Planetary gearbox fault diagnosis method based on improved MobileNetV3-AHFF network

Abstract

The invention discloses a planetary gearbox fault diagnosis method based on an improved MobileNetV3-AHFF network, and belongs to the technical field of fault diagnosis, and the method comprises the following steps: collecting one-dimensional vibration acceleration signals of a planetary gearbox under different working conditions, then employing Chirplet continuous transformation to generate a two-dimensional time-frequency diagram, and making a planetary gearbox fault data set; an SWD-HNNE module is integrated into a MobileNetV3-AHFF network, a measurement mode based on an SWD distance and a semi-supervised learning thought are further introduced, and the improved MobileNetV3-AHFF network, namely the MobileNetV3-AHFF-SWD-HNNE network, is obtained; and carrying out multiple rounds of iterative training on the MobileNetV3-AHFF-SWD-HNNE network, storing an optimal model and taking the optimal model as a fault diagnosis model, and carrying out planetary gearbox fault diagnosis by using the model to obtain a diagnosis result. According to the method, a manifold dimension reduction module is integrated into a MobileNetV3-AHFF network architecture, and feature screening and dimension reduction processing of a high-dimensional fault data set are realized in combination with a measurement mode based on an SWD distance and a semi-supervised learning thought, so that the fault identification precision of the planetary gearbox is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a planetary gearbox fault diagnosis method based on an improved MobileNetV3-AHFF network. Background Art

[0002] As a critical component of rotating machinery, the operating status of planetary gearboxes is directly linked to the performance, reliability, and safety of the entire equipment. Given that planetary gearboxes often operate under high loads, high speeds, and strong vibrations, they face significant operating pressures and wear risks. A fault can significantly impact the operational stability of the transmission system, potentially reducing transmission efficiency. If the fault is not promptly addressed or is severe, it can disrupt system control logic, trigger protective shutdown commands, or cause physical damage to critical equipment components. In extreme cases, a chain reaction of such faults can spread throughout the entire production chain, disrupting and stalling the production process. This carries a high risk of safety incidents caused by equipment loss of control, misoperation, or structural failure, posing a serious threat to personnel safety, production facilities, and the surrounding environment. Therefore, timely and accurate fault diagnosis is crucial to ensure the high reliability and stability of planetary gearboxes.

[0003] With the rapid development of artificial intelligence (AI) technology, fault diagnosis methods based on deep learning network models have garnered widespread attention and research. These methods, by constructing neural network structures, can efficiently extract multi-level features from fault signals, enabling end-to-end intelligent diagnosis from data to features to decision-making. However, fault datasets extracted using deep learning networks typically have high dimensionality, significantly increasing computational complexity and potentially introducing redundant information, impacting model training efficiency and fault identification accuracy. Therefore, employing appropriate feature screening and data dimensionality reduction methods is crucial. The MobileNetV3-AHFF network uses pooling layers to reduce dimensionality and filter features. However, during the data dimensionality reduction process, the pooling layer only retains the maximum or average value within each pooling window. When the pooling window is large or the pooling layer is deep, this can lead to the loss of critical information during network transmission. Manifold learning, a classic dimensionality reduction algorithm, can simultaneously reduce the dimensionality of high-dimensional data while reconstructing the data in a low-dimensional space, preserving key information. Common manifold dimensionality reduction methods include isometric feature mapping (Isomap), multidimensional scaling (MDS), locally linear embedding (LLE), and local tangent space regularization (LTSA). However, these algorithms have difficulty in preserving both the global and local structures of the data when processing high-dimensional data, and are overly dependent on specific hyperparameters, making it impossible to achieve satisfactory dimensionality reduction results.

[0004] The above problems need to be solved urgently. To this end, the present invention proposes a planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to solve the problems of complex high-dimensional data calculation, feature redundancy, and lack of sufficient labeled data in the existing planetary gearbox fault diagnosis. A planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network is provided. The method integrates the manifold dimensionality reduction module into the MobileNetV3-AHFF network architecture, and combines the SWD distance-based measurement method and semi-supervised learning ideas to realize feature screening and dimensionality reduction processing of high-dimensional feature data, retaining the core information of the features while screening the data.

[0006] The present invention solves the above technical problems through the following technical solutions, which include the following steps:

[0007] S1: Collect one-dimensional vibration acceleration signals of the planetary gearbox under different working conditions;

[0008] S2: The collected one-dimensional vibration acceleration signal data is sampled and then Chirplet continuous transformation is applied to generate a two-dimensional time-frequency graph to create a planetary gearbox fault data set, which is then divided into a training set and a test set.

[0009] S3: Integrate the SWD-HNNE module into the MobileNetV3-AHFF network, and then introduce the SWD distance-based measurement method and semi-supervised learning ideas to obtain the improved MobileNetV3-AHFF network, namely the MobileNetV3-AHFF-SWD-HNNE network;

[0010] S4: Perform multiple rounds of iterative training on the MobileNetV3-AHFF-SWD-HNNE network, save the optimal model and use it as the fault diagnosis model, use the model to diagnose planetary gearbox faults and obtain diagnostic results.

[0011] Furthermore, in step S1, the fault status types of the planetary gearbox include crack, normal, missing tooth, broken tooth and wear.

[0012] Furthermore, in step S2, in the MobileNetV3-AHFF-SWD-HNNE network, the planetary gearbox fault dataset is first feature extracted and learned to establish a high-dimensional dataset, and then the SWD-HNNE module is used to introduce a SWD distance-based measurement method and semi-supervised learning ideas to screen and reduce the dimensionality of the high-dimensional dataset.

[0013] Furthermore, when extracting and learning features from the planetary gearbox fault dataset, the AHFF module is used to adaptively fuse local features from different levels, global features, and feature information from previous hierarchical fusion based on the input features. The specific processing process is as follows:

[0014]

[0015] Among them, SE represents the channel attention mechanism, SA represents the spatial attention mechanism, stands for element-by-element multiplication, C n ′ represents the feature output by the SE attention mechanism, S n ′ represents the feature output by the spatial attention mechanism, F n Represents the features generated after the previous level fusion, Represents the features of the down-sampled output by the previous stage of the AHFF module, F n ′ represents the result of fusion of channel and spatial features with the previous stage.

[0016] Then C n ′、S n ′、F n ′ performs adaptive fusion to generate feature F n+1 And input into the inverse residual perceptron.

[0017] Furthermore, in step S2, the SWD distance is used to construct a neighborhood graph during the dimensionality reduction process of the high-dimensional dataset using the SWD-HNNE module. The steps of the SWD distance algorithm are as follows:

[0018] S201: Initial test variables X, Y in a given high-dimensional feature dataset, where X = {X i |i=1,…,n},Y={Y j |j=1,2…,n}, and obtain a new data set by randomly selecting multiple directions for projection:

[0019]

[0020] Here, “·” represents the dot product operation of the vectors.

[0021] S202: Calculate the SWD distance between vector X and vector Y. The formula is as follows:

[0022]

[0023] where π is a rearrangement operation that minimizes the calculated SWD distance.

[0024] Furthermore, in step S2, when the SWD-HNNE module uses the SWD distance to construct the neighborhood graph, it introduces the idea of ​​semi-supervised learning, using the information of labeled data to influence the construction of the neighborhood graph, enhancing the influence of labeled data while effectively utilizing the information of unlabeled data. The specific processing process of semi-supervised learning is as follows:

[0025] S211: For each projection direction θ i , set the weight ω(X i ,Y j )as follows:

[0026]

[0027] Among them, λ>1 means giving higher weight to labeled data;

[0028] S212: Update the calculation of Wasserstein distance. The weighted Wasserstein distance is expressed as:

[0029]

[0030] in, Indicates that in the i-th projection direction, the k-th sample point is The position in Indicates that in the i-th projection direction, the k-th sample point is The position in is the sample pair calculated based on the optimal matching.

[0031] S213: The weighted Wasserstein distance in each projection direction reflects the similarity of the data points in that direction, and the weight of the distance is adjusted according to the label information; the weighted Wasserstein distance is averaged according to the results of each projection direction, and finally the semi-supervised SWD distance measurement method is obtained. The process is as follows:

[0032]

[0033] Where M is the number of projection directions.

[0034] S214: In the process of constructing the neighborhood graph, the weighted Wasserstein distance is used to determine the adjacency relationship between data points. For each pair of data points, if their weighted Wasserstein distance satisfies the set threshold ε, they are considered to belong to the same neighborhood, that is:

[0035] ifW1 semi (X i ,X j )≤ε,thenX i andXj areconnectedintheneighborhoodgraph

[0036] Among them, ε is the set threshold, which is used to control whether the data points are similar enough to be considered as the same neighborhood.

[0037] Compared with the existing technology, the present invention has the following advantages: the planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network integrates the manifold dimensionality reduction module into the MobileNetV3-AHFF network architecture, and combines the SWD distance-based measurement method and semi-supervised learning ideas to realize feature screening and dimensionality reduction processing of high-dimensional fault data sets, effectively improving the fault identification accuracy of planetary gearboxes. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a flow chart of a planetary gearbox fault detection method based on the MobileNetV3-AHFF-SWD-HNNE model in an embodiment of the present invention;

[0039] Figure 2 1 is a waveform diagram of a planetary gearbox under different fault conditions in an embodiment of the present invention;

[0040] Figure 3 Schematic diagram of the Chirple time-frequency transformation process in an embodiment of the present invention;

[0041] Figure 4 Schematic diagram of the AHFF module in MobileNetV3-AHFF in an embodiment of the present invention;

[0042] Figure 5 Schematic diagram of the process of the SWD-HNNE module in an embodiment of the present invention;

[0043] Figure 6 Figure 3 shows the three-dimensional visualization effects of the SWD-HNNE module and other manifold dimensionality reduction algorithms after dimensionality reduction of high-dimensional feature sets in an embodiment of the present invention, where (a) is Isomap, (b) is LLE, (c) is MDS, (d) is LTSA, (e) is HNNE, and (f) is SWD-HNNE. DETAILED DESCRIPTION

[0044] The following is a detailed description of an embodiment of the present invention. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process. However, the protection scope of the present invention is not limited to the following embodiment.

[0045] Example 1

[0046] like Figure 1As shown, this embodiment provides a planetary gearbox fault diagnosis method based on the MobileNetV3-AHFF-SWD-HNNE model, and the specific steps include:

[0047] Step 1: Collect the one-dimensional vibration acceleration signal of the planetary gearbox under different working conditions, such as Figure 2 As shown, the collected different fault status types of the planetary gearbox include: crack, normal, missing tooth, broken tooth and wear, among which normal fault is regarded as a special fault.

[0048] Step 2: Sample the collected one-dimensional vibration acceleration signal data, and then apply Chirplet continuous transformation to generate a two-dimensional time-frequency diagram to create a planetary gearbox fault data set, and divide the fault data set into a training set and a test set.

[0049] Step 3: Build a MobileNetV3-AHFF network to extract features from the fault dataset, and then use the SWD-HNNE algorithm to reduce the dimensionality of the high-dimensional dataset obtained through feature extraction to obtain low-dimensional fault features.

[0050] Step 4: Perform multiple rounds of iterative training on the MobileNetV3-AHFF-SWD-HNNE network, save the optimal model and use it as the fault diagnosis model, and use this model to diagnose planetary gearbox faults.

[0051] As a more specific example, in step 1, a total of 5 types of faults are collected, with a sampling frequency of 20480 Hz, a sample length of 2048, and an overlap sampling rate of 0.5. Five folders are created and named after each type of fault. Then, Chirple transform is used to perform time-frequency transformation on each type of fault sample collected to generate 224×224×3 time-frequency samples. The transformation process is as follows: Figure 3 As shown, finally save the generated time-frequency sample graph to the corresponding folder.

[0052] In this embodiment, the two-dimensional time-frequency diagrams of the five types of faults obtained in step 2 are used as the planetary gearbox fault dataset. The total number of dataset samples is 5000, of which 4000 are training set samples and 1000 are test set samples. The specific sample distribution is shown in Table 1.

[0053] Table 1 Distribution of planetary gearbox fault samples

[0054]

[0055] In this embodiment, the MobileNetV3-AHFF network is used in step 3 to extract features of samples and establish a high-dimensional fault data set. The MobileNetV3-AHFF network mainly uses the AHFF module to effectively fuse features from different levels, extract multi-level and multi-scale feature information, and can automatically adjust the weights of features at different levels according to the characteristics of the data, optimize the feature representation, thereby improving the model's sensitivity to fault features and enhancing the model's ability to identify planetary gearbox faults. In the process of constructing a high-dimensional fault data set, the MobileNetV3-AHFF network generates rich feature representations through the efficient integration of inter-layer information, which can effectively capture complex fault modes and changing patterns. Using the extracted high-dimensional feature data, a more comprehensive fault data set is further constructed, providing a high-quality input basis for subsequent dimensionality reduction processing and fault mode classification.

[0056] like Figure 4 As shown in Figure 1, the AHFF module can adaptively fuse local features, global features, and feature information from different levels of the input features. The specific processing of the AHFF module is as follows:

[0057]

[0058]

[0059] Among them, SE represents the channel attention mechanism, SA represents the spatial attention mechanism, stands for element-by-element multiplication, C n ′ represents the feature output by the SE attention mechanism, S n ′ represents the feature output by the spatial attention mechanism, F n Represents the features generated after the previous level fusion, Represents the features of the down-sampled output by the previous stage of the AHFF module, F n ′ represents the result of fusion of channel and spatial features with the previous stage. Then, C n ′、S n ′、F n ′ performs adaptive fusion to generate feature F n+1 And input into the inverse residual perceptron IRMLP.

[0060] In this embodiment, the SWD-HNNE module in step 3 is constructed based on the semi-supervised Sliced ​​Wasserstein Distance (SWD). HNNE is an integrated learning algorithm that captures the local structure and global characteristics of data at different scales by constructing a tree hierarchy based on 1-NNG, overcoming the limitations of traditional dimensionality reduction algorithms in processing complex high-dimensional data, where local structures are easily distorted and global topology is difficult to maintain. At the same time, the SWD-HNNE module introduces a metric based on SWD distance and a semi-supervised learning idea to optimize the neighborhood graph construction process, making the features after dimensionality reduction more significant in terms of inter-class differences and intra-class consistency. The specific steps of the SWD-HNNE module are as follows:

[0061] (1) Given a high-dimensional feature dataset, the initial variables X and Y are given, where X = {X i |i=1,…,n},Y={Yj|j=1,2…,n},by randomly selecting multiple directions for projection, we can obtain a new data set:

[0062]

[0063] Here, “·” represents the dot product operation of the vectors.

[0064] (2) The SWD distance between vector X and vector Y is calculated as follows:

[0065]

[0066] Here, π is a rearrangement operation that minimizes the calculated distance.

[0067] The construction of the neighborhood graph incorporates semi-supervised learning, leveraging information from labeled data to influence the construction of the neighborhood graph, enhancing the influence of labeled data while effectively utilizing information from unlabeled data. This approach allows for a better fusion of labeled and unlabeled data, thereby enhancing the expressive and inference capabilities of the neighborhood graph in manifold learning and effectively improving the performance of the fault diagnosis model.

[0068] The steps of semi-supervised learning are as follows:

[0069] (1) For each projection direction θ i , set the weight ω(X i ,Y j )as follows:

[0070]

[0071] Among them, λ>1 means giving higher weight to labeled data.

[0072] (2) Update the calculation of Wasserstein distance. The weighted Wasserstein distance can be expressed as:

[0073]

[0074] in, Indicates that in the i-th projection direction, the k-th sample point is The position in Indicates that in the i-th projection direction, the k-th sample point is The position in is the sample pair calculated based on the optimal matching.

[0075] (3) The weighted Wasserstein distance in each projection direction reflects the similarity of the data points in that direction, and the weight of the distance is adjusted by the label information; these weighted Wasserstein distances are averaged according to the results of each projection direction, and finally a semi-supervised SWD distance measurement method is obtained. The process is as follows:

[0076]

[0077] Where M is the number of projection directions.

[0078] (4) In the process of constructing the neighborhood graph, the weighted Wasserstein distance is used to determine the adjacency relationship between data points. For each pair of data points, if their weighted Wasserstein distance satisfies the set threshold ε, they are considered to belong to the same neighborhood, that is:

[0079] ifW1 semi (X i ,X j )≤ε,thenX i andX j areconnectedintheneighborhoodgraph

[0080] Where ε is a set threshold that controls whether data points are similar enough to be considered as belonging to the same neighborhood.

[0081] Based on the above steps, the SWD-HNNE module leverages semi-supervised learning principles and a SWD distance-based metric to effectively preserve the local structure of the data by adaptively adjusting the neighborhood relationships of data points. The algorithm optimizes the similarity assessment between data points by incorporating the SWD distance metric, ensuring that the resulting low-dimensional representation not only accurately reflects the global structure of the data but also effectively preserves intra-class similarity. Furthermore, the introduction of semi-supervised learning enables the algorithm to more accurately partition and classify data with the aid of partial label information, thereby improving the model's robustness and accuracy.

[0082] Based on the above, this embodiment provides a planetary gearbox fault diagnosis method, including collecting status signals of the planetary gearbox under different operating conditions; then using Chirplet transform to convert them into a two-dimensional time-frequency graph with distinct features, and creating a planetary gearbox fault dataset; then building a MobileNetV3-AHFF network, completing network parameter initialization, learning and training the planetary gearbox fault dataset, performing feature extraction, and constructing a high-dimensional dataset; finally, using the SWD-HNNE module to implement feature screening and dimensionality reduction, quantify the dimensionality reduction performance indicators between low-dimensional sample points, and construct a three-dimensional visualization graph.

[0083] Example 2

[0084] like Figure 1 As shown, this embodiment provides a planetary gearbox fault diagnosis method based on the MobileNetV3-AHFF-SWD-HNNE model, and the specific steps include:

[0085] Step 1: Use the Drivetrain Dynamic Simulator (DDS) test bench to obtain the original one-dimensional vibration acceleration signal of the first-stage sun gear under different input shaft speeds and load torque conditions, that is, to obtain the one-dimensional vibration signal of the planetary gearbox.

[0086] Step 2: If Figure 3 As shown in Figure 1, the one-dimensional vibration signal of the planetary gearbox is converted into a two-dimensional time-frequency diagram using Chirplet transform, and the fault data set is divided.

[0087] Step 3: Use the MobileNetV3-AHFF network model to extract features from the planetary gearbox fault dataset. Integrate the SWD-HNNE module into the network and introduce semi-supervised learning to perform data screening and dimensionality reduction. Finally, build the MobileNetV3-AHFF-SWD-HNNE model and iteratively train the fault dataset to achieve fault classification. The number of iterations is 30, and the optimal network model weights are retained. The SGD optimizer is used to update the model parameters. A cosine annealing strategy is used to adjust the algorithm, and the learning rate of the SGD optimizer is dynamically adjusted to adjust the learning rate parameter.

[0088] Step 4: Set the dimension of dimensionality reduction to 3 dimensions, and measure the advantages and disadvantages of the method of the present invention and other dimensionality reduction methods by the data clustering effect after dimensionality reduction. The number of neighbor points of Isomap, LLE, LTSA, and HNNE is set to 150, 100, 120, and 30 respectively, the neighbor point parameter of SWD-HNNE is set to 40, the distance measurement method is SWD measurement method, the data dimension after dimensionality reduction is 3 dimensions, and the number of iterations is 100. At the same time, in order to quantify the dimensionality reduction effect, the inter-class divergence S of the data after dimensionality reduction is calculated. b and intra-class divergence S w The ratio of is used as an indicator to evaluate dimensionality reduction performance. A larger ratio indicates a greater concentration of similar samples and a better dimensionality reduction effect. As shown in Table 2, the dimensionality reduction performance of the SWD-HNNE module is significantly higher than that of Isomap, LLE, MDS, LTSA, and HNNE. It also achieves the highest recognition accuracy for the planetary gearbox fault dataset, demonstrating that the proposed method has certain advantages over other dimensionality reduction methods.

[0089] Table 2 Dimensionality reduction performance index results and accuracy

[0090]

[0091]

[0092] Step 5: According to the three-dimensional clustering effect of Isomap, LLE, MDS, LTSA, HNNE and SWD-HNNE ( Figure 6 shown), by Figure 6 It can be seen that although Isomap basically achieves the separation of the five types of faults, it has the defect of low aggregation; LLE and MDS both show different degrees of sample aliasing after dimensionality reduction; HNNE has relatively better classification and aggregation effects, but still mixes some crack and tooth-chip samples together; only SWD-HNNE not only effectively separates the five fault types, but also has the best aggregation performance of each type of sample.

[0093] In summary, the planetary gearbox fault diagnosis method based on the MobileNetV3-AHFF-SWD-HNNE model in the above embodiment uses Chirplet transform to convert the one-dimensional vibration signal of the planetary gearbox into a two-dimensional time-frequency graph, and produces and divides the fault data set; then, a MobileNetV3-AHFF network is built, deep feature extraction is performed on the two-dimensional time-frequency graph, and the manifold dimensionality reduction module (SWD-HNNE module) is integrated into the network structure. The semi-supervised learning idea is used to realize effective screening and dimensionality reduction of high-dimensional features. Finally, the distribution characteristics of different fault categories are intuitively displayed by constructing a three-dimensional visualization graph, and the Softmax classifier is combined to realize accurate identification of different fault states of the planetary gearbox.

[0094] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network, characterized in that: The following steps are involved: S1: Collect the one-dimensional vibration acceleration signal of the planetary gearbox under different working conditions; S2: Sample the collected one-dimensional vibration acceleration signal data, and then apply Chirplet continuous transformation to generate a two-dimensional time-frequency diagram to make a planetary gearbox fault data set, and divide it into a training set and a test set; S3: Integrate the SWD-HNNE module into the MobileNetV3-AHFF network, and then introduce the SWD distance-based measurement method and semi-supervised learning idea to obtain the improved MobileNetV3-AHFF network, namely the MobileNetV3-AHFF-SWD-HNNE network; S4: Perform multiple rounds of iterative training on the MobileNetV3-AHFF-SWD-HNNE network, save the optimal model and use it as the fault diagnosis model, use the model to diagnose the planetary gearbox fault and obtain the diagnosis result.

2. According to claim 1, a planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network is characterized in that: In step S1, the fault status types of the planetary gearbox include crack, normal, missing teeth, broken teeth and wear.

3. According to claim 1, a planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network is characterized in that: In the step S2, in the MobileNetV3-AHFF-SWD-HNNE network, the planetary gearbox fault data set is firstly feature extracted and learned to establish a high-dimensional data set, and then the SWD-HNNE module is used to introduce a SWD distance-based measurement method and a semi-supervised learning idea to screen and reduce the dimension of the high-dimensional data set.

4. According to claim 3, a planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network is characterized in that: When extracting and learning features from the planetary gearbox fault dataset, the AHFF module is used to adaptively fuse local features, global features, and feature information from different levels according to the input features. The specific processing process is as follows: Among them, SE represents the channel attention mechanism, SA represents the spatial attention mechanism, represents element-by-element multiplication, C n ′ represents the features output by the SE attention mechanism, S n ′ represents the feature output by the spatial attention mechanism, F n Represents the features generated after the previous level fusion, represents the features of the downsampled output by the previous stage of the AHFF module, F n ′ represents the result of fusion of channel and spatial features with the previous stage; Then C n ′、S n ′、F n ′ performs adaptive fusion to generate feature F n+1 And input into the inverse residual perceptron.

5. A planetary gearbox fault diagnosis method based on an improved MobileNetV3-AHFF network according to claim 4, characterized in that: In step S2, the SWD distance is used to construct a neighborhood graph in the process of reducing the dimensionality of the high-dimensional data set using the SWD-HNNE module. The steps of the SWD distance algorithm are as follows: S201: Initial test variables X, Y in a given high-dimensional feature data set, where X = {X i |i=1,…,n},Y={Y j |j=1,2…,n}, by randomly selecting multiple directions for projection, a new data set is obtained: Where "·" represents the dot product operation of the vector; S202: Calculate the SWD distance between vector X and vector Y. The formula is as follows: where π is a rearrangement operation that minimizes the calculated SWD distance.

6. A planetary gearbox fault diagnosis method based on the improved MobileNetV3-AHFF network according to claim 5, characterized in that: In step S2, when the SWD-HNNE module uses the SWD distance to construct the neighborhood graph, the semi-supervised learning idea is introduced, and the information of the labeled data is used to influence the construction of the neighborhood graph, thereby enhancing the influence of the labeled data and effectively utilizing the information of the unlabeled data. The specific processing process of the semi-supervised learning is as follows: S211: For each projection direction θi, set the weight ω(X i ,Y j )as follows: Among them, λ>1 means giving higher weight to labeled data; S212: Update the calculation of Wasserstein distance. The weighted Wasserstein distance is expressed as: in, Indicates that in the i-th projection direction, the k-th sample point is The position in Indicates that in the i-th projection direction, the k-th sample point is The position in is the sample pair calculated based on the optimal match; S213: The weighted Wasserstein distance in each projection direction reflects the similarity of the data points in that direction, and adjusts the weight of the distance through the label information; the weighted Wasserstein distance is averaged according to the results of each projection direction, and finally the semi-supervised SWD distance measurement method is obtained. The processing method of this process is as follows: Where M is the number of projection directions; S214: In the process of constructing the neighborhood graph, the weighted Wasserstein distance is used to determine the adjacency relationship between data points; for each pair of data points, when the weighted Wasserstein distance satisfies the set threshold ε, they are considered to belong to the same neighborhood, that is: ifW1 semi (X i ,X j )≤ε,thenX i andX j areconnectedintheneighborhoodgraph Among them, ε is the set threshold, which is used to control whether the data points are similar enough to be considered as the same neighborhood.

Images