Motor bearing fault diagnosis method and system based on 1D-XFormer model

Through the adaptive lightweight 1D-Xception-Transformer network, the problem of fault diagnosis of motor bearing vibration signals in noise environments is solved, and the fault detection of high accuracy and stability in complex operating conditions is achieved, especially in low signal-to-noise ratio Excellent performance in the environment.

CN118936886APending Publication Date: 2024-11-12WUXI UNIV
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202410956833.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

When the motor is running, the bearing vibration signal is easily contaminated by background noise, making it difficult to extract and identify the fault signal, and it is difficult for the prior art to accurately diagnose faults under complex working conditions.

Method used

Adaptive lightweight 1D-Xception-Transformer network is adopted to build a 1D-XFormer network for motor bearing fault diagnosis through improved activation function AdaptGELU, deformable convolution and 1D-Transformer model to enhance nonlinear feature extraction and denoising ability.

Benefits of technology

The accuracy and stability of fault detection are significantly improved under the conditions of noise and sample scarcity, and the high accuracy of fault diagnosis is achieved, especially in low signal-to-noise ratio environments, which maintain prediction accuracy above 90%.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118936886A_ABST
    Figure CN118936886A_ABST
Patent Text Reader

Abstract

The invention discloses a motor bearing fault diagnosis method and system based on a 1D-XFormer model. The method comprises the following steps: collecting and preprocessing a motor bearing vibration signal; an adaptive lightweight 1D-XFormer network is constructed for motor bearing fault diagnosis; inputting the preprocessed bearing vibration signal into a 1D-XFormer network to obtain a diagnosis result; wherein the 1D-XFormer network is composed of one-dimensional Xception and Transform, is used for motor bearing fault diagnosis, and comprises an input channel, a middle channel and an output channel. The motor fault diagnosis method has high motor fault diagnosis capability under complex working conditions, and the complexity of the method is low.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The patent of this invention relates to the field of motor fault diagnosis, and specifically to a motor bearing fault diagnosis method and system based on the 1D-XFormer model. Background Art

[0002] The motor is an important electromechanical device that can convert electrical energy and mechanical energy into each other. It plays a very important role in many industrial fields such as electric transmission, transportation, servo control, etc. Therefore, it is of great significance to ensure the safety and reliability of motor operation. A normally operating motor will generate various characteristic signals. For example, physical quantities such as current, temperature, and vibration can be used as representations of motor performance and operating conditions. Among them, the bearing vibration signal can most directly reflect the specific operating conditions of mechanical equipment such as motors compared with other signals. Therefore, collecting and analyzing the characteristic information contained in the vibration signal is an effective fault diagnosis method. With the continuous development of artificial intelligence and deep learning, intelligent diagnosis technologies based on vibration signals continue to emerge.

[0003] However, in actual working conditions, the bearing vibration signal is inevitably contaminated by background noise, especially the strong interference from the motor. The bearing is usually installed near the motor or even in the motor. Therefore, the vibration generated when the motor is running always makes it very difficult to extract the bearing fault characteristics. At the same time, its fault signal is modulated by a complex nonlinear transmission path and is affected by non-Gaussian noise. In this case, the bearing fault is not easy to be found. Therefore, a more accurate and intelligent method is needed for fault detection of rotating machinery. Summary of the invention

[0004] Aiming at the complex working conditions where fault feature extraction is difficult under the above sample imbalance and strong noise conditions, this paper proposes an adaptive lightweight 1D-Xception-Transformer network for motor bearing fault diagnosis, which provides a new solution for the field of motor fault diagnosis:

[0005] A motor bearing fault diagnosis method based on 1D-XFormer model includes the following steps:

[0006] Collect motor bearing vibration signals and pre-process them;

[0007] Construct an adaptive lightweight 1D-XFormer network for motor bearing fault diagnosis;

[0008] The preprocessed bearing vibration signal is input into the 1D-XFormer network to obtain the diagnosis result.

[0009] Furthermore, the 1D-XFormer network is based on the Xception network architecture, which is composed of three channels: input channel, intermediate channel and output channel;

[0010] Among them, in the input channel, the adaptive activation function AdaptGELU is used to replace the ReLU activation function in the Xception network architecture, and the deformable convolution DeformConv is used to replace the convolution Conv in the Xception network architecture; in the intermediate channel, the 1D-Transformer model is used as the intermediate channel structure; in the output channel, the adaptive activation function AdaptGELU is used to replace the ReLU activation function in the Xception network architecture, and the deformable convolution DeformConv is used to replace the convolution Conv in the Xception network architecture, and the depth separable convolution module SeparableConv after the residual connection is replaced.

[0011] Furthermore, the specific calculation formula of the adaptive activation function AdaptGELU is:

[0012]

[0013] Here, x represents the input, G represents a Gaussian random variable with zero mean and unit variance, erf(·) represents the Gaussian error function, and a represents a trainable parameter.

[0014] Furthermore, it is characterized in that the specific steps of deformable convolution include:

[0015] From the input feature map The sampling pixel grid is in the middle, where C is the number of feature map channels, H is the feature map height, and W is the feature map width. The two-dimensional K×K convolution is used for sampling, and the position of the pixel in the receptive field is recorded as P n , the specific formula is:

[0016] P n = {P ij}(0≤i,j≤K-1;0≤n≤K 2 -1; i, j, n∈N) (2)

[0017] Offset point P on the feature map n The position of the feature map is determined by adding a weight coefficient to each sampling point to reduce the interference of irrelevant information on the feature map. The specific formula is:

[0018]

[0019] Among them, Υ(P out) represents the feature map output after convolution operation, ω(·) represents the value learned by the network in convolution, x(·) represents the value after grid sampling on the input feature map, P out represents the output position after the convolution operation, ΔΥ n represents the feature weight, ΔP n Indicates P n The offset of

[0020] Among them, x(·) uses bilinear interpolation method, the specific formula is:

[0021]

[0022] Among them, P represents the offset position, B(·,·) represents the two-dimensional bilinear interpolation kernel, and q represents the selection in the feature map X position.

[0023] Furthermore, the 1D-Transformer model is a one-dimensional Transformer model including L layers of sequential transformer encoders;

[0024] Among them, each layer of the temporal transformer encoder contains two sub-layers: a multi-head self-attention layer and a position-intelligent fully connected feed-forward network layer;

[0025] Among them, the multi-head self-attention layer includes a multi-head attention layer and a residual connection normalization layer; the position intelligent fully connected feedforward network layer includes a feedforward network layer and a residual connection normalization layer.

[0026] Furthermore, the diagnosis results include: normal state, inter-turn short circuit, rotor rod breakage, eccentricity fault, bearing inner ring fault and bearing outer ring fault.

[0027] A motor bearing fault diagnosis system based on 1D-XFormer model, comprising:

[0028] The sensor, specifically a triaxial vibration acceleration sensor, is used to collect the vibration signal of the bearing;

[0029] A transmission module, used for transmitting the collected signals;

[0030] Pre-training module, used to build 1D-XFormer network and pre-train the network using existing datasets;

[0031] A calculation module, used to input the signal into the 1D-XFormer network to obtain the diagnosis result;

[0032] The display module is used to display and remind the obtained diagnostic results.

[0033] According to one aspect of the present invention, a storage medium is provided, wherein instructions are stored in the storage medium. When a computer reads the instructions, the computer executes any one of the above-mentioned motor bearing fault diagnosis methods based on the 1D-XFormer model.

[0034] According to another aspect of the present invention, there is provided an electronic device, comprising a processor and the above-mentioned storage medium, wherein the processor executes instructions in the storage medium.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. This paper proposes an adaptive lightweight backbone network composed of one-dimensional Xception and Transformer, named 1D-XFormer, for motor bearing fault diagnosis. The network is mainly composed of three channels: input, intermediate and output.

[0037] 2. In response to the noise contained in actual industrial data, the present invention proposes a combined framework of an improved adaptive activation function AdaptGELU, an inverted depthwise separable convolution module and a deformable convolution residual connection to enhance the nonlinear feature extraction and expression capability of the model, thereby achieving the denoising effect of the nonlinear transmission path.

[0038] 3. Aiming at small sample fault diagnosis, the present invention introduces a 1D-Transformer structure in the middle channel of the model. The Encoder module in the structure has obvious advantages in global feature extraction, and combined with the local feature extraction advantages of deformable convolution in the input and output channels, it can improve the fault detection accuracy when the number of samples is scarce. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0040] Figure 1 It is a schematic diagram of a flow chart of an embodiment of the present invention;

[0041] Figure 2 A structural diagram of a sequential transformer encoder according to an embodiment of the present invention;

[0042] Figure 3 1D-XFormer network structure diagram of an embodiment of the present invention;

[0043] Figure 4 A comparison diagram of diagnostic results under sparse working conditions in an embodiment of the present invention;

[0044] Figure 5 This is a comparison diagram of diagnosis results in a noisy environment according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] like Figure 1 As shown, a motor bearing fault diagnosis method based on the 1D-XFormer model includes the following steps:

[0047] Collect motor bearing vibration signals and pre-process them;

[0048] Construct an adaptive lightweight 1D-XFormer network for motor bearing fault diagnosis;

[0049] The preprocessed bearing vibration signal is input into the 1D-XFormer network to obtain the diagnosis result.

[0050] In response to the challenges brought by the existing model parameters, the present invention adopts the inverted depthwise separable convolution Xception network as the basic framework of the backbone network and strengthens its optimization. The Xception network represents Google's improvement on the Inceptionv3 network. It fundamentally decouples the standard convolutions and replaces them with inverted depthwise separable convolutions, thereby increasing the depth of the model while reducing the number of parameters. This architectural modification helps to enhance network performance. Compared with standard convolutions, depthwise separable convolutions completely separate channel correlation and spatial correlation. It initially completes the learning of spatial correlation through depthwise convolutions, and then captures the information between channels through pointwise convolutions, thereby promoting the improvement of the channel dimension of the model.

[0051] The Xception network structure can be divided into three parts: inlet flow, intermediate flow and outlet flow, including 14 modules. The detailed structure is shown in Table 1. It can be seen that the Xception network consists of independent convolutional network blocks.

[0052] Table 1 Xception network structure table

[0053]

[0054] like Figure 2As shown, the proposed technique adopts a time series transformer encoder part, which is responsible for processing the input data and converting it into a set of meaningful feature vectors. It consists of L layers of encoders, where L is a positive integer, and each encoder contains two sublayers: a multi-head self-attention and a position-intelligent fully connected feed-forward network. In order to enhance information flow and alleviate the gradient vanishing problem, the transformer uses residual connections for each sublayer. This is accompanied by layer normalization. Therefore, the output of each sublayer can be expressed as LayerNorm(x+Sublayer(x)), where Sublayer(x) refers to the function performed by the sublayer itself. In order to achieve smooth integration of these residual connections, all sublayers in the model, including the embedding layer, produce outputs of dimension dmodel. The attention function acts as a mechanism that enables machine learning models to focus on specific fragments of the input sequence. It takes a query vector and a set of key-value pairs as input. The query vector represents the current focus of the model, while the key-value pairs represent different segments of the input sequence. The attention function then estimates the weighted sum of the value vectors, where the weight assigned to each value vector is determined by the compatibility between the query vector and the corresponding key vector. In essence, the attention function enables the model to focus on the most relevant parts of the input sequence based on the query. Here, the query vector is a representation of the model’s current focus, typically computed by the model’s encoder, and the key-value pairs are collections of vectors that describe different segments of the input sequence. These keys compute the compatibility between the query vector and the value vectors, which contain information about the various parts of the input sequence. The weights assigned to the value vectors determine the influence of each segment on the output. The output of the attention function is a weighted sum of the value vectors. The weights are computed using a compatibility function that measures the similarity between the query vector and the key vector.

[0055] Instead of applying a single attention function that operates on the model dimensions key, value, and query, multi-head attention employs h linear projections. These projections transform queries, keys, and values ​​into d k d q and d v dimensional, each projection is learned independently. Subsequently, the attention function is performed on each projected version simultaneously, producing d v dimensional output values. These values ​​are then concatenated and subjected to another projection to obtain the final value. The introduction of multi-head attention enables the model to effectively consider information from different representation subspaces and different positions simultaneously.

[0056] The Temporal Transformer Encoder module takes the extracted features as input. The self-attention mechanism allows the model to focus on different positions in the feature sequence, capturing the interactions between different elements. The self-attention layer allows the encoder to learn long-term dependencies between features in the vibration signal. By focusing on the relevant context of the entire sequence, the model gains a global understanding of the underlying patterns and structures in the vibration data. This ability is particularly valuable in capturing temporal dependencies, as vibrations are fundamentally dynamic and show complex relationships over time. The resulting features are then combined with the original input features via residual connections and layer normalization. Following the self-attention step, the algorithm employs a feed-forward network to enable the model to learn nonlinear relationships between input tokens. The feed-forward network uses a combination of linear transformations, activation functions, and normalization techniques to enhance the representation of the input sequence. By introducing nonlinearities and modeling complex interactions, the feed-forward network is able to extract high-level features that are discriminative for subsequent classification tasks.

[0057] Obviously, the standard fixed-size convolution kernel is not well suited for processing irregular bearing vibration signals and random noise generated by vibration of the vibration signal motor; therefore, the present invention uses deformable convolution, which adapts the structure of the convolution kernel by learning the offset.

[0058] From the input feature map The sampling pixel grid is in the form of a grid, where C is the number of feature map channels, H is the feature map height, and W is the feature map width. The two-dimensional K×K convolution is used for sampling, and the position of the pixel in the sensory field is recorded as P n , the specific formula is:

[0059] P n = {P ij}(0≤i,j≤K-1;0≤n≤K 2 -1; i, j, n∈N) (2)

[0060] Offset point P on the feature map n At each sampling point, a weight coefficient is added to reduce the interference of irrelevant information on the feature map. The specific formula is:

[0061]

[0062] Among them, Υ(P out ) represents the feature map output after convolution operation, ω(·) represents the value learned by the network in convolution, x(·) represents the value after grid sampling on the input feature map, ΔP n Indicates P n The offset, P out represents the output position after the convolution operation, ΔΥ n Represents feature weight.

[0063] The offset value predicted by the convolution operation is often a very small number and cannot be directly sampled from the feature map. Bilinear interpolation is used here to ensure that the offset feature map can be sampled. The specific formula is:

[0064]

[0065] Where P represents the offset position, B(·,·) represents the two-dimensional bilinear interpolation kernel, q represents the selection of the feature map X position, and B(q,P)=b(q x ,P x )·b(q y ,P y ), b(e,f)=max(0,1-|ef|).

[0066] Residual connection is a technology used in deep neural networks. Its function is to add the output of the previous layer directly to the input of the next layer, thus forming a cross-layer connection method. The purpose of this is to solve the problems of gradient vanishing and information attenuation in deep networks, and at the same time accelerate network training and improve network performance. Residual connection is equivalent to introducing a shortcut path in the network, so that information can be transmitted to the next layer without loss. The present invention deploys a deformable convolution module into the residual connection to form a deformable convolution residual connection, which can better capture the local characteristics of the vibration signal while solving the degradation phenomenon and improving network performance.

[0067] Since GELU is not optimal in some cases, the present invention proposes an adaptive GELU (AdaptGELU) activation function to further improve the performance of the network by introducing a trainable parameter a. AdaptGELU replaces the original activation function in the backbone network and replaces the original ReLU activation function in the Xception feedforward network with Sigmoid, ReLU, GELU and AdaptGELU functions. The specific calculation formula is:

[0068]

[0069] Here, x represents the input, G represents a Gaussian random variable with zero mean and unit variance, erf(·) represents the Gaussian error function, and a represents a trainable parameter.

[0070] like Figure 3As shown, the present invention proposes an adaptive lightweight 1D-Xception-Transformer (1D-XFormer) network for motor bearing fault diagnosis. The network is composed of an inverted depth separable convolution Xception network as the basic framework of the backbone network, and is composed of three channels: input, intermediate, and output. The main network structure of the input and output channels uses an improved adaptive activation function AdaptGELU, combined with an inverted depth separable convolution module, and deformable convolution for residual connection, and is repeatedly stacked three times to enhance the local nonlinear feature extraction and expression ability of the model, thereby achieving a nonlinear transmission path denoising effect. The intermediate channel is mainly composed of a 1D-Transformer model, and the Encoder module is used as the channel structure. When the number of fault samples is scarce, the fault features with irregular changes are better extracted, and the data mutations caused by the fault are better grasped as a whole.

[0071] like Figure 4 and Figure 5 As shown, the present invention uses real bearing damage samples generated by accelerated life tests at the University of Paderborn and artificial bearing damage samples generated by a comprehensive mechanical failure simulation test bench to perform verification under complex working conditions. A total of six different deep learning models are selected for fault diagnosis as benchmark models for comparison, including 1D-CNN, 1D-Inception, 1D-Inception-SE, 1D-Xception, 1D-Transformer, 1D-CNN-Transformer and 1D-Iformer. All models use the original time domain vibration signal as input.

[0072] like Figure 4 As shown, the 1D-XFormer proposed in the present invention performs best under a limited amount of training samples, with accuracies reaching 97.69%, 97.95% and 98.85%, respectively, which are 7.22%, 6.12% and 3.75% higher than the 1D-Xception model, respectively, and outperforms other models in terms of stability and accuracy.

[0073] like Figure 5 As shown in the figure, the diagnostic accuracy will decrease regardless of the method used due to the decrease in signal-to-noise ratio. Although the signal-to-noise ratio dropped to -4, the method proposed in the present invention maintained a prediction accuracy of more than 90%. Obviously, the robustness of this method to noise is better than other benchmark models, especially when the signal-to-noise ratio is low and the noise amplitude is large. Under different noise conditions, this method is less disturbed and the diagnostic results are more consistent.

[0074] Xception uses the ReLU global activation function. In the present invention, Sigmoid, SiLU, GELU and improved adaptive AdaptGELU activation functions are used to replace the original ReLU activation function in the 1D-XFormer network. The comparison results of the detection experiment are shown in Table 2.

[0075] Table 21 Comparative experimental results of different activation functions in D-XFormer-ReLU for fault diagnosis

[0076] Model Accuracy Params GFLOPs 1D-XFormer-ReLU 98.60% 0.91M 0.201 1D-XFormer-Sigmoid 98.09% 0.91M 0.201 1D-XFormer-SiLU 98.52% 0.91M 0.201 1D-XFormer-GELU 98.59% 0.91M 0.201 1D-XFormer-AdaptGELU 98.85% 0.91M 0.201

[0077] As can be seen from the table, the improved GELU activation function AdaptGELU proposed in the present invention has more advantages than other activation functions in the fault classification environment of this article. When 1D-XFormer-ReLU, 1D-XFormer-Sigmoid, 1D-XFormer-SiLU and 1D-XFormer-GELU are used for fault diagnosis, the Accuracy value of 1D-XFormer-ReLU is the highest, reaching 91.89%, and the Accuracy value of 1D-XFormer-AdaptGELU is 0.25% higher than that of 1D-XFormer-ReLU. After comparing the experimental results, the present invention selects AdaptGELU as the activation function of 1D-XFormer, and uses 1D-XFormer-AdaptGELU as the backbone network of the final fault diagnosis model.

[0078] The results show that the hybrid network that combines adaptive activation function, variable convolutional residual connection and transformer structure significantly improves the accuracy of the model in fault diagnosis under complex working conditions.

[0079] The experimental results show that the proposed adaptive lightweight 1D-Xception-Transformer (1D-XFormer) network is an effective method for classifying motor fault vibration signals. The proposed method has strong motor fault diagnosis capabilities under complex working conditions and the complexity of this method is lower than most other models.

[0080] The examples described in the present invention are merely descriptions of the preferred implementation modes of the present invention, and are not intended to limit the concept and scope of the present invention. Without departing from the design concept of the present invention, various modifications and improvements made to the technical solutions of the present invention by engineers and technicians in this field should all fall within the protection scope of the present invention.

Claims

1. A motor bearing fault diagnosis method based on 1D-XFormer model, characterized in that: The steps include: Collect motor bearing vibration signals and pre-process them; Construct an adaptive lightweight 1D-XFormer network for motor bearing fault diagnosis; The preprocessed bearing vibration signal is input into the 1D-XFormer network to obtain the diagnosis result.

2. The method according to claim 1, characterized in that The 1D-XFormer network is based on the Xception network architecture and is composed of three channels: input channel, intermediate channel, and output channel; Among them, in the input channel, the adaptive activation function AdaptGELU is used to replace the ReLU activation function in the Xception network architecture, and the deformable convolution DeformConv is used to replace the convolution Conv in the Xception network architecture; in the intermediate channel, the 1D-Transformer model is used as the intermediate channel structure; in the output channel, the adaptive activation function AdaptGELU is used to replace the ReLU activation function in the Xception network architecture, and the deformable convolution DeformConv is used to replace the convolution Conv in the Xception network architecture, and the depth separable convolution module SeparableConv after the residual connection is replaced.

3. The method according to claim 2, characterized in that The specific calculation formula of the adaptive activation function AdaptGELU is: Here, x represents the input, G represents a Gaussian random variable with zero mean and unit variance, erf(·) represents the Gaussian error function, and a represents a trainable parameter.

4. The method according to claim 2, characterized in that: The specific steps of deformable convolution include: From the input feature map The sampling pixel grid is in the middle, where C is the number of feature map channels, H is the feature map height, and W is the feature map width. The two-dimensional K×K convolution is used for sampling, and the position of the pixel in the receptive field is recorded as P n , the specific formula is: P n ={P ij }(0≤i,j≤K-1;0≤n≤K 2 -1;i,j,n∈N) (2) Offset point P on the feature map n The position of the feature map is determined by adding a weight coefficient to each sampling point to reduce the interference of irrelevant information on the feature map. The specific formula is: Among them, Υ(P out ) represents the feature map output after convolution operation, ω(·) represents the value learned by the network in convolution, x(·) represents the value after grid sampling on the input feature map, P out represents the output position after the convolution operation, ΔΥ n represents the feature weight, ΔP n Indicates P n The offset of Among them, x(·) uses bilinear interpolation method, the specific formula is: Among them, P represents the offset position, B(·,·) represents the two-dimensional bilinear interpolation kernel, and q represents the selection in the feature map X position.

5. The method according to claim 2, characterized in that: The 1D-Transformer model is a one-dimensional Transformer model, including L layers of sequential transformer encoders; Among them, each layer of the temporal transformer encoder contains two sub-layers: a multi-head self-attention layer and a position-intelligent fully connected feed-forward network layer; Among them, the multi-head self-attention layer includes a multi-head attention layer and a residual connection normalization layer; the position intelligent fully connected feedforward network layer includes a feedforward network layer and a residual connection normalization layer.

6. The method according to claim 1, characterized in that The diagnostic results include: normal state, inter-turn short circuit, rotor rod breakage, eccentricity fault, bearing inner ring fault and bearing outer ring fault.

7. A motor bearing fault diagnosis system based on 1D-XFormer model, characterized in that: include: The sensor, specifically a triaxial vibration acceleration sensor, is used to collect the vibration signal of the bearing; A transmission module, used for transmitting the collected signals; Pre-training module, used to build 1D-XFormer network and pre-train the network using existing datasets; A calculation module, used to input the signal into the 1D-XFormer network to obtain the diagnosis result; The display module is used to display and remind the obtained diagnostic results.

8. A storage medium, characterized in that: The storage medium stores instructions, and when a computer reads the instructions, the computer executes the motor bearing fault diagnosis method based on the 1D-XFormer model as described in any one of claims 1 to 6.

9. An electronic device, characterized in that: The invention comprises a processor and the storage medium as claimed in claim 8, wherein the processor executes instructions in the storage medium.

Citation Information

Cited By

  • Integrated power generation equipment health state supervision system

    CN120067837A