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

Abstract

The present invention belongs to the field of machine learning technology and proposes a harmonic reducer fault diagnosis method that combines a deep migration network with fusion samples. The method comprises: using a multi-channel sensor to collect the dynamic characteristic signal of the harmonic reducer and divide it into full cycles; using the dragonfly optimization algorithm to decompose the dynamic characteristic signal after full cycle division and extract a set of intrinsic mode functions; performing a Hilbert transform to obtain the Hilbert spectrum to obtain three axial time-frequency images, and then performing multi-channel image fusion on them using an image integration method in the wavelet domain to construct fused image samples. The fused image samples are divided into training and test sets and then labeled; using the labeled training set to train a CBAM-based fault diagnosis model, while considering domain migration loss and cross entropy loss for reverse parameter adjustment; inputting the test set into the trained CBAM-based fault diagnosis model and outputting the fault diagnosis results. The present invention can accurately obtain fault diagnosis under different working conditions.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a harmonic reducer fault diagnosis method combining a deep migration network with fusion samples. Background Art

[0002] Harmonic reducers are a crucial component of industrial robots. Unlike conventional reducers, they offer advantages such as a large transmission ratio, strong load-bearing capacity, compact size, lightweight, high transmission precision, and high efficiency. They are widely used in aerospace, precision medical devices, industrial robotics, and other fields. The operating status of harmonic reducers is directly related to the efficiency and safety of industrial robots. Given that operating conditions of mechanical equipment may fluctuate or evolve during operation, fault identification of harmonic reducers under different operating conditions is crucial.

[0003] In recent years, numerous research results have demonstrated that data-driven intelligent fault diagnosis methods can achieve high-precision fault detection. Based on the state monitoring signals of the diagnosed object and utilizing pattern recognition algorithms, data-driven intelligent fault diagnosis methods can achieve fault diagnosis by mining and learning the monitoring data of the diagnosed object's operating status without accurately describing the physical model of the diagnosed object. Among them, harmonic reducer fault diagnosis combined with deep learning technology is currently a hot topic in the research field.

[0004] Deep learning models require extensive labeling of samples and extensive datasets for model training, consuming significant human and material resources. While diagnostic results are generally good when the training and test sets meet the same distribution assumption, in actual operation, the samples obtained may lack labels or have different distributions, resulting in suboptimal diagnostic results.

[0005] Due to the complex internal structure of the harmonic reducer and the involvement of multiple nonlinear factors during operation, the signals collected usually exhibit significant non-stationary characteristics, which are manifested as dynamic changes in frequency components over time. This non-stationarity reflects the complexity and diversity of the equipment's operating status, posing a major challenge to signal analysis and feature extraction. A single signal processing method is usually unable to fully capture the multi-dimensional characteristics of the equipment during operation, 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 depiction and diagnosis of the overall operating status 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 non-stationary signal characteristics to comprehensively extract key features and improve the robustness and accuracy of diagnosis. Summary of the Invention

[0006] The purpose of the present invention is to provide a harmonic reducer fault diagnosis method that combines a deep migration network with fusion samples, which can adapt to non-stationary signal characteristics, can diagnose faults under conditions with large differences in different working conditions, and significantly improve the adaptability and accuracy of the diagnostic system.

[0007] The present invention solves the technical problem and adopts the following technical solution:

[0008] The harmonic reducer fault diagnosis method combining deep migration network and fusion samples includes the following steps:

[0009] The dynamic characteristic signals of the harmonic reducer are collected using a multi-channel sensor, and the collected dynamic characteristic signals are divided into full cycles;

[0010] 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;

[0011] 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 the three axes;

[0012] The three-axis time-frequency images are fused using a wavelet domain image integration method to construct fused image samples. The fused samples are then divided into a training set and a test set, and label calibration is performed. The calibrated label training set is used to train a CBAM-based fault diagnosis model, while considering domain transfer loss and cross entropy loss for reverse parameter adjustment.

[0013] The test set is input into the trained CBAM-based fault diagnosis model and the fault diagnosis results are output.

[0014] As a further optimization, the method of collecting the dynamic characteristic signal of the harmonic reducer using a multi-channel sensor includes:

[0015] A harmonic reducer operation test bench was built, and an experiment for collecting dynamic characteristic signals of the harmonic reducer was designed. Harmonic reducers in different conditions were operated under different working conditions, and the dynamic characteristic signals of the harmonic reducer were collected using a multi-channel sensor.

[0016] As a further optimization, when the collected dynamic characteristic signal is divided into full cycles, the cycle sample length is determined by the sampling frequency and the operating conditions of the experimental platform;

[0017] The calculation formula for the entire period is:

[0018] ,

[0019] in, is the sampling frequency; is the rotation speed.

[0020] As a further optimization, 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.

[0021] As a further optimization, 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:

[0022] The dragonfly optimization algorithm uses the information entropy and orthogonality of the signal as the objective function to optimize the number of modes and penalty factor of VMD. After the iteration is completed, the number of modes and penalty factor with the minimum objective function in the iterative process are selected, and the dynamic characteristic signal after period division is decomposed by VMD to obtain the set of intrinsic mode functions.

[0023] As a further optimization, the three axial time-frequency images are fused using the image integration method in the wavelet domain to construct a fused image sample, including:

[0024] Perform wavelet decomposition on the time-frequency images of different channels to obtain the high-frequency and low-frequency components of each sub-image;

[0025] The high-frequency and low-frequency components of each sub-image are fused using corresponding rules. The high-frequency channel adopts the maximum fusion rule, and the low-frequency channel adopts the energy-weighted fusion rule.

[0026] 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.

[0027] 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 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.

[0028] As a further optimization, 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 cross entropy loss are simultaneously considered during the network reverse parameter adjustment process.

[0029] As a further optimization, the output fault diagnosis results include: fault diagnosis results, test set diagnosis confusion matrix effect diagram and feature T-SNE visualization effect diagram.

[0030] The beneficial effects of the present invention are: through the above-mentioned harmonic reducer fault diagnosis method combining deep migration network and fusion sample, on the one hand, since the dynamic signal of the harmonic reducer often shows local non-stationarity, the spectral components of the signal vary greatly over time, and although the traditional wavelet transform method has a certain time-frequency localization capability, for non-stationary signals, the scale selection of the wavelet basis will lead to undesirable time-frequency distribution, affecting the accurate representation of the signal. In view of the complexity of the harmonic reducer structure, and the wavelet transform relies on a preset fixed wavelet basis, it lacks sufficient adaptability to changes and mutations in instantaneous frequency during the signal processing process. Therefore, in view of the complex spectral characteristics of the output signal of the harmonic reducer, the present invention combines the variational mode decomposition (VMD) of the dragonfly optimization algorithm (DA) with the time-frequency analysis strategy of the Hilbert transform, and optimizes the hyperparameters of VMD through the dragonfly optimization algorithm to improve the accurate extraction of the inherent frequency components of VMD in the process of decomposing non-stationary signals, thereby enhancing the energy concentration of the time-frequency spectrum and optimizing the characterization effect of the signal time-frequency characteristics. Therefore, after obtaining the intrinsic mode function set of the dynamic signal, the present invention performs time-frequency image mapping, and then fuses the three-axis time-frequency images through wavelet transform, instead of directly performing time-frequency conversion on the complex signal to construct samples.

[0031] On the other hand, since traditional neural networks pay equal attention to every part of the input data during training, it is difficult to effectively identify and focus on the key features in the input, resulting in unsatisfactory diagnostic results. Moreover, when faced with data with distribution differences, the diagnostic 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, a joint distribution adaptation difference measurement method is added to the fully connected layer and the output layer, which can reduce the data distribution differences between different domains and enhance the model's domain migration diagnostic capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flowchart of a harmonic reducer fault diagnosis method combining a deep migration network and fusion samples in an embodiment of the present invention;

[0033] Figure 2 Schematic diagram of a harmonic speed reduction fault diagnosis test bench in an embodiment of the present invention;

[0034] Figure 3 Schematic diagram of the arrangement of the three-axis acceleration sensors in an embodiment of the present invention;

[0035] Figure 4 In the embodiment of the present invention, an image schematic diagram is generated, wherein: Figure 4 (a) is the channel X time-frequency image, 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;

[0036] Figure 5 This is a schematic diagram of an image generated by wavelet transformation in an embodiment of the present invention, wherein: Figure 5 (a) is the channel X time-frequency image, 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;

[0037] Figure 6 This is a flow chart of the model built in an embodiment of the present invention;

[0038] Figure 7 This is a comparison diagram of migration model diagnosis in an embodiment of the present invention;

[0039] Figure 8 The following are the diagnosis confusion matrix effect diagrams of six models in the embodiment of the present invention, among which: Figure 8 (a) is a confusion matrix effect diagram of the method proposed in this embodiment of the present invention, Figure 8 (b) is the DAN confusion matrix effect diagram, Figure 8 (c) is the DANN confusion matrix effect diagram, Figure 8 (d) is the AFN confusion matrix effect diagram, Figure 8 (e) is the CBAM-Resnet confusion matrix effect diagram, Figure 8 (f) is the Resnet confusion matrix effect diagram;

[0040] Figure 9 This is a comparison chart of the T-SNE feature visualization effects of six models in the embodiment of the present invention, where: Figure 9 (a) is the T-SNE feature visualization effect of the method proposed in the embodiment of the present invention, Figure 9 (b) is the visualization effect of DAN T-SNE features. Figure 9 (c) is the DANN T-SNE feature visualization effect. Figure 9 (d) is the visualization effect of AFN T-SNE features. Figure 9 (e) is the visualization effect of CBAM-Resnet T-SNE features. Figure 9 (f) is the Resnet T-SNE feature visualization effect. DETAILED DESCRIPTION

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0042] Example

[0043] This embodiment provides a harmonic reducer fault diagnosis method that combines a deep migration network with fusion samples. It is mainly divided into a sample generation part and a fault diagnosis part. The specific implementation steps are as follows:

[0044] Step 1: Build a harmonic reducer operating test bench, design a harmonic reducer dynamic characteristic signal acquisition experiment, operate harmonic reducers in different conditions under different working conditions, and use a multi-channel sensor to collect the dynamic characteristic signals of the harmonic reducer, where the dynamic characteristic signals can be vibration signals. Figure 2 In this embodiment, signal acquisition can be performed according to the scheme shown in Table 1.

[0045]

[0046] This embodiment collects vibration signals of harmonic reducers in four different working environments. Each collection can obtain vibration signals of three channels. The sampling frequency is 10KHz and the collection time is 2 minutes. Figure 3 shown.

[0047] Step 2: Divide the collected signal into full cycles. The cycle sample length is determined by the sampling frequency and the operating conditions of the test bench. The full cycle calculation formula is:

[0048] ,

[0049] Where, is the sampling frequency; is the speed. By substituting the experimental conditions in step 1 into the above formula, we can obtain the periodic signal length under each working condition.

[0050] Calculations of the entire cycle indicate that the maximum number of sampling points in the signal cycle is 600. To ensure the integrity and accuracy of the sampling information, the number of sampling points per sample was set to 5000 in the experiment. This setting not only covers the entire cycle of the signal but also ensures high-resolution capture of signal details, providing sufficient data support for subsequent feature extraction and analysis.

[0051] Step 3: Use DA-VMD to extract the intrinsic mode function set of the periodic signal in step 2. First, perform VMD decomposition on the sample signal obtained in step 2 and construct the VMD Lagrangian function as follows:

[0052] ,

[0053] ,

[0054] Where: is the kth eigenmode function, is the center frequency of the kth IMF, is the Lagrange multiplier, is the penalty factor, is the number of modes.

[0055] Based on step 3, the dragonfly optimization algorithm, or DA, is introduced. The information entropy and orthogonality of the signal are used as the objective function to optimize the penalty factor and the number of modes of VMD. 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 in a local range through attraction, repulsion, and cohesion 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, and global exploration is enhanced in the early stage to avoid falling into local extreme values. In the later stage, local development 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 formula of the dragonfly optimization algorithm is as follows:

[0056] Attraction behavior: Each dragonfly tends to move to the center of its neighbors. The expression is as follows:

[0057] ,

[0058] Where: For the dragonfly position; is the number of dragonflies;

[0059] Repulsive behavior: Each dragonfly will stay away from neighbors that are too close to prevent overlap. The expression is as follows:

[0060] ,

[0061] Cohesive behavior: Dragonflies tend to maintain the overall consistency of the group, as expressed by:

[0062] ,

[0063] Migration behavior: Dragonflies are attracted to the target point and move closer to the optimal solution or the global optimal solution. The expression is as follows:

[0064] ,

[0065] Where: is the target point position.

[0066] Inertial behavior: The dragonfly maintains its own movement trend to prevent frequent adjustments that would reduce efficiency. The expression is as follows:

[0067] ,

[0068] The formula for updating the dragonfly's speed and position is as follows:

[0069] ,

[0070] ,

[0071] Where: is the current speed of the dragonfly, is the current position of the dragonfly, and s, a, c, m, and e control the weight factors of attraction, repulsion, cohesion, migration, and inertia behaviors, respectively.

[0072] This embodiment combines the orthogonality of modal components and information entropy as objective functions to comprehensively evaluate the decomposition effect. Orthogonality is used to measure the independence between modal components and indicates whether there is redundancy between components, while information entropy is used to determine the uncertainty of the signal and reflects the concentration and complexity of the component energy distribution. The expression is as follows:

[0073] ,

[0074] Where: 、 are the trade-off coefficients between orthogonality and information entropy. pi is the probability of the modal component, expressed as follows:

[0075] ,

[0076] DA-VMD decomposes non-stationary signals into a set of intrinsic mode functions. These mode functions can effectively separate the different frequency components in the signal through adaptive decomposition, while avoiding the modal aliasing phenomenon that may be caused by traditional decomposition methods. In addition, each mode function maintains the local time-varying characteristics of the signal and accurately describes the frequency variation trend of the signal in different time domain ranges. This decomposition method not only retains the detailed characteristics of the signal, but also provides a high-resolution foundation for subsequent time-frequency analysis, which helps to more deeply explore the dynamic characteristics and physical meaning of the signal.

[0077] After the dragonfly optimization algorithm is used to optimize, the VMD decomposition parameters of each channel signal can be obtained. The VMD optimization parameters are as follows:

[0078]

[0079] Step 4: Substitute the optimized parameters obtained in step 3 into VMD to obtain the modal function of each channel signal sample. Here, the modal components are Hilbert transformed to construct the Hilbert spectrum. All Hilbert spectra are combined and the joint distribution is obtained to obtain the time-frequency image. The modal components obtained in step 3 are Hilbert transformed. The calculation formula is as follows:

[0080] ,

[0081] Where: is the integral variable, representing Position on the timeline;

[0082] Based on the probability formula of the modal component, the Hilbert spectrum is constructed and the calculation formula is as follows:

[0083] ,

[0084] Where: is the modal component; Represents the projection of instantaneous frequency onto the time-frequency plane.

[0085] According to the Hilbert spectrum calculation formula, each modal component is Hilbert transformed to obtain its corresponding Hilbert spectrum. Subsequently, the Hilbert spectra of all modal components are combined and a joint distribution method is used to construct a time-frequency image of the single-channel signal. 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 fully 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.

[0086] Step 5: The time-frequency images obtained in step 4 are fused by wavelet transform to construct fused image samples. The fused samples are divided into source domain dataset, target domain dataset and test set. The target domain dataset is not labeled and has the same working conditions as the test set but different samples. The source domain dataset and the test set are labeled and have different working conditions. The specific process of wavelet transform fusion image is as follows:

[0087] (1) The time-frequency image of the three channels (X, Y, Z) is decomposed by wavelet to obtain the high-frequency and low-frequency components of each sub-image.

[0088] (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 fusion rule and select the maximum value of the three axial high-frequency components as the value of the fused high-frequency component. Low channel fusion rule: Calculate the energy of each axial low-frequency component, assign a weight to each low-frequency component according to the energy size, add the weighted low-frequency components, and obtain the fused low-frequency component. The weight calculation formula is as follows:

[0089] ,

[0090] Where: 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.

[0091] (3) Perform inverse wavelet transform on the fused high-frequency and low-frequency components to obtain the final fused image.

[0092] contrast Figure 4 and Figure 5 ,pass Figure 4 (a) to (d) and Figure 5 As can be seen from (a) to (d), the wavelet-transformed time-frequency image contains a large amount of interference information, especially in the high-frequency portion, resulting in a relatively divergent time-frequency distribution, which causes the fused image to exhibit a diffuse frequency distribution. Due to the complexity of the harmonic reducer structure, the acquired signal is typically non-stationary, with its frequency components varying significantly over time. Since the wavelet transform relies on a fixed wavelet basis, it cannot flexibly respond to instantaneous frequency variations and mutations during signal processing, resulting in a divergent frequency distribution in the generated time-frequency image. The method proposed in this embodiment uses DA-VMD to decompose the non-stationary signal into a set of intrinsic mode functions. These mode functions can effectively separate the signal's different frequency components and preserve the signal's local characteristics. Subsequently, by applying a Hilbert transform to each mode function, its instantaneous frequency information is accurately extracted, generating a high-resolution time-frequency spectrum. Compared with the wavelet transform, this process significantly improves the concentration of the energy distribution, presenting a clearer time-frequency image and providing stronger analytical capabilities and physical interpretability for the signal's local characteristics and dynamic changes.

[0093] After the fusion samples are completed, the samples are divided and the fault diagnosis tasks are formulated as shown in Table 3.

[0094]

[0095] Step 6: Input the source domain and target domain features from step 5 into the diagnostic model. In network training, both domain transfer loss and cross entropy loss are considered, and the network parameters are adjusted inversely.

[0096] The CBAM-Resnt model structure is constructed based on Figure 6 ,The model includes input convolution layer, residual module, CBAM attention module, fully connected layer and output layer;

[0097] The specific processes of each network layer are as follows:

[0098] The input convolution layer uses a convolution kernel to extract local features of the input image. It includes three parts: convolution operation, batch normalization, and activation function. The convolution kernel size is set to 3×3, the stride is 1×1, and the padding is 1×1. The convolution process is expressed as follows:

[0099] ,

[0100] Where: 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; relu represents the activation function, which transforms the normalized result through a nonlinear transformation.

[0101] The residual module consists of three parts: main branch calculation, residual connection, and residual superposition. The main branch consists of two convolution blocks. The first convolution kernel size is set to 3×3, the stride is 2×2, and the padding is 1×1. The second convolution kernel size is set to 3×3, the stride is 1×1, and the padding is 1×1. The expressions of the three processes of the residual module are as follows:

[0102] ,

[0103] ,

[0104] ,

[0105] The channel attention module re-adjusts the feature response of each channel by weighting each channel of the feature map calculated by the residual module, so that the network pays more attention to channels that contain more information. The principle is as follows: As shown. The spatial attention module weights the spatial position in the feature map. Its principle is as follows shown.

[0106] ,

[0107] ,

[0108] Where: represents the sigmoid function, AvgPool(·) represents the feature pooling operation, and f7×7 represents the convolution operation of size 7×7.

[0109] The CBAM calculation process is shown in the following formula:

[0110] ,

[0111] ,

[0112] Where: represents channel attention operation; represents spatial attention operation, Element-wise matrix multiplication.

[0113] After the CBAM module is operated, the initial feature set is obtained. After that, the extracted high-dimensional features are mapped to the target space through the fully connected layer. The calculation expression is as follows:

[0114] ,

[0115] Finally, the output layer is responsible for outputting the network prediction results.

[0116] Due to the different operating conditions of samples in the source and target domains, there are significant distribution differences between the two. This results in poor generalization performance of deep neural networks trained in the source domain in the target domain, resulting in poor diagnostic effectiveness. Transfer learning leverages source domain knowledge to reduce the reliance of target domain data on annotations, overcoming the limitations of traditional machine learning's assumption of identical data distribution and enabling cross-domain knowledge application. Therefore, based on the proposed model, a transfer learning strategy is introduced to achieve feature alignment of cross-domain samples.

[0117] Set the source domain sample set to: , the target domain sample set is: , 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 of MMD is as follows:

[0118] ,

[0119] Where: H represents the regenerated Hilbert space; f(.) represents the feature mapping function; EP[f(.)] represents the mathematical expectation of the mapped samples.

[0120] MMD aligns the marginal distributions of single-dimensional features, which can lead to insufficient feature alignment in high-dimensional data. However, this method ignores the finer-grained inconsistencies in high-order features between the source and target domains. In particular, in complex high-dimensional data, simple mean comparisons cannot fully capture the differences between domains.

[0121] To improve the model's adaptability and enhance the flexibility and expressiveness of feature distribution alignment, this paper introduces a multi-kernel JMMD transfer learning strategy. By combining information and cross-domain alignment mechanisms, combined with bilateral distribution differences, it measures the joint distribution difference between the two domains. On this basis, a multi-kernel function mapping is introduced. By combining multiple different kernel functions, JMMD can calculate the distribution difference between the source and target domains at multiple feature scales and levels. The calculation process is as follows:

[0122] The JMMD expression is as follows:

[0123] ,

[0124] Where: Represents the i-th dimension joint feature of the source domain; Represents the i-th dimension joint feature of the target domain;

[0125] Based on the above formula, the multi-core feature map is introduced and the calculation expression is as follows:

[0126] ,

[0127] Where: Hi represents the i-th kernel function mapping; βi represents the i-th kernel weight;

[0128] Combining the above formula with the cross entropy loss function, the migration loss function expression is constructed as follows:

[0129] ,

[0130] Where: α is the domain migration adaptation factor, N is the number of samples, C is the number of categories, yij∈{0,1} is whether the i-th sample belongs to the j-th category, is the model's predicted probability that the i-th sample belongs to the j-th class.

[0131] The model iterator is Adam, the number of iterations is 50, the learning rate is set to 0.01, and the model parameters are saved after the number of iterations is reached.

[0132] Step 7: Input the test set into the model trained in step 6 and output the fault diagnosis results. To verify the diagnostic effect of the present invention, comparative experiments were conducted on images generated by wavelet transform and the method proposed in the present invention. The experimental results are shown in Tables 3 and 4.

[0133]

[0134]

[0135] Comparing the results in Tables 4 and 5, we can see that when the operating condition gap is 200, both methods demonstrate high diagnostic accuracy, but the proposed method achieves superior performance. Under conditions of large operating condition gaps, the diagnostic accuracy of the proposed method for single-channel samples remains above 80%, and the diagnostic accuracy of the fused samples reaches over 90%. In contrast, the diagnostic performance of the wavelet transform method decreases significantly with increasing operating condition gaps, and the diagnostic performance of its fused samples is significantly lower than that of the proposed method. This further demonstrates the significant superiority of the proposed method in responding to changes in operating condition gaps.

[0136] In order to verify the generalization ability of the method proposed in this example, a diagnostic comparison experiment of six models was designed and carried out. The experimental results are shown in Table 6.

[0137]

[0138] It can be seen from the diagnostic results in Table 6 that the diagnostic results of the method of this embodiment can all reach more than 90%, and the average diagnostic result can reach 95%. Overall, 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 differences in working conditions, the method of this embodiment shows better generalization effect than other methods. When DAN adapts to the alignment of the two domains, it uses the maximum mean difference to measure the distribution difference of the feature layers of the two domains, and it is more effective in diagnosing tasks with smaller differences. However, when it comes to tasks with large difference distributions, 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, and the diagnostic effect is poor. DANN introduces a domain classifier during the training process to determine whether the feature comes from the source domain or the target domain. At the same time, the feature extractor and the task classifier learn together, so that the feature extractor can extract beneficial features. The domain classification loss function is introduced during the network backpropagation process to reduce the distribution difference between the two domains. However, when the working condition difference is 500-1000r / min, the diagnostic effect of this method is not stable. Due to the large working condition difference, the domain classifier has not learned effective features and cannot fully capture the complex relationship between the source domain and the target domain, resulting in unstable diagnostic effect. 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 scarce labels in the target domain and poor adaptability, resulting in poor diagnostic effect. CBAM-Resnet uses the attention mechanism to guide the model to focus on key features and dynamically adjust feature weights to achieve selective enhancement of significant information, thereby improving the model's recognition and extraction capabilities of feature expressions. Its diagnostic effect is better than Resnet.

[0139] Taking Task 10 as an example, confusion matrix effect diagrams and T-SNE feature visualization effects are produced for six models, such as Figure 8 、 Figure 9 As shown, by comparison Figure 8 (a)~(f) correspond to Figure 9 (a)~(f).

[0140] Therefore, this embodiment combines the deep migration network with the harmonic reducer fault diagnosis method of fusion samples, which has the following technical effects:

[0141] (1) The sample generation method proposed in this paper effectively improves the concentration of time-frequency image features by optimizing the time-frequency decomposition strategy, ensuring a clear distribution of signal energy. In addition, combined with the multi-resolution analysis characteristics of the wavelet transform, it fully integrates the time-frequency features of different channels, further enriching the sample's feature information expression capability. Therefore, this method not only improves the sample representation accuracy but also lays the foundation for the accurate diagnosis of subsequent fault diagnosis models.

[0142] (2) The constructed diagnostic model guides the model to focus on key features through the attention mechanism and dynamically adjusts the feature weights to achieve selective enhancement of significant information, thereby improving the model's ability to recognize and extract feature expressions.

[0143] (3) During the model training phase, a domain transfer loss function based on the multi-core JMMD metric is adopted. By introducing this strategy, the model can adaptively adjust the feature space during the training process to minimize the joint distribution difference between the source domain and the target domain, thereby effectively reducing the impact of working condition changes on the diagnostic performance.

[0144] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A harmonic reducer fault diagnosis method combining deep migration network and fusion samples is characterized by: The steps include: The dynamic characteristic signals of the harmonic reducer are collected using a multi-channel sensor, and the collected dynamic characteristic signals are divided into full 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 the three axes; The three axial time-frequency images are fused using the image integration method in the wavelet domain to construct fused image samples. The fused samples are divided into training and test sets, and label calibration is performed. The CBAM-based fault diagnosis model is trained using the labeled training set, while 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 samples 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 the harmonic reducer was designed. Harmonic reducers in different conditions were operated under different working conditions, and the dynamic characteristic signals of the harmonic reducer were collected using a multi-channel sensor.

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 full cycles, the cycle sample length is determined by the sampling frequency and the operating conditions of the test bench; The calculation formula for the full 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 samples 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 to optimize the number of modes and penalty factor of VMD. After the iteration is completed, the number of modes and penalty factor with the minimum objective function in the iterative process are selected, and the dynamic characteristic signal after period division is decomposed by VMD to obtain the set of intrinsic mode functions.

6. The harmonic reducer fault diagnosis method combining deep migration network and fusion samples 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. The high-frequency channel adopts the maximum fusion rule, and the low-frequency 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 samples 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 transfer loss of the training set samples and the test set samples. The sample domain transfer loss and 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 samples 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