Bearing fault detection method and system based on recurrent graph and deep adaptive transformer

Abstract

The application provides a bearing fault detection method based on a recurrent graph and a deep adaptive Transformer, relates to the technical field of bearing detection, and comprises bearing vibration signal training data acquisition; bearing vibration signal data is subjected to enhancement pretreatment; a one-dimensional vibration signal is converted into a two-dimensional texture image to form a feature map through a recurrent graph coding technology, and is divided into three data sets, namely, a training set, a verification set and a test set; a bearing fault detector based on a deep adaptive Transformer is established, model training, parameter optimization and model testing are carried out by using data set data, bearing fault detection is completed, and the problem of low bearing fault detection precision in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of bearing fault detection technology, specifically to a bearing fault detection method and system based on recursive graphs and depth adaptive Transformers. Background Technology

[0002] With industrial development, the demand for high-precision automated equipment is constantly increasing. The more precise these devices are, the more difficult it is to monitor their operating status. Their operating status often directly determines the quality of the equipment. Bearing failures will lead to high maintenance costs and a lot of maintenance time. Therefore, how to reduce the probability of bearing failures and improve fault diagnosis capabilities has become the mainstream research direction.

[0003] Currently, bearing fault diagnosis technologies based on vibration signals are mainly divided into signal processing-based methods and machine learning-based methods. Signal processing-based methods rely on human intervention to extract feature information for bearing diagnosis, making it easier to identify the type of bearing fault. However, this often depends on human experience, leading to low diagnostic accuracy. Machine learning-based bearing fault diagnosis, on the other hand, can better capture information about different bearing faults and more accurately monitor the bearing's operating status. When using machine learning methods, convolutional networks are used for bearing vibration signal feature extraction. In practical engineering, CNN algorithms are commonly used. CNN algorithms extract features from bearing vibration signals and then classify them using a classifier to determine whether the fault is in the outer or inner race. However, CNN convolution cannot capture global features and cannot extract multiple vibration signals in real time. Recurrent networks also need to consider temporal information, leading to the loss of more and more features as information propagates forward. The accuracy of both methods needs further improvement. Summary of the Invention

[0004] Technical problems to be solved

[0005] To address the shortcomings of existing technologies, this invention provides a bearing fault detection method and system based on recursive graphs and depth adaptive Transformers, which solves the problem of low bearing fault detection accuracy mentioned in the background art.

[0006] Technical solution

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A bearing fault detection method based on recursive graph and depth adaptive Transformer includes: acquisition of bearing vibration signal training data;

[0009] Enhancement preprocessing is performed on bearing vibration signal data;

[0010] Recursive graph coding technology is used to generate two-dimensional texture images from one-dimensional vibration signals to form feature maps, which are then divided into three datasets: training set, validation set, and test set.

[0011] A bearing fault detector based on deep adaptive Transformer was built, and the model was trained, the parameters were optimized, and the model was tested using the dataset to complete the bearing fault detection.

[0012] Preferably, the specific method for acquiring the bearing vibration signal training data is as follows: acquiring the vibration signal of the drive end bearing on the device through a vibration signal acquisition device;

[0013] The pre-prepared bearing to be tested must include three types of defect locations: rolling element damage, outer ring damage, and inner ring damage, as data required for training.

[0014] Accelerometers were set up in different locations to collect vibration signals, which included one normal signal and nine fault signal types.

[0015] Preferably, the process of establishing and constructing a bearing fault detector based on a depth adaptive Transformer includes:

[0016] First, we need to obtain the recursively encoded two-dimensional texture image. Based on the processed two-dimensional texture image, we can obtain the category of each fault from the input image.

[0017] For a two-dimensional texture image with only one type of fault, a convolutional network F is applied to diagnose the fault points of the bearing, resulting in a point heatmap. Where k is the number of predicted fault points, and h and w are the height and width of the feature map, respectively. In form:

[0018] H represents all fault points on the feature map, F represents the convolution function, and I is the input fault type; each pixel in H represents the probability of the encoded fault point at that pixel. This is to obtain the joint coordinates. These methods use a maximum value operation to obtain the location with peak activation. Formally, let p be a spatial location on H, it can be written as:

[0019] for For each fault point, in the heatmap-based method, the accuracy of p reaches the resolution of H. During model training, J is used as a k-length sequence to directly map the input I to the coordinates J of the bearing fault point.

[0020] F consists of three main components: a standard CNNBackbone for extracting multi-level feature representations, a Transformer Encoder for capturing and fusing multi-level features, and a Transformer Decoder for generating keypoint coordinate sequences.

[0021] Preferably, the process of establishing and constructing a bearing fault detector based on a deep adaptive Transformer, and using dataset data for model training, parameter optimization, and model testing to complete bearing fault detection, also includes: a deep adaptive algorithm, which can apply the training method to fault diagnosis under different working conditions by changing the maximum mean difference and using multi-core MMD for transfer learning.

[0022] The present invention also provides a bearing fault detection system based on recursive graph and depth adaptive Transformer, the system comprising: a rolling motor bearing device, a vibration signal acquisition device, and a terminal server;

[0023] The rolling motor bearing device includes: a pre-prepared bearing to be tested, which is required to contain three types of defect locations, namely rolling element damage, outer ring damage and inner ring damage, as data required for training;

[0024] The vibration signal acquisition device is used for acquiring bearing vibration signal training data.

[0025] The terminal server is used to receive bearing signal data obtained by the vibration signal acquisition device, and includes a processor and a memory. The memory stores at least one instruction or at least one program. The at least one instruction or at least one program is loaded and executed by the processor to realize a bearing fault detection based on recursive graph and depth adaptive Transformer as described in any of the preceding claims.

[0026] Beneficial effects

[0027] This invention provides a bearing fault detection method and system based on recursive graphs and depth-adaptive Transformers. It offers the following advantages:

[0028] This invention provides a bearing fault detection method and system based on recursive graphs and depth-adaptive Transformers. It introduces Transformers into bearing fault diagnosis and combines a depth-adaptive mechanism with convolution to optimize the accuracy of extracting different features before convolutional extraction of low-level features. The acquired images have features from different angles, and the network depth can be adjusted according to these image features, only calculating images with complex features. A bearing fault detector based on depth-adaptive Transformers is constructed. Convolution is used to extract basic features at the low level, while Transformers capture global dependencies at higher levels. Texture maps are flattened as input, and the output is restored to a two-dimensional heatmap, resulting in more accurate fault classification and detection. Attached Figure Description

[0029] Figure 1 A flowchart of a bearing fault detection method based on recursive graph and depth adaptive Transformer provided in an embodiment of the present invention;

[0030] Figure 2 This is a structural diagram of a bearing fault detection system based on recursive graphs and depth adaptive Transformers, provided as an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0032] like Figure 1 As shown, this embodiment of the invention provides a bearing fault detection method based on recursive graphs and depth adaptive Transformers, including:

[0033] S1 bearing vibration signal training data acquisition;

[0034] S2 performs enhanced preprocessing on bearing vibration signal data;

[0035] S3 uses recursive graph coding technology to generate two-dimensional texture images from one-dimensional vibration signals to form feature maps, and divides them into three datasets: training set, validation set, and test set.

[0036] S4 was used to build a bearing fault detector based on deep adaptive Transformer, and the model was trained, the parameters were optimized and the model was tested using the dataset to complete the bearing fault detection.

[0037] Preferably, the specific method for acquiring the bearing vibration signal training data is as follows: acquiring the vibration signal of the drive end bearing on the device through a vibration signal acquisition device;

[0038] The pre-prepared bearing to be tested must include three types of defect locations: rolling element damage, outer ring damage, and inner ring damage, as data required for training.

[0039] Accelerometers were set up in different locations to collect vibration signals, which included one normal signal and nine fault signal types.

[0040] In one embodiment, the system sampling frequency is 12 kHz.

[0041] In one embodiment, the bearing vibration signal data is enhanced through preprocessing. Since the collected data is abundant and accompanied by different signal characteristics, data enhancement is a very important step that can better extract features.

[0042] Preferably, the process of establishing and constructing a bearing fault detector based on a depth adaptive Transformer includes:

[0043] First, we need to obtain the recursively encoded two-dimensional texture image. Based on the processed two-dimensional texture image, we can obtain the category of each fault from the input image.

[0044] For a two-dimensional texture image with only one type of fault, a convolutional network F is applied to diagnose the fault points of the bearing, resulting in a point heatmap. Where k is the number of predicted fault points, and h and w are the height and width of the feature map, respectively. In form:

[0045] H represents all fault points on the feature map, F represents the convolution function, and I is the input fault type; each pixel in H represents the probability of the encoded fault point at that pixel. This is to obtain the joint coordinates. These methods typically use a maximum value operation to obtain the location with peak activation. Formally, let p be the spatial location on H, it can be written as:

[0046] for For each fault point, in the heatmap-based method, the accuracy of p reaches the resolution of H. During model training, J is used as a k-length sequence to directly map the input I to the coordinates J of the bearing fault point.

[0047] F consists of three main components: a standard CNNBackbone for extracting multi-level feature representations, a Transformer Encoder for capturing and fusing multi-level features, and a Transformer Decoder for generating keypoint coordinate sequences.

[0048] In one embodiment, a one-dimensional vibration signal is used to generate a two-dimensional texture image through recursive graph encoding. These texture images are then input into a CNN network. Leveraging the CNN network's excellent feature extraction capabilities for two-dimensional image data, bearing fault diagnosis is performed. However, while the CNN network performs well in extracting image features, it still fails to extract some high-level features. Therefore, a Transformer model is introduced to extract high-level features from the texture image, significantly improving accuracy.

[0049] Preferably, the process of establishing and constructing a bearing fault detector based on a deep adaptive Transformer, and using dataset data for model training, parameter optimization, and model testing to complete bearing fault detection, further includes: a deep adaptive algorithm, which uses multi-core MMD for transfer learning by changing the maximum mean difference, so that the training method can be applied to fault diagnosis under different working conditions.

[0050] In one embodiment, the construction of a bearing fault detector based on a depth adaptive Transformer is essentially a structure of a CNN backbone network and a Transformer Encoder-Decoder. The Transformer is a model based on a multi-head attention mechanism. This model can solve the problem of long-distance dependence, has better memory, and reduces lengthy and complex calculations to achieve more accurate recognition results. However, it also loses the ability to capture local features. Therefore, in the original model, a CNN convolutional network is introduced to extract low-level features and better capture local features.

[0051] In one embodiment, the waveforms of vibration signals acquired by the vibration signal acquisition device are constantly changing, and the fault characteristics are also different. There will be certain faults in the frequency diagram of each vibration signal waveform, so it is difficult to make accurate diagnosis and it affects the diagnostic classification effect. That is, the depth adaptive mechanism can make different input image data more suitable for model training, select a specific Decoder layer for output, reduce the amount of computation, and improve accuracy.

[0052] For a deep network, as the number of layers increases, the network's dependency gradually decreases, becoming more dependent on a specific task. However, shallow layers only learn a general set of features; these features are generally universal across different tasks, but the features differ in higher layers. Therefore, a more adaptive network is needed to adapt to the higher layers and extract more accurate features. This application employs a multi-kernel MMD and multi-layer adaptive high-level network method, called the DAN algorithm MK-MMD, which emphasizes multiple kernels to construct the overall kernel for two probability distributions. The distance between their multi-core MMD is

[0053] For mapping functions, , These are two sample distribution vectors. Mapping these two sample distributions into a space and then applying the mean and expected value... Perform a maximum difference calculation.

[0054] The total kernel defined by these multiple kernels is

[0055] Among them It is a fixed function, and now it is weighted using m different kernels, where the weights are... Its expressive power is far superior to that of a single-core MMD.

[0056] Regarding the DAN method, it is based on the AlexNet network and explores the fit between the source and target. The optimization objective of the DAN algorithm mainly consists of the loss function and the distribution distance, which is the MK-MMD distance. Therefore, the optimization objective is...

[0057] in It is the collection of all label data in the target domain. It is the set of all tags in the source domain. It is a penalty coefficient, however It is the defined loss function. The numbers 6 and 8 indicate that the network adaptation extends from layer 6 to layer 8, which corresponds to layers 1 to 3 after the Transformer decoder. This represents the collection of all labeled data in the source and target.

[0058] Learning weights The main purpose is to determine the weights of the multi-core system and ensure that the variance of the MMD distance generated by each core is minimized.

[0059] This invention decomposes the DAN algorithm into 5 convolutional layers and 3 fully connected layers, placing them before the Transformer encoder and after the decoder, respectively. The 5 convolutional layers extract low-level features, and the Transformer outputs are divided into Source and Target domains. The compatibility between these two domains is explored, and finally, the Target domain output is compared with the original image to determine the fault point data. This invention transforms the one-dimensional signal into a two-dimensional feature texture map using recursive graph technology, distributing fault points across every pixel of the image for better fault feature representation. Furthermore, by combining a depth adaptive algorithm and using multi-kernel MMD for transfer learning by changing the maximum mean difference, the training method can be applied to fault diagnosis under different operating conditions. By combining transfer learning with natural language processing, two different domains, to identify and diagnose bearing fault texture maps, the diagnostic results are more accurate.

[0060] This invention also provides a bearing fault detection system based on recursive graphs and depth adaptive Transformers, the system comprising: a rolling motor bearing device, a vibration signal acquisition device, and a terminal server;

[0061] The rolling motor bearing device includes: a pre-prepared bearing to be tested, which is required to contain three types of defect locations, namely rolling element damage, outer ring damage and inner ring damage, as data required for training;

[0062] The vibration signal acquisition device is used for acquiring bearing vibration signal training data.

[0063] The terminal server is used to receive bearing signal data obtained by the vibration signal acquisition device. It also includes a processor and a memory. The memory stores at least one instruction or at least one program. The at least one instruction or at least one program is loaded and executed by the processor to realize a bearing fault detection based on recursive graph and depth adaptive Transformer as described in the preceding claim. After the signal is acquired from the vibration signal acquisition device, it is transmitted to the terminal server and processed into a frequency graph in the time domain or frequency domain. The bearing signal is then subjected to feature analysis of time series data through recursive graph (RP), and the training accuracy is improved by combining an improved version of the Transformer model.

[0064] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing fault detection method based on recursive graphs and depth-adaptive Transformers, characterized in that, include: Bearing vibration signal training data acquisition; Enhancement preprocessing is performed on bearing vibration signal data; Recursive graph coding technology is used to generate two-dimensional texture images from one-dimensional vibration signals to form feature maps, which are then divided into three datasets: training set, validation set, and test set. A bearing fault detector based on deep adaptive Transformer was built, and the model was trained, the parameters were optimized, and the model was tested using the dataset to complete the bearing fault detection. The establishment and construction of a bearing fault detector based on depth adaptive Transformer includes: First, we need to obtain the recursively encoded two-dimensional texture image. Based on the processed two-dimensional texture image, we can obtain the category of each fault from the input image. For a two-dimensional texture image with only one type of fault, a convolutional network F is applied to diagnose the fault points of the bearing, resulting in a point heatmap. Where k is the number of predicted fault points, and h and w are the height and width of the feature map, respectively. In form: H represents all fault points on the feature map, F represents the convolution function, and I is the input fault type; each pixel in H represents the probability of the encoded fault point at that pixel. This is to obtain the joint coordinates. These methods use a maximum value operation to obtain the location with peak activation. Formally, let p be a spatial location on H, it can be written as: for For each fault point, in the heatmap-based method, the accuracy of p reaches the resolution of H. During model training, J is used as a k-length sequence to directly map the input I to the coordinates J of the bearing fault point. F consists of three main components: a standard CNNBackbone for extracting multi-level feature representations, a Transformer Encoder for capturing and fusing multi-level features, and a Transformer Decoder for generating keypoint coordinate sequences. The method establishes a bearing fault detector based on deep adaptive Transformer, and uses dataset data for model training, parameter optimization, and model testing to complete bearing fault detection. It also includes a deep adaptive algorithm, which uses multi-core MMD for transfer learning by changing the maximum mean difference, so that the training method can be applied to fault diagnosis under different working conditions.

2. The bearing fault detection method based on recursive graph and depth adaptive Transformer according to claim 1, characterized in that, The specific method for acquiring the bearing vibration signal training data is as follows: the vibration signal of the drive end bearing on the device is acquired through a vibration signal acquisition device; The pre-prepared bearing to be tested must include three types of defect locations: rolling element damage, outer ring damage, and inner ring damage, as data required for training. Accelerometers were set up in different locations to collect vibration signals, which included one normal signal and nine fault signal types.

3. A bearing fault detection system based on recursive graphs and depth-adaptive Transformers, characterized in that, The system includes: a rolling motor bearing device, a vibration signal acquisition device, and a terminal server; The rolling motor bearing device includes: a pre-prepared bearing to be tested, which is required to contain three types of defect locations, namely rolling element damage, outer ring damage and inner ring damage, as data required for training; The vibration signal acquisition device is used for acquiring bearing vibration signal training data. The terminal server is used to receive bearing signal data obtained by the vibration signal acquisition device, and includes a processor and a memory. The memory stores at least one instruction or at least one program. The at least one instruction or at least one program is loaded and executed by the processor to implement the bearing fault detection method based on recursive graph and depth adaptive Transformer as described in any one of claims 1 to 2.

Images