizing the local maximum mean discrepancy (LMMD), thereby improving the generalization performance of the transfer learning model. However, when facing the actual fault diagnosis problem of engineering, there is often a case that the fault category of the test data is less than that of the training data, that is, the label inconsistency problem, which is manifested as the number of sub-domains in the target domain being less than that in the source domain. In this case, the transfer is called partial transfer, and the use of the sub-domain alignment method will lead to incorrect sub-domain alignment and reduce the generalization ability of the model. Therefore, a new sub-domain alignment technology is needed to solve the partial transfer problem caused by the label distribution inconsistency under the condition of changing working conditions, which reduces the performance of the model. SUMMARY

[0005] The application provides a mechanical fault classification and identification method under variable working conditions and partial transfer conditions, which can perform transfer learning under variable working conditions and partial transfer conditions, uses an improved local maximum mean discrepancy as an index for measuring the distribution difference between each sub-domain of the source domain and the target domain, aligns the distributions of the relevant sub-domains of the same category in the source domain and the target domain, overcomes the problem of reduced fault identification accuracy of the intelligent fault diagnosis model caused by label inconsistency, and better realizes cross-domain fault diagnosis. On the basis of the existing local sub-domain adaptation method, the accuracy of fault classification and identification of the fault diagnosis model is further improved, and finally the fault classification and identification of different fault positions of engineering equipment under different working conditions are realized. This provides a new idea for subsequent transfer learning algorithm research and engineering application.

[0006] The technical scheme of the application is a mechanical fault classification and identification method under variable working conditions and partial transfer conditions, and the specific steps are as follows:

[0007] Step 1: A single-axis acceleration sensor collects vibration signals of known faults as source domain data, and labels are marked according to fault categories; working conditions and vibration signals of unknown faults are taken as target domain data;

[0008] Step 2: The vibration signals in the source domain data and the target domain data are collected in the form of a sliding window, and data samples are obtained after preprocessing;

[0009] Step three: build a multi-kernel cascaded residual network model, input data samples to the multi-kernel cascaded residual network model, extract high-dimensional features of the source domain data and the target domain data, and predict the target domain data through the trained classifier to obtain pseudo labels;

[0010] Step four: calculate the corrected local maximum mean difference between the source domain data and the target domain data through the extracted high-dimensional features of the source domain data and the target domain data, the source domain real label annotation and the target domain pseudo label, as a domain adaptation loss term, and measure the distribution difference between the source domain and the target domain subdomains;

[0011] Step five: the domain adaptation loss term and the classification loss term are jointly optimized as a target function, and after a specified number of iterations of training, a trained multi-kernel cascaded residual network model is obtained as a fault diagnosis model of the engineering machinery equipment under variable working conditions, which is used for fault diagnosis.

[0012] The target domain test sample set is input into the trained fault diagnosis model, the fault diagnosis result of the test sample is compared with its real label, and the accuracy of the fault diagnosis model for engineering machinery equipment fault diagnosis is obtained to measure the performance of the fault diagnosis model.

[0013] The multi-kernel cascaded residual network model (Multi-kernel cascaded ResNet, MK ResNet) is improved based on the existing residual network model ResNet, and three 3x3 convolution cascades are used instead of 7x7 convolution. The path of the 1x1 convolution in the residual block of the residual network model ResNet is improved as follows: change the step of the convolution layer to 1, and add an average pooling layer with a kernel of 2 and a step of 2 before it to replace the downsampling work.

[0014] The corrected local maximum mean difference is improved based on the local maximum mean difference in the subdomain adaptation method;

[0015] Through the dynamic feedback mechanism, an inverse weight coefficient determined by a proportion coefficient r is added The calculation formula of the corrected local maximum mean difference ReLMMD is:

[0016]

[0017] Where p represents the distribution of the source domain data samples, q represents the distribution of the target domain data samples; c represents the class, and C represents the total number of classes; represents the i-th source domain data sample, represents the j-th target domain data sample; n s represents the number of source domain data samples, n tN represents the number of target domain data samples; G represents a feature extractor, and k is a kernel function; is a weight coefficient, is a feedback weight coefficient for the target domain data samples; the loss function after adding the feedback coefficient is called Rectify Local Max Mean Distance (RLMMD), and a multi-kernel cascaded residual network using RLMMD as a domain adaptation loss term is called a subdomain adaptation rectified network (SARN).

[0018] r is a proportion coefficient, representing the proportion of the pseudo labels of each class of samples in the target domain in the total samples in the target domain in Δt iterations; r and The calculation formula of is as follows:

[0019]

[0020]

[0021]

[0022] wherein k is the number of iterations of the multi-kernel cascaded residual network model; represents the pseudo label of the jth target domain sample in the kth iteration, obtained after inputting the multi-kernel cascaded residual network, 1 when not defined, and α, β, and ε are hyperparameters, all of which are between 0 and 1.

[0023] In the process of calculating , every 3Δt iterations form a cycle, and the feedback weight coefficient is not effective during the 0th-tth iteration. When k=t, the cycle is entered, and during the tth-t+Δtth iteration, the pseudo label generated through Δt iterations is used to calculate the proportion coefficient r. During the t+Δth-t+2Δth iteration, the feedback weight coefficient is effective, and the feedback weight coefficient r is determined by the proportion coefficient calculated before: r c represents the proportion of the cth class of samples predicted by the model in the total samples in the target domain, and the theoretical value is the reciprocal of the number of sample categories. When r c is greater than the theoretical value by 10%-20%, it indicates that other class samples are incorrectly attributed to this class, and should decrease to attenuate the weight coefficient of this class; when r c is too small, less than 40% of the theoretical value, it indicates that the samples of this class are incorrectly attributed to other classes or do not exist, and should increase If the weight coefficient of the class is increased, the proportion of the class is obviously increased in the following iteration, that is, the increase is more than 30%, the class exists, the sample of the class is wrongly attributed to other classes in the previous iteration, and in the iteration process from t+2DeltaT to t+3DeltaT, If the weight coefficient of the class is not increased, the proportion of the class is not obviously increased, which indicates that the class does not exist in the target domain, 0 is set in the iteration process from t+2DeltaT to t+3DeltaT, 0 is set, and the cycle or stop is performed according to the rule.

[0024] The target function is

[0025]

[0026] Lambda is a domain adaptation weight coefficient, and J is a classification loss term, and the expression is

[0027]

[0028] Y i Indicates the label of the sample i, z i Indicates the probability that the sample is correctly classified.

[0029] For the rolling bearing data set of Case Western Reserve University, t is 30, DeltaT is 10, alpha is 0.3, beta is 0.3, and epsilon is 0.003; for the gear box data set of Southeast University, t is 30, 60, and other parameters are the same as above.

[0030] The beneficial effects of the present application are as follows: the present application aims at the partial migration problem caused by inconsistent label distribution and model performance decline under the condition of changing working conditions, and proposes a mechanical fault classification and identification method under the condition of variable working conditions and partial migration. The existing residual network model ResNet is improved by decomposing the large convolution kernel of the first convolution layer into multiple small convolution kernels and using average pooling in the residual block to replace 1*1 convolution for down sampling; the local maximum mean difference is improved by combining the feedback mechanism, and the problem of reducing the network prediction accuracy of the model under the condition of partial migration is solved. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 It is a whole process block diagram of the sub-domain adaptation correction network mechanical fault classification and identification;

[0032] Figure 2 It is a network structure of the existing residual network;

[0033] Figure 3 It is a network structure diagram of the multi-core cascaded residual network;

[0034] Figure 4 It is a schematic diagram of domain adaptation and sub-domain adaptation;

[0035] Fig. 5(a) is a fault identification result of the atomic domain adaptive network;

[0036] Fig. 5(b) is a fault identification result of the improved sub-domain adaptive correction network. DETAILED DESCRIPTION

[0037] In order to enable the person skilled in the art to better understand the technical solutions and advantages of the technical solutions of the present application, the embodiments of the present application are further described in detail in combination with the drawings. The overall flowchart of the sub-domain adaptive correction network mechanical fault classification and identification under variable working conditions and partial migration conditions is shown in Fig. 1. Figure 1

[0038] Step one: the single-axis acceleration sensor collects the vibration signals of known faults to make a source domain data set, and the vibration signals of unknown working conditions and faults are made into a target domain data set. Specifically, in the experiment, two data sets were used to verify the effectiveness of the method, which were the rolling bearing data set of Case Western Reserve University and the gearbox data set of Southeast University.

[0039] The rolling bearing fault test bench of Case Western Reserve University mainly consists of three parts: motor, torque sensor / decoder and power test meter. The bearing to be detected is connected to the rotating shaft of the motor, the driving end bearing is model SKF6205, and the fan end bearing is model SKF6203. The bearing faults are point corrosion faults generated by an electric spark machine on the inner ring, rolling body and outer ring of the bearing. The damage diameter of each fault position includes 0.1778mm, 0.3556mm and 0.5334mm, plus the normal state, which can be divided into 10 categories, corresponding to 0-9 ten labels. The experimental data is collected under four different loads of 1hp, 2hp, 3hp and 0hp (1hp≈0.75kW), and these data sets are named A, B, C and D respectively.

[0040] The gearbox data set obtained by the transmission dynamics simulator of the gearbox fault simulation test bench of Southeast University contains bearing and gear fault data. The bearing fault types include three single faults and one composite fault: crack in the ball, crack in the inner ring, crack in the outer ring, and crack in the inner and outer rings. The gear fault types include crack in the gear, crack in the tooth root, wear on the tooth surface and broken tooth on the gear, plus the healthy state, which can be divided into 10 categories. The experimental data is collected under two working conditions (rotating speed-load), which are 20HZ-0V (0Nm) and 30HZ-2V (7.32Nm), named E and F.

[0041] ​Step two: the collected vibration signal is intercepted by a sliding window, and the data sample is obtained after preprocessing. Specifically, the entire data set is traversed by a sliding window, and there is no overlap between samples. Then the data is preprocessed by fast Fourier transform algorithm to obtain the source domain data sample set and the target domain data sample set. In the experiment, the data set under one working condition is used as the source domain data, and the data set under another working condition is used as the target domain data. The data sample of one fault state is randomly removed from the target domain data sample set. The unlabeled target domain data and the source domain data are used as the training data set of the network model together. After the trained model is obtained, the target domain data not participating in the training is used for model testing to obtain the accuracy of the model fault diagnosis, which is used to measure the performance of the model.

[0042] Step three: a multi-core cascade residual network model is constructed to extract high-dimensional features of the source domain data and the target domain data, and a pseudo label is obtained by using the trained classifier to predict the target domain data. Specifically, the residual network solves the problem of network degradation with the increase of network layers by adding an identity mapping between the network layers of the convolutional neural network. The residual block is the basic unit of the residual network, the input is x, and the output is H(x). The residual refers to the difference between the output value H(x) and the input value identity mapping x, that is:

[0043] f(x) = H(x) - x

[0044] Where f(x) is the learning object of the residual network. If the residual block needs to complete the data downsampling work, the identity mapping in it is replaced by a convolution layer, which is called a downsampling block. Figure 2 In the formula, ① is a simplified representation of the residual network model structure, ② is the specific structure of the first convolution module in ①, ③ is the specific structure of the residual block, and ④ is the specific structure of the downsampling block.

[0045] The first convolution layer of the convolution network model usually uses a large convolution kernel, but the large size of the convolution kernel will cause the parameter quantity and the calculation quantity of the model to increase sharply. In order to solve the problem of calculation overhead, the first layer convolution module of the residual network model is improved, and the improved structure is as follows: Figure 3As shown in Figure ②, in the improved multi-kernel cascaded residual network, the 7×7 convolution is replaced by a cascade of three 3×3 convolutions, significantly reducing the number of parameters and computational cost of the model, while increasing the network depth to increase network capacity and complexity. On the other hand, a convolutional kernel with a stride of 2 and a size of 1 cannot traverse all the information on the input feature map during its movement, resulting in a decrease in model performance due to the loss of half or more of the input information during training. Therefore, the path containing the 1×1 convolution in the downsampling block of the residual network model is improved as follows: the stride of this convolutional layer is changed to 1, and an average pooling layer with a size of 2 and a stride of 2 is added before it to take over the downsampling work. This improvement allows the path to traverse all the information on the input feature map while having a small impact on computational cost. The improved downsampling block is shown in Figure ②. Figure 3 As shown in ④.

[0046] Step 4: Using the high-dimensional features of the source and target domains extracted by the network, the true labels of the source domain, and the pseudo labels of the target domain, calculate the local maximum mean difference between the source and target domains. This difference serves as the domain adaptation loss term, measuring the distributional differences between the various subdomains of the source and target domains. Specifically, in transfer learning, given a set of n... s The source domain of a labeled sample and a containing n t The target domain of an unlabeled sample in, It is the same as the i-th source domain sample Corresponding tags, This represents the j-th unlabeled target domain sample. Traditional domain adaptation methods perform a global domain transformation on the source and target domains to make the feature distributions of the transformed source and target domains as similar as possible, extracting globally applicable domain-invariant features. The maximum mean difference, as a metric for measuring the global distribution difference between the source and target domains, is widely used in transfer learning methods, and its calculation formula is:

[0047]

[0048] Where H represents the regenerating kernel Hilbert space, This represents mapping the original sample data to the feature map of H.

[0049] The method of aligning the feature distribution of source domain data and target domain data by minimizing MMD only considers the global feature alignment, and ignores the alignment between each sub-domain in the feature space, that is, the global source domain and target domain distribution is aligned without considering the relationship between each sub-domain in the two domains, which may lose the fine-grained information of each class. As shown in FIG. 5, after global domain adaptation, the overall distribution of the source domain and the target domain tends to be consistent, but it can be seen that the sub-domain distribution on the source domain is discrete, and the corresponding sub-domain on the target domain is not well aligned. The deep sub-domain adaptation network aligns the feature distribution of each sub-domain by minimizing LMMD, which not only matches the local distribution of the source domain and target domain feature samples, but also realizes the matching of the global distribution, thereby improving the generalization performance of the transfer learning model. The calculation formula of LMMD is:

[0050]

[0051] where p and q represent the distributions of the source domain and the target domain samples, respectively, and p≠q; c represents a class, and C represents the total number of classes; n s , n t represent the number of samples of the source domain and the target domain, respectively. G represents a feature extractor, and k is a kernel function. is a weight coefficient, and its calculation formula is as follows:

[0052]

[0053] where y ic is the c-th element of vector y i .

[0054] When considering actual engineering application scenarios, fault data is difficult to contain all fault types, which leads to the situation that the transfer learning under different working conditions is not consistent in labels, that is, the label types of the target domain dataset are less than those of the source domain. At this time, the network will still predict a part of the target domain samples as missing labels, and this misclassification will seriously reduce the accuracy of fault prediction. FIG. 5(a) is the classification result obtained by the network model trained using the traditional LMMD when tested. In order to solve the problem of reduced prediction accuracy of the network under the partial transfer condition, the sub-domain adaptation correction network improves the local maximum mean difference LMMD by adding a feedback weight coefficient The calculation formula of the modified local maximum mean difference ReLMMD is:

[0055]

[0056] wherein, Here are the feedback weighting coefficients for the target domain. r is a proportionality coefficient, representing the proportion of pseudo-labels for each type of sample in the target domain to the total samples in the target domain during Δt iterations. r and The calculation formula is:

[0057]

[0058]

[0059]

[0060] Where k is the number of iterations of the network model; This indicates that the j-th target domain sample is at the k-th iteration. The pseudo-tags obtained after inputting into the network, When undefined, take 1, α, β, ε as hyperparameters, α, β, ε∈(0,1).

[0061] In calculation During the process, each iteration consists of 3Δt iterations. In the first 0 to t iterations, the feedback weight coefficient is ineffective. When k = t, the loop begins. In the t to t+Δt iterations, the proportional coefficient r is calculated using the pseudo-labels generated in the Δt iterations. In the t+Δt to t+2Δt iterations, the feedback weight coefficient... Effective, the feedback weight coefficient is determined by the previously calculated proportional coefficient: r c This represents the proportion of the c-th class samples predicted by the model to the total number of samples in the target domain. Its theoretical value is the reciprocal of the number of sample classes. c If the percentage is 10% to 20% higher than the theoretical value, it indicates that samples from other classes have been incorrectly classified into this class, and the value should be reduced. To attenuate such weighting coefficients; when r c If the value is too small, less than 40% of the theoretical value, it indicates that samples of this type have been incorrectly classified into other classes or do not exist. Increase the value first. To increase the weighting coefficient of this class, in subsequent iterations, if the proportion of this class increases significantly, i.e., the increase exceeds 30%, then this class exists. Samples of this class were incorrectly classified into other classes in previous iterations. During the iterations from t+2Δt to t+3Δt, Set to 1; if the proportion of this class does not increase significantly, it indicates that this class does not exist in the target domain. During the (t+2Δt) to (t+3Δt)th iteration, Set to 0, and then repeat or stop according to this pattern. For the CWRU dataset, t is 30, Δt is 10, α is 0.3, β is 0.3, and ε is 0.003; for SEU, t is 30 and 60, and other parameters are the same as above. The fault identification results obtained by the network model trained by ReLMMD during testing are shown in Figure 5(b).

[0062] Step five: the domain adaptation loss term and the cross-entropy loss term of the residual network are jointly optimized as the objective function, and after a specified number of iterations of training, the fault diagnosis model of the engineering mechanical equipment under the variable working condition is obtained. Specifically, the domain adaptation loss term and the cross-entropy loss term of the residual network are jointly optimized as the objective function, and the expression of the objective function is:

[0063]

[0064] Where λ is the domain adaptation weight coefficient, which is a trade-off parameter, and J is the cross-entropy loss function, and its expression is:

[0065]

[0066] y i represents the label of sample i, z i represents the probability that the sample is correctly classified.

[0067] The method adopts the ReLU function as the activation function, and for the rolling bearing data set of Case Western Reserve University, the total number of iterations is set to 50 and the learning rate is set to 0.001; for the gearbox data set of Southeast University, the total number of iterations is set to 100 and the learning rate is set to 0.0001. The domain adaptation weight coefficient is set to 0.5, and each experiment is repeated 3 times, and the average value is taken as the final result. The hardware environment used in the experiment is: the CPU model is Intel(R) Core(TM) i7-11800H; the memory is 16GB; the GPU model is NVIDIA GeForce RTX 3050Ti.

[0068] The target domain test set is input into the trained fault diagnosis model to obtain the accuracy of the fault diagnosis model for engineering mechanical equipment fault diagnosis, so as to measure the performance of the fault diagnosis model. Specifically, the ablation experiment is carried out with the residual network Resnet as the network skeleton. Source Only indicates that only the source domain data is used to train the residual network. The CWRU data set is used for experiment, and the results are shown in Table 1.

[0069] Table 1 Experimental results on CWRU data set (%)

[0070]

[0071] Table 1 (continued)

[0072]

[0073] From the experimental results, the fault diagnosis accuracy of the network model using MMD is 92.82% on average, and the fault diagnosis accuracy of the network model using LMMD is 95.94% on average. By improving MMD to LMMD, the fault diagnosis accuracy of the network model is improved, which verifies the effectiveness of LMMD. After improving the network structure, the fault diagnosis accuracy of the model is 98.73% on average; after using ReLMMD, the fault diagnosis accuracy of the model is 100% on average. Even without any improvement, the network model Resnet performs well in the fault classification task of the CWRU data set, which is due to the large correlation between the data of different working conditions in the CWRU data set. However, the improved Resnet network structure and the improved LMMD still further improve the performance of the fault diagnosis model.

[0074] In order to further verify the method constructed, the gear box data set of Southeast University is used for test, and the result is shown in Table 2.

[0075] Table 2 SEU data set (%)

[0076]

[0077] From the test results, since the gear box data set has large differences between different working conditions and obvious differences in data distribution, neither MMD nor LMMD can make the network model have satisfactory diagnostic effect on part of the transferred variable working condition. The MKResnet network structure and the ReLMMD both further improve the performance of the fault diagnosis model. The improvement of the network structure improves the fault diagnosis accuracy of the network model by 12.99%, and the improvement of the LMMD further improves the accuracy by 14.41%, which proves the effectiveness of the improved method.

[0078] The application proposes a mechanical fault classification and identification method under variable working condition and partial migration condition. The results show that the multi-kernel cascade residual network model and the modified local maximum mean difference can effectively improve the fault recognition rate under variable working condition and partial migration condition. In summary, the model and method proposed in the application achieve good results.

Claims

1. A method for classifying and identifying mechanical faults under varying operating conditions and partial migration conditions, characterized in that, The specific steps are as follows: Step 1: The single-axis accelerometer collects vibration signals of known faults as source domain data and labels them according to the fault type; the vibration signals of the working conditions and corresponding unknown faults are used as target domain data. Step 2: Collect vibration signals from the source and target domain data using a sliding window, and obtain data samples after preprocessing; Step 3: Construct a multi-kernel cascaded residual network model, input data samples into the multi-kernel cascaded residual network model, extract high-dimensional features of source domain data and target domain data, and use the trained classifier to predict the target domain data to obtain pseudo-labels; Step 4: Calculate the corrected local maximum mean difference between the source domain data and the target domain data using the extracted high-dimensional features, source domain true label annotations, and target domain pseudo labels. This difference is used as the domain adaptation loss term to measure the distribution difference between each subdomain of the source domain and the target domain. Step 5: Optimize the domain adaptation loss term and classification loss term together as the objective function. After training for a specified number of iterations, obtain the trained multi-core cascaded residual network model, which serves as a fault diagnosis model for engineering machinery under varying operating conditions for fault diagnosis. The modified local maximum mean difference is an improvement on the local maximum mean difference in the subdomain adaptation method; By using a dynamic feedback mechanism, a feedback weight coefficient determined by a scaling factor r is added. The formula for calculating the corrected local maximum mean difference (ReLMMD) is as follows: ； Where p represents the distribution of the source domain data samples, q represents the distribution of the target domain data samples, c represents the category, and C represents the total number of categories; This represents the i-th source domain data sample. This represents the j-th target domain data sample; Number of source domain data samples G represents the number of data samples in the target domain; G represents the feature extractor. For kernel functions; These are the weighting coefficients. Here, r represents the feedback weighting coefficient for the target domain data samples; r is the proportionality coefficient, indicating... The proportion of pseudo-labels for each type of sample in the target domain in the total samples of the target domain during the next iteration; r and The calculation formula is: k∈(t,t+Δt): ; k∈(t+Δt, t+2Δt): ; k∈(t+2Δt, t+3Δt): ; Where k is the number of iterations of the multi-core cascaded residual network model; This indicates that the j-th target domain sample is at the k-th iteration. The pseudo-labels obtained after inputting a multi-core cascaded residual network When undefined, the value is 1, and α, β, ε are hyperparameters, with α, β, and ε all ranging from 0 to 1; In calculation During the process, each iteration consists of 3Δt iterations. In the first 0 to t iterations, the feedback weight coefficient is ineffective. When k=t, the loop begins. In the t to t+Δt iterations, the proportional coefficient r is calculated using the pseudo-labels generated in the Δt iterations. In the t+Δt to t+2Δt iterations, the feedback weight coefficient... Effective; the feedback weighting coefficient is determined using the previously calculated proportional coefficient. This represents the proportion of the c-th class samples predicted by the model to the total number of samples in the target domain. Its theoretical value is the reciprocal of the number of sample classes. If the percentage is 10% to 20% higher than the theoretical value, it indicates that samples from other classes have been incorrectly classified into this class, thus reducing the likelihood of misclassification. To attenuate such weighting coefficients; when If the value is less than 40% of the theoretical value, it indicates that some samples of this type have been incorrectly classified into other classes or do not exist. Increase the value first. To increase the weighting coefficient of this class, if the increase in the proportion of this class exceeds 30% in subsequent iterations, then this class exists. This class of samples was incorrectly classified into other classes in previous iterations. During the (t+2Δt) to (t+3Δt)th iterations... Set to 1; if the increase in the proportion of this class does not exceed 30%, it indicates that this class does not exist in the target domain. During the (t+2Δt) to (t+3Δt)th iteration, Set to 0, and then repeat or stop according to this pattern.

2. The method for classifying and identifying mechanical faults under varying operating conditions and partial migration conditions according to claim 1, characterized in that, The multi-core cascaded residual network model is an improvement on the existing residual network model ResNet. It replaces the 7x7 convolution with a cascade of three 3x3 convolutions and makes the following improvements to the path where the 1×1 convolution of the residual block is located in the ResNet residual network model: the stride of the convolutional layer is changed to 1, and an average pooling layer with a kernel of 2 and a stride of 2 is added before it to take over the downsampling work.

3. The method for classifying and identifying mechanical faults under varying operating conditions and partial migration conditions according to claim 1, characterized in that, The objective function is: ； For domain-adaptive weight coefficients, The classification loss term is expressed as follows: ； The label representing sample i, This represents the probability that a sample is correctly classified.

Images