Bearing fault diagnosis method and system based on twin neural network under small sample

Abstract

The invention provides a bearing fault diagnosis method and system based on a twin neural network under a small sample, and relates to the technical field of mechanical fault diagnosis and artificial intelligence. The method comprises the following steps: collecting X-axis, Y-axis and Z-axis vibration signals of a bearing, slicing, and generating a time-frequency grey-scale map through short-time Fourier transform to enhance feature expression; the method comprises the following steps: constructing same-class and different-class sample pair training sets, training by adopting a weight-shared twin neural network model which comprises two same sub-networks, and optimizing model parameters by calculating the Euclidean distance of sample pair features and combining a cross entropy loss function; in the test stage, unknown samples are matched based on One-shot and N-shot strategies, and the fault category is judged with the maximum similarity probability. According to the method, the accuracy of 95% or above is achieved under 70 training samples, the noise immunity and generalization ability under the small sample condition are remarkably improved, and the method is suitable for low-cost intelligent diagnosis of industrial equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical fault diagnosis and artificial intelligence, and specifically relates to a bearing fault diagnosis method and system based on a siamese neural network under small samples, which is applicable to the intelligent monitoring and fault identification of the health status of bearings in industrial equipment. Background Art

[0002] As a core component of rotating machinery, the failure of bearings may lead to equipment shutdown or even safety accidents. Traditional fault diagnosis methods rely on a large amount of labeled data to train models. However, in actual industrial scenarios, fault samples are scarce, resulting in poor generalization ability and low diagnostic accuracy of the models. Existing deep learning methods, such as CNN and RNN, are prone to overfitting problems under small sample conditions. Although multi-sensor data fusion technology can improve the richness of features, it lacks effective processing means for small samples. Therefore, there is an urgent need for a method that can achieve high-precision bearing fault diagnosis under small sample conditions. Summary of the Invention

[0003] The purpose of the present invention is to provide a bearing fault diagnosis method and system that can still achieve high precision, strong generalization ability, and good robustness when the number of samples is extremely small, so as to make up for the defects of low recognition accuracy and poor anti-noise performance of the existing technology under small sample conditions.

[0004] In response to the recognition requirements of various fault states of bearings under complex working conditions, the present invention proposes an end-to-end diagnosis model based on a siamese neural network. By constructing two convolutional sub-networks with shared weights to extract the time-domain features of sample pairs and using feature similarity to judge the fault categories, the classification difficulties caused by insufficient samples are effectively overcome.

[0005] The main contents of the present invention include:

[0006] A bearing fault diagnosis method and system based on a siamese neural network under small samples, comprising the following steps:

[0007] (1) Collect the X, Y, and Z-axis vibration signals of the bearing, slice the collected original vibration signals, and divide them into a training set and a test set;

[0008] (2) Convert single-channel, double-channel, and triple-channel vibration signals into time-frequency grayscale images through short-time Fourier transform;

[0009] (3) Construct a training data set including pairs of like samples and pairs of unlike samples;

[0010] (4) Train using a siamese neural network model with shared weights, the model includes two identical sub-networks, and each sub-network includes a convolutional layer, a pooling layer, and a fully connected layer;

[0011] (5) By calculating the Euclidean distance of the features of the samples and combining with the cross-entropy loss function to optimize the model parameters;

[0012] (6) Adopt the One-shot and N-shot strategies to classify the faults of the test samples.

[0013] In the step (2), the time-frequency grayscale map is generated by the short-time Fourier transform to enhance the time-frequency expression ability of the features. The short-time Fourier transform is defined as:

[0014]

[0015] where x(m) is the input signal, ω(m) is the window function, and X(n,ω) is the two-dimensional function of time n and frequency ω.

[0016] In the step (3), the construction ratio of the sample pairs is 1:1 for the same-class sample pairs and different-class sample pairs.

[0017] The sub-network in the step (4) specifically includes: the input layer receives the time-frequency grayscale map. Each twin sub-network contains 5 convolutional layers, and each convolutional layer is followed by a ReLU activation function and a max-pooling layer. The fully connected layer outputs the low-dimensional feature vector. ReLU is defined as:

[0018]

[0019] where m is the input to this activation function and q is the output of this activation function.

[0020] In the step (5), calculate the Euclidean distance of the features of the sample pairs and combine with the cross-entropy loss function to optimize the model parameters until the model converges. The Euclidean distance is defined as:

[0021]

[0022] where x1 and x2 represent the input sample pairs, and f(x1) and f(x2) represent the low-dimensional features extracted from the input sample pairs by the sub-network. represents the Euclidean distance of the low-dimensional features.

[0023] The loss function for a single training is defined as:

[0024] Loss(x1,x2,y)=ylog(P(x1,x2))+(1-y)log((1-P(x1,x2)))

[0025] where y is the sample label, P(x1,x2) represents the difference degree of the sample pairs. The training objective of the twin network is that when the input sample x1 and the input x2 belong to the same health state, the output difference degree is as close to 0 as possible; when the input sample x1 and the input x2 do not belong to the same health state, the output difference degree is as close to 1 as possible.

[0026] In step (6), the One-shot strategy specifically constructs sample pairs by pairing a test sample with one sample of each class, and the N-shot strategy constructs sample pairs with N samples of each class.

[0027] A bearing fault diagnosis system includes: a multi-channel vibration signal acquisition module for acquiring the X, Y, and Z-axis vibration signals of the bearing; a time-frequency grayscale map conversion module for converting the vibration signals into time-frequency grayscale maps through short-time Fourier transform; a sample pair construction module for generating same-class sample pairs and different-class sample pairs; a siamese neural network training and diagnosis module for training the siamese neural network model according to any one of claims 1-6 and classifying the faults of the test samples.

[0028] The output of the siamese neural network training and diagnosis module is the similarity probability of the sample pairs, and the fault class of the test sample is determined by the maximum value of the probability.

[0029] The main innovations of the present invention include:

[0030] 1. Fault diagnosis modeling idea for small sample scenarios: Break through the bottleneck of the traditional neural network's dependence on large-scale data, use the similarity between sample pairs as the supervision signal, and effectively expand the number of training sample combinations;

[0031] 2. Optimized design of the siamese neural network structure: Integrate a five-layer convolutional pooling network to improve the feature extraction and discrimination ability without increasing the model complexity, and adapt to the structural features of the original vibration signals. The following table shows the main structural parameters of the siamese neural network:

[0032]

[0033] 3. Classifier-free discrimination mechanism: The output is the similarity probability values of two samples, avoiding the limitations of traditional "softmax classification" and improving the versatility and deployment flexibility of the model;

[0034] 4. Multi-channel data fusion mechanism: Introduce the joint analysis of signals in the X, Y, and Z directions and process them through short-time Fourier transform to enhance the time-frequency expression ability of features, improve the sensitivity of the model to the essential features of faults, and significantly improve the recognition accuracy.

[0035] Through the above innovative designs, the present invention can achieve high-accuracy recognition of multiple bearing fault types on the premise of using only a very small number of training samples. Experimental results show that when the number of training samples is only 140, the recognition accuracy can reach more than 90%, far superior to the existing 1D-CNN and SVM models.

[0036] The bearing small-sample fault diagnosis method provided by the present invention realizes the identification of high precision and multiple fault states without the need for a large number of samples, significantly improving the practicability and generalization ability of the diagnosis model. By constructing a siamese neural network structure and taking feature similarity as the core discrimination basis, the problems of easy overfitting and poor diagnosis performance of traditional models under small-sample conditions are effectively overcome. The introduction of a multi-channel vibration signal fusion strategy further enhances the model's ability to perceive fault information, showing good advantages in terms of anti-noise performance, interpretability, and scalability, and having broad industrial application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a schematic diagram of the diagnosis process of the fault diagnosis model of the present invention.

[0038] Figure 2 is the division process of the training set and test set of the small-sample data set.

[0039] Figure 3 is the structural diagram of the siamese neural network fault diagnosis model of the present invention.

[0040] Figure 4 is the t-SNE visualization clustering effect diagram after feature dimensionality reduction. DETAILED DESCRIPTION OF THE INVENTION

[0041] The present invention provides a bearing fault diagnosis method and system based on a siamese neural network under small samples. The specific implementation manners thereof include the following technical details:

[0042] Build a bearing fault simulation system on the experimental platform. The vibration signal acquisition system consists of triaxial acceleration sensors installed on the bearing end seats, and the vibration signals in the X, Y, and Z directions are collected respectively. The sensor selection has industrial-grade anti-interference ability, and a data acquisition card with high sampling accuracy is used to synchronously sample the signals. The sampling frequency is set to 25.6 kHz to ensure that high-frequency impacts and characteristic frequency components during the bearing operation are captured.

[0043] During the signal acquisition process, the sensors are connected to the acquisition card through shielded twisted pair wires, and all signal channels are preprocessed by an anti-aliasing filter to suppress high-frequency noise and improve the signal quality.

[0044] In the signal processing link, the length of the collected original vibration signal is 61,440 sampling points, which are obtained from the X, Y, and Z channels respectively. The original signal is windowed, the slice length is 1024, and the sliding step is 50 sampling points. The sample data is converted into a time-frequency grayscale map through short-time Fourier transform for constructing training samples and test samples.

[0045] In terms of network architecture, the twin neural network employed consists of two fully symmetrical sub-networks with shared weights. These networks are primarily used for deep feature extraction and similarity calculation of input vibration signal sample pairs. Each sub-network employs a one-dimensional convolutional neural network structure, consisting of five stacked convolutional and pooling layers, resulting in strong time-domain feature extraction capabilities. The early layers of the network use larger strides and wider convolution kernels to extract macroscopic features, while subsequent layers employ smaller convolution kernels to gradually capture localized details, thereby enhancing the ability to identify subtle fault signatures.

[0046] Each convolutional layer uses the ReLU activation function, introducing nonlinear mapping to enhance network expressiveness. Each convolution operation is followed by a max pooling operation to reduce feature dimensionality, enhance translation invariance, and prevent overfitting. Following the convolutional pooling layer, a fully connected layer containing 10 neurons compresses the feature map into a fixed-dimensional embedding vector, which represents the position of the input sample in the low-dimensional feature space. The two sub-networks extract features from each input sample pair and calculate the Euclidean distance between their embedding vectors, which serves as a similarity metric between the samples. Finally, a sigmoid activation function is used to map the distance to a probability value between 0 and 1, which is used to determine whether two samples are from the same health state. The overall network structure is compact, and the parameter sharing mechanism effectively improves the model's training efficiency and generalization ability, making it particularly suitable for fault diagnosis scenarios where samples are scarce.

[0047] During the model training phase, samples in the training set are paired according to their labels to generate "similar sample pairs" and "heterogeneous sample pairs" as model input. A cross-entropy loss function is used to construct the training objective, calculating the error between the model output and the true label (1 or 0). The Adam optimizer is used to update weights. During training, the validation set accuracy is monitored, and an early termination mechanism is implemented to prevent overfitting.

[0048] After model training is complete, the testing phase begins. Vibration signals of unknown health states are collected in real time and sliced to generate the test sample x′. One or N reference samples are randomly selected from each training sample category and paired with x′ for one-shot or N-shot matching. These sample pairs are then fed into the twin neural network, where similarity probabilities are calculated. The category with the highest probability is selected as the judgment label for x′. A multi-channel fusion strategy is implemented, and single-channel X, dual-channel XY, and triple-channel XYZ input methods are tested under small sample conditions. Results show that triple-channel fusion significantly improves recognition accuracy, achieving over 95% accuracy with just 70 training samples.

[0049] The present invention realizes high-precision bearing fault recognition with extremely few training samples by combining a small-sample learning strategy, a Siamese neural network architecture design, and multi-channel data fusion, and is applicable to the fault diagnosis requirements of high reliability and low cost in industrial fields.

Claims

1. A bearing fault diagnosis method and system based on twin neural network under small sample conditions, characterized by The following steps are involved: (1) Collect the X, Y, and Z axis vibration signals of the bearing, slice the collected original vibration signals, and divide them into training sets and test sets; (2) Convert the single-channel, dual-channel, and triple-channel vibration signals into time-frequency grayscale images through short-time Fourier transform; (3) Construct a training dataset containing similar sample pairs and heterogeneous sample pairs; (4) A weight-sharing twin neural network model is used for training, wherein the model comprises two identical sub-networks, each of which includes a convolutional layer, a pooling layer, and a fully connected layer; (5) Optimize model parameters by calculating the Euclidean distance between sample pairs and features and combining the cross entropy loss function; (6) One-shot and N-shot strategies are used to classify faults of test samples.

2. The method according to claim 1, characterized in that, In step (2), the time-frequency grayscale image is generated by short-time Fourier transform to enhance the time-frequency expression capability of the feature. The short-time Fourier transform is defined as: Where x(m) is the input signal, ω(m) is the window function, and X(n,ω) is a two-dimensional function of time n and frequency ω.

3. The method according to claim 1, wherein The sample pairs constructed in step (3) are constructed with a ratio of 1:1 between similar sample pairs and heterogeneous sample pairs.

4. The method according to claim 1, wherein The sub-network in step (4) specifically includes: an input layer receiving a time-frequency grayscale image, each twin sub-network comprising five convolutional layers, each convolutional layer being followed by a ReLU activation function and a maximum pooling layer, and a fully connected layer outputting a low-dimensional feature vector, where ReLU is defined as: Where m is the input to the activation function and q is the output of the activation function.

5. The method according to claim 1, wherein The step (5) calculates the Euclidean distance of the sample pair features and optimizes the model parameters in combination with the cross entropy loss function until the model converges. The Euclidean distance is defined as: where x1 and x2 represent input sample pairs, and f(x1) and f(x2) represent the low-dimensional features extracted from the input sample pairs by the sub-network. represents the Euclidean distance of the low-dimensional features. The loss function for a single training run is defined as: Loss(x1,x2,y)=ylog(P(1,x2))+(1-y)log((1-P(x1,x2))) Where y is the sample label, P(x1,x2) represents the difference between the sample pairs, and the training goal of the twin network is that when the input sample x1 and the input x2 belong to the same health state, the output difference is as close to 0 as possible; when the input sample x1 and the input x2 do not belong to the same health state, the output difference is as close to 1 as possible.

6. The method according to claim 1, characterized in that, In step (6), the one-shot strategy specifically constructs a sample pair with the test sample and one sample from each class, and the N-shot strategy constructs a sample pair with N samples from each class.

7. A bearing fault diagnosis system, characterized in that, include: Multi-channel vibration signal acquisition module, used to collect X, Y, and Z axis vibration signals of the bearing; A time-frequency grayscale image conversion module is used to convert the vibration signal into a time-frequency grayscale image through short-time Fourier transform; a sample pair construction module is used to generate similar sample pairs and heterogeneous sample pairs; a twin neural network training and diagnosis module is used to train the twin neural network model described in any one of claims 1-6 and perform fault classification on the test samples.

8. The system according to claim 7, characterized in that, The output of the twin neural network training and diagnosis module is the similarity probability of the sample pairs, and the fault category of the test sample is determined by the maximum probability.

Images