High-precision and high-reliability harmonic reducer machining and testing method

Abstract

The invention discloses a high-precision and high-reliability harmonic reducer machining and testing method, and belongs to the technical field of mechanical fault diagnosis. The method comprises the steps that vibration signals under different working conditions are collected and divided into source domain data and target domain data; the method comprises the following steps: extracting migratable features through an improved deep residual network (ResNet), replacing a convolution kernel, adding a batch normalization layer and an SE attention module, and adopting an ELU activation function; based on a multi-kernel maximum mean difference (MK-MMD) and a multi-domain discriminator, jointly aligning edge distribution and condition distribution, and dynamically balancing the importance of the edge distribution and the condition distribution through an adaptive factor; a plurality of label classifiers are used for carrying out adversarial optimization, the Wasserstein distance between the classifiers is minimized, and class-level domain adaptation is achieved; and finally, performing classification diagnosis on target domain data by using the trained model. According to the method, the domain transfer problem is effectively solved, and the precision and reliability of fault diagnosis of the cross-working-condition harmonic reducer are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a processing and testing method for a harmonic reducer with high precision and high reliability. Background Art

[0002] Timely and accurate fault diagnosis of rotating machinery is crucial in industrial practice, which helps to improve operation safety, reduce maintenance costs and enhance equipment reliability. In recent years, intelligent fault diagnosis methods have been significantly developed and usually rely on the latest data-driven methods (such as artificial neural network (ANN), support vector machine (SVM), random forest (RF), etc.) to solve the fault diagnosis task. However, the main limitation of most existing methods is the assumption that the training data and the test data come from the same distribution. In the actual industrial environment, due to factors such as changes in operating conditions and additional environmental noise, this assumption often does not hold. This problem, known as domain shift, significantly reduces the effectiveness of machine learning in fault diagnosis, especially when applying the labeled training data (source domain) to the unlabeled test data (target domain). Therefore, it is of great significance to develop more powerful algorithms that can effectively handle the problem of unlabeled target data and the distribution difference between the source domain and the target domain.

[0003] Generally, the labeled data in the source domain can be used to establish an effective source classifier to classify the machine health status. However, due to the difference in data distribution, the learned source classifier often performs poorly in the target domain, thus affecting the generalization ability of the model. To solve the domain shift problem, transfer learning technology has been widely developed in recent years, aiming to relax the restriction that the training data and the test data must follow the same distribution, so as to achieve the transfer of knowledge from the source domain to the target domain. Specifically, domain adaptation methods have received attention in fault diagnosis research, focusing on exploring potential domain-invariant features to bridge the distribution difference in domain adaptation.

[0004] Domain adaptation (DA), as a subfield of transfer learning (TL), provides a promising solution to the domain shift problem. The goal of DA is to transfer knowledge from the source domain to the target domain by minimizing the difference in data distributions between the two domains. Therefore, DA-based fault diagnosis methods have been widely studied. The mainstream domain adaptation methods mainly include difference-based adaptation methods and adversarial-based adaptation methods. The difference-based adaptation methods adopt difference metric functions, including Correlation Alignment (Coral) and Maximum Mean Discrepancy (MMD), to reduce the distribution differences between the source and target domains. Yang et al. developed a cross-domain fault identification method for bearings, which uses MMD and multi-layer domain adaptation with pseudo-label learning to narrow the distribution gap between the two domains. Li et al. applied multi-kernel MMD (MK-MMD) to multiple convolutional layers to reduce the distribution differences in the two domains. Wang et al. established a fault identification method of hierarchical deep domain adaptation (HDDA), which uses Coral for multi-layer adaptation to reduce the distribution gap between the two domains. Chen et al. proposed an intelligent fault identification method across multiple domains, which combines Local Maximum Mean Discrepancy (LMMD) and residual network to minimize the distribution differences between the source domain and multiple target domains. In difference-based DA methods, MMD is often used to reduce the distribution differences between the two domains. However, MMD requires the selection of appropriate kernel or multiple kernels. At the same time, due to the existence of the kernel matrix, the calculation of MMD is very complex and time-consuming.

[0005] In recent years, adversarial-based adaptation methods have received extensive attention from experts and scholars. Different from the difference-based adaptive DA methods, the adversarial-based adaptive methods use the mutual adversarial learning between the feature generator and the domain discriminator, and the feature generator extracts domain-invariant features that help identify cross-domain faults. Han et al. incorporated adversarial learning into DCN to identify mechanical faults under various operating conditions by obtaining transferable features. Chai et al. proposed using a feature generator to learn domain-invariant features by competing with multiple domain discriminators. Although the adversarial-based adaptation methods can achieve high accuracy, their convergence performance is poor. The convergence of the difference-based adaptation methods is better than that of the adversarial-based adaptation methods. In addition, most DA-based methods only align the global distribution or sub-domain distribution between the source domain and the target domain. The performance of the model with global or sub-domain adaptation can still be improved by simultaneously performing global and sub-domain adaptation. Generally speaking, DA-based fault diagnosis methods still need to be improved.

[0006] Regardless of the gains obtained by these methods, most of them only consider relatively simple domain adaptation scenarios where the two domains differ only in the marginal distribution. More generally, conditional distribution differences may also coexist in realistic diagnostic scenarios, and poor diagnostic performance will be obtained when only marginal distribution adaptation is considered. Therefore, it is of great significance and necessity to jointly handle marginal distribution differences and conditional distribution differences when solving the fault diagnosis problem.

[0007] Recently, methods based on conditional distribution and joint distribution adaptation have been proposed. Pei et al. used multiple domain discriminators to achieve fine-grained alignment of different data distributions. Han et al. used joint distribution adaptation (JDA) to perform artificial damage diagnosis tasks under different operating conditions. Jiao et al. introduced the adversarial domain method into JDA and successfully achieved bearing fault diagnosis in the target domain. It is proved that the method using joint distribution obtains higher accuracy than the method using only conditional distribution, and two points are found. First, these JDA methods focus on using methods based on statistical distance metrics or adversarial-based methods. Although adversarial-based methods have higher accuracy, they are not easy to converge. The method based on statistical distance metrics has better convergence performance than the method based on adversarial learning, but the diagnostic accuracy is lower. Second, most JDA methods do not quantitatively analyze the importance of the two distributions, but there are many differences in the contributions of the two distributions in different cross-domain scenarios. Therefore, it is necessary to use different domain adaptation strategies in further research to obtain better diagnostic accuracy.

[0008] In actual fault diagnosis, due to different working loads of harmonic reducers, different vibration signal acquisition positions, and different equipment structures, there are large differences in the distribution of vibration signals. Aiming at the problems of manual feature extraction and insufficient adaptability of previous models, therefore, a deep convolutional multi-adversarial domain adaptation model DCMAAN for harmonic reducer fault diagnosis is proposed. Summary of the Invention

[0009] The purpose of the present invention is to provide a high-precision and high-reliability harmonic reducer processing and testing method, which solves the problem of the decline in cross-domain fault diagnosis accuracy caused by the differences in data distributions between the source domain and the target domain (especially the marginal distribution and conditional distribution differences).

[0010] To achieve the above purpose, the present invention provides a high-precision and high-reliability harmonic reducer processing and testing method, including the following steps: S1. Collect the vibration signals of the harmonic reducer under different working conditions, and divide the vibration signals into source domain data and target domain data; S2. Extract the transferable features in the vibration signals through an improved deep residual network ResNet; S3. Based on the multi-kernel maximum mean discrepancy (MK-MMD) and multiple domain discriminators, jointly align the marginal distributions and conditional distributions of the source domain and the target domain, and dynamically balance the importance of the two through an adaptive factor; S4. Use multiple label classifiers for adversarial optimization to minimize the Wasserstein distance between classifiers and achieve class-level domain adaptation; S5. Use the trained model to classify the data in the target domain and output the diagnostic results.

[0011] Among them, the feature extraction module: learns transferable features through an improved residual feature extractor F. Input the vibration signals under different working conditions into the feature extractor to extract transferable features.

[0012] The domain adaptation module: constructs a distribution difference metric and multiple domain discriminators to help the feature extractor learn domain-invariant features. Specifically, first input the extracted high-level features into multiple classifiers to obtain the average probability to guide the input of the domain discriminator, and use the loss value of the domain discriminator as the difference index of the conditional distribution. Secondly, use MK-MMD to calculate the inter-domain distribution distance and the intra-class distribution distance respectively, and use their ratio as the adaptive factor to guide the adaptation of the marginal distribution and the conditional distribution. Finally, train the model to extract domain-invariant features.

[0013] The fault identification module: uses the classifier H to classify the extracted domain-invariant features, thereby effectively realizing the identification of different fault types.

[0014] Preferably, in step S1, the data preprocessing includes: S11. Segment the vibration signal, and the length of each segment is 1024 sample points; S12. Eliminate the difference in signal dimension through standardization; S13. Divide the training set, validation set, and test set in a ratio of 3:1:1.

[0015] Preferably, in step S2, the improved ResNet includes the following optimization steps: S21. Replace the 7×1 convolutional kernel in the original ResNet with multiple 3×1 convolutional kernels; S22. Add a batch normalization layer (BN) and a SE attention module in the residual block; the structure of the SE module is mainly composed of an average pooling layer, two 1×1 Convs, and a multiplication operation. The SE module belongs to the channel attention mechanism and can adaptively learn the dependence between different channels of the input feature map.

[0016] S23. Use the ELU activation function to replace the ReLU function to retain negative value information.

[0017] Among them, after the feature extractor, a label classifier H is adopted to achieve fault classification. The classifier uses the Softmax function to output the probabilities of each category. During the model training process, the output probabilities of the label classifier will be used to guide the input of the multi-domain discriminator to achieve the alignment of sub-domains belonging to the same category in the source domain and the target domain. Assume that the dataset contains C health conditions. Assume that the output of the th classifier is , where . The th classifier outputs the probability defined by the following formula: ; where refers to the parameter to be learned.

[0018] The average of the output probability values of all classifiers is calculated to obtain the average probability of each category: ; A new probability distribution is obtained to represent the average prediction probability of the data in all classifiers, and then this average probability is used as the input of the domain discriminator to guide the alignment between the source domain and the target domain, especially the alignment of sub-domains of the same category.

[0019] For the outputs of multiple classifiers, the classification loss considers the output of each classifier and averages the results of multiple classifiers to obtain the final classification loss. Assume there are classifiers, and the output of each classifier is denoted as , where represents classifiers; to calculate the final classification loss, first take the average of the outputs of all classifiers, and then calculate the classification loss based on the average output. The specific calculation is as follows: ; where refers to the cross-entropy loss function, means its value is 1 if ; is the probability value output by the softmax function of a certain classifier.

[0020] The Wasserstein distance is used to evaluate the predicted distribution of the target samples. The Wasserstein distance is a tool for effectively measuring the difference between two probability distributions and is often used in domain adaptation tasks to compare different distributions.

[0021] Use multiple classifiers. Consider the output of each classifier (such as softmax probability) as a distribution, and then use the Wasserstein distance to measure the difference between these distributions. Specifically, the Wasserstein distance measures the minimum "transportation workload" required to transform one distribution into another. Therefore, it can effectively capture the geometric differences between the Softmax probability distributions of multiple classifiers.

[0022] The output of each classifier is a softmax probability distribution , representing the classification prediction of this classifier for the input sample .

[0023] The Wasserstein distance W is used to measure the difference between the prediction distributions of classifier and : ; where represents all possible joint distributions from to . The Wasserstein distance represents the minimum cost of moving mass under this joint distribution.

[0024] For the case of multiple classifiers, the Wasserstein distance can be extended to the average Wasserstein distance between pairs of multiple classifiers: ; where is the number of samples in the target domain, represents the softmax probability output of the -th classifier for the sample . This loss function represents the average of the Wasserstein distance differences between all pairs of classifiers.

[0025] Using the Wasserstein distance to measure the difference between multiple classifiers can more precisely capture the geometric differences in the classifier output distributions. In adversarial training, first maximize the Wasserstein distance between classifiers to increase the diversity of different classifiers, and then update the feature extractor by minimizing the Wasserstein distance to ensure that the predictions of classifiers for the target domain tend to be consistent. This can better achieve class-level alignment between the source domain and the target domain and improve the performance of cross-domain fault diagnosis.

[0026] Preferably, in step S3, the specific method for aligning the marginal distribution is as follows: By calculating the MK-MMD distance between the source domain and target domain features in the Reproducing Kernel Hilbert Space (RKHS) and minimizing this distance, the reduction of the global distribution difference is achieved.

[0027] Preferably, in step S3, the specific method for aligning the conditional distribution is as follows: Introduce multiple domain discriminators, where each domain discriminator corresponds to a fault category, and through adversarial training, make the feature distributions of the same type of fault samples consistent in the source domain and the target domain.

[0028] Preferably, in step S3, the dynamic adaptive factor combines the MK-MMD distance with the quantitative sum of the multi-adversarial domain losses to form a joint distribution distance. Add MK-MMD to the second FC layer of the domain discriminator to calculate the difference in the fault feature distributions within the same category. Its calculation formula is as follows: ; where represents the weight for adapting to the marginal distribution, and the weight for adapting to the conditional distribution is .

[0029] Preferably, in step S4, the specific implementation of optimizing the adversarial multi-classifier includes: S41. Enhance the diversity of the classifier by maximizing the classification difference of unlabeled target samples; S42. Optimize the feature extractor to extract domain-invariant features by minimizing the Wasserstein distance output by the classifier.

[0030] Preferably, in step S4, the number of label classifiers is 3. Each classifier outputs a probability distribution through the Softmax function, and the input of the domain discriminator is guided by the average probability.

[0031] Preferably, in step S4, the training of the model uses a Gradient Reversal Layer (GRL) to reverse the gradient sign of the domain discriminator during backpropagation.

[0032] Therefore, the harmonic reducer processing and testing method with high precision and high reliability adopting the above structure has the following beneficial effects: (1) Adaptive factor dynamically balances distribution alignment: Introduce the adaptive factor , quantitatively combine the adaptation of the marginal distribution and the conditional distribution, effectively weigh the relative importance of the two, and achieve more precise distribution alignment.

[0033] (2) Improve the calculation of the joint distribution: Use a multi-domain discriminator combined with the MK-MMD distance to replace the traditional single-domain discriminator, greatly improve the calculation effect of the joint distribution, and reduce the distribution difference between the source domain and the target domain.

[0034] (3) Optimization of adversarial multi-classifier: Multiple classifiers are used to perform adversarial optimization with the extracted features, and the average probability of the classifier output is used to guide the input of the multi-domain discriminator to achieve class-level alignment, improving cross-domain diagnosis performance and robustness.

[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Brief Description of the Drawings

[0036] Figure 1 It is a structural diagram of a deep convolutional multi-adversarial domain adaptation model; Figure 2 It is a schematic structural diagram of a feature extractor; Figure 3 It is a schematic structural diagram of an SE module; Figure 4 It is a fault diagnosis flow chart based on the DCMAAN model; Figure 5 It is a schematic diagram of the vibration signal of the CWRU rolling bearing under the 3HP working load in Embodiment 1; Figure 6 It is a radar chart of the classification results of different methods in Embodiment 1; Figure 7 It is a schematic diagram of the diagnostic accuracy of six models in the 3hp→0hp transfer task in Embodiment 1; (a) is the accuracy rate; (b) is the loss; Figure 8 It is the visualization of six models by t-SNE in the 3 hp → 0 hp transfer task in Embodiment 1; (a) is DCC, (b) is DANN, (c) is DCTLN, (d) is MADA, (e) is MAAN, (f) is DCMAAN; Figure 9 It is a schematic diagram of the 3-0 HP confusion matrix in Embodiment 1; Figure 10 It is a schematic diagram of the test system; Figure 11 It is a radar chart of the classification results of different methods in Embodiment 2; Figure 12 It is a schematic diagram of the diagnostic accuracy of six models in the S5L4-S11L0 transfer task in Embodiment 2; (a) is the accuracy rate; (b) is the loss; Figure 13 It is a schematic diagram of the S5L4-S11L0 confusion matrix in Embodiment 2; (a) is DCC, (b) is DANN, (c) is DCTLN, (d) is MADA, (e) is MAAN, (f) is DCMAAN; Figure 14Schematic diagram of visualizing six models by t-SNE under the S5L4-S11L0 migration task in Example 2; (a) is DCC, (b) is DANN, (c) is DCTLN, (d) is MADA, (e) is MAAN, and (f) is DCMAAN. Detailed implementation manners

[0037] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] Unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the ordinary meanings understood by those of ordinary skill in the art to which the present invention belongs. The "first", "second" and similar terms used in the present invention do not denote any order, quantity or importance, but are only used to distinguish different components. The terms such as "comprising" or "including" mean that the elements or objects appearing before this term cover the elements or objects listed after this term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0039] As Figure 1-4 shown, the present invention provides a high-precision and high-reliability harmonic reducer processing and testing method, including the following steps: S1. Collect vibration signals of the harmonic reducer under different working conditions, and divide the vibration signals into source domain data and target domain data; S2. Extract transferable features in the vibration signals through an improved deep residual network ResNet; S3. Based on the multi-kernel maximum mean discrepancy MK-MMD and multiple domain discriminators, jointly align the marginal distributions and conditional distributions of the source domain and the target domain, and dynamically balance the importance of the two through an adaptive factor; S4. Adopt multiple label classifiers for adversarial optimization to minimize the Wasserstein distance between the classifiers and achieve class-level domain adaptation; S5. Use the trained model to classify the faults of the target domain data and output the diagnostic results.

[0040] Among them, the feature extraction module: extracts transferable features through an improved residual feature extractor F. Input the vibration signals under different working conditions into the feature extractor to extract transferable features.

[0041] Domain adaptation module: construct a distribution difference metric and multiple domain discriminators To help the feature extractor learn domain-invariant features. Specifically, first, the extracted high-level features are input into multiple classifiers to obtain the average probability to guide the input of the domain discriminator, and the loss value of the domain discriminator is used as an index of the difference in conditional distributions. Second, MK-MMD is used to calculate the inter-domain distribution distance and the inter-class distribution distance respectively, and their ratio is used as an adaptive factor to guide the adaptation of the marginal distribution and the conditional distribution. Finally, the model is trained to extract domain-invariant features.

[0042] Fault identification module: Use the classifier H to classify the extracted domain-invariant features, so as to effectively realize the identification of different fault types.

[0043] Embodiment 1 Based on the bearing fault signal dataset provided by Case Western Reserve University (CWRU), test and compare the performance of different models under various noise and working conditions. Then, use the data collected by the built harmonic reducer fault test bench to further verify and analyze the diagnostic ability of the model. Table 1 shows the network structure parameters of the DCMAAN model. The feature extractor in the feature extraction module is an improved ResNet18, mainly composed of 8 residual blocks. Both the classifier and the domain discriminator are composed of two fully connected (FC) layers, and the Softmax function is used for multi-classification in the final layer. To reduce the risk of overfitting, Dropout is added after each layer, and the retention rate is 0.5. In all experiments, the model is trained using the Adam optimizer, the learning rate is set to 0.001, the momentum value is 0.9, and the batch size is 128. The number of training epochs is set to 100, and each experiment is repeated five times, and the average value is taken as the final result.

[0044] Table 1 Network structure parameters of the DCMAAN model ;

[0045] The rolling bearing dataset used in the experiment is provided by the Bearing Data Center of Case Western Reserve University (CWRU), and the vibration signals are collected by driving the SKF6205 rolling bearing at the motor drive end at a sampling frequency of 12 kHz under a load of 0 - 3 HP.

[0046] To make the collected fault data as real and effective as possible, a single-point damage is processed by using an electric discharge machining machine to obtain a faulty bearing. The data contains four different states: normal (N), ball fault (BF), inner ring fault (IF), and outer ring fault (OF). Each fault state has three damage sizes: 0.007, 0.014, and 0.021 inches (1 inch = 25.4 mm) respectively. Therefore, there are nine fault states and one normal state.

[0047] The vibration signals under 0 - 3HP loads are expanded into datasets A, B, C, and D through overlapping sampling. Each dataset (A, B, C, and D) represents ten working conditions of rolling bearings, and the detailed information is shown in Table 2. In the experiment, the length of each sample is 1024 sample points, and 250 samples are set for each type of fault, that is, each dataset has a total of 2500 samples. The datasets in each domain are divided into a training set, a validation set, and a test set according to the ratio of 3:1:1.

[0048] Table 2 Information of experimental datasets ;

[0049] Randomly select vibration signals under different health conditions at 3HP working load as shown in Figure 5 It can be seen from the figure that there are obvious distribution differences in the vibration samples under different health states.

[0050] To address the domain shift problem caused by different working loads, 12 transfer tasks are set: 0hp→1hp / 2hp / 3hp, 1hp→0hp / 1hp / 2hp, 2hp→0hp / 1hp / 3hp, 3hp→0hp / 1hp / 2hp. For example, 1hp→2hp means training the network model with source domain samples under the 1hp working load condition and testing the performance of the network model with target domain samples under the 2hp working load. To increase the sample size, overlapping sampling is adopted with an overlapping rate of 40%. When a certain working load is used as the target domain, its corresponding sample labels do not participate in model training.

[0051] To further evaluate the effectiveness of the proposed DCMAAN model, five different models are compared on two datasets. Each model is implemented based on the same architecture benchmark as DCMAAN, but differs according to its own required modules. The experimental settings of all comparison models are similar to those of DCMAAN. Table 3 shows the transfer methods and adaptation strategies used by these six fault diagnosis models. These models transfer knowledge in the source domain to the target domain through different strategies to achieve effective cross - domain fault diagnosis.

[0052] Table 3 Transfer methods and adaptation strategies used by various fault diagnosis methods ;

[0053] Table 4 Diagnostic accuracies of six models for transfer tasks under different working loads ;

[0054] Table 3 and Figure 6Shows the diagnostic accuracy of the transfer tasks of 6 models under different workloads. The average diagnostic accuracies of DDC, DANN, DCTLN, MADA, MAAN, and the proposed DCMAAN are 96.49%, 96.02%, 99.24%, 96.25%, 97.94%, and 99.89% respectively. Obviously, the proposed DCMADA model outperforms several other models in all tasks. From the comparison results, we obtain the following three observations.

[0055] It can be seen that the DCMAAN model performs significantly better than other models in all tasks, with the best domain adaptation ability, especially outstanding in tasks with large load changes (such as 0hp → 3hp, 3hp → 2hp). Although DCTLN is slightly inferior to the DCMAAN model, its stability and high precision make it a good choice. The performances of MADA and MAAN models are relatively close, but slightly insufficient compared to other models. DANN and DDC show certain instability when dealing with some tasks, especially ineffective in tasks with large load differences (such as 3hp → 0hp).

[0056] In summary, the proposed DCMAAN model performs best among the six models, and its superiority can be attributed to the following three aspects: (1) Introduction of the adaptive factor : The adaptive factor quantitatively combines the adaptation of the marginal distribution and the conditional distribution, effectively weighing the relative importance of the two, thus achieving more precise distribution alignment.

[0057] (2) Improved joint distribution calculation: The DCMAAN model uses a multi-domain discriminator combined with the MK-MMD distance instead of the traditional single-domain discriminator, greatly improving the calculation effect of the joint distribution, enabling the model to more effectively reduce the distribution difference between the source domain and the target domain.

[0058] (3) Adversarial multi-classifier optimization: Multiple classifiers are used to perform adversarial optimization with the extracted features, and the average probability of the classifier output is used to guide the input of multiple domain discriminators, thus achieving class-level alignment between the two domains, significantly improving the performance and robustness of cross-domain diagnosis.

[0059] Randomly select the 3hp → 0hp transfer task to compare the classification accuracy and loss of rolling bearing fault diagnosis models using different domain adaptation strategies. Figure 7 (a)-(b) show the diagnostic performances of DDC, DANN, DCTLN, MADA, MAAN, and the proposed DCMAAN models under the 3hp → 0hp migration task. In contrast, the improvement makes the DCMAAN model have stronger reliability and performance in cross-domain fault diagnosis tasks.

[0060] The fault classification results of DDC, DCTLN, DANN, MADA, MAAN and the proposed DCMAAN model under the 3hp→0hp transfer task were visualized using the common t-SNE algorithm. As can be seen from Figure 8 Figures (a)-(f), the proposed DCMAAN model is significantly superior to other methods in terms of classification effect. In addition, it was also observed that three different degrees of outer ring faults were difficult to distinguish in all models, and these three types of faults were almost completely confused together, mainly because of their high similarity in the feature space. In the models DCTLN, MADA and MAAN, there were individual overlaps between the inner ring fault of 0.014 inches and the outer ring fault of 0.021 inches, and the proposed DCMADA model accurately distinguished these various types of faults. This indicates that the proposed DCMADA model can better learn transferable features and obtain satisfactory diagnostic accuracy in the unlabeled target domain.

[0061] Figure 9 Figures (a)-(f) show the evaluation of the respective performances by comparing the confusion matrices of the DDC, DCTLN, DANN, MADA, MAAN and DCMAAN models in the 3hp→0hp migration task. As Figure 9 shown in (a), half of the 50 test samples of the outer ring fault with a diameter of 0.014 inches were misclassified as the inner ring fault of 0.014 inches and the outer ring fault of 0.021. As Figure 9 shown in (b), almost all of the outer ring faults of 0.0021 inches were misclassified as the outer ring faults of 0.007 inches, which led to a significant reduction in the fault diagnosis accuracy of the DANN model in the bearing fault classification task. There were also significant overlaps in other fault models in distinguishing the inner ring and outer ring fault types. However, the proposed DCMAAN model could completely and correctly distinguish various faults, showing extremely high classification accuracy. The results show that the DCMAAN model has significant advantages in rolling bearing fault classification, can effectively handle complex fault diagnosis tasks, and maintain strong generalization ability under different working conditions.

[0062] Example 2 To verify the effectiveness of the method under actual working conditions, a harmonic reducer fault test bench was established. The operating states of the harmonic reducer under different working conditions and fault types were obtained. The test system consists of a control unit, a drive unit, a workpiece under test unit, a sensing unit, a data acquisition unit and a load unit. The control unit refers to the controller of the servo motor, the drive unit refers to the servo motor, the workpiece under test unit refers to the harmonic reducer with various faults, the sensing unit is an acceleration sensor, the data acquisition unit refers to the data acquisition system, and the load unit refers to the magnetic powder brake.

[0063] The structure of the test system is as Figure 10 shown. The servo motor drives the harmonic reducer through a connecting shaft, and the output shaft is connected to the load mechanism to simulate the load condition of the reducer in actual applications. The acceleration sensor monitors the vibration signal of the reducer in real time to ensure accurate data acquisition. Through the host computer software, the motor control program controls the drive motor and records the vibration data during the entire test process.

[0064] The detailed process of data acquisition is as follows: 1) The motor control program was written and deployed in the computer software. This program is responsible for starting and stopping the motor, and at the same time adjusts the rotation speed and operation mode of the motor according to the preset parameters to ensure that the test system works as required.

[0065] 2) Start the data collector and initialize the parameters of the vibration sensor in the software, including sensor calibration, sensitivity adjustment, and data sampling frequency setting, to ensure that the working state of the sensor meets the test requirements.

[0066] 3) Load the motor drive program and start the motor control module, and the measurement and control system starts to run. The acceleration sensor collects the vibration signal of the reducer in real time, and the data is transmitted to the host computer software through the data collector for processing and analysis. During the whole process, the measurement and control system monitors and saves the key data of each test in real time to ensure the accuracy and repeatability of the test.

[0067] In order to deeply study the signal performance of harmonic reducers with different fault types under complex variable working conditions, a total of 16 different working conditions were designed in this experiment at different rotation speeds and different loads. The motor rotation speed is divided into 4 levels, which are 295 r / min, 590 r / min, 885 r / min, and 1180 r / min respectively. The load has 4 levels, which are 0 N.m (no load), 4 N.m, 8 N.m, and 12 N.m respectively. The specific detailed working conditions are summarized in Table 6.

[0068] Table 6 Details of different working conditions of harmonic reducers ;

[0069] The process of collecting various fault data of the harmonic reducer is as follows: (1) Develop the motor control program and build the test bench. Check the assembly of each component, and then debug each module system to ensure stable state and normal operation.

[0070] (2) Under the condition that the load is set to 0 N.m, adjust the rotation speed of the drive motor to 295 rpm, record the monitoring data, that is, obtain the vibration sensor data under the working condition S2L0, and repeat the test three times.

[0071] (3) Ensure that the load remains unchanged, and successively adjust the rotational speed to 590 rpm, 885 rpm, and 1190 rpm, and record the sensor data under the working conditions of S5L0, S8L0, and S 11 L0.

[0072] (4) Continuing in the same manner, under the conditions of loads of 4 N.m, 8 N.m, and 12 N.m respectively, complete the fault tests for the above 4 rotational speeds.

[0073] (5) After completing the fault test of the external teeth wear of the flexspline, stop the machine to replace the faulty components and prepare for subsequent tests (6) After completing all the tests of the faulty components, stop the machine to tidy up the test bench and package and process the data to prepare for subsequent data analysis and fault classification.

[0074] During the experiment, 3 acceleration sensors were used to collect the working condition data of the harmonic reducer, which were respectively installed in the X-axis direction, Y-axis direction, and Z-axis direction.

[0075] The measuring point distribution scheme was obtained through enumeration and comparative experiments, and the fault detection accuracy was used as the evaluation index. According to the different health states and working conditions of the harmonic reducer, the collected vibration signals were made into a data set. The conditional acquisition scheme is shown in Table 7.

[0076] Table 7 Summary Table of Artificial Damage Experiments of Harmonic Reducer LSG-25-100-U-Ⅱ ;

[0077] Table 8 Diagnostic Accuracy of Six Models for Transfer Tasks under Different Working Loads and Rotational Speeds ;

[0078] To verify the effectiveness of the proposed model, 8 working condition migration experiments were set up. In the experiment, the length of each sample was 1024 sample points, and 1000 samples were set for each type of fault, that is, each working condition data set had 12000 samples. These samples were divided into a training set, a validation set, and a test set according to the ratio of 3:1:1.

[0079] As shown in Table 8 and Figure 11As shown, eight migration experiments were carried out at different rotational speeds and loads. First, DCTLN is more effective than DANN and DDC in dealing with distribution differences, with an average accuracy of 92.92%, but it is still lower than the method proposed in the present invention. This is mainly because DCTLN only considers the marginal distribution and does not consider the conditional distribution. Second, the domain adaptation strategy of MAAN is similar to the DCMAAN method, and its diagnostic accuracy is better than other domain adaptation methods except DCMAAN, verifying the superiority of the proposed domain adaptation strategy. This strategy not only considers the marginal distribution and conditional distribution, but also introduces an adaptive factor to adjust the adaptability of the two. In addition, by using multiple classifiers to perform adversarial optimization with the extracted features, the obtained average probability is used to guide the input of the domain discriminator features, enabling the proposed model to better learn transferable features and fault separation features, thereby improving the diagnostic accuracy. Compared with the other five methods, DCMAAN achieved higher diagnostic accuracy in each migration task, and the average accuracy in 8 fault diagnosis tasks reached 95.94%. Especially in the S 11 L4-S8L0 migration task, the model obtained the highest accuracy of 98.6%. In addition, in the migration task S 11 L8-S8L 12 (97.63%) and S5L4-S 11 L0 (97.35%), DCMAAN performed excellently, showing strong stability and adaptability. This is mainly attributed to the integration of multiple classifiers, which makes the model more sensitive to subtle changes in distribution differences and can more effectively adapt to distribution changes under different working conditions.

[0080] In the randomly selected S5L4-S 11 L0 transfer task, the classification accuracy and loss of the rolling bearing fault diagnosis model under different domain adaptation strategies were compared, and the results are as Figure 12 shown in (a) of Figure 12 and (b) of

[0081] In the loss curve, the loss value of the method of the present invention drops rapidly in the first few rounds and quickly tends to be stable, remaining at a low level, indicating that the method has good convergence and robustness. In contrast, the loss curves of other methods fluctuate greatly and are not stable enough during the training process. Generally speaking, the DCMAAN model is significantly superior to other models in terms of accuracy, convergence speed, and loss stability.

[0082] As Figure 13 shown in (a)-(f) of 11 the confusion matrices of the DDC, DCTLN, DANN, MADA, MAAN, and DCMAAN models in the S5L4-S

[0083] In categories 5, 6, 10, and 11, the accuracy of the DDC model drops significantly and there are many misclassifications; the DANN model has more misclassifications in categories 5 and 6, especially in category 6; in addition, it also performs poorly in the compound fault categories 10 and 11, and many samples are misclassified as single fault categories, indicating that it is difficult to handle the combination of multiple fault features. Compared with the first two models, DCTLN has improved accuracy, especially in the compound fault categories 10 and 11, and the misclassifications have decreased, but the overall performance is still not ideal, showing its limitations in identifying complex combined faults. Since the MADA model adopts the conditional distribution strategy, its misclassification situation in categories 5, 6, 10, and 11 has improved; the MAAN model combines the marginal distribution and the conditional distribution, adopts the same domain adaptation strategy as the present invention, and the misclassifications in various fault categories have been significantly reduced, second only to DCMAAN. The DCMAAN model dynamically adjusts the marginal and conditional distributions by introducing an adaptive adjustment factor and combines multiple classifiers for optimization, demonstrating excellent performance in dealing with complex combined fault categories. It can accurately capture the characteristics of different types of faults, especially in the categories including complex fault combinations such as tooth breakage, wear, and fatigue fracture, showing strong adaptability and accuracy. In summary, the DCMAAN model has excellent classification ability for various harmonic reducer faults, especially showing strong adaptability and accuracy in dealing with complex fault diagnosis tasks.

[0084] The t-SNE algorithm commonly used was adopted to visualize the fault classification results of the DDC, DCTLN, DANN, MADA, MAAN, and the proposed DCMAAN models under the S5L4-S 11 L0 transfer task, as Figure 14As shown in (a)-(f) therein. It can be seen that there is a large amount of overlap between the DDC and DANN models in the feature space, especially in complex composite fault categories (such as categories 10 and 11), resulting in a low recognition accuracy. In contrast, the DCMAAN model shows the best class separation effect in the feature space, with clear and independent distributions for each class. Especially for complex composite fault categories, precise separation can be achieved. The MAAN model ranks second, but there is still a small amount of overlap when dealing with complex combined fault categories. The results show that by introducing the adaptive factor and the adversarial optimization of the multi-classifier, DCMAAN significantly improves the discrimination between classes. Especially for complex combined faults, there is almost no overlap in the classification effect. The effectiveness of DCMAAN in the domain adaptation fault diagnosis task is verified. It shows excellent robustness and adaptability when identifying complex fault categories, demonstrating its significant advantages in cross-domain fault diagnosis.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention. However, such modifications or equivalent replacements do not make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A processing and testing method for a harmonic reducer with high precision and high reliability, characterized in that: It includes the following steps: S1. First, construct the DCMAAN model, and then collect the vibration signals of the harmonic reducer under different working conditions, and divide the vibration signals into source domain data and target domain data; S2. Extract the transferable features in the vibration signals through the improved deep residual network ResNet; S3. Based on the multi-kernel maximum mean discrepancy MK-MMD and multiple domain discriminators, jointly align the marginal distributions and conditional distributions of the source domain and the target domain, and dynamically balance the importance of the two through an adaptive factor; S4. Use multiple label classifiers for adversarial optimization to minimize the Wasserstein distance between classifiers and achieve class-level domain adaptation; S5. Use the trained DCMAAN model to classify the faults of the target domain data and output the diagnostic results.

2. The high-precision and high-reliability harmonic reducer processing and testing method according to claim 1, characterized in that: In step S1, the data preprocessing includes: S11. Segment the vibration signals, and the length of each segment is 1024 sample points; S12. Eliminate the difference in signal dimension through standardization; S13. Divide the training set, validation set and test set according to the ratio of 3:1:

1.

3. A high-precision and high-reliability harmonic reducer processing and testing method according to claim 1, characterized in that: In step S2, the improved ResNet includes the following optimization steps: S21. Replace the 7×1 convolutional kernel in the original ResNet with multiple 3×1 convolutional kernels; S22. Add a batch normalization layer BN and an SE attention module in the residual block; S23. Use the ELU activation function to replace the ReLU function to retain negative value information.

4. A high-precision and high-reliability harmonic reducer processing and testing method according to claim 1, characterized in that: In step S3, the specific method for aligning the marginal distributions is: calculate the MK-MMD distance between the source domain and target domain features in the reproducing kernel Hilbert space RKHS, and minimize this distance to reduce the global distribution difference.

5. A high-precision and high-reliability processing and testing method for a harmonic reducer according to claim 1, characterized in that: In step S3, the specific method for aligning the conditional distributions is: introduce multiple domain discriminators, each domain discriminator corresponds to a fault category, and make the feature distributions of the same type of fault samples in the source domain and the target domain consistent through adversarial training.

6. A high-precision and high-reliability processing and testing method for a harmonic reducer according to claim 1, characterized in that: In step S3, the dynamic adaptive factor combines the MK-MMD distance with the quantitative sum of the multi-adversarial domain losses to form a joint distribution distance. Add MK-MMD to the second FC layer of the domain discriminator to calculate the difference in the fault feature distributions within the same category. The calculation formula is as follows: ； where represents the weight for adapting to the marginal distribution, and the weight for adapting to the conditional distribution is .

7. A high-precision and high-reliability harmonic reducer processing and testing method according to claim 1, characterized in that: In step S4, the specific implementation of the adversarial multi-classifier optimization includes: S41. Enhance the classifier diversity by maximizing the classification difference of unlabeled target samples; S42. Optimize the feature extractor to extract domain-invariant features by minimizing the Wasserstein distance output by the classifier.

8. A high-precision and high-reliability processing and testing method for a harmonic reducer according to claim 1, characterized in that: In step S4, the number of label classifiers is 3. Each classifier outputs a probability distribution through the Softmax function, and the average probability guides the input of the domain discriminator.

9. A high-precision and high-reliability harmonic reducer processing and testing method according to claim 1, characterized in that: In step S4, the gradient reversal layer GRL is used in the training of the model, and the gradient sign of the domain discriminator is reversed during backpropagation.

Images