Gear fault diagnosis method based on enhanced relational network

Abstract

This invention proposes a gear fault diagnosis method based on an enhanced relational network. The method involves: 1) acquiring multiple sample gear vibration signals and dividing them into a support set and a query set; 2) converting the gear vibration signals in the support and query sets into two-dimensional video images using continuous wavelet transform; 3) enhancing the relational network's ability to obtain strong features by incorporating a CA attention mechanism; 4) constructing and training a relational network model for gear fault diagnosis using the acquired data; 5) monitoring and collecting gearbox vibration signals in real time and preprocessing the vibration signals; and 6) inputting the preprocessed vibration signals into the relational network classification model, and obtaining the gearbox's operating status through analysis and comparison, thereby completing the gearbox fault diagnosis. This invention fully utilizes the CA attention mechanism to enhance the relational network's ability to extract features, reducing the problem of low diagnostic accuracy due to a small sample size.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis, and specifically to a gear fault diagnosis method based on an enhanced relational network. Background Technology

[0002] Gears are indispensable key components in modern rotating machinery. Failures can severely impact equipment production and even lead to safety accidents, causing significant economic losses and personal injury. Traditional deep learning-based fault diagnosis techniques face challenges such as limited fault sample numbers, significant variations in operating conditions, and unclear fault characteristics, resulting in poor model generalization performance and low diagnostic accuracy. Existing methods for addressing small sample problems often rely on simulation expansion based on already collected data; however, the effectiveness of data expansion becomes less pronounced under varying operating conditions. Furthermore, applying a model more suitable for gearbox fault matching accuracy is crucial.

[0003] The attention mechanism (CoordAttention, CA) not only captures cross-channel information but also direction-aware and position-aware information, helping the model to more accurately locate and identify targets of interest. CoordAttention decomposes channel attention into two 1D feature encoding processes that aggregate features along different directions. This allows for the capture of long-range dependencies along one spatial direction while preserving precise positional information along the other. The generated feature maps are then encoded separately, forming a pair of direction-aware and position-sensitive feature maps that can be complementaryly applied to the input feature map to enhance the representation of the target of interest. By abandoning data augmentation methods and focusing the network on fault features when the need for labeled data is low, it reduces attention to unnecessary information and maximizes the capture of key fault feature information, thereby improving fault classification capabilities.

[0004] Relation networks are a few-shot learning method that compares the similarity of images in a given dataset by learning a deep, non-linear distance metric between input images. Relation networks do not require augmentation of sample data and do not suffer from training non-convergence or gradient vanishing problems, while exhibiting strong generalization performance. Compared to traditional linear metrics such as Euclidean and cosine measures, relation networks learn a metric themselves by constructing a neural network, training a learnable non-linear distance metric classifier to calculate the matching degree between samples.

[0005] Therefore, it is essential to propose a small-sample gear fault diagnosis method based on a relational network of coordinated attention. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a gear fault diagnosis method based on enhanced relational networks, which introduces a CA mechanism to enhance the feature extraction capability of the relational network, make up for the problem of insufficient sample size, and improve the diagnostic accuracy of the model.

[0007] The embodiments of this application are implemented as follows:

[0008] This application provides a gear fault diagnosis method based on an enhanced relational network, characterized by the following steps:

[0009] Step a: By building a gearbox experimental platform, various types of gear vibration signal data are acquired, preprocessed, and divided into training sets x. i and test set x j ;

[0010] Step b: Using continuous wavelet transform, the obtained dataset is transformed from a one-dimensional time series signal into a two-dimensional time-frequency diagram.

[0011] Step c: Add CA attention mechanism, use convolutional neural network as feature extraction module, use neural network as relation module, and use feature extraction module and relation module to build relation network model for gear fault diagnosis;

[0012] Step d, the training set x i and test set x i Using the spectrum data as input, the test set x is obtained through a relational network model. j Samples and training set x i The training loss parameters are obtained by analyzing the relationship scores between different categories and the loss function value of the relationship network model. These parameters are then implanted into the relationship network model to train it.

[0013] Step e: Monitor and acquire the gear vibration signal of the gearbox in real time, and preprocess the gear vibration signal;

[0014] Step f: Use the trained relational network model to classify the vibration signal of the gear under test, and determine the fault condition of the gear under test based on the classification results.

[0015] In some optional implementations, the specific content of the CA attention mechanism described in step c includes:

[0016] Step c1, Coordinate information embedding:

[0017] For input X, each channel is first encoded using pooling kernels of size (H, 1) and (1, W) along the horizontal and vertical coordinate directions. Therefore, the output of the c-th channel with height h is expressed as:

[0018]

[0019] The output of the c-th channel with width w is expressed as:

[0020]

[0021] The two transformations described above are used to aggregate features along two spatial directions, returning a pair of direction-aware attention feature maps;

[0022] Step c2, coordinate attention generation:

[0023] First, the two feature maps generated in cascaded step c1 are combined, and then a shared 1x1 convolution is used to perform an F1 transformation, as expressed in the following formula:

[0024]

[0025] generated It is an intermediate feature map of spatial information in the horizontal and vertical directions, where r represents the downsampling ratio;

[0026] Divide f into two separate tensors along the spatial dimension. and Then use two 1x1 convolutions F h and F w feature map f h and f w Transforming the input X to have the same number of channels yields the following result:

[0027] g h =σ(F h (f h ))

[0028] g w =σ(F w (f w )),

[0029] Then for g h and g w Extending this further, as attention weights, the final output of the CA module can be expressed as follows:

[0030]

[0031] In some alternative implementations, the relational network model RN comprises two modules: an embedding function module consisting of four convolutional blocks. and a correlation calculation function module consisting of two convolutional blocks: g φ First, use the embedded module for mapping to obtain... and Then use a join operator: By directly concatenating the two vectors and sending the result to the relation module to calculate the relevance score, a relevance score r between 0 and 1 is generated. i，j To indicate x i With x j The similarity score has N possible values:

[0032]

[0033] The relationship network uses mean squared error to supervise the similarity scores, and the optimization objective function is as follows:

[0034]

[0035] When y i ==y j The output is 1 when it is established. By continuously training and updating the parameters, the diagnostic accuracy of the relational network model is improved.

[0036] In some alternative implementations, the gear vibration signals include a fault-free state, gearbox surface spalling, tooth root fault, broken tooth fault, and tooth tip wear or pitting.

[0037] In some alternative implementations, 70% of the gear vibration signal data is divided into a training set, which is fully labeled; and 30% is divided into a test set, which is partially labeled.

[0038] In some alternative implementations, the gear vibration signal data is acquired by sampling using accelerometers attached to the gearbox surface and bearing housing.

[0039] The beneficial effects of this application are as follows: This application provides a gear fault diagnosis method based on enhanced relational networks, which introduces a CA mechanism to enhance the feature extraction capability of the relational network, make up for the problem of insufficient sample size, and improve the diagnostic accuracy of the model; at the same time, the relational network calculates the distance between two samples by constructing a neural network to analyze the degree of matching. Compared with Siamese networks and prototype networks, the relational network can be regarded as providing a learnable nonlinear classifier for judging relationships, unlike Siamese networks and prototype networks, where the distance is only a linear relational classifier. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of an embodiment of this application;

[0042] Figure 2 This is a flowchart illustrating the operation of the CA attention mechanism in an embodiment of this application.

[0043] Figure 3 This is a schematic diagram of the relationship network in an embodiment of this application. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0045] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0046] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0047] The features and performance of this application will be further described in detail below with reference to the embodiments.

[0048] like Figure 1 As shown, the present invention provides a gear fault diagnosis method based on an enhanced relational network, characterized by comprising the following steps:

[0049] Step a: By building a gearbox experimental platform, various types of gear vibration signal data are acquired, preprocessed, and divided into training sets x. i and test set x j ;

[0050] Step b: Using continuous wavelet transform, the obtained dataset is transformed from a one-dimensional time series signal into a two-dimensional time-frequency diagram.

[0051] Step c: Add CA attention mechanism, use convolutional neural network as feature extraction module, use neural network as relation module, and use feature extraction module and relation module to build relation network model for gear fault diagnosis;

[0052] Step d, the training set x iand test set x j Using the spectrum data as input, the test set x is obtained through a relational network model. j Samples and training set x i The training loss parameters are obtained by analyzing the relationship scores between different categories and the loss function value of the relationship network model. These parameters are then implanted into the relationship network model to train it.

[0053] Step e: Monitor and acquire the gear vibration signal of the gearbox in real time, and preprocess the gear vibration signal;

[0054] Step f: Use the trained relational network model to classify the vibration signal of the gear under test, and determine the fault condition of the gear under test based on the classification results.

[0055] like Figure 2 As shown, the specific content of the CA attention mechanism includes:

[0056] Step c1, Coordinate information embedding:

[0057] For input X, each channel is first encoded using pooling kernels of size (H, 1) and (1, W) along the horizontal and vertical coordinate directions. Therefore, the output of the c-th channel with height h is expressed as:

[0058]

[0059] The output of the c-th channel with width w is expressed as:

[0060]

[0061] The two transformations described above are used to aggregate features along two spatial directions, returning a pair of direction-aware attention feature maps;

[0062] Step c2, coordinate attention generation:

[0063] First, the two feature maps generated in cascaded step c1 are combined, and then a shared 1x1 convolution is used to perform an F1 transformation, as expressed in the following formula:

[0064] f=δ(F1([z h , z w ])),

[0065] generated It is an intermediate feature map of spatial information in the horizontal and vertical directions, where r represents the downsampling ratio;

[0066] Divide f into two separate tensors along the spatial dimension. and Then use two 1x1 convolutions F h and Fw feature map f h and f w Transforming the input X to have the same number of channels yields the following result:

[0067] g h =σ(F h (f h ))

[0068] g w =σ(F w (f w )),

[0069] Then for g h and g w Extending this further, as attention weights, the final output of the CA module can be expressed as follows:

[0070]

[0071] In some alternative implementations, the relational network model RN comprises two modules: an embedding function module consisting of four convolutional blocks. And the correlation calculation function module, gφ, consisting of two convolutional blocks, is first mapped using an embedding module to obtain... and Then use a join operator: By directly concatenating the two vectors and sending the result to the relation module to calculate the relevance score, a relevance score r between 0 and 1 is generated. i，j To indicate x i With x j Similarity (see) Figure 3 There are N scores in total:

[0072]

[0073] The relationship network uses mean squared error to supervise the similarity scores, and the optimization objective function is as follows:

[0074]

[0075] When y i ==y j The output is 1 when it is established. By continuously training and updating the parameters, the diagnostic accuracy of the relational network model is improved.

[0076] In some alternative implementations, gear vibration signals, including fault-free conditions, gearbox surface spalling, tooth root faults, broken tooth faults, and tooth tip wear or pitting, are acquired by sampling using accelerometers attached to the gearbox surface and bearing housings.

[0077] Example 1

[0078] 1. Acquire several gear vibration signals and perform data normalization processing, dividing 70% of them into a training set x. i 30% is allocated to the test set x j The training set is fully labeled, while the test set is partially labeled.

[0079] 2. Using MATLAB, perform continuous wavelet transform on the acquired dataset to convert it from a one-dimensional time series plot to a two-dimensional time-frequency plot.

[0080] 3. Perform coordinated attention mechanism processing on the two-dimensional time-frequency graph:

[0081] 3.1 Construct a CA module to encode channel relationships and long-range dependencies using precise location information. This involves two steps: coordinate information embedding and coordinate attention generation.

[0082] 3.2 Coordinate Information Embedding. Specifically, for input X, each channel is first encoded along the horizontal and vertical coordinate directions using pooling kernels of sizes (H, 1) and (1, W). Therefore, the output of the c-th channel with height h is expressed as follows:

[0083]

[0084] Similarly, the output of the c-th channel with width w is expressed as follows:

[0085]

[0086] The two transformations above perform feature aggregation along two spatial directions, returning a pair of direction-aware attention maps.

[0087] 3.3 Coordinate Attention Generation. First, the two feature maps generated by the previous modules are concatenated, then a shared 1x1 convolution is used for F1 transformation, expressed as follows, generating... It is an intermediate feature map of spatial information in the horizontal and vertical directions, where r represents the downsampling ratio.

[0088]

[0089] Next, f is divided into two separate tensors along the spatial dimension. and Then use two 1x1 convolutions F h and F w feature map f h and fw Transforming the input X to have the same number of channels yields the following result:

[0090] g h =σ(F h (f h ))

[0091] g w =σ(F w (f w ))

[0092] Then for g h and g w Extending this further, as attention weights, the final output of the CA module can be expressed as follows:

[0093]

[0094] 4. Build a relationship network.

[0095] 4.1 Construct a fault classification model for relational networks, including an input layer, convolutional layer, pooling layer, fully connected layer, and sigmoid output layer.

[0096] 4.2 Input the signal set processed by the CA mechanism into the relational network fault classification model, train the relational network fault classification model, and obtain the trained relational network fault classification model.

[0097] 5. The relational network RN contains two modules: an embedding module consisting of four convolutional blocks. and a relational module consisting of two convolutional blocks: g φ First, use an embedded module for mapping to obtain... and Then use a join operator: The two vectors are concatenated directly. The result is then fed into the relation module to calculate the relevance score, ultimately producing a relation score between 0 and 1 to indicate the relationship between x and x. i With x j The similarity score has N possible values.

[0098]

[0099] The relationship network uses mean squared error to supervise the similarity scores, and the optimization objective function is as follows:

[0100]

[0101] The parameters are continuously trained and updated to improve the diagnostic accuracy of the network model.

[0102] in For embedded functions, For the join function, g φ r is the correlation calculation function. i，j For the relevance score, and when y i ==y j The output is 1 when it is true.

Claims

1. A gear fault diagnosis method based on reinforced relational networks, characterized in that, Includes the following steps: Step a: By building a gearbox experimental platform, various types of gear vibration signal data are acquired, preprocessed, and divided into training sets. and test set ; Step b: Using continuous wavelet transform, the obtained dataset is transformed from a one-dimensional time series signal into a two-dimensional time-frequency diagram. Step c: Add CA attention mechanism, use convolutional neural network as feature extraction module, use neural network as relation module, and use feature extraction module and relation module to build relation network model for gear fault diagnosis; Step d, the training set and test set Using spectral data as input, a test set is obtained through a relational network model. Samples and training set The training loss parameters are obtained by analyzing the relationship scores between different categories and the loss function value of the relationship network model. These parameters are then implanted into the relationship network model to train it. Step e: Monitor and acquire the gear vibration signal of the gearbox in real time, and preprocess the gear vibration signal; Step f: Use the trained relational network model to classify the vibration signal of the gear under test, and determine the fault condition of the gear under test based on the classification results.

2. The gear fault diagnosis method based on reinforced relational networks according to claim 1, characterized in that, The specific content of the CA attention mechanism described in step c includes: Step c1, Coordinate information embedding: For input X, each channel is first encoded using pooling kernels of size (H,1) and (1,W) along the horizontal and vertical coordinate directions. Therefore, the output of the c-th channel at height h is expressed as: ， The output of the c-th channel with width w is expressed as: ， The two transformations described above are used to aggregate features along two spatial directions, returning a pair of direction-aware attention feature maps; Step c2, coordinate attention generation: First, the two feature maps generated in cascaded step c1 are combined, and then a shared 1x1 convolution is used to perform an F1 transformation, as expressed in the following formula: ， generated It is an intermediate feature map of spatial information in the horizontal and vertical directions, where r represents the downsampling ratio; Along the spatial dimension Divided into two separate tensors and Then use two 1x1 convolutions and feature map Transforming the input X to have the same number of channels yields the following result: ， Then to and Extending this further, as attention weights, the final output of the CA module can be expressed as follows: 。 3. The gear fault diagnosis method based on reinforced relational networks according to claim 2, characterized in that, The relational network model comprises two modules: an embedding function module consisting of four convolutional blocks. And a correlation calculation function module consisting of two convolutional blocks: First, use the embedded module for mapping to obtain... and Then use a join operator: The two vectors are concatenated directly, and the result is sent to the relation module to calculate the relevance score, which ultimately produces a relevance score between 0 and 1. To indicate and The similarity score has N possible values: ， The relationship network uses mean squared error to supervise the similarity scores, and the optimization objective function is as follows: ， when The output is 1 when it is established. By continuously training and updating the parameters, the diagnostic accuracy of the relational network model is improved.

4. The gear fault diagnosis method based on reinforced relational networks according to claim 1 or 3, characterized in that, The gear vibration signals include fault-free state, gearbox surface peeling, tooth root fault, broken tooth fault, and tooth tip wear or pitting.

5. The gear fault diagnosis method based on reinforced relational networks as described in claim 4, characterized in that, 70% of the gear vibration signal data was divided into a training set, all of which were labeled; 30% was divided into a test set, some of which were labeled.

6. The gear fault diagnosis method based on reinforced relational networks as described in claim 5, characterized in that, The gear vibration signal data is obtained by sampling through accelerometers attached to the surface of the gearbox and the bearing housing.

Images