Harmonic reducer fault diagnosis method combining deep migration network and fusion sample

Abstract

The invention belongs to the technical field of machine learning, and provides a harmonic reducer fault diagnosis method combining a deep migration network and a fusion sample, which comprises the following steps: collecting dynamic characteristic signals of a harmonic reducer by using a multi-channel sensor, and carrying out full-period division; performing signal decomposition on the dynamic characteristic signals after the whole period division by using a dragonfly optimization algorithm, and extracting an intrinsic mode function set; carrying out Hilbert transform to obtain a Hilbert spectrum, obtaining time-frequency images in three axial directions, carrying out multichannel image fusion on the time-frequency images through an image integration method in a wavelet domain, constructing a fused image sample, dividing the fused image sample into a training set and a test set, and carrying out label calibration; training a CBAM-based fault diagnosis model by using the training set of the calibration label, and performing reverse parameter adjustment by considering domain migration loss and cross entropy loss; and inputting the test set into the trained CBAM-based fault diagnosis model, and outputting a fault diagnosis result. According to the invention, fault diagnosis under different working conditions can be accurately obtained.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and particularly to a fault diagnosis method for harmonic reducers that combines a deep transfer network and fusion samples. Background Art

[0002] The harmonic reducer is an important component of industrial robots. Different from ordinary reducers, it has the advantages of a large transmission ratio, strong load-bearing capacity, small volume, light weight, high transmission accuracy, and high transmission efficiency. It is widely used in fields such as aerospace, precision medical devices, and industrial robots. The operating state of the harmonic reducer is directly related to the working efficiency and safety performance of industrial robots. Given that during the operation of mechanical equipment, the working conditions may fluctuate or evolve. Therefore, it is of great significance to identify faults in harmonic reducers under different working conditions.

[0003] In recent years, numerous research results have proven that data-driven intelligent fault diagnosis methods can perform high-precision fault detection. Based on the state monitoring signals of the diagnostic object, data-driven intelligent fault diagnosis methods can use pattern recognition algorithms to achieve fault diagnosis by mining and learning the monitoring data of the operating state of the diagnostic object without precisely describing the physical model of the diagnostic object. Among them, the fault diagnosis of harmonic reducers combined with deep learning technology is a hot topic in the current research field.

[0004] Deep learning models require a large amount of label calibration for samples and need a large number of data sets to train the model, consuming a large amount of manpower and material resources. A better diagnostic effect can be obtained when the training set and the test set satisfy the same distribution assumption. However, in the actual operation of equipment, the obtained samples may lack labels or the sample distributions may be different, resulting in an unsatisfactory diagnostic effect.

[0005] Due to the complex internal structure of the harmonic reducer and the involvement of various non-linear factors during operation, the signals collected from it usually exhibit significant non-stationary characteristics, manifested as the frequency components changing dynamically over time. This non-stationarity reflects the complexity and diversity of the equipment operating state, posing great challenges to signal analysis and feature extraction. A single signal processing method usually has difficulty comprehensively capturing the multi-dimensional characteristics during the operation of the equipment, which may lead to the loss of key information or one-sided feature description, thereby reducing the accuracy and reliability of data analysis. At the same time, this deficiency will further affect the accurate description and diagnosis of the overall operating state of the equipment, bringing potential risks to equipment health management and fault prediction. Therefore, there is an urgent need for a comprehensive signal processing solution that can efficiently adapt to the characteristics of non-stationary signals to comprehensively extract key features and improve the robustness and accuracy of diagnosis. Summary of the Invention

[0006] The object of the present invention is to provide a harmonic reducer fault diagnosis method combining a deep transfer network and a fusion sample, which can adapt to the characteristics of non-stationary signals, can perform fault diagnosis under large differences in different working conditions, and significantly improve the adaptability and accuracy of the diagnosis system.

[0007] To solve its technical problems, the present invention adopts the following technical solutions: A harmonic reducer fault diagnosis method combining a deep transfer network and a fusion sample includes the following steps: Collect the dynamic characteristic signals of the harmonic reducer using a multi-channel sensor, and perform a full-cycle division on the collected dynamic characteristic signals; Use the dragonfly optimization algorithm to decompose the dynamically characteristic signals after full-cycle division, and extract the set of intrinsic mode functions; Perform Hilbert transform on the set of intrinsic mode functions to obtain the Hilbert spectrum, and take the joint distribution of all Hilbert spectra to obtain the time-frequency images of three axes; Perform multi-channel image fusion on the time-frequency images of the three axes by the image integration method in the wavelet domain to construct a fused image sample, divide the fused sample into a training set and a test set, and perform label calibration; use the training set with calibrated labels to train the CBAM-based fault diagnosis model, and perform reverse parameter adjustment considering both the domain transfer loss and the cross-entropy loss; Input the test set into the trained CBAM-based fault diagnosis model to output the fault diagnosis result.

[0008] As a further optimization, the collection of the dynamic characteristic signals of the harmonic reducer using a multi-channel sensor includes: Build an operating experimental platform for the harmonic reducer, design an experiment for collecting the dynamic characteristic signals of the harmonic reducer, operate the harmonic reducer in different conditions respectively, and collect the dynamic characteristic signals of the harmonic reducer using a multi-channel sensor.

[0009] As a further optimization, when performing full-cycle division on the collected dynamic characteristic signals, the length of the cycle sample is determined by the sampling frequency and the operating conditions of the experimental platform; The calculation formula for the full cycle is: , where is the sampling frequency; is the rotational speed.

[0010] As a further optimization, during the operation of the harmonic reducer, the multi-channel sensor can simultaneously collect the vibration signals of three axes, namely X, Y, and Z, and use the vibration signals as the dynamic characteristic signals, and set the working conditions by adjusting the operating speed of the experimental platform.

[0011] As a further optimization, the step of using the dragonfly optimization algorithm to decompose the dynamic feature signal after the whole cycle division and extract the set of intrinsic mode functions means that: The dragonfly optimization algorithm takes the information entropy and orthogonality of the signal as the objective function, optimizes and solves the number of modes and the penalty factor of VMD. After the iteration is completed, the number of modes and the penalty factor when the objective function is the smallest during the iteration process are selected, and the VMD decomposition is performed on the dynamic feature signal after the period division to obtain the set of intrinsic mode functions.

[0012] As a further optimization, the multi-channel image fusion of the time-frequency images of the three axes is performed by the image integration method in the wavelet domain to construct a fused image sample, including: Perform wavelet decomposition on the time-frequency images of different channels to obtain the high-frequency and low-frequency components of each sub-image; Fuse the high-frequency and low-frequency components of each sub-image according to corresponding rules. Among them, the maximum value fusion rule is used for the high channel, and the energy-weighted fusion rule is used for the low channel; Perform inverse wavelet transform on the fused low-frequency and high-frequency components to obtain the fused image, which is used as a sample.

[0013] As a further optimization, the model structure of the CBAM-based fault diagnosis model is as follows: The first part of the network structure is the input convolutional layer, the second part is the residual module, and the third part is the CBAM attention mechanism. The first three parts are used for initial feature extraction. Two fully connected layers are set at the end of the network. The input size of the first fully connected layer depends on the output feature size of the backbone network, and the second fully connected layer is the output layer.

[0014] As a further optimization, during the training process of the CBAM-based fault diagnosis model, the joint distribution distance of the output feature of the last two fully connected layers of the training set samples and the test set samples is measured by the JMMD metric to obtain the domain transfer loss between the training set samples and the test set samples. During the network backpropagation and parameter adjustment process, both the sample domain transfer loss and the cross-entropy loss are considered.

[0015] As a further optimization, the output fault diagnosis results include: the fault diagnosis result, the diagnostic confusion matrix effect diagram of the test set, and the feature T-SNE visualization effect diagram.

[0016] The beneficial effects of the present invention are as follows: Through the above harmonic reducer fault diagnosis method combining a deep transfer network and fusion samples, on the one hand, since the dynamic signals of harmonic reducers often exhibit local non-stationarity and the spectral components of the signals vary greatly with time. Although the traditional wavelet transform method has certain time-frequency localization ability, for non-stationary signals, the scale selection of the wavelet basis will result in an unsatisfactory time-frequency distribution, affecting the accurate representation of the signals. Given the complexity of the harmonic reducer structure, and also, the wavelet transform depends on a preset fixed wavelet basis and lacks sufficient adaptability to the changes and mutations of the instantaneous frequency during the signal processing. Therefore, aiming at the complex spectral characteristics of the harmonic reducer output signals, the present invention combines the variational mode decomposition (VMD) with the dragonfly optimization algorithm (DA) and the time-frequency analysis strategy of the Hilbert transform. The dragonfly optimization algorithm is used to optimize the hyperparameters of VMD to improve the accurate extraction of the inherent frequency components by VMD during the decomposition of non-stationary signals, thereby strengthening the energy concentration of the time-frequency spectrogram and optimizing the representation effect of the signal time-frequency characteristics. Therefore, after obtaining the set of intrinsic mode functions of the dynamic signals, the present invention performs time-frequency image mapping, and then fuses the three-axis time-frequency images through wavelet transform, rather than directly performing time-frequency conversion on the complex signals to construct samples.

[0017] On the other hand, since the traditional neural network equally focuses on each part of the input data during training, it is difficult to effectively identify and focus on the key features in the input, resulting in an unsatisfactory diagnosis effect. Also, when facing data with distribution differences, the diagnosis effect of the model is often poor. For this reason, the present invention introduces the CBAM attention mechanism based on the residual network to screen important features, thereby improving the model's attention to key features. At the same time, in the fully connected layer and the output layer part, a joint distribution adaptation difference metric method is added, which can reduce the data distribution differences between different domains and enhance the domain transfer diagnosis ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flowchart of the harmonic reducer fault diagnosis method combining a deep transfer network and fusion samples in an embodiment of the present invention; Figure 2 is a schematic diagram of a harmonic reducer fault diagnosis test bench in an embodiment of the present invention; Figure 3 is a schematic diagram of the arrangement of three-axis acceleration sensors in an embodiment of the present invention; Figure 4 is a schematic diagram of the generated image in an embodiment of the present invention, where Figure 4 (a) is the time-frequency image of channel X, Figure 4 (b) is the time-frequency image of channel Y, Figure 4 (c) is the time-frequency image of channel Z, Figure 4 (d) is the fused time-frequency image; Figure 5 Schematic diagram of the image generated by wavelet transform in the embodiment of the present invention, where Figure 5 (a) is the time-frequency image of channel X, Figure 5 (b) is the time-frequency image of channel Y, Figure 5 (c) is the time-frequency image of channel Z, Figure 5 (d) is the fused time-frequency image; Figure 6 Flow chart of the model established in the embodiment of the present invention; Figure 7 Diagnostic comparison chart of the transfer model in the embodiment of the present invention; Figure 8 Effect diagrams of the diagnostic confusion matrices of six models in the embodiment of the present invention, where Figure 8 (a) is the effect diagram of the confusion matrix of the method proposed in this embodiment of the present invention, Figure 8 (b) is the effect diagram of the confusion matrix of DAN, Figure 8 (c) is the effect diagram of the confusion matrix of DANN, Figure 8 (d) is the effect diagram of the confusion matrix of AFN, Figure 8 (e) is the effect diagram of the confusion matrix of CBAM-Resnet, Figure 8 (f) is the effect diagram of the confusion matrix of Resnet; Figure 9 Comparison chart of the visualization effects of the T-SNE features of six models in the embodiment of the present invention, where Figure 9 (a) is the visualization effect of the T-SNE features of the method proposed in the embodiment of the present invention, Figure 9 (b) is the visualization effect of the T-SNE features of DAN, Figure 9 (c) is the visualization effect of the T-SNE features of DANN, Figure 9 (d) is the visualization effect of the T-SNE features of AFN, Figure 9 (e) is the visualization effect of the T-SNE features of CBAM-Resnet, Figure 9 (f) is the visualization effect of the T-SNE features of Resnet. Detailed implementation manners

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0020] Embodiment This embodiment provides a harmonic reducer fault diagnosis method that combines a deep transfer network and fusion samples, which is mainly divided into a sample generation part and a fault diagnosis part. The specific implementation steps are as follows: Step 1: Build a harmonic reducer operation test bench, design a harmonic reducer dynamic characteristic signal acquisition experiment, operate harmonic reducers in different conditions under different working conditions, and use multi-channel sensors to collect the dynamic characteristic signals of the harmonic reducer. The dynamic characteristic signals can be vibration signals. The built harmonic reducer test bench is as Figure 2 shown. In this embodiment, the signal acquisition can be carried out according to the scheme shown in Table 1.

[0021]

[0022] In this embodiment, harmonic reducers in four conditions are collected in four working condition environments. Each collection can obtain vibration signals of three channels. The sampling frequency is 10KHz, and the collection time for each time is 2 minutes. The sensor layout is as Figure 3 shown.

[0023] Step 2: Divide the collected signals into complete cycles. The length of the cycle sample is determined by the sampling frequency and the operating conditions of the test bench; the complete cycle calculation formula is: , where, is the sampling frequency; is the rotational speed. By substituting the experimental conditions in Step 1 into the above formula, the cycle signal lengths under each working condition can be obtained.

[0024] Through the complete cycle calculation, it is obtained that the maximum number of sampling points in a cycle of the signal is 600. To ensure the integrity and accuracy of the sampling information, the number of sampling points for each sample is set to 5000 in the experiment. This setting can not only cover the complete cycle information of the signal but also ensure the high-resolution capture of the signal details, thus providing sufficient data support for subsequent feature extraction and analysis.

[0025] Step 3: Extract the set of intrinsic mode functions from the cycle signals in Step 2 through DA-VMD. First, perform VMD decomposition on the sample signals obtained in Step 2 and construct the VMD Lagrangian function as follows: , , where: is the k-th intrinsic mode function, is the central frequency of the k-th IMF, is the Lagrange multiplier, is the penalty factor, is the number of modes.

[0026] Based on Step 3, the dragonfly optimization algorithm, i.e., DA, is introduced. Taking the information entropy and orthogonality of the signal as the objective function, the penalty factor and the number of modes of VMD are optimized. At the beginning of the optimization, the positions of a group of dragonflies are randomly initialized, and the position of each dragonfly represents a set of VMD parameters. Individuals interact within a local range through attraction, repulsion, and aggregation behaviors to explore the current area; at the same time, they approach the global optimal solution through inertia and migration behaviors. During the optimization process, the behavior weights are dynamically adjusted. In the initial stage, global exploration is enhanced to avoid falling into local extrema, and in the later stage, local exploitation is strengthened to improve the accuracy of the solution. By continuously updating the position, speed, and optimal solution, the global optimal solution is finally output. The behavior calculation formulas of the dragonfly optimization algorithm are as follows: Attraction behavior: Each dragonfly individual tends to move to the center position of neighboring individuals. The expression is as follows: , In the formula: is the position of the dragonfly; is the number of dragonflies; Repulsion behavior: Each dragonfly individual will move away from neighbors that are too close to prevent overlap. The expression is as follows: , Aggregation behavior: Dragonfly individuals tend to maintain the overall consistency of the group. The expression is as follows: , Migration behavior: Dragonfly individuals are attracted by the target point and approach the optimal solution or the global optimal solution. The expression is as follows: , In the formula: is the position of the target point.

[0027] Inertia behavior: Dragonflies maintain their own motion trend to prevent efficiency reduction caused by frequent adjustments. The expression is as follows: , The update formulas for the speed and position of the dragonfly are as follows: , , In the formula: is the current speed of the dragonfly, is the current position of the dragonfly, and s, a, c, m, and e are the weight factors controlling the attraction, repulsion, aggregation, migration, and inertia behaviors respectively.

[0028] In this embodiment, the orthogonality of modal components and information entropy are combined as the objective function to comprehensively evaluate the decomposition effect. Among them, orthogonality is used to measure the independence between modal components, indicating whether there is redundancy between components, while information entropy is used to determine the uncertainty of the signal, reflecting the concentration degree and complexity of the component energy distribution. The expression is as follows: , In the formula: and are the trade-off coefficients of orthogonality and information entropy respectively. pi is the probability of the modal component, and the expression is as follows: , By using DA-VMD, the non-stationary signal is decomposed into a group of intrinsic mode functions. These mode functions can effectively separate different frequency components in the signal through an adaptive decomposition method, while avoiding the modal aliasing phenomenon that may occur in traditional decomposition methods. In addition, each mode function maintains the local time-varying characteristics of the signal, accurately describing the frequency change trend of the signal in different time domain ranges. This decomposition method not only retains the detailed features of the signal, but also provides a basis with high resolution for subsequent time-frequency analysis, which helps to more deeply explore the dynamic characteristics and physical meaning of the signal.

[0029] After optimizing by the dragonfly optimization algorithm, the VMD decomposition parameters of each channel signal can be obtained. The VMD optimization parameters are as follows:

[0030] Step 4: Substitute the optimized parameters obtained in Step 3 into VMD to obtain the mode functions of each channel signal sample. Here, the Hilbert spectrum is constructed by performing the Hilbert transform on the modal components. By synthesizing all Hilbert spectra and taking the joint distribution, a time-frequency image is obtained. The calculation formula for performing the Hilbert transform on the modal components obtained in Step 3 is as follows: , In the formula: is the integration variable, representing the position on the time axis; Based on the probability formula of the modal component, the Hilbert spectrum is constructed, and the calculation formula is as follows: , In the formula: is the modal component; represents projecting the instantaneous frequency onto the time-frequency plane.

[0031] According to the Hilbert spectrum calculation formula, Hilbert transform is performed on each modal component to obtain its corresponding Hilbert spectrum. Subsequently, the Hilbert spectra of all modal components are synthesized, and the time-frequency image of the single-channel signal is constructed in a joint distribution manner. By applying the Hilbert transform to each modal function, the instantaneous frequency and amplitude information of the signal can be accurately extracted, thereby generating a time-frequency spectrum with high resolution and high focus. Therefore, this embodiment can not only comprehensively reflect the local dynamic characteristics of the signal, but also effectively avoid the frequency aliasing phenomenon that may occur in traditional methods, providing a more intuitive and accurate representation for the time-frequency analysis of complex signals.

[0032] Step 5: Perform multi-channel image fusion on the time-frequency image obtained in Step 4 through wavelet transform to construct a fused image sample. Divide the fused sample into a source domain dataset, a target domain dataset, and a test set. The target domain dataset has no label calibration, has the same working condition as the test set but different samples. The source domain dataset and the test set are labeled, and their working conditions are different. Among them, the specific process of wavelet transform for fusing images is as follows: (1) Perform wavelet decomposition on the time-frequency images of the three channels (X, Y, Z) to obtain the high-frequency and low-frequency components of each sub-image.

[0033] (2) For the high-frequency and low-frequency components of each sub-image obtained by decomposition, the following rules are adopted for fusion: High-channel fusion rule: Adopt the maximum value fusion rule, and select the maximum value among the high-frequency components of the three axes as the value of the fused high-frequency component. Low-channel fusion rule: Calculate the energy of each axial low-frequency component, assign weights to each low-frequency component according to the energy size, and add the weighted low-frequency components to obtain the fused low-frequency component. The weight calculation formula is as follows: , In the formula: is the weight of the low-frequency component of the i-th image, Ei is the energy of the i-th axial low-frequency component, and N is the total number of images.

[0034] (3) Perform inverse wavelet transform on the fused high-frequency and low-frequency components to obtain the final fused image.

[0035] Compare Figure 4 and Figure 5 through Figure 4 (a)~(d) and Figure 5It can be seen from (a) to (d) that there is a large amount of interference information in the time-frequency image of wavelet transform. Especially in the high-frequency part, the time-frequency distribution is relatively divergent, resulting in a diffusion-like frequency distribution in the fused image. Due to the complexity of the harmonic reducer structure, the collected signals are usually non-stationary, and the frequency components of the signals change significantly over time. Since wavelet transform depends on a fixed wavelet basis, it cannot flexibly handle instantaneous frequency changes and mutations during signal processing, resulting in a divergent frequency distribution in the generated time-frequency image. The method proposed in this embodiment decomposes the non-stationary signal into a set of intrinsic mode functions through DA-VMD. These mode functions can effectively separate different frequency components of the signal and retain the local characteristics of the signal. Subsequently, by applying the Hilbert transform to each mode function, its instantaneous frequency information is accurately extracted to generate a high-resolution time-frequency spectrum. Compared with wavelet transform, this process significantly improves the concentration of energy distribution, presents a clearer time-frequency image, and provides stronger analytical ability and physical interpretability for the local characteristics and dynamic changes of the signal.

[0036] After the fusion samples are completed, the samples are divided, and a fault diagnosis task is formulated as shown in Table 3.

[0037]

[0038] Step 6: Input the source domain and target domain features in Step 5 into the diagnostic model. During network training, both the domain transfer loss and the cross-entropy loss are considered, and the network parameters are adjusted backward.

[0039] The constructed CBAM-Resnt model structure refers to Figure 6 , and the model includes an input convolutional layer, a residual module, a CBAM attention module, a fully connected layer, and an output layer; The specific processes of each layer of the network are as follows: The input convolutional layer uses a convolutional kernel to extract the local features of the input image, including three parts: convolution operation, batch normalization, and activation function. The convolution part sets the convolution kernel size to 3×3, the stride to 1×1, and the padding to 1×1. The following is the expression of the convolution process: , In the formula: conv represents the convolution operation, which includes the dot product of the input feature map and the convolution kernel and the addition of the bias. bn represents the batch normalization operation, and relu represents the activation function, which performs a non-linear transformation on the normalized result.

[0040] The residual module part includes three parts: main branch calculation, residual connection, and residual superposition. The main branch part consists of two convolutional blocks. The size of the first convolutional kernel is set to 3×3, the stride is 2×2, and the padding is 1×1. The size of the second convolutional kernel is 3×3, the stride is 1×1, and the padding is 1×1. The following are the expressions of three processes in sequence for the residual module part: , , , The channel attention module weights each channel of the feature map calculated by the residual module, readjusts the feature response of each channel, and enables the network to pay more attention to those channels containing more information. Its principle is as shown in Equation . The spatial attention module weights the spatial positions in the feature map. Its principle is as shown in Equation .

[0041] , , In the formula: represents the sigmoid function, AvgPool(·) represents the average pooling operation on the feature, and f7×7 represents the convolutional operation with a size of 7×7.

[0042] The CBAM calculation process is as shown in the following formula: , , In the formula: represents the channel attention operation; represents the spatial attention operation, is the element-wise multiplication of matrices.

[0043] After the operation of the CBAM module, the initial feature set is obtained. Then, through the fully connected layer, the high-dimensional features extracted are mapped to the target space. The calculation expression is as follows: , Finally, the output layer is responsible for outputting the network prediction result.

[0044] Since the working conditions of the samples in the source domain and the target domain are different, there are significant distribution differences between the two, resulting in a decline in the generalization performance of the deep neural network trained on the source domain in the target domain and poor diagnostic effects. Transfer learning utilizes the knowledge of the source domain, reduces the dependence of the target domain data on annotations, overcomes the limitation of the data i.i.d. assumption in traditional machine learning, and realizes the cross-domain application of knowledge. Therefore, a transfer learning strategy is introduced based on the proposed model to achieve feature alignment of cross-domain samples.

[0045] Set the source domain sample set as: , and the target domain sample set as: . The two domains come from distributions P and Q respectively. MMD measures the marginal distribution distance between data by measuring the mean difference between P and Q in the high-dimensional feature space. The definition formula of MMD is as follows: , In the formula: H represents the reproducing Hilbert space; f(.) represents the feature mapping function; EP[f(.)] represents taking the mathematical expectation of the mapped samples.

[0046] MMD aligns the marginal distribution of single-dimensional features, and it is easy to have the problem of insufficient feature alignment in high-dimensional data. However, this method ignores the finer-grained high-order feature inconsistencies between the source domain and the target domain. Especially in complex high-dimensional data, simple mean comparison cannot fully capture the differences between domains.

[0047] To improve the adaptation ability of the model and enhance the flexibility and performance of feature distribution alignment, the present invention introduces a multi-kernel JMMD transfer learning strategy. Through the joint information and cross-domain alignment mechanism, combined with the bilateral distribution difference, the joint distribution difference between the two domains is measured. On this basis, a multi-kernel function mapping is introduced. By combining multiple different kernel functions, JMMD can calculate the distribution differences between the source domain and the target domain from multiple feature scales and levels. The calculation process is as follows: The JMMD expression is as follows: , In the formula: represents the i-th dimensional joint feature of the source domain; represents the i-th dimensional joint feature of the target domain; Based on the above formula, a multi-kernel feature mapping is introduced, and the calculation expression is as follows: , In the formula: Hi represents the i-th kernel function mapping; βi represents the size of the i-th kernel weight; Combining the above formula with the cross-entropy loss function, the expression of the transfer loss function is constructed as follows: , In the formula: α is the domain transfer adaptation factor, N is the number of samples, C is the number of categories, yij∈{0,1} indicates whether the i-th sample belongs to the j-th category, is the prediction probability that the model assigns the i-th sample to the j-th category.

[0048] The model iterator is Adam, the number of iterations is 50 times, the learning rate is set to 0.01, and after reaching the number of iterations, the model parameters are saved.

[0049] Step 7: Input the test set into the model trained in Step 6 to output the fault diagnosis results. To verify the diagnosis effect of the present invention, a comparative experiment on the images generated by wavelet transform and the method proposed in the present invention was carried out respectively. The experimental results are compared as shown in Tables 3 and 4.

[0050]

[0051]

[0052] By comparing the results in Tables 4 and 5, it can be seen that when the working condition difference is 200, both methods show high diagnostic accuracy, but the method proposed in this study has a better performance. Under the condition of a large working condition difference, the diagnostic accuracy of the single-channel samples of this method remains above 80%, and the diagnostic accuracy of the fused samples reaches above 90%. In contrast, the diagnostic performance of the wavelet transform method decreases significantly with the increase of the working condition difference, and the diagnostic effect of its fused samples is also significantly lower than that of the method proposed in this study. This further verifies the significant superiority of this method in dealing with the change of working condition difference.

[0053] To verify the generalization ability of the method proposed in this embodiment, a diagnostic comparison experiment of six models was designed and carried out. The experimental results are shown in Table 6.

[0054]

[0055] As can be seen from the diagnostic results in Table 6, the diagnostic results of the method in this embodiment can all reach over 90%, and the average diagnostic result can reach 95%. Generally speaking, the stability of the method of the present invention is better than that of other methods. When facing Tasks 2, 7, and 10 with large working condition differences, the method in this embodiment shows better generalization effect compared with other methods. When DAN adapts to the alignment of two domains, it uses the maximum mean difference to measure the distribution difference of the feature layers in the two domains, and it has a high effect in diagnosing tasks with small differences. However, when dealing with tasks with a large difference in distribution, only the marginal probability distributions of the two domains are aligned, and the feature differences between the source domain and the target domain cannot be fully utilized, resulting in poor diagnostic effects. DANN introduces a domain classifier during the training process to judge whether the features come from the source domain or the target domain. At the same time, the feature extractor and the task classifier jointly learn, enabling the feature extractor to extract beneficial features. A domain classification loss function is introduced during the backpropagation of the network to reduce the distribution difference between the two domains. However, when the working condition difference is 500 - 1000 r / min, the diagnostic effect of this method is not stable. Due to the large working condition difference, the domain classifier fails to learn effective features and cannot fully capture the complex relationship between the source domain and the target domain, resulting in unstable diagnostic effects. Under large working condition differences, AFN fails to effectively eliminate the global distribution differences between the source domain and the target domain due to local feature normalization. At the same time, AFN has poor adaptability due to the scarcity of labels in the target domain, resulting in poor diagnostic effects. While CBAM-Resnet guides the model to focus on key features and dynamically adjusts the feature weights through the attention mechanism, realizing the selective enhancement of significant information, thereby improving the model's discrimination and extraction ability for feature expressions, and its diagnostic effect is better than that of Resnet.

[0056] Taking Task 10 as an example, the confusion matrix effect diagrams and T-SNE feature visualization effects of six models are made, as Figure 8 , Figure 9 shown. Among them, by comparing Figure 8 (a) - (f) respectively correspond to Figure 9 (a) - (f).

[0057] Therefore, the harmonic reducer fault diagnosis method combining the deep transfer network and the fusion samples in this embodiment has the following technical effects: (1) The sample generation method proposed by the present invention effectively improves the concentration of time-frequency image features and ensures the clear distribution of signal energy by optimizing the time-frequency decomposition strategy. In addition, combining the multi-resolution analysis characteristics of wavelet transform, the time-frequency features of different channels are fully fused, further enriching the feature information expression ability of the samples. Therefore, this method not only improves the representation accuracy of the samples but also lays a foundation for the accurate diagnosis of subsequent fault diagnosis models.

[0058] (2)The constructed diagnostic model uses the attention mechanism to guide the model to focus on key features and dynamically adjust the feature weights, achieving selective enhancement of significant information, thereby improving the model's discrimination and extraction ability of feature expressions.

[0059] (3)In the model training stage, a domain transfer loss function based on the multi-kernel JMMD metric is adopted. By introducing this strategy, the model can adaptively adjust the feature space during training, minimizing the difference in the joint distribution between the source domain and the target domain, thereby effectively reducing the impact of working condition changes on the diagnostic performance.

[0060] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various changes and modifications can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A harmonic reducer fault diagnosis method combining a deep migration network with fusion samples, characterized in that: The steps include: The dynamic characteristic signals of the harmonic reducer are collected by using a multi-channel sensor, and the collected dynamic characteristic signals are divided into whole cycles; The dragonfly optimization algorithm is used to decompose the dynamic characteristic signal after the whole period is divided, and the intrinsic mode function set is extracted; The intrinsic mode function set is Hilbert transformed to obtain the Hilbert spectrum, and the joint distribution of all Hilbert spectra is taken to obtain the time-frequency images of three axes; The three axial time-frequency images are fused by multi-channel image integration method in wavelet domain to construct fused image samples, and the fused samples are divided into training set and test set, and label calibration is performed; The CBAM-based fault diagnosis model is trained using the labeled training set, and the domain transfer loss and cross entropy loss are considered for reverse parameter adjustment. The test set is input into the trained CBAM-based fault diagnosis model and the fault diagnosis results are output.

2. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 1 is characterized in that: The method of collecting the dynamic characteristic signal of the harmonic reducer by using a multi-channel sensor includes: A harmonic reducer operation test bench was built, and an experiment for collecting dynamic characteristic signals of harmonic reducers was designed. Harmonic reducers in different conditions were operated under different working conditions, and multi-channel sensors were used to collect the dynamic characteristic signals of the harmonic reducers.

3. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 2 is characterized in that: When the collected dynamic characteristic signal is divided into whole cycles, the cycle sample length is determined by the sampling frequency and the operating conditions of the experimental platform; The calculation formula for the entire cycle is: ， in, is the sampling frequency; is the rotation speed.

4. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 2 is characterized in that: During the operation of the harmonic reducer, the multi-channel sensor can simultaneously collect vibration signals in the three axes of X, Y and Z, and use the vibration signals as dynamic characteristic signals to set the working conditions by adjusting the operating speed of the test bench.

5. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 1 is characterized in that: The dragonfly optimization algorithm is used to decompose the dynamic characteristic signal after the whole period is divided, and the intrinsic mode function set is extracted, which means: The dragonfly optimization algorithm uses the information entropy and orthogonality of the signal as the objective function, and optimizes the number of modes and penalty factors of VMD. After the iteration is completed, the number of modes and penalty factors with the smallest objective function in the iteration process are selected, and the dynamic characteristic signal after period division is decomposed by VMD to obtain the set of inherent mode functions.

6. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 1 is characterized in that: The method of performing multi-channel image fusion on the three axial time-frequency images by an image integration method in the wavelet domain to construct a fused image sample includes: Perform wavelet decomposition on the time-frequency images of different channels to obtain the high-frequency and low-frequency components of each sub-image; The high-frequency and low-frequency components of each sub-image are fused using corresponding rules, where the high channel adopts the maximum value fusion rule and the low channel adopts the energy-weighted fusion rule; The fused low-frequency and high-frequency components are subjected to inverse wavelet transform to obtain a fused image, which is used as a sample.

7. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 1 is characterized in that: The model structure of the CBAM-based fault diagnosis model is as follows: the first part of the network structure is the input convolution layer, the second part is the residual module, and the third part is the CBAM attention mechanism. The first three parts are used for initial feature extraction. Two fully connected layers are set at the end of the network. The input size of the first fully connected layer depends on the output feature size of the backbone network, and the second fully connected layer is the output layer.

8. The harmonic reducer fault diagnosis method combining deep migration network and fusion samples according to claim 7 is characterized in that: During the training process of the CBAM-based fault diagnosis model, the joint distribution distance of the output features of the two fully connected layers after the training set samples and the test set samples is measured by JMMD to obtain the domain migration loss of the training set samples and the test set samples. The sample domain migration loss and the cross entropy loss are simultaneously considered during the network reverse parameter adjustment process.

9. The harmonic reducer fault diagnosis method combining deep migration network and fusion sample according to claim 1 is characterized in that: The output fault diagnosis results include: fault diagnosis results, test set diagnosis confusion matrix effect diagram and feature T-SNE visualization effect diagram.

Images