Aero-engine bearing fault diagnosis method based on deep prototype network

Abstract

The application discloses an aero-engine bearing fault diagnosis method based on a deep prototype network, acquires an aero-engine bearing fault data set, and is pretreated by a two-stage signal decomposition module composed of grey wolf optimization variational mode decomposition and complete ensemble empirical mode decomposition with adaptive noise; a fault diagnosis model is constructed, including a hierarchical stacked multi-scale feature extraction module and an attention enhancement module; the multi-scale feature extraction module divides the pretreated data into different branches, performs independent parallel dilated convolution operations in each branch, and fuses the features of the multiple branches to different degrees; the attention enhancement module recalibrates the channel attention and spatial attention after enhancing the multi-scale feature channels; finally, the fault diagnosis result is output through the output layer of the fault diagnosis model; the fault diagnosis model is trained using the data set, and the trained model is used for fault diagnosis of bearing signals.

Description

Technical Field

[0001] This invention relates to the application of deep learning and fault diagnosis technology in the field of rotating machinery, and particularly to a fault diagnosis method for aero-engine bearings based on deep prototype networks. Background Technology

[0002] As a core supporting component of the dual-rotor system of aero-engines, the operating status of inter-shaft bearings directly affects the reliability, safety, and service life of the aero-engine, making them a key monitoring target for aero-engine health management and fault early warning. Under actual aero-engine operating conditions, inter-shaft bearings operate in a complex environment of high speed, high temperature, and high load. They are affected by multiple factors, including variable speed switching, strong coupled vibration of the rotor system, strong background noise from the mechanical structure, and the superposition of vibration signals from multiple components. This results in vibration signals exhibiting significant non-stationarity, multi-component aliasing, and weak fault characteristic representation. Effective fault features are masked by a large amount of redundant noise, significantly increasing the difficulty of feature extraction. Furthermore, aero-engine bearings are precision core components. Conducting fault simulation tests in engineering practice is extremely costly and time-consuming, and the collection of fault samples is easily limited by test conditions and equipment precision. This leads to a high difficulty and limited quantity of effective samples with accurate fault labels, making the small sample size a core pain point in the field of bearing fault diagnosis. Traditional deep neural network models rely on a large number of labeled samples for parameter training, making them prone to overfitting in small-sample scenarios. Their generalization ability and robustness are poor, making them unsuitable for the complex and variable operating conditions of aero-engines. Traditional fault diagnosis methods, such as frequency domain analysis based on Fourier transform and empirical mode decomposition, are insufficiently adaptable to non-stationary signals, easily leading to modal aliasing and endpoint effects. The accuracy and completeness of fault feature extraction are difficult to guarantee, and they often involve single-scale feature analysis, failing to fully exploit fault correlation information at different scales within vibration signals. Furthermore, existing fault diagnosis models that integrate attention mechanisms or multi-scale features are often simple module stacks, lacking end-to-end joint optimization of signal preprocessing, feature extraction, and fault classification. Poor adaptability between modules hinders effective synergy. For the complex diagnostic scenarios of aero-engine bearings—characterized by small samples, strong interference, and non-stationarity—they still suffer from low diagnostic accuracy, poor stability, and insufficient engineering applicability, failing to meet the practical engineering requirements of accurate, efficient, and real-time bearing fault diagnosis for aero-engines. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a fault diagnosis method for aero-engine bearings based on deep prototype networks.

[0004] The objective of this invention is achieved through the following technical solution: A method for diagnosing aero-engine bearing faults based on deep prototype networks, comprising the following steps: Acquire a dataset of aero-engine bearing failures and perform preprocessing; A fault diagnosis model is constructed, which includes a hierarchical stacked multi-scale feature extraction module and an attention enhancement module; The multi-scale feature extraction module divides the preprocessed data into different branches, performs independent parallel dilated convolution operations in each branch, and fuses the features of multiple branches to different degrees. The attention enhancement module enhances the multi-scale feature channels and then recalibrates the channel attention and spatial attention; finally, the fault diagnosis result is output through the output layer of the fault diagnosis model. The fault diagnosis model is trained using a dataset, and the trained model is then used to diagnose bearing signals.

[0005] Furthermore, the construction of the aero-engine bearing fault dataset includes: collecting vibration signal data under different bearing faults, and labeling the bearings according to different fault types, including normal bearings, bearings with outer ring faults, and two types of bearings with inner ring faults.

[0006] Furthermore, the preprocessing process includes: first, format normalization and standardization of the vibration signal; then, a two-stage signal decomposition consisting of Grey Wolf optimized variational mode decomposition and fully integrated empirical mode decomposition with adaptive noise is performed. Specifically, the optimal VMD penalty factor is adaptively searched using the GWO algorithm to complete the signal decomposition. The most representative IMF component is selected by sample entropy evaluation and then subjected to CEEMDAN supplementary decomposition to obtain fine-grained, low-noise sub-band components. Subsequently, the key information sub-bands are spliced ​​into a multi-channel time-series input. Finally, the signal is clipped to a fixed length according to the input dimension requirements of the CBAM-MsePNet model to unify the data scale and complete the entire data preprocessing process.

[0007] Furthermore, the network structure of the fault diagnosis model includes: a multi-scale feature extraction module and an attention enhancement module stacked in layers, with a normalization layer and an activation layer added in the middle, a convolutional layer and a pooling layer added at the front of the model, and a pooling layer and a fully connected layer added at the rear of the model.

[0008] Furthermore, the operation steps of the multi-scale feature extraction module specifically include: zero-padding the input signal, dividing the vibration signal features into three branches, and then... In each branch, independent parallel dilated convolution operations are performed, where... Indicates channel i Separating convolution kernels; It is a passage i The offset of the separating convolution kernel; It is the output of the multi-scale separation convolution kernel on each branch. After obtaining the feature vector of each branch, the features of multiple branches are fused to obtain the convolution features of feature maps at different scales of each branch. The generated multi-scale convolution features are concatenated and added to the original data to obtain the final output feature map.

[0009] Furthermore, the operation steps of the attention enhancement module specifically include: The relevant information of each channel in the vibration signal is obtained by max pooling and average pooling. The global information mapped by the input features is obtained by a multilayer perceptron consisting of two convolutional layers and by addition operation. The global information is then activated by the Sigmoid activation function. The activated global information is used as the channel attention weight, and multiplied with the vibration signal data channel by channel to obtain refined features; The refined features are subjected to max pooling and average pooling along the channel dimension to generate two features. The generated features are then concatenated along the channel dimension, and spatial attention weights are generated by passing them through a convolutional layer and a sigmoid activation function. The spatial attention weights are multiplied digit by digit by the rearranged channel features to obtain the final output features.

[0010] Furthermore, training the fault diagnosis model using the dataset includes: The training set is divided into a support set and a query set, which are then fed into the model for training. According to the formula and Obtain prototype data for each category of the support set and feature vectors for the query set; where, i It is the index of the image vector. q i It is the encoded feature vector; k It is a category; C k It is a category k The prototype; S k It is a category k Supported episode count. It is an encoding function; According to the formula Use Euclidean functions to compute the query set. q i With prototype C k Distance between d ; The distance for each query set is calculated using this formula. Calculated as class probability; Use the model's loss function Update the basic model parameters; repeat the training steps until the maximum number of iterations is reached.

[0011] According to another aspect of the specification, an aero-engine bearing fault diagnosis device based on a deep prototyping network is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the aforementioned aero-engine bearing fault diagnosis method based on a deep prototyping network.

[0012] According to another aspect of the specification, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the aforementioned method for diagnosing aero-engine bearing faults based on deep prototyping networks.

[0013] The beneficial effects of this invention are: it proposes a multi-scale separable convolutional prototype network method (CBAM-MsePNet) that integrates convolutional attention mechanism with two-stage mode decomposition enhancement.

[0014] A two-stage signal decomposition module is constructed, consisting of Grey Wolf Optimized Variational Mode Decomposition (GWO-VMD) and Fully Integrated Empirical Mode Decomposition Adaptive Noise (CEEMDAN). The original vibration signal is stably decomposed and multiple IMF components are obtained by adaptively searching for key penalty factors in VMD using GWO. Subsequently, a sample entropy evaluation strategy is introduced to screen the most representative IMF, and it is further decomposed by CEEMDAN to obtain fine-grained, low-noise sub-band components with weaker mode mixing. Finally, the key information sub-bands are spliced ​​into a multi-channel time-series input to improve feature separability from the source.

[0015] A multi-scale depth-separable convolutional feature extractor is designed. Parallel branches with different receptive fields (e.g., K=3 / 5 / 7) are used to perform hierarchical representation of multi-channel signals. Cross-scale information integration is achieved through channel splicing and convolution fusion, thereby simultaneously capturing high-frequency details of impact-type signals and mid-to-low-frequency structural features of steady-state modulation-type signals.

[0016] Based on the above, a lightweight convolutional block attention module (CBAM) is introduced to sequentially perform channel attention and temporal (spatial) attention recalibration, enabling the network to adaptively emphasize key modalities and key time segments containing fault information, suppress redundant noise interference, and enhance discriminative representation.

[0017] A few-shot classifier is constructed using the metric learning mechanism of prototype networks: prototype vectors of each class are calculated from the support set, and the Euclidean distance between the query sample and each class prototype is measured. Furthermore, the distance is mapped to the posterior probability through Softmax, thereby achieving robust decision-making based on "class center distance".

[0018] By using end-to-end training, joint optimization decomposition to enhance input, and multi-scale separable convolutional feature encoder and attention module parameters, the model can still obtain stable and reliable bearing fault identification results under conditions of scarce samples and operating disturbances, providing a diagnostic method with engineering feasibility for aero-engine health management. Attached Figure Description

[0019] Figure 1 This is a diagram of the multi-scale module feature extraction structure provided in an embodiment of the present invention; Figure 2 This is a diagram of the CBAM attention weight allocation structure provided in an embodiment of the present invention; Figure 3 The category prototype and query vector graph provided for embodiments of the present invention; Figure 4 The first test result image of the CBAM-MsePNet model provided in the embodiment of the present invention; Figure 5 The second result image of the CBAM-MsePNet model test provided in the embodiment of the present invention; Figure 6 The third result of the test of the CBAM-MsePNet model provided in the embodiment of the present invention is shown in the figure. Figure 7 The fourth result figure for testing the CBAM-MsePNet model provided in this embodiment of the invention; Figure 8 This is a schematic diagram of an aero-engine bearing fault diagnosis device based on a deep prototype network, provided in an embodiment of the present invention. Detailed Implementation

[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown, the present invention provides a method for diagnosing aero-engine bearing faults based on deep prototype networks, comprising: Data collection. Vibration signal data under different bearing faults were collected by publicly available bearing fault vibration signal data. A total of 4 fault types were identified.

[0022] Fault type marking. Based on different bearing fault types, including normal bearing, bearing with outer ring fault depth of 0.5mm, bearing with inner ring fault depth of 0.5mm, and bearing with inner ring fault depth of 1mm, the corresponding labels are 1, 2, 3, and 4, respectively.

[0023] The dataset was randomly divided from the bearing fault signals collected in step one into training data, validation data, and test data. The training data comprised 60% (180 samples), the validation data comprised 20% (60 samples), and the test data comprised 20% (60 samples). These data were used to train and test the CBAM-MsePNet model.

[0024] Data preprocessing, for the divided training, validation, and test datasets, first involves format normalization and standardization of the vibration signals. Then, a two-stage signal decomposition is performed, consisting of Grey Wolf Optimized Variational Mode Decomposition (GWO-VMD) and Fully Ensemble Empirical Mode Decomposition Adaptive Noise (CEEMDAN). Specifically, the GWO algorithm adaptively searches for the optimal penalty factor for VMD and completes signal decomposition. The most representative IMF components are selected based on sample entropy evaluation, and then further decomposed using CEEMDAN to obtain fine-grained, low-noise sub-band components. Subsequently, key information sub-bands are concatenated into a multi-channel time-series input. Finally, the signal is pruned to a fixed length according to the input dimension requirements of the CBAM-MsePNet model to unify the data scale, completing the entire data preprocessing process. This improves the separability of fault features from the source, providing standardized, high-quality signal data for subsequent feature extraction. The CBAM-MsePNet fault diagnosis model is constructed from multi-scale modules, CBAM attention, and fully connected layers. The multi-scale modules and CBAM attention are stacked in layers, with a normalization layer and activation layer added in the middle, totaling four blocks. Additionally, a convolutional layer and a pooling layer are added at the beginning of the model, and a pooling layer and a fully connected layer are added at the end. Batch size is set. B batch =64, maximum number of training iterations n epoch =120. Use Dropout to avoid overfitting. The default Dropout rate is set to 0.5.

[0025] The multi-scale feature extraction module divides the preprocessed vibration signal data into three branches for multi-scale feature extraction and finally merges the results. The specific steps are as follows: Step 1: Assume the preprocessed vibration signal data is as follows: , . Representing data points X, C These represent the length and width of the signal data, respectively. Zero padding was used to maintain consistency in the magnitude of vibration signal characteristics at different scales. According to the formula... The vibration signal characteristics are divided into three branches. X and C represent the length and width of the signal data, respectively. H represents the input data.

[0026] Step 2: According to the formula In each branch, independent parallel dilated convolution operations are performed, where... This represents three different scales of separate convolution kernels. These are the offsets of the three different scales of the separated convolution kernels. It is the output of the multi-scale separating convolution kernel on each branch.

[0027] Step 3: After obtaining the feature vectors of each branch, the features of multiple branches can be fused to varying degrees, as shown in the following formula. To obtain richer features. Among them, T (M) This indicates the degree of fusion of feature maps at different scales in each branch.

[0028] Step 4: According to the formula The generated multi-scale convolutional features are concatenated, where Feature fusion calculation representing features at different scales. F output This represents the output of multi-scale feature fusion. Concat It is a feature fusion function.

[0029] Step 5: T 1 , T 2 and T 3 These represent the calculation results for the first, second, and third fusion blocks, respectively; the three branches... T 1 , T 2 and T 3 This corresponds to different combinations of dilation rates (D) and different convolution kernel sizes (K). The signal... H join in F output You can get .Depend on The final output of the multi-scale module is obtained. X It is the original feature of the previous layer. This is the final output of the feature map. Multi-scale structures, such as... Figure 1 As shown.

[0030] The attention enhancement module (CBAM) introduces channel attention and temporal (spatial) attention after fusing features from the multi-scale feature extraction module. It adaptively reweights key modalities and key time segments to suppress redundant noise interference and highlight fault-sensitive features.

[0031] The steps are as follows: Step 1: Assume the input vibration signal data is First, the channel attention module is used to obtain the channel attention value through max pooling and average pooling. The relevant information for each channel. and .

[0032] Step 2: Then, the global information mapped by the input feature F is obtained through a multilayer perceptron (MLP) consisting of two convolutional layers and an addition operation. Then, select the Sigmoid activation function to activate the global information and obtain... .

[0033] Step 3: Channel attention weights can be generated by activating global information. P c and F in By multiplying each channel, a refined feature can be obtained. .

[0034] Step 4: The enhanced features are further input into the spatial attention module. First, max pooling and average pooling are performed along the channel dimension to generate two features, resulting in... and .

[0035] Step 5: Connect them along the channel dimension, and then generate spatial attention weights through convolutional layers and a sigmoid activation function to obtain... Finally, the spatial attention weights are multiplied bit by bit with the channel rearranged features to obtain the final output features. The structure is as follows: Figure 2 As shown.

[0036] Model training: Input the training data into the constructed CBAM-MsePNet model for training to obtain the trained CBAM-MsePNet. This includes: The training data is input, and the training set is divided into a support set and a query set, which are then input into the CBAM-MsePNet model for training.

[0037] According to the formula and Obtain prototype data for each category of the support set and feature vectors for the query set. Among these, i It is the index of the image vector. q i It is the encoded feature vector. k It is a category. C k It is a category k The prototype. Sk It is a category k Supported episode count. It is an encoding function. The prototype category and query vector point are as follows: Figure 3 As shown.

[0038] According to the formula Use Euclidean functions to compute the query set. q i With prototype C k Distance between d .

[0039] The distance calculated for each query set can be obtained using this formula. Calculated as class probability.

[0040] Calculate the loss function of each model. The trained CBAM-MsePNet fault diagnosis model is obtained.

[0041] Update the basic model parameters. Return to the input training data and repeat the training until the maximum number of iterations is reached, finally obtaining the trained CBAM-MsePNet model.

[0042] like Figure 4 , Figure 5 , Figure 6 and Figure 7 As shown, the CBAM-MsePNet fault diagnosis model is tested by inputting test data into the trained CBAM-MsePNet model to obtain the probability distribution of each label and the confusion matrix, thus obtaining the fault identification result.

[0043] Corresponding to the aforementioned embodiment of an aero-engine bearing fault diagnosis method based on deep prototype networks, the present invention also provides an embodiment of an aero-engine bearing fault diagnosis device based on deep prototype networks.

[0044] See Figure 8 The present invention provides an aero-engine bearing fault diagnosis device based on a deep prototype network, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an aero-engine bearing fault diagnosis method based on a deep prototype network as described in the above embodiment.

[0045] The embodiment of the aero-engine bearing fault diagnosis device based on deep prototyping networks provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, such as... Figure 8 The diagram shown is a hardware structure diagram of any data processing-capable device, including the aero-engine bearing fault diagnosis device based on deep prototyping networks provided by this invention. (Except for...) Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0046] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0047] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0048] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements a method for diagnosing aero-engine bearing faults based on a deep prototype network as described in the above embodiments.

[0049] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0050] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method for diagnosing aero-engine bearing faults based on deep prototyping networks.

[0051] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0052] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for diagnosing aero-engine bearing faults based on deep prototype networks, characterized in that, Includes the following steps: Acquire a dataset of aero-engine bearing failures and perform preprocessing; A fault diagnosis model is constructed, which includes a hierarchical stacked multi-scale feature extraction module and an attention enhancement module; The multi-scale feature extraction module divides the preprocessed data into different branches, performs independent parallel dilated convolution operations in each branch, and fuses the features of multiple branches to different degrees. The attention enhancement module enhances the multi-scale feature channels and then recalibrates the channel attention and spatial attention; finally, the fault diagnosis result is output through the output layer of the fault diagnosis model. The fault diagnosis model is trained using a dataset, and the trained model is then used to diagnose bearing signals.

2. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The construction of the aero-engine bearing fault dataset includes: collecting vibration signal data under different bearing faults, and labeling the bearings according to different fault types, including normal bearings, bearings with outer ring faults, and two types of bearings with inner ring faults.

3. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The preprocessing process includes: first, format normalization and standardization of the vibration signal; then, a two-stage signal decomposition consisting of Grey Wolf optimized variational mode decomposition and fully integrated empirical mode decomposition with adaptive noise; that is, adaptively searching for the optimal VMD penalty factor through the GWO algorithm and completing the signal decomposition; selecting the most representative IMF component through sample entropy evaluation; then performing CEEMDAN supplementary decomposition to obtain fine-grained, low-noise sub-band components; subsequently, splicing the key information sub-bands into a multi-channel time-series input; and finally, pruning the signal to a fixed length according to the input dimension requirements of the CBAM-MsePNet model to unify the data scale and complete the entire data preprocessing process.

4. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The network structure of the fault diagnosis model includes: a multi-scale feature extraction module and an attention enhancement module stacked in layers, with a normalization layer and an activation layer added in the middle, a convolutional layer and a pooling layer added at the front of the model, and a pooling layer and a fully connected layer added at the rear of the model.

5. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The operation steps of the multi-scale feature extraction module specifically include: zero-padding the input signal, dividing the vibration signal features into three branches, and then, according to the formula... In each branch, independent parallel dilated convolution operations are performed, where... Indicates channel i Separating convolution kernels; It is a passage i The offset of the separating convolution kernel; It is the output of the multi-scale separation convolution kernel on each branch. After obtaining the feature vector of each branch, the features of multiple branches are fused to obtain the convolution features of feature maps at different scales of each branch. The generated multi-scale convolution features are concatenated and added to the original data to obtain the final output feature map.

6. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The specific operation steps of the attention enhancement module include: The relevant information of each channel in the vibration signal is obtained by max pooling and average pooling. The global information mapped by the input features is obtained by a multilayer perceptron consisting of two convolutional layers and by addition operation. The global information is then activated by the Sigmoid activation function. The activated global information is used as the channel attention weight, and multiplied with the vibration signal data channel by channel to obtain refined features; The refined features are subjected to max pooling and average pooling along the channel dimension to generate two features. The generated features are then concatenated along the channel dimension, and spatial attention weights are generated by passing them through a convolutional layer and a sigmoid activation function. The spatial attention weights are multiplied digit by digit by the rearranged channel features to obtain the final output features.

7. The method for diagnosing aero-engine bearing faults based on deep prototype networks according to claim 1, characterized in that, The training of the fault diagnosis model using the dataset includes: The training set is divided into a support set and a query set, which are then fed into the model for training. According to the formula and Obtain prototype data for each category of the support set and feature vectors for the query set; where, i It is the index of the image vector. q i It is the encoded feature vector; k It is a category; C k It is a category k The prototype; S k It is a category k Supported episode count; It is an encoding function; According to the formula Use Euclidean functions to compute the query set. q i With prototype C k Distance between d ; The distance for each query set is calculated using this formula. Calculated as class probability; Use the model's loss function Update the basic model parameters; repeat the training steps until the maximum number of iterations is reached.

8. A fault diagnosis device for aero-engine bearings based on deep prototype networks, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements a method for diagnosing aero-engine bearing faults based on a deep prototype network as described in any one of claims 1-7.

9. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for diagnosing aero-engine bearing faults based on a deep prototype network as described in any one of claims 1-7.

Images