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

Abstract

The invention discloses a high-precision and high-reliability harmonic reducer system testing method, and belongs to the technical field of mechanical fault diagnosis. The method comprises the following steps: constructing an auxiliary classifier Wasserstein generative adversarial network model based on gradient penalty and spectrum normalization, and generating a minority class of fault samples by using an unbalanced training set; preprocessing the vibration signals to form an initial training set; the effectiveness of the generated sample is evaluated through the Pearson's correlation coefficient and the cosine similarity; adding the sample into the original set and adjusting the equilibrium ratio to 1: 1; and a one-dimensional convolutional neural network is adopted for classification. The ACWGAN-SG integrates a Wasserstein distance, gradient penalty and spectrum normalization, a generator comprises four stages of deconvolution layers, and a discriminator is of a double-branch structure. Experiments are verified in a CWRU data set and a harmonic reducer test bed, and results show that the method can achieve stable convergence, the similarity between a generated sample and an original sample is high, the diagnosis accuracy after balance is remarkably superior to those of methods such as DCGAN and ACGAN, the problem of data imbalance is effectively solved, and the fault diagnosis precision and reliability are improved.

Description

Technical Field

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

[0002] With the development of modern industry, as an important part of intelligent devices, the safety performance of rotating machinery has attracted wide attention. Among them, the key transmission components of the harmonic reducer: bearings and gears, operate under extreme conditions such as high-speed rotation and overload for a long time, and are extremely prone to various faults. These faults will not only reduce the operating reliability of rotating machinery, but may also cause serious economic losses and safety hazards. Therefore, conducting research on the health state diagnosis and evaluation of bearings and gears has important engineering significance.

[0003] In recent years, intelligent fault diagnosis methods based on deep learning have become a research hotspot in this field due to their powerful feature extraction ability and end-to-end diagnosis characteristics. Typical deep learning models include convolutional neural network (CNN), long short-term memory network (LSTM), recurrent neural network (RNN), autoencoder (AE), and deep belief network (DBN), etc. However, these methods generally adopt supervised learning, and their model training heavily relies on a large number of labeled samples. In practical engineering applications, the acquisition cost of rotating machinery fault samples is high and the labeling difficulty is large, resulting in insufficient training sample quantity, and further leading to the overfitting problem of the model. Therefore, constructing a high-precision and high-reliability deep learning fault diagnosis model under the condition of limited labeled samples has become a key scientific problem to be solved urgently.

[0004] Due to the scarcity of labeled data for fault samples in actual engineering, unsupervised learning models have shown greater research value and application potential in the field of fault diagnosis compared to supervised learning models. Among them, the Generative Adversarial Networks (GAN) proposed by Goodfellow, as an innovative unsupervised generation model, can generate new samples with the same distribution as the original samples without label information, and its various variants have received extensive attention in multiple fields. In recent years, significant progress has been made in the research on rotating machinery fault diagnosis based on GAN: Lee et al. were the first to construct a GAN framework using fully connected layers, successfully generating rotor and bearing fault samples, effectively solving the problem of unbalanced diagnostic samples; Zhou et al. innovatively designed an acoustic emission structure, improving the generation quality of rolling bearing fault diagnosis data; Gao et al. significantly improved the generation stability of vibration signal fault samples by introducing Wasserstein GAN with gradient penalty; Dai et al. proposed a GAN model combined with the Autoencoder (AE) structure, further optimizing the sample generation effect by inputting training samples into both the discriminator and the generator simultaneously; Liu et al. innovatively used kernel maximum mean discrepancy as the penalty term for the generator loss function, achieving precise quantification of the difference between generated samples and real samples; Pei et al. proposed a new GAN framework based on the Reptile meta-learning strategy and the autoencoder structure, providing a new solution for rolling bearing fault diagnosis under unbalanced datasets.

[0005] Although unsupervised GAN shows good application prospects in rotating machinery fault diagnosis, it still has obvious limitations in multi-modal data augmentation. When using multi-modal fault diagnosis samples for unsupervised GAN training, it mainly faces the following three technical challenges: First, the difference in feature distributions of different health states may lead to a decline in the binary classification performance of the discriminator; Second, since the input of the generator is limited to random noise, it is difficult to accurately fit the data distributions of various health states, thus affecting the authenticity of the generated samples; Finally, the generation process lacks an effective health state control mechanism, which mainly stems from the fact that the generator only receives random noise without label information as input. The above problems result in the need to train independent GAN models for each health state separately in practical applications, which not only increases the computational complexity but also reduces the overall efficiency of fault diagnosis.

[0006] Auxiliary Classifier GAN (ACGAN) is an improved supervised learning algorithm. By embedding label information into the input noise of the generator, it can not only judge the authenticity of the sample, but also accurately identify the sample category, thereby achieving the generation of multi-modal high-quality samples. This feature enables ACGAN to handle the training tasks of multi-modal samples at the same time, showing significant advantages in the field of data enhancement. In recent years, the application research of ACGAN in different fields has achieved fruitful results: in image processing, Zou et al. proposed a multi-scale ACGAN based on Wasserstein distance, which effectively improved the resolution of ship slice images; Jin et al. developed an ACGAN with a multi-layer branch structure, which was successfully applied to signaling data generation; Ren et al. designed an ACGAN model that can learn statistical and structural features for synthetic aperture radar image classification tasks; Chen et al. achieved diversified style conversion in the field of image style transfer by introducing auxiliary classifiers.

[0007] It is worth noting that ACGAN has made breakthrough progress in the field of fault diagnosis in the past two years. Shao et al. successfully applied ACGAN to generate motor vibration signals and achieved synchronous sample generation of six working conditions; Guo et al. constructed a new ACGAN with a fully convolutional layer structure, which effectively generated spectral samples of bearing vibration signals; Huang et al. proposed a robust ACGAN framework, which significantly improved the fault diagnosis stability of wind turbine gearboxes under noise interference; Dijiang et al. innovatively combined meta-learning strategies with ACGAN, providing an effective solution to the problem of insufficient data in data-driven fault diagnosis.

[0008] Although ACGAN has achieved remarkable results in many fields, its inherent defects still restrict its further application in multimodal sample generation and fault diagnosis. Existing research has the following three main limitations: First, the discriminator of ACGAN undertakes the dual tasks of sample classification and authenticity discrimination at the same time. This functional coupling may lead to the accumulation of discriminator output errors, which in turn affects the quality of generated samples; second, traditional ACGAN generally adopts a loss function design based on JS divergence. Due to the discrete characteristics of JS divergence, problems such as gradient vanishing and training instability are prone to occur during model training; finally, when the discriminator performance is too strong, the generator tends to produce highly similar samples, that is, the model collapse phenomenon occurs. Therefore, developing an efficient ACGAN improved algorithm to achieve reliable multimodal sample generation and rotating machinery fault diagnosis has become a key scientific issue that needs to be solved urgently. Summary of the invention

[0009] The object of the present invention is to provide a high-precision and high-reliability testing method for a harmonic reducer system, which solves the problems of data imbalance, insufficient quality of generated samples, insufficient stability in the training process, and insufficient support for multi-modal faults.

[0010] To achieve the above object, the present invention provides a high-precision and high-reliability testing method for a harmonic reducer system, including the following steps: S1. Construct a fault diagnosis model of an auxiliary classifier Wasserstein generative adversarial network ACWGAN-SG based on gradient penalty and spectral normalization, and use an imbalanced training set to generate minority-class fault samples; S2. Perform data preprocessing on the vibration signals of the harmonic reducer to form an initial training set; S3. Generate samples for the imbalanced training set through the fault diagnosis model of ACWGAN-SG, evaluate the usability of the generated samples, and quantify the similarity between the generated samples and the original samples in time-domain and frequency-domain features through the Pearson correlation coefficient PCC and cosine similarity CS to ensure the effectiveness of the generated samples; S4. Add the generated samples to the original training set, adjust the balance ratio of the data set to a preset threshold, and monitor the change in model performance in real time during the balancing process; S5. Use a deep learning classification model to classify the fault modes of the balanced data set and output the health status diagnosis result of the harmonic reducer.

[0011] Preferably, the principle of the ACWGAN-SG model is as follows: Spectral normalization (SN): The application of spectral normalization in generative adversarial networks was initially in SNGAN and is used to improve the training stability of GANs. Spectral normalization is applied to the transposed convolutional layer of the generator and the convolutional layer of the discriminator. Its brief mathematical principle is as follows: ; This formula calculates the gradient of the eigenvalue matrix X after being activated by the activation function, converts it into the relationship between the eigenvalue matrix X and the weight matrix, and finally limits the value of within 1, thereby constraining the range of the gradient. is the gradient value matrix of the weight matrix and represents the largest eigenvalue (singular value) of W, also known as the spectral norm of W. This helps prevent the gradient from becoming too large or too small during training, making the model easier to converge. 1 is also called the 1-Lipschitz constraint. Spectral normalization acts on each convolutional layer of ACWGAN-SG to improve the training stability, generate more uniform and high-quality images, and thus enhance the overall robustness of ACWGAN-SG.

[0012] Gradient Penalty (GP): WGAN satisfies the 1-Lipschitz condition by adding gradient clipping. However, this method still has problems of difficult training and slow convergence speed. To further improve the quality of the generated images, gradient penalty from WGAN-GP is introduced. As shown in the following formula: ; Among them, the gradient penalty ensures that the two-norm is calculated between the gradient value of the weight matrix W and 1, making the gradient value around 1 instead of all less than 1. This helps to enhance the distinguishability of different generated fault data and is conducive to generating high-quality images. The purpose of the gradient penalty term is to make all gradient terms of the discriminator close to 1 instead of being compressed within 1. Therefore, the discriminator obtains richer gradients and can generate higher-quality data samples.

[0013] New CNN Fault Classifier: The present invention designs a new convolutional neural network (1D-CNN) fault classifier. This classifier enables the fault data generated by different generation models to be evaluated under the same criteria, ensuring the fairness of the experiment.

[0014] Preferably, the ACWGAN-SG model is constructed by introducing the Wasserstein distance and gradient penalty into ACGAN. The generator uses the random noise Z and specific labels to generate new samples. Then, the new samples and real samples are imported into the discriminator for adversarial training together. For each sample, the discriminator determines whether it is true or false and identifies its category. After meeting the similarity criteria, the generated samples will be used for augmenting the original dataset to achieve fault diagnosis.

[0015] Preferably, the ACWGAN-SG model consists of three parts: a generator, a discriminator, and a classifier. Considering the excellent performance of CNN in feature extraction and classification, both the generator and the discriminator are constructed based on 2D-CNN.

[0016] The input layer of the generator integrates the random noise vector and category label information. Its input is first passed to a fully connected layer with 1024 neuron nodes. Then, the output of the fully connected layer undergoes a reshape operation and is converted into a feature map with a dimension of 2×2×256, where the spatial dimension is 2×2 and the number of channels is 256. The network structure includes four levels of transposed convolution operations, and each level is composed of a transposed convolution layer, a batch normalization layer, and a ReLU activation layer connected in sequence. In addition, each transposed convolution layer integrates spectral normalization processing, uses a 5×5 convolution kernel, and operates with a stride of 2. After four levels of transposed convolution operations, the data is converted into a two-dimensional feature matrix. Before being input into the discriminator, this matrix will be reconstructed into a one-dimensional vector and normalized.

[0017] The discriminator adopts a symmetric architecture design with the generator and consists of four levels of convolutional operation modules. Each convolutional module is composed of a convolutional layer and a LeakyReLU activation layer connected in sequence, and spectral normalization processing is integrated in the convolutional layer. All convolutional layers use a 5×5 convolutional kernel, and a gradient penalty term is adopted in the objective loss function to improve the stability of the training process. The output layer adopts a dual-branch structure: the Sigmoid function is used to realize the discrimination of sample authenticity, and the Softmax function is used to complete the prediction of class probabilities.

[0018] Preferably, the training process of the ACWGAN-SG model is as follows: A random noise vector that follows the distribution is input together with the label set into the generator to generate synthetic samples whose data distribution is similar to that of the real sample data.

[0019] The new sample s is mixed with the original sample and used as the input of the discriminator for authenticity discrimination. The generator and the discriminator are alternately trained until the Nash equilibrium is reached.

[0020] The objective function expression of CWGAN-SG is as follows: ; ; where is a random sample interpolated on the line between and , is the gradient penalty coefficient, represents the conditional probability distribution on the class labels.

[0021] Preferably, the experimental results and the analysis process are as follows: To verify the effectiveness of the proposed method, it is verified on the CWRU dataset. First, the ACWGAN-SG model is trained to generate new samples. Then, the quality of the new samples is evaluated to determine whether they can be used for fault diagnosis. After evaluation, the generated samples are gradually added to the original dataset for the fault diagnosis task. Multiple balance ratios (BalanceRatio, BR) are set for testing, where BR is defined as follows: ; where and represent the number of samples of the majority class and the minority class, respectively.

[0022] Therefore, the present invention adopts a test method for a harmonic reducer system with high precision and high reliability having the above structure, and has the following beneficial effects: (1) By integrating the Wasserstein distance, spectral normalization, and gradient penalty into the ACGAN, the proposed model ACWGAN-SG can converge stably, avoiding gradient vanishing and mode collapse.

[0023] (2) ACWGAN-SG can generate high-quality samples to increase the unbalanced s dataset of the harmonic reducer. Training the fault diagnosis model using the augmented dataset, the diagnostic accuracy is gradually improved, proving that this method can effectively solve the data imbalance problem in fault diagnosis.

[0024] (3) By comparing with DCGAN, ACGAN, WGAN, and SMOTE and ADASYN methods, ACWGAN-SG can generate samples with a higher similarity distribution and performs better in the unbalanced fault diagnosis experiment.

[0025] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0026] Figure 1 Schematic diagram of the fault diagnosis algorithm framework based on ACWGAN-SG for a test method of a harmonic reducer system with high precision and high reliability of the present invention; Figure 2 Schematic diagram of the overall architecture of a test method of a harmonic reducer system with high precision and high reliability of the present invention; Figure 3 Schematic diagram of the generator of a test method of a harmonic reducer system with high precision and high reliability of the present invention; Figure 4 Schematic diagram of the discriminator of a test method of a harmonic reducer system with high precision and high reliability of the present invention; Figure 5 Schematic diagram of the comparison of the time-domain images of the original vibration signal (left) and the generated vibration signal (right) in Example 1, (a) is label 0, (b) is label 3, (c) is label 4, (d) is label 6, (e) is label 8; Figure 6 Schematic diagram of the fault diagnosis accuracy of six methods in the unbalanced case in Example 1; Figure 7 Schematic diagram of the confusion matrix of multi-class unbalanced fault diagnosis of bearings with different BRs using 1D-CNN in Example 1, (a) is 1:100, (b) is 1:20, (c) is 1:5, and (d) is 1:1; Figure 8Schematic diagram of the feature distribution of the test set data when the six algorithms in Example 1 expand the imbalanced samples to the balanced state. (a) is the ADASYN model, (b) is the SMOTE model, (c) is the DCGAN, (d) is the ACGAN model, (e) is the WGAN model, and (f) is the ACWGAN-SG model; Figure 9 Schematic diagram of the test system structure in Example 2; Figure 10 Schematic diagram of the fault diagnosis results of the four methods in the imbalanced case in Example 2; Figure 11 Schematic diagram of the confusion matrix analysis of the comparison methods in Example 2. (a) is the DCGAN model, (b) is the ACGAN model, (c) is the WGAN model, and (d) is the ACWGAN-SG model. Detailed implementation manners

[0027] The technical solutions of the present invention will be further described below with reference to the accompanying drawings and examples.

[0028] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meanings understood by those with ordinary skills in the field 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 "including" or "comprising" 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.

[0029] As Figures 1-4 shown, the present invention provides a high-precision and high-reliability test method for a harmonic reducer system, including the following steps: S1. Construct a fault diagnosis model of an auxiliary classifier Wasserstein generative adversarial network ACWGAN-SG based on gradient penalty and spectral normalization, and use the imbalanced training set to generate minority-class fault samples; The adversarial network ACWGAN-SG fault diagnosis model integrates the Wasserstein distance, gradient penalty term and spectral normalization technology. The input of the generator is a random noise vector and a class label. The discriminator constrains the Lipschitz continuity through the gradient penalty term, and its gradient penalty term is defined as: ; The generator of the ACWGAN-SG fault diagnosis model consists of four levels of transposed convolution layers, each integrated with spectral normalization processing. The size of the transposed convolution kernel is 5×5, the stride is 2, and the activation function is ReLU; the discriminator consists of four levels of convolution layers, the size of the convolution kernel is 5×5, the stride is 2, and the activation function is LeakyReLU. The output layer adopts a dual-branch structure. The first branch discriminates the authenticity of the sample through the Sigmoid function, and the second branch predicts the class probability through the Softmax function. Its objective function is: ; ; where, is a random sample obtained by online interpolation between and . is the gradient penalty coefficient, represents the conditional probability distribution on the class labels.

[0030] S2. Perform data preprocessing on the vibration signals of the harmonic reducer to form an initial training set. The data preprocessing adopts a sliding window sampling method with a window length of 1024 and a stride of 400, extract time-domain features and label the fault class labels; S3. Generate samples for the imbalanced training set through the fault diagnosis model of ACWGAN-SG, evaluate the usability of the generated samples, and quantify the similarity between the generated samples and the original samples in time-domain and frequency-domain features through the Pearson correlation coefficient PCC and cosine similarity CS to ensure the effectiveness of the generated samples. The Pearson correlation coefficient PCC between the generated samples and the real samples is ≥0.5, the cosine similarity CS is ≥0.8, and the frequency-domain energy distribution error does not exceed 10%.

[0031] Meanwhile, the initial balance ratio of the imbalanced training set is 1:100, and it is optimized to 1:1 by gradually adding generated samples, and the number of samples for each type of fault increases to be equal to the number of normal samples.

[0032] S4. Add the generated samples to the original training set, adjust the balance ratio of the data set to the preset threshold, and monitor the change of the model performance in real time during the balancing process; The balance ratio is defined as: ; where, and are the number of samples in the majority class and the minority class respectively, and the preset threshold is 1:1.

[0033] S5. Adopt a deep learning classification model to classify the fault modes of the balanced data set and output the diagnosis result of the health state of the harmonic reducer.

[0034] The deep learning classification model is a one-dimensional convolutional neural network 1D-CNN, which includes four levels of convolutional modules. The size of the convolutional kernel at each level is 5×1, the stride is 2, and the Softmax function is used in the output layer. The input of the 1D-CNN classifier is a one-dimensional vibration signal. After data normalization, it is input into the convolutional layer, and the ReLU activation function and batch normalization layer are connected after each convolutional module.

[0035] Example 1 The rolling bearing dataset used in this example is provided by the Bearing Data Center of Case Western Reserve University (CWRU). The test bench includes two motors, a torque sensor, a dynamometer and other control devices. The fault bearing data collected at the driving end with a sampling frequency of 12 kHz and a load of 2 hp is used as the experimental data. The dataset contains four fault modes (normal state, ball damage, inner race damage and outer race damage) and three damage degrees (0.007, 0.014 and 0.021 inches), forming a total of 10 healthy state categories. The sliding window sampling method is adopted, the sliding window is set to 1024, and the sliding step is 400. 100 samples are selected from each fault category, and 500 samples are selected from the normal category to form the training set. At the same time, 200 samples of each category are selected as the test set. In the training data, the normal category samples are used as the majority class, and the remaining fault category samples form the minority class. The detailed information is shown in Table 1.

[0036] Table 1 Details of the bearing dataset ;

[0037] After the ACWGAN-SG model is trained, new samples are generated for each category. First, to evaluate the data generation effect, this paper randomly selects the generated signals in 5 states for comparison with the original signals, and analyzes the time-domain images of the generated data.

[0038] Figure 5 (a)- Figure 5 (e) show the time-domain signals of the real samples and the generated samples in five states, indicating that the change curves of the generated vibration signals are similar to those of the real vibration signals, demonstrating that the ACWGAN-SG network proposed in this chapter has good data generation ability.

[0039] To further evaluate the generated samples, PCC and CS are calculated to quantitatively measure the similarity between the generated samples and the original samples. PCC represents the correlation between the generated samples and the original samples. Generally speaking, a PCC greater than 0.5 means a significant correlation. Similarly, CS evaluates the similarity of data distribution by calculating the cosine value of the angle between two sample vectors. The value ranges of both PCC and CS are from 0 to 1, and the higher the value, the higher the similarity. As shown in Table 2, the PCC of all ten categories of samples is higher than 0.5, and the CS of all categories is higher than 0.8, indicating that the generated data has a highly similar distribution to the original data.

[0040] Table 2 PCC and CS between the generated samples and the original samples of the bearing dataset ;

[0041] To simulate a multi-class data imbalance scenario, 200 samples are selected from the normal category, and only 2 samples are selected from each fault category to form the training set. At the same time, 100 test samples are prepared for each category. The detailed composition information of the dataset is shown in Table 3.

[0042] Table 3 Bearing samples for multi-level unbalanced fault diagnosis ;

[0043] To verify the performance of the generated samples in multi-class unbalanced fault diagnosis, the generated fault samples are gradually added to the unbalanced training set to expand it, as shown in Table 4.

[0044] Table 4 Methods for enhancing the training set ;

[0045] By comparing ACWGAN-SG with several common methods for solving the data imbalance problem, including DCGAN, ACGAN, WGAN, as well as SMOTE and ADASYN. To avoid contingency, 5 independent fault diagnosis experiments are carried out respectively at each imbalance rate, and the average accuracy is calculated. The results are as Figure 6 shown.

[0046] From Figure 6It can be seen that the diagnostic performance of all six algorithms is improved as the number of generated samples in the training set increases. At an imbalance ratio of 1:25, the classification accuracy of the present invention reaches 57.82%. Compared with the other five comparison methods, the performance advantage is not obvious at this time. However, when the sample number reaches the balanced state, the present invention achieves an identification accuracy of 98.4%, which is 5.8%, 4.96%, 9.62%, 8.74%, and 11.02% higher than the ACGAN, WGAN, DCGAN, ADASYN, and SMOTE methods, respectively. This experimental result fully demonstrates the significant advantage of ACWGAN-SG in dealing with the data imbalance fault diagnosis problem.

[0047] To effectively evaluate the diagnostic performance of the proposed method, confusion matrices at different balance rates are plotted. As Figure 7 (a)- Figure 7 (d) shown. It can be seen from Figure 7 (a) that when BR = 1:100, due to the extremely imbalanced training set being unable to provide sufficient feature information of the minority class for the diagnostic model, almost all minority class samples are misclassified. When BR is increased to 1:20, the diagnostic accuracy of the minority class samples is significantly enhanced. Further increasing BR to 1:5, the diagnostic model can learn richer fault features and more effectively distinguish minority class samples from majority class samples. When the training set reaches balance, all samples are basically correctly identified.

[0048] To further verify the effectiveness of the proposed method, t-SNE algorithm is used for visualization analysis. Figure 8 It shows the feature distribution of the six algorithms under the sample balance condition. The analysis results show that compared with Figure 8 (a)- Figure 8 (e) the feature overlap phenomenon existing in various algorithms, the present invention shows significant advantages. Among them Figure 8 (c) there are obvious overlaps in the data of labels 2, 7, 4, and 8, and in addition, similar problems exist in the ACGAN and WGAN methods. In contrast, through Figure 8 (f) it can be observed that the feature distribution of the samples expanded by the present method shows excellent inter-class discrimination and intra-class aggregation. The experiment proves that the ACWGAN-SG method can generate richer feature representations, thus effectively improving the performance of the unbalanced data fault diagnosis task.

[0049] Embodiment 2 To verify the effectiveness of the method under actual working conditions, a harmonic reducer fault test bench is established. The operating states of the harmonic reducer under different working conditions and fault types are obtained. (The harmonic reducer test bench is shown in the figure).

[0050] 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 harmonic reducers 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 particle brake. The structure of the test system is as shown in Figure 9 the following figure. 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.

[0051] 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 adjusting the speed and operation mode of the motor according to the preset parameters to ensure that the test system works as required.

[0052] 2) Start the data acquisition device 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.

[0053] 3) Load the motor drive program and start the motor control module. 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 acquisition device 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.

[0054] In order to deeply study the signal performance of harmonic reducers with different fault types under complex variable working conditions, 16 different working conditions were designed in this experiment at different speeds and different loads. The motor speeds are divided into 4 levels, which are 295 r / min, 590 r / min, 885 r / min, and 1180 r / min respectively. The loads are 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 5 as follows.

[0055] Table 5 Details of Different Working Conditions of Harmonic Reducers ;

[0056] 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.

[0057] (2) With the load set to 0 N·m, adjust the driving motor speed 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.

[0058] (3) Ensure that the load remains unchanged, and sequentially adjust the 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.

[0059] (4) Continue in the same way to complete the fault tests at the above four speeds under the conditions of loads of 4 N·m, 8 N·m, and 12 N·m respectively.

[0060] (5) After completing the fault test of the external teeth wear of the flexspline, stop the machine to replace the faulty parts and prepare for subsequent tests (6) After completing all the faulty part tests, stop the machine to tidy up the test bench and package and process the data to prepare for subsequent data analysis and fault classification.

[0061] During the experiment, 3 acceleration sensors were used to collect the working condition data of the harmonic reducer. The measurement points were arranged in the X-axis direction, Y-axis direction, and Z-axis direction respectively.

[0062] The measuring point distribution scheme was obtained through enumeration and comparison 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 6.

[0063] Table 6 Summary Table of Artificial Damage Experiments of Harmonic Reducer LSG-25-100-U-Ⅱ ;

[0064] The sliding window sampling method was used for data preprocessing. The window length was set to 1024 and the step size was 400. After preprocessing, 600 samples were randomly selected from each fault category, and 1200 samples were selected from the normal category to jointly form the training set. The test set consisted of 400 samples in each state. The data distribution in the training set showed an unbalanced characteristic, where the normal samples were the majority category and the various fault samples constituted the minority category. Table 7 details the composition structure and specific parameters of the data set.

[0065] Table 7 Data Division of the Fault Platform of the Harmonic Reducer ;

[0066] The Pearson correlation coefficient (PCC) is introduced as a quantitative evaluation index for sample quality, which can effectively measure the linear correlation between the generated data and the source data. When the PCC value exceeds 0.5, it indicates that there is a significant correlation between the data. During the evaluation process, the time-domain signal is first transformed into the frequency-domain signal through Fourier transform, and then the PCC value between the generated samples and the real samples is calculated. As shown in Table 8, compared with the other four comparison methods, the ACWGAN-SG method proposed in the present invention performs better in terms of the PCC index, indicating that the samples generated by it have a higher correlation with the real samples. The experimental results prove that this method has obvious advantages in terms of sample generation quality.

[0067] Table 8 PCC between the samples generated by five methods and the original samples of the harmonic reducer dataset ;

[0068] To evaluate the quality of the samples generated by the proposed ACWGAN-SG method, a progressive data augmentation strategy is adopted, and the generated samples are gradually added to the original imbalanced training set until the number of various fault samples and normal samples reaches balance. The original training set contains 1200 majority-class samples and 12 minority-class samples, and the test set contains 400 samples for each class. Labels 4, 7, and 10 are randomly selected as the minority classes, and other labels are used as the majority classes. The detailed information of the dataset is shown in Table 9, and the augmentation method of the minority classes is shown in Table 10.

[0069] Table 9 Original samples for multi-class imbalanced fault diagnosis ;

[0070] Table 10 Number of samples in the imbalanced state ;

[0071] To comprehensively evaluate the effectiveness of the proposed method, three advanced GAN variants, namely DCGAN, ACGAN, and WGAN networks, are used as comparison methods in this section. To ensure the reliability of the experimental results, five repeated experiments are conducted for each imbalanced ratio, and the final results are averaged, as shown in Figure 10 and Table 11: Table 11 Fault diagnosis accuracy of four methods under imbalanced conditions ;

[0072] Figure 10As can be seen from Table 11, the fault diagnosis accuracy rate is positively correlated with the sample size, which confirms that the data generation technology has a positive impact on the improvement of the classification accuracy rate. When the imbalance rate is in the range of 1:100 to 1:20, due to the limited degree of sample expansion, the performance differences of each method are small, but the method provided by the present invention still shows a relative advantage, and the accuracy rate is increased by 8.292%. As the imbalance rate is further increased to 1:1, the superiority of ACWGAN-SG becomes more significant. In the completely balanced state, the accuracy rate of the proposed method reaches 97.627%, which is 3.51%, 4.57% and 7.84% higher than that of WGAN, ACGAN and DCGAN respectively. To sum up, ACWGAN-SG has significant advantages in dealing with the problem of unbalanced data classification and can effectively improve the accuracy of fault diagnosis.

[0073] To further evaluate the classification results, the confusion matrices of the four algorithms in the sample balanced state are plotted, as Figure 11 shown in (a)-(d) of. The analysis results show that there are significant classification biases in the fault diagnosis of the three comparison methods of DCGAN, ACGAN and WGAN, and their accuracy rates show large fluctuations. In contrast, ACWGAN-SG performs outstandingly in terms of diagnostic accuracy and shows excellent stability at the same time. In addition, since the number of samples of label 11 (healthy state) is sufficient, all methods have obtained ideal results in this category.

[0074] 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, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A test method for a harmonic reducer system with high precision and high reliability, characterized in that: It includes the following steps: S1. Construct a fault diagnosis model of the auxiliary classifier Wasserstein generative adversarial network ACWGAN-SG based on gradient penalty and spectral normalization, and use the imbalanced training set to generate minority-class fault samples; S2. Perform data preprocessing on the vibration signals of the harmonic reducer to form an initial training set; S3. Use the fault diagnosis model of ACWGAN-SG to generate samples for the imbalanced training set, evaluate the usability of the generated samples, and quantify the similarity between the generated samples and the original samples in time-domain and frequency-domain features through the Pearson correlation coefficient PCC and cosine similarity CS to ensure the effectiveness of the generated samples; S4. Add the generated samples to the original training set, adjust the balance ratio of the data set to a preset threshold, and monitor the change of the model performance in real time during the balancing process; S5. Use a deep learning classification model to classify the fault modes of the balanced data set and output the diagnosis results of the health state of the harmonic reducer.

2. The testing method of a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S1, the adversarial network ACWGAN-SG fault diagnosis model integrates the Wasserstein distance, gradient penalty term, and spectral normalization technology. The input of the generator is a random noise vector and a class label. The discriminator constrains the Lipschitz continuity through the gradient penalty term, and its gradient penalty term is defined as: 。 3. A method for testing a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S1, the generator of the ACWGAN-SG fault diagnosis model contains four levels of deconvolution layers, each layer integrating spectral normalization processing. The deconvolution kernel size is 5×5, the stride is 2, and the activation function is ReLU; the discriminator contains four levels of convolution layers, the convolution kernel size is 5×5, the stride is 2, and the activation function is LeakyReLU. The output layer adopts a two-branch structure. The first branch discriminates the authenticity of the sample through the Sigmoid function, and the second branch predicts the class probability through the Softmax function. Its objective function is: ； ； Among them, is the random sample obtained by and online interpolation, is the gradient penalty coefficient, represents the conditional probability distribution on the class labels.

4. A test method for a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S2, the sliding window sampling method is used for data preprocessing. The window length is 1024, the stride is 400, and the time-domain features are extracted and the fault class labels are marked.

5. A method for testing a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S3, the Pearson correlation coefficient PCC between the generated samples and the real samples is ≥0.5, the cosine similarity CS is ≥0.8, and the frequency-domain energy distribution error does not exceed 10%.

6. A test method for a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S4, the balance ratio is defined as: ； wherein, and are the number of samples of the majority class and the minority class respectively, and the preset threshold is 1:

1.

7. A method for testing a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S5, the deep learning classification model is a one-dimensional convolutional neural network 1D-CNN, which contains four levels of convolutional modules. The convolutional kernel size of each level is 5×1, the stride is 2, and the output layer adopts the Softmax function; the input of the 1D-CNN classifier is a one-dimensional vibration signal. After data normalization, it is input into the convolutional layer, and the ReLU activation function and batch normalization layer are connected after each level of convolutional module.

8. A method for testing a harmonic reducer system with high precision and high reliability according to claim 1, characterized in that: In step S3, the initial balance ratio of the imbalanced training set is 1:100, and it is optimized to 1:1 by gradually adding generated samples, and the number of each type of fault sample increases to be equal to the number of normal samples.

Images