A fault diagnosis method for a planetary gearbox

Abstract

The present invention discloses a fault diagnosis method for planetary gearboxes. This method first converts the fault time series signal into a time-frequency diagram fault sample, and then establishes an improved local fusion generative adversarial network model to make full use of the existing very few fault samples. The network uses the encoding and decoding structure and the local fusion module to flexibly mine sample features, and constructs a generator module embedded with a multi-head attention mechanism, which effectively improves the network feature extraction performance and optimizes the local generation details of the generated samples; and proposes a set of generation quality optimization and evaluation strategies based on SSIM and FID, which further improves the diversity and similarity of the generated samples. In the case of only a very small number of planetary gearbox fault samples, this method can generate a large number of high-quality samples, which are superior to the existing mainstream generative adversarial networks in terms of the diversity and similarity of the generation quality. The generated samples are then used to expand the original fault data set, which significantly improves the accuracy of the fault diagnosis task.

Description

Technical Field

[0001] The present invention relates to the technical field of data-enhanced fault diagnosis, and in particular to a fault diagnosis method for a planetary gearbox. Background Art

[0002] Planetary gearboxes are widely used in critical equipment such as wind turbines, ships, and helicopters. They are prone to failure during long-term operation, impacting the overall performance of the equipment. Therefore, fault diagnosis for planetary gearboxes is of great significance. Thanks to the rapid development of data-driven technologies, numerous researchers are attempting to exploit the rich information contained in planetary gearbox vibration signals to conduct intelligent fault diagnosis research. Due to its superior end-to-end feature learning capabilities, deep learning-based fault diagnosis research has attracted greater attention than traditional data-driven technologies.

[0003] However, the success of deep learning models often depends on a large amount of fault sample training data. However, in real industrial scenarios, effective fault information data is extremely scarce, and it is very difficult to obtain naturally developed mechanical equipment failure data, resulting in data imbalance problems, thereby reducing the accuracy of deep learning-based fault diagnosis.

[0004] Data generation technology aims to capture features from the original data from the perspective of data distribution and then generate new samples with similar distribution. Typical representative algorithms include generative adversarial networks, Wasserstein generative adversarial networks, deep convolutional generative adversarial networks, auxiliary classifier generative adversarial networks and variational autoencoder-generative adversarial networks, which are widely used in the field of fault diagnosis. However, when faced with more demanding situations with few sample data, the following problems still need to be solved in order to further improve the fault diagnosis performance: (1) When the current generative adversarial network is used for small sample fault diagnosis, the input of the generator is mostly random noise or a small amount of additional label information, which fails to fully utilize the different information of each training sample and makes it difficult to extract diverse and representative feature representations; (2) The current generative adversarial network mostly generates original time domain vibration signals or their spectral signals, and rarely explores the extraction of related features of the time-frequency energy distribution of vibration signals; (3) In the current research on fault diagnosis based on generative adversarial networks, the number of each category in the existing fault samples used to train GAN is generally more than 20. However, in the actual engineering field, due to the increasingly expensive cost of labeled data and the long-term accumulation of massive monitoring data, the fault samples that can be obtained will become increasingly limited.

[0005] Therefore, the intelligent fault diagnosis method of planetary gearboxes based on very few samples needs further research. Summary of the Invention

[0006] To solve the above problems, the present invention discloses a fault diagnosis method for a planetary gearbox. The technical solution of the present invention is as follows:

[0007] A fault diagnosis method for a planetary gearbox comprises the following steps:

[0008] Step S1: collecting vibration signals of the planetary gearbox under different fault categories, and using continuous wavelet transform to convert each vibration time series signal segment into a corresponding time-frequency spectrum matrix, and visualizing the time-frequency spectrum matrix into an RGB three-channel time-frequency graph in the form of a heat map to form a time-frequency graph dataset;

[0009] Step S2: Establish a local fusion generative adversarial network model, and use the local fusion generative adversarial network as the basic structure to construct an adversarial network model embedded in a multi-head attention mechanism generator; the adversarial network model embedded in the multi-head attention mechanism generator is used to generate a preset number of generated samples for each fault category;

[0010] Step S3: establishing a generation quality optimization and evaluation strategy based on SSIM and FID, and using the generation quality optimization and evaluation strategy based on SSIM and FID to evaluate and screen the generated samples, thereby obtaining screened generated samples;

[0011] Step S4: Use the time-frequency graph dataset to train the sample generator to obtain a trained adversarial network model embedded in a multi-head attention mechanism generator, and use the trained adversarial network model embedded in a multi-head attention mechanism generator to generate a preset number of generated samples for each fault category, and use the generation quality optimization and evaluation strategy based on SSIM and FID to evaluate and screen the generated samples to obtain screened generated samples;

[0012] Step S5: Combine the filtered samples with the original samples and input them into the intelligent diagnosis network model to perform model optimization training to obtain a trained intelligent diagnosis network model, and use test samples to verify its fault classification effect;

[0013] Step S6: Convert the vibration signal of the planetary gearbox failure into the original time-frequency graph, then input the original time-frequency graph into the trained adversarial network model embedded with the multi-head attention mechanism generator to obtain generated samples, and then evaluate and screen the generated samples through the generation quality optimization and evaluation strategy screening based on SSIM and FID to obtain screened generated samples, input the screened generated samples and the original time-frequency graph set into the trained intelligent diagnosis network model to identify the fault type of the planetary gearbox.

[0014] Further improvements: In step S2, the construction process of the adversarial network model embedded in the multi-head attention mechanism generator is as follows:

[0015] The local fusion generative adversarial network model includes a generator G and a discriminator D. The discriminator D uses four residual blocks as feature extractors. Each residual block contains two convolutional layers and an average pooling layer with a residual connection. Finally, two fully connected layers are used to evaluate the authenticity and classification results of the image respectively. The generator G is a generator embedded in a multi-head attention mechanism, thus constructing a sample generator with a generator embedded in a multi-head attention mechanism.

[0016] The generator G uses the encoding and decoding structure as the main framework. When inputting data, it converts k pictures X={x1,...,x k} are input into the encoder for decoding, and then the local fusion module LFM and random coefficient α as input and generate the fused feature vector Then this fused feature vector Input decoder to get generated picture Finally, the image will be generated Input the discriminator together with the real image X for adversarial training;

[0017] in is the picture decoding vector, LFM represents the local fusion module, represents the fused feature vector, and H represents the decoder;

[0018] The local fusion module LFM first selects one of the feature vectors as the basic feature f base , the remaining k-1 feature vector sets are used as reference features Its purpose is to be the basic feature f base Provide fusion features; then in f base Randomly select n local positions in the , thus obtaining a benchmark local feature φ base ; Then build f ref With φ base The similarity map M is as follows:

[0019]

[0020] Where k represents the number of eigenvectors; f ref express Each reference feature vector in; i∈{1,...,n},j∈{1,...,h*w}, is a similarity matrix, represents the i-th benchmark local feature, represents the jth reference eigenvector, M (i,j) express and The cosine similarity of h represents the image height and w represents the image width; according to the similarity map M, in f ref Find the value corresponding to φ base The most similar corresponding local feature representation φ at each position ref , and then use a random coefficient vector α = [α i ,...,α k ] represents the selected local feature φ ref With the reference local feature φ base Fusion is performed to obtain the local fusion feature φ fuse :

[0021]

[0022] in, And t=1,...,n; Finally, the local fusion feature φ fuse In f base Replace the corresponding position of , so as to obtain a fused feature vector As the output of the LFM module; fuse represents the local fusion feature, α base Represents the base local feature φ base The coefficient vector of represents the tth benchmark eigenvector, α i represents the coefficient corresponding to the i-th local eigenvector, represents the tth local feature representation selected from the i-th reference feature vector.

[0023] Further improvement: In step S2,

[0024] The generator of the multi-head attention mechanism includes an encoder and a decoder; the encoder consists of a multi-head attention module and six convolution blocks; each convolution block contains a convolution layer, Leaky-ReLU activation and batch normalization. The multi-head attention module is used to enable the encoder to quickly obtain the global information of the fault sample and pay more attention to the local key information of the energy distribution in the time-frequency graph, so that when the network performs local feature fusion, it can more accurately find the corresponding position of local features between the baseline image and the reference image; the structure of the decoder is symmetrical with that of the encoder, including a multi-head attention mechanism block, two convolution blocks and four upsampling convolution blocks. The multi-head attention module of the decoder is used to enable the encoder to focus on the generation details of local features, thereby achieving the purpose of improving the generation quality.

[0025] Further improvement: In step S4, the sample generator is optimized using the following loss function:

[0026]

[0027]

[0028] in:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] G represents the generator, D represents the discriminator, X represents the input image, c(X) represents the image category, D(X) represents the discriminator's true or false discrimination result for the real image, D(z) represents the discriminator's true or false discrimination result for the generated image, z = G(X, α) represents the generated image, represents the adversarial loss of the discriminator, represents the adversarial loss of the generator, represents the classification loss of the discriminator, represents the classification loss of the generator, represents the calculation method of the generator reconstruction loss, λ local They are the regularization parameters of the classification loss of the corresponding classifier, the classification loss of the generator, and the local reconstruction loss of the generator; Represents the adversarial loss of the generator Generator classification loss and the generator reconstruction loss The overall optimization function of the generator composed of represents the adversarial loss of the discriminator and the classification loss of the discriminator The overall optimization function of the discriminator composed of

[0035] Further improvement: In step S3, the generation quality optimization and evaluation strategy based on SSIM and FID includes the SSIM index and the FID index;

[0036] The specific calculation formula of SSIM is:

[0037]

[0038] Where μ represents the mean, σ represents the variance, and C1 = (0.01L) 2 , C2=(0.03L) 2, C3=C2 is a constant to maintain the stability of the operation, L is the dynamic range of the pixel value; x represents the original sample, y represents the generated sample, μ x 、μ y , σ xy , σ x and σ y are statistical parameters of x and y pixel values;

[0039] Using the SSIM index, the generated samples in each category are compared with the original samples in terms of brightness, contrast, and structure. When the SSIM index is less than the first threshold, it indicates that the quality of the generated sample is poor and it is removed, thereby improving the similarity of the generated samples. At the same time, the SSIM is also used to compare the similarities between the generated samples in each category. When the SSIM index is greater than the second threshold, one of the generated samples is removed, thereby improving the diversity of the generated samples and finally obtaining the preliminarily screened generated samples.

[0040] The generated samples of the preliminary screening are evaluated using the FID indicator.

[0041] The specific calculation formula of the FID indicator is as follows:

[0042]

[0043] Among them, μ represents the mean, Σ represents the covariance, r represents the real image set, g represents the generated image set, Tr represents the trace of the matrix, μ r represents the mean of the real image set, μ g represents the mean of the generated image set, Σ r represents the covariance of the real image set, Σ g represents the covariance of the generated image set;

[0044] When the FID is higher than the preset third threshold, the first threshold and the second threshold are adjusted to re-screen to obtain a preliminarily screened generated sample, and the FID indicator is evaluated again until the FID is no higher than the preset third threshold.

[0045] Further improvements: the first threshold is 0.25, the second threshold is 0.95, and the third threshold is 100.

[0046] The fault diagnosis method for a planetary gearbox is characterized in that the intelligent diagnosis network model includes a resnet50 intelligent diagnosis network model, a mobilenet intelligent diagnosis network model and a googlenet intelligent diagnosis network model.

[0047] The present invention provides a fault diagnosis method for planetary gearboxes with the following beneficial effects: In step S1, a continuous wavelet transform is used to convert the time-domain vibration signal into a corresponding RGB three-channel time-frequency diagram, clearly describing the relationship between the signal frequency and time, and highlighting the inherent characteristics of the fault signal. In step S2, an improved local fusion generative adversarial network model is established. This network fully utilizes the training image information through the local fusion module, flexibly mines the local features of the data, and constructs a generator module embedded with a multi-head attention mechanism to obtain different local features of the data and assign attention weights. This assists the local fusion module in more accurately matching the local feature locations between images, ultimately achieving the goal of optimizing reconstruction quality. In step S3, a generation quality optimization and evaluation strategy based on SSIM and FID is proposed, which can scientifically evaluate and optimize the similarity and diversity of generated samples from multiple aspects at the individual and overall distribution levels. The proposed method can flexibly mine sample features, fully utilize existing fault samples, and accurately match the local feature locations between images, significantly improving the feature extraction performance of the generative adversarial network and the local generation quality of samples. It generates high-quality fault samples and assists the training of deep learning intelligent fault diagnosis models, thereby improving the stability and accuracy of planetary gearbox fault diagnosis with very few samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. Those skilled in the art can also derive other drawings based on these drawings without inventive work, among which:

[0049] Figure 1 is a flow chart of a fault diagnosis method for a planetary gearbox according to the present invention;

[0050] Figure 2 Schematic diagram of converting a time-domain vibration signal into a time-frequency diagram according to an embodiment of the present invention;

[0051] Figure 3 This is the improved local fusion generative adversarial network framework model diagram in the present invention; DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention.

[0053] Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] In addition, the technical solutions of various examples of the present invention may be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0055] See also Figure 1 As shown, the fault diagnosis method for a planetary gearbox according to the present invention mainly includes the following steps:

[0056] Step S1: collect the vibration signals of the planetary gearbox under different fault modes, and use continuous wavelet transform to convert each vibration time series signal segment into a corresponding time spectrum matrix, and visualize the time spectrum matrix into a RGB three-channel time spectrum diagram in the form of a heat map, see Figure 2 As shown;

[0057] Step S2: Establish an improved local fusion generative adversarial network model. For the specific structure of the model, please refer to Figure 3 As shown. Use the local fusion generative adversarial network as the basic structure and build a generator embedded with a multi-head attention mechanism. The specific details of step S2 are as follows:

[0058] The generative adversarial network mainly consists of a generator G and a discriminator D. For the discriminator, it uses four residual blocks as feature extractors. Each residual block contains two convolutional layers and an average pooling layer with a residual connection. Finally, two fully connected layers are used to evaluate the authenticity of the image and the classification results respectively.

[0059] The generator uses a new encoding and decoding structure as the main framework. When inputting data, k pictures X={x1,...,x k} are input into the encoder for decoding, and then the local fusion module LFM and random coefficient α as input and generate the fused feature vector Then this fused feature vector Input decoder to get generated picture Finally, the image will be generated Together with the real image X, it is input into the discriminator D for adversarial training.

[0060] The local fusion module first selects one of the feature vectors as the basic feature f base , the remaining k-1 feature vectors are used as reference features Its purpose is to be the basic feature f base Provide fusion features. Then in f base Randomly select n local positions in the , thus obtaining a benchmark local feature φ baseThen build f ref With φ base The similarity map M is as follows:

[0061]

[0062] Among them, i∈{1,...,n},j∈{1,...,h*w},g is a similarity matrix. According to this similarity map M, we can ref Find the value corresponding to φ base The most similar corresponding local feature representation φ at each position ref Then use a random coefficient vector α = [α i ,...,α k ] represents the selected local feature φ ref With the reference local feature φ base Fusion is performed to obtain the local fusion feature φ fuse :

[0063]

[0064] in, And t=1,...,n. Finally, we will fusion local features φ fuse In f base Replace the corresponding position of , so as to obtain a fused feature vector As the output of the LFM module.

[0065] The encoder consists of a multi-head attention module and six convolutional blocks. Each convolutional block contains a convolutional layer, Leaky-ReLU activation, and batch normalization (BN). The addition of the multi-head attention module enables the encoder to quickly obtain global information about the fault sample and pay more attention to the local key information of the energy distribution in the time-frequency graph. This facilitates the network to more accurately find the corresponding positions of local features between the baseline image and the reference image when performing local feature fusion. The decoder has a symmetrical structure with the encoder, including a multi-head attention mechanism block, two convolutional blocks, and four upsampling convolutional blocks. The multi-head attention module enables the encoder to focus on the generation details of local features, thereby achieving the goal of improving generation quality.

[0066] The entire improved local fusion generative adversarial network is optimized using the following loss function:

[0067]

[0068]

[0069] in:

[0070]

[0071]

[0072]

[0073]

[0074]

[0075] Where X represents the input image, c(X) represents the image category, and z=G(X,α) represents the generated image. represents the adversarial loss of the discriminator, represents the adversarial loss of the discriminator, represents the classification loss of the discriminator, represents the classification loss of the generator, represents reconstruction losses, λ local They are the regularization parameters of the classification loss of the classifier, the classification loss of the generator, and the local reconstruction loss of the generator.

[0076] Step S3: A set of generation quality optimization and evaluation strategies based on SSIM and FID is proposed to further screen and evaluate the generated samples of the network to improve the diversity and similarity of the generated samples. Step S3 specifically includes the following steps:

[0077] Step 31: Use the SSIM index to compare the comprehensive similarity between the generated samples in each category and the original samples in terms of brightness, contrast, and structure. When the SSIM index is less than 0.25, it means that the quality of the generated sample is poor and it is removed, thereby improving the similarity of the generated samples. At the same time, the SSIM is also used to compare the similarity between the generated samples in each category. When the SSIM index is greater than 0.95, one of the generated samples is removed, thereby improving the diversity of the generated samples. The specific calculation formula of SSIM is:

[0078]

[0079] Where μ represents the mean, σ represents the variance, and C1 = (0.01L) 2 , C2=(0.03L) 2 , C3 = C2 is a constant that maintains operational stability, and L is the dynamic range of pixel values. This step compares the brightness, contrast, and structure of two image samples x and y, primarily to describe the similarity between the two images. The SSIM value range is [-1, 1]. When the two images are identical, the SSIM value is 1.

[0080] Step 32: Use the FID indicator to evaluate the diversity and similarity of the generated sample set after processing. If the generated quality of the processed sample set is lower than that of the original sample set, return to step 31 and adjust the relevant threshold to optimize the process again. The specific calculation formula of FID is:

[0081]

[0082] Here, μ represents the mean, Σ represents the covariance, r represents the real image set, g represents the generated image set, and Tr represents the trace of the matrix. By calculating the Fréchet distance between the generated image set g and the real image set r, the diversity and similarity of the generated images are evaluated simultaneously. A lower FID means that the two distributions are closer, that is, the generated images have higher quality and better diversity.

[0083] Step S4: Use a very small amount of time-frequency graph dataset for network model training, and use the trained generator to generate a large number of generated samples for each of the few fault categories. The quality of the generated samples is optimized using the above optimization and evaluation strategies.

[0084] Step S5: Combine the generated samples with the original samples and input them into the intelligent diagnosis model based on the convolutional neural network to perform model optimization training, thereby obtaining a trained intelligent diagnosis network model, and use test samples to verify its fault classification effect;

[0085] Step S6: using the intelligent diagnosis network model to identify different fault types of the planetary gearbox.

[0086] Example 1

[0087] In Example 1, the planetary gearbox fault data expansion method is used for verification. The experimental equipment mainly includes a motor, a motor controller, a two-stage planetary gearbox and a brake. The two-stage planetary gearbox mainly includes a 32-tooth pinion and an 80-tooth gear on the first-stage input shaft, and a 48-tooth pinion and a 64-tooth gear on the second stage. The speed of the input shaft is measured by a tachometer, and the vibration signal of the gear is measured using an accelerometer, the accelerometer model is PCB608A11. The signal acquisition system is a dSPACE system, and the sampling frequency is set to 20kHz. The experiment introduces 9 different gear conditions on the pinion on the input shaft to collect vibration signal sets of various faults, including health status, missing teeth, root cracks, spalling and tip cutting of five different gears. Then, continuous wavelet transform is used to convert various time-series vibration signals into corresponding time-frequency graphs. To verify the effectiveness of the present invention for data augmentation of a very small number of sample datasets, eight fault categories other than the healthy category were selected as the few-sample categories. To simulate the very small number of samples in real industrial scenarios, six images were used for generative network training in each category. Table 1 lists the specific distribution of the dataset.

[0088] Table 1. Details of the time-frequency graph dataset

[0089]

[0090] Next, we established a boosted local fusion generative adversarial network model and used a very small set of time-frequency graph samples for network model training. Table 2 shows the boosted local fusion generative adversarial network structure and parameters used in this example. The network was stochastically gradient optimized using the Adam optimizer. We first trained the network for 50,000 iterations using a fixed learning rate of 0.0001, then for 50,000 iterations using a linearly decaying learning rate. Gradient penalty regularization was also used to improve training stability.

[0091] Table 2 Network structure and parameters of the method of the present invention

[0092]

[0093]

[0094] In step 4, the trained generator is used to generate a large number of generated samples for each of the minority fault categories. Using a generation quality optimization and evaluation strategy based on SSIM and FID, the network's generated samples are further screened and evaluated. First, the generated samples in each category are compared pairwise with the original samples using the SSIM metric. When the SSIM metric is less than 0.25, the generated sample is considered of poor quality and is removed, thereby improving the similarity of the generated samples. Simultaneously, the generated samples in each category are also compared pairwise using SSIM. When the SSIM metric is greater than 0.95, one of the generated samples is removed, thereby increasing the diversity of the generated samples.

[0095] Then use the FID indicator to evaluate the diversity and similarity of the generated sample set after processing. If the generation quality of the processed sample set is lower than that of the original sample set, it is necessary to adjust the relevant threshold of SSIM and re-optimize the process.

[0096] In step 5, the generated samples are combined with the original samples and fed into a convolutional neural network-based intelligent diagnosis model for model optimization and training, resulting in a more accurate diagnosis model. Test samples are then used to verify the fault classification performance. As shown in Table 3, for the three intelligent diagnosis models (ResNet50, MobileNet, and GoogleNet), the classification accuracy of each model gradually improves as the number of generated samples increases.

[0097] Table 3 Classification results of each diagnostic model after data augmentation

[0098]

[0099]

[0100] In step 6, a deep learning intelligent diagnosis model assisted by generated samples is used to identify different fault types of planetary gearboxes, and ultimately the fault diagnosis task for planetary gearboxes is achieved with very few samples.

[0101] Reference Figure 1 As shown, the content of the present invention is mainly divided into four parts. The first part is to obtain the time-domain vibration signals of the planetary gearbox under different operating conditions and obtain the corresponding time-frequency graph samples through continuous wavelet transform, among which the time-frequency graph samples of the fault signal are extremely rare. The second part is to establish an improved local fusion generative adversarial network model, using the local fusion generative network as the basic structure. At the same time, a multi-head attention mechanism module is added to the generator, thereby improving the image generation quality of the model in the case of very few samples. The third part is to train the generative model with a processed dataset and use the trained generator to generate a large number of generated samples for each minority fault category. The generated samples of the network are further screened and evaluated using a generation quality optimization and evaluation strategy based on SSIM and FID. The fourth part is to combine the generated samples with the original samples and input them into an intelligent diagnosis model based on convolutional neural network for optimization training. Finally, the diagnostic model is used to identify different fault types of the planetary gearbox.

[0102] Reference Figure 2 As shown in the figure, a planetary gearbox fault sample is first obtained through the signal acquisition system. Each sample is a vibration time series signal segment containing 3600 data points. Continuous wavelet transform is then used to convert each vibration time series signal segment into a corresponding time-frequency spectrum matrix. The time-frequency spectrum is then visualized as a heat map with three RGB channels and 64*64 pixels.

[0103] Reference Figure 3 As shown in the figure, the improved local fusion generative adversarial network mainly consists of a generator and a discriminator. The generator mainly includes a local fusion module, an encoder, and a decoder. The encoder consists of a multi-head attention module and six convolutional blocks. The decoder has a symmetrical structure with the encoder, mainly including a multi-head attention mechanism block, two convolutional blocks, and four upsampling convolutional blocks. As for the discriminator, it uses four residual blocks as feature extractors. Each residual block contains two convolutional layers and an average pooling layer with residual connections. Finally, two fully connected layers are used to evaluate the image authenticity and classification results respectively.

[0104] The beneficial effects of this planetary gearbox fault diagnosis method are as follows: the method fully utilizes existing fault samples, flexibly mines sample features, and, by embedding a multi-head attention mechanism, accurately matches local feature locations between images. This significantly improves the feature extraction performance of the generative adversarial network and the quality of local sample generation. It also effectively enhances the accuracy and stability of deep learning fault diagnosis models with extremely small sample sizes.

[0105] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and the embodiments. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A fault diagnosis method for a planetary gearbox, characterized in that: The steps include: Step S1: collecting vibration signals of the planetary gearbox under different fault categories, and using continuous wavelet transform to convert each vibration time series signal segment into a corresponding time-frequency spectrum matrix, and visualizing the time-frequency spectrum matrix into an RGB three-channel time-frequency graph in the form of a heat map to form a time-frequency graph dataset; Step S2: Establish a local fusion generative adversarial network model, and use the local fusion generative adversarial network as the basic structure to construct an adversarial network model embedded in a multi-head attention mechanism generator; the adversarial network model embedded in the multi-head attention mechanism generator is used to generate a preset number of generated samples for each fault category; Step S3: establishing a generation quality optimization and evaluation strategy based on SSIM and FID, and using the generation quality optimization and evaluation strategy based on SSIM and FID to evaluate and screen the generated samples, thereby obtaining screened generated samples; Step S4: Use the time-frequency graph dataset to train the sample generator to obtain a trained adversarial network model embedded in a multi-head attention mechanism generator, and use the trained adversarial network model embedded in a multi-head attention mechanism generator to generate a preset number of generated samples for each fault category, and use the generation quality optimization and evaluation strategy based on SSIM and FID to evaluate and screen the generated samples to obtain screened generated samples; Step S5: Combine the filtered samples with the original samples and input them into the intelligent diagnosis network model to perform model optimization training to obtain a trained intelligent diagnosis network model, and use test samples to verify its fault classification effect; Step S6: Convert the vibration signal of the planetary gearbox failure into the original time-frequency graph, then input the original time-frequency graph into the trained adversarial network model embedded with the multi-head attention mechanism generator to obtain generated samples, and then evaluate and screen the generated samples through the generation quality optimization and evaluation strategy screening based on SSIM and FID to obtain screened generated samples, input the screened generated samples and the original time-frequency graph set into the trained intelligent diagnosis network model to identify the fault type of the planetary gearbox.

2. The fault diagnosis method for a planetary gearbox according to claim 1, characterized in that: In step S2, the construction process of the adversarial network model embedded in the multi-head attention mechanism generator is as follows: The local fusion generative adversarial network model includes a generator G and a discriminator D. The discriminator D uses four residual blocks as feature extractors. Each residual block contains two convolutional layers and an average pooling layer with a residual connection. Finally, two fully connected layers are used to evaluate the authenticity and classification results of the image respectively. The generator G is a generator embedded in a multi-head attention mechanism, thus constructing a sample generator with a generator embedded in a multi-head attention mechanism. The generator G uses the encoding and decoding structure as the main framework. When inputting data, it converts k pictures X={x1,...,x k } are input into the encoder for decoding, and then the local fusion module LFM and random coefficient α as input and generate the fused feature vector Then this fused feature vector Input decoder to get generated picture Finally, the image will be generated Input the discriminator together with the real image X for adversarial training; in is the picture decoding vector, LFM represents the local fusion module, represents the fused feature vector, and H represents the decoder; The local fusion module LFM first selects one of the feature vectors as the basic feature f base , the remaining k-1 feature vector sets are used as reference features Its purpose is to be the basic feature f base Provide fusion features; then in f base Randomly select n local positions in the , thus obtaining a benchmark local feature φ base ; Then build f ref With φ base The similarity map M is as follows: Where k represents the number of eigenvectors; f ref express Each reference feature vector in; i∈{1,...,n},j∈{1,...,h*w}, is a similarity matrix, represents the i-th benchmark local feature, represents the jth reference eigenvector, M (i,j) express and The cosine similarity of h represents the image height and w represents the image width; according to the similarity map M, in f ref Find the value corresponding to φ base The most similar corresponding local feature representation φ at each position ref , and then use a random coefficient vector α = [α i ,...,α k ] represents the selected local feature φ ref With the reference local feature φ base Fusion is performed to obtain the local fusion feature φ fuse : in, And t=1,...,n; Finally, the local fusion feature φ fuse In f base Replace the corresponding position of , so as to obtain a fused feature vector As the output of the LFM module; fuse represents the local fusion feature, α base Represents the base local feature φ base The coefficient vector of represents the tth benchmark eigenvector, α i represents the coefficient corresponding to the i-th local eigenvector, represents the tth local feature representation selected from the i-th reference feature vector.

3. The fault diagnosis method for a planetary gearbox according to claim 2, characterized in that: In step S2, The generator of the multi-head attention mechanism includes an encoder and a decoder. The encoder consists of a multi-head attention module and six convolutional blocks. Each convolutional block contains a convolutional layer, Leaky-ReLU activation, and batch normalization. The multi-head attention module enables the encoder to quickly obtain global information about the fault sample and pay more attention to the local key information of the energy distribution in the time-frequency graph. This facilitates the network to more accurately find the corresponding local features between the baseline image and the reference image when performing local feature fusion. The decoder has a symmetrical structure with the encoder, consisting of a multi-head attention mechanism block, two convolution blocks, and four upsampling convolution blocks. The decoder's multi-head attention module is used to enable the encoder to focus on the generation details of local features, thereby improving the generation quality.

4. The fault diagnosis method for a planetary gearbox according to claim 2, characterized in that: In step S4, the sample generator is optimized using the following loss function: in: G represents the generator, D represents the discriminator, X represents the input image, c(X) represents the image category, D(X) represents the discriminator's true or false discrimination result for the real image, D(z) represents the discriminator's true or false discrimination result for the generated image, z = G(X, α) represents the generated image, represents the adversarial loss of the discriminator, represents the adversarial loss of the generator, represents the classification loss of the discriminator, represents the classification loss of the generator, represents the calculation method of the generator reconstruction loss, λ local They are the regularization parameters of the classification loss of the corresponding classifier, the classification loss of the generator, and the local reconstruction loss of the generator; Represents the adversarial loss of the generator Generator classification loss and the generator reconstruction loss The overall optimization function of the generator composed of represents the adversarial loss of the discriminator and the classification loss of the discriminator The overall optimization function of the discriminator composed of 5. The fault diagnosis method for a planetary gearbox according to claim 1, characterized in that: In step S3, the generation quality optimization and evaluation strategy based on SSIM and FID includes the SSIM index and the FID index; The specific calculation formula of SSIM is: Where μ represents the mean, σ represents the variance, and C1 = (0.01L) 2 , C2=(0.03L) 2 , C3=C2 is a constant to maintain the stability of the operation, L is the dynamic range of the pixel value; x represents the original sample, y represents the generated sample, μ x 、μ y , σ xy , σ x and σ y are statistical parameters of x and y pixel values; Using the SSIM index, the generated samples in each category are compared with the original samples in terms of brightness, contrast, and structure. When the SSIM index is less than the first threshold, it indicates that the quality of the generated sample is poor and it is removed, thereby improving the similarity of the generated samples. At the same time, the SSIM is also used to compare the similarities between the generated samples in each category. When the SSIM index is greater than the second threshold, one of the generated samples is removed, thereby improving the diversity of the generated samples and finally obtaining the preliminarily screened generated samples. The generated samples of the preliminary screening are evaluated using the FID indicator. The specific calculation formula of the FID indicator is as follows: Among them, μ represents the mean, Σ represents the covariance, r represents the real image set, g represents the generated image set, Tr represents the trace of the matrix, μ r represents the mean of the real image set, μ g represents the mean of the generated image set, Σ r represents the covariance of the real image set, Σ g represents the covariance of the generated image set; When the FID is higher than the preset third threshold, the first threshold and the second threshold are adjusted to re-screen to obtain a preliminarily screened generated sample, and the FID indicator is evaluated again until the FID is no higher than the preset third threshold.

6. The fault diagnosis method for a planetary gearbox according to claim 5, characterized in that: The first threshold is 0.25, the second threshold is 0.95, and the third threshold is 100.

7. The fault diagnosis method for a planetary gearbox according to claim 1, characterized in that: The intelligent diagnosis network models include the resnet50 intelligent diagnosis network model, the mobilenet intelligent diagnosis network model and the googlenet intelligent diagnosis network model.

Images