Double-convolution bearing fault diagnosis method based on extrusion excitation

Abstract

The invention relates to the technical field of bearing fault diagnosis, in particular to a double-convolution bearing fault diagnosis method based on extrusion excitation, and the method comprises the steps: collecting an original fault signal of a rolling bearing, and carrying out the preprocessing of the original signal; a CNN-LSTM model is constructed, local feature extraction and global feature extraction are carried out on the time domain signals of the original signals through a two-channel convolutional neural network, and fusion is carried out by using the local features and the global features; and classifying the fused features by using the improved double LSTM network, and outputting a fault type. The problems that an existing model is complex, and the processing efficiency and the processing precision need to be further improved are solved.

Description

Technical Field

[0001] The invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method based on double convolution of squeeze excitation. Background Art

[0002] Rolling bearings are indispensable components in rotating machinery, and their performance stability is crucial to the reliability of the entire system. Currently, rolling bearing fault diagnosis mainly relies on the analysis of its vibration signal.

[0003] Existing methods usually need to convert time domain vibration signals into frequency domain signals, and then extract local or global features to reveal the intrinsic characteristics of the signal through methods such as amplitude spectrum, power spectrum analysis, envelope spectrum analysis, short-time Fourier transform, empirical mode decomposition distribution, s transform or wavelet packet transform. In this process, not all statistical features are useful for fault diagnosis, and some features may increase the computational burden and reduce the accuracy of classification.

[0004] For example, the particle swarm optimization (PSO) algorithm is used to perform feature selection and decision fusion strategies to improve the classification performance of SVM. ANN is widely used in classification problems due to its ability to fit nonlinear models.

[0005] For example, a hierarchical deep learning algorithm containing three long short-term memory (LSTM) modules was proposed, and a dropout layer was added to each module to avoid overfitting; in addition, in order to detect rolling bearing damage in real time; another example is a deep convolutional neural network (CNN), which consists of three convolutional layers, three pooling layers, and one fully connected layer, and finally performs fault prediction through a softmax function;

[0006] However, the above methods have complex models, and their processing efficiency and accuracy need to be further improved. Summary of the invention

[0007] In view of the shortcomings of the existing methods, the present invention solves the problems that the existing models are complex, and the processing efficiency and processing accuracy need to be further improved.

[0008] The technical solution adopted by the present invention is: a bearing fault diagnosis method based on double convolution of squeeze excitation comprises the following steps:

[0009] Step 1: Collect the original signal of the rolling bearing fault and pre-process the original signal;

[0010] As a preferred embodiment of the present invention, the preprocessing includes Wiener filtering to remove noise.

[0011] As a preferred implementation manner of the present invention, the original signal includes: fault type, fault defect and normal signal.

[0012] As a preferred embodiment of the present invention, the fault types include: ball, inner ring and outer ring faults.

[0013] Step 2: Build a CNN-LSTM model, extract local features and global features of the original signal’s time domain signal through a dual-channel convolutional neural network, and fuse the local and global features; use the improved dual LSTM network to classify the fused features and output the fault type;

[0014] As a preferred embodiment of the present invention, a dual-channel convolutional neural network includes: a first channel and a second channel, the first channel is a stacked convolution module, the stacked convolution module is composed of two improved convolution layers and a pooling layer, the two improved convolution layers are composed of 1D convolution, a normalization layer and a squeeze excitation module; the convolution kernel sizes of the two 1D convolutions are 20×20 and 10×10 respectively; the second channel includes: 2 stacked convolution modules, the convolution kernel sizes of the two 1D convolutions in the stacked convolution module are both 6×6.

[0015] As a preferred embodiment of the present invention, the improved dual LSTM network includes: a first LSTM network, an attention mechanism, a second LSTM network, a regularization and a fully connected layer.

[0016] As a preferred embodiment of the present invention, accuracy, precision, recall and F1-score are used to verify the CNN-LSTM model.

[0017] As a preferred implementation of the present invention, mean square error is used as the loss function of the CNN-LSTM model.

[0018] As a preferred embodiment of the present invention, a bearing fault diagnosis system based on double convolution of squeeze excitation includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a bearing fault diagnosis method based on double convolution of squeeze excitation.

[0019] As a preferred embodiment of the present invention, a computer readable medium stores a computer program code, and when the computer program code is executed by a processor, a bearing fault diagnosis method based on double convolution of squeeze excitation is implemented.

[0020] Beneficial effects of the present invention:

[0021] 1. The present invention uses a dual convolutional neural network to extract time-domain features and combines it with an improved LSTM network to capture temporal dependencies. It improves training stability and convergence speed by adding normalization processing during feature extraction, and the squeeze excitation module improves the expressiveness of key features by weighting the features.

[0022] 2. The advantage of the present invention is that it provides a network model with a compact structure, low resource consumption, easy integration and deployment, has efficient data processing capabilities and provides high-precision fault diagnosis results, and can maintain high accuracy in noisy environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a bearing fault diagnosis model diagram based on double convolution of squeeze excitation of the present invention;

[0024] Figure 2 It is a schematic diagram of the improved 1-dimensional convolution of the present invention;

[0025] Figure 3 It is a confusion matrix effect diagram of the present invention. DETAILED DESCRIPTION

[0026] The present invention is further described below in conjunction with the accompanying drawings and embodiments. This figure is a simplified schematic diagram, which only illustrates the basic structure of the present invention in a schematic manner, and therefore it only shows the components related to the present invention.

[0027] like Figure 1 As shown, the bearing fault diagnosis method based on double convolution of squeeze excitation includes the following steps:

[0028] Step 1: Collect the original signal of the rolling bearing fault and pre-process the original signal;

[0029] The original signal includes: normal vibration signal and abnormal signal;

[0030] Preprocessing includes: the original signal sample contains 240 feature dimensions, each dimension has 2000 sampling points; Wiener filtering is used to reduce noise interference; the window size of Wiener filtering is set to 3; the filtered data is downsampled with a sampling rate of 1 / 8, that is, 1 sampling point is taken for every 8 sampling points, to reduce the data dimension and improve processing efficiency;

[0031] Wiener filtering is used to denoise the signal based on the minimum mean square error (MMSE) criterion. It tries to retain the original characteristics of the signal while denoising.

[0032] The transfer function H(f) in the frequency domain is defined as follows:

[0033]

[0034] Where H(f) is the frequency response of the Wiener filter; S xx (f) is the power spectral density (PSD) of the original signal; S nn (f) is the power spectral density of the noise.

[0035] The output in the time domain is defined as follows:

[0036]

[0037] in, is the output signal of the filter, that is, the estimated original signal; h(t) is the impulse response of the Wiener filter (that is, the coefficient of the filter); y(t) is the observed noisy signal; * represents the convolution operation.

[0038] Step 2: Construct a dual-channel convolutional neural network to extract local features and global features of the time domain signal of the original signal, and fuse the local and global features; use the improved dual LSTM network to classify the fused features and output the fault type;

[0039] A dual-channel convolutional neural network feature extractor is constructed, including two CNN networks with different kernel sizes, to extract the feature representation of the rolling bearing fault vibration signal; the extracted features are then input into the improved dual LSTM network classifier (i.e., stacked LSTM network) for bearing fault assessment;

[0040] The dual-channel convolutional neural network uses two parallel one-dimensional convolutional encoders with different kernel sizes and depths to extract features from the input signal respectively;

[0041] like Figure 2 , where the first channel CNN_1 extracts global features from the original time domain signal, including: 1 stacked convolution module, the stacked convolution module consists of two improved convolution layers and a pooling layer; the two improved convolution layers both include: 1D convolution, normalization layer and squeeze excitation module; the convolution kernel sizes of the two 1D convolutions are 20×20 and 10×10 respectively;

[0042] The second channel CNN_2 extracts local features from the original time domain signal, including: 2 stacked convolution modules, where the convolution kernel size of the two 1D convolutions in the stacked convolution module is 6×6;

[0043] The feature vectors of the first channel CNN_1 and the second channel CNN_2 are fused element-wise through the Multiply layer;

[0044] One-dimensional CNN has several advantages as a feature extractor: it can automatically learn the potential meaning of different vibration signals; the use of a shared weight strategy greatly reduces the dimensional input vector and the number of parameters; and the output of the previous layer and the input of the next layer are limited by the kernel size; both feature extractors are composed of multiple convolutional layers, normalization layers, squeeze excitation modules, and maximum pooling layers; the convolutional layer extracts features from time series signals; the normalization layer helps speed up training and stabilize model performance; the squeeze excitation module dynamically adjusts the weights of channels according to the average values ​​of different channels, emphasizing important features and suppressing unimportant parts; the maximum pooling layer downsamples the output to reduce the data dimension, which can prevent overfitting;

[0045] Two independent convolutional neural network channels (CNN_1 and CNN_2) are introduced, each of which is specially optimized for feature extraction at different scales. Specifically, CNN_1 and CNN_2 use different sizes of convolution kernels, depths of convolution layers, and feature extraction strategies to effectively capture local and global features in vibration signals. CNN 1 focuses on extracting global features of signals. It uses larger convolution kernels (20×20 and 10×10) and a larger stride 2 to capture features and dynamic trends over longer time spans in vibration signals. CNN 2 focuses on extracting local features of signals. It uses a smaller convolution kernel (6×6) and a smaller stride 1, combined with more convolution layers, to deeply capture subtle changes (i.e., local features) in short time scales in signals. The formula for one-dimensional convolution (1D-CNN) is as follows:

[0046]

[0047] in, and are the weights and biases of the i-th convolution kernel and the l-th layer, respectively, and x l (j) is the jth local input of the lth layer.

[0048] Operation of the squeeze excitation module: The process of a squeeze excitation module is divided into two steps: squeeze (Squeeze) and excitation (Excitation). Squeeze obtains the global compressed feature vector of the current feature layer by performing global average pooling on the feature map. The squeeze operation obtains the weight of each channel in the feature map through two layers of full connection, and uses the weighted feature layer as the input of the next layer of the network.

[0049] Squeeze:

[0050] The role of the squeeze part is to obtain the global information embedding (feature vector) of each channel of U. In the squeeze excitation module, this step is achieved by introducing global average pooling (GAP), that is, by finding the average value of the feature map of each channel c, c∈{1,C}:

[0051]

[0052] Excitation:

[0053] The role of the Excitation part is to pass z c To learn the feature weights of each channel in C, a gate mechanism consisting of two layers of full connections is used. The calculation method of the gate control unit s is expressed as:

[0054] s=F ex (z, W) = σ(g(z, W)) = σ(W2δ(W1z)) (5) where δ represents the ReLU activation function and σ represents the sigmoid activation function. are the weight matrices of the two fully connected layers, and r is the number of hidden nodes in the middle layer.

[0055] After obtaining the gate control unit s, the final output Expressed as the vector product of s and U, that is, F scale (·,·)operate:

[0056]

[0057] in yes A Feature Map of a feature channel, s c is a scalar value in the gated cell s.

[0058] The improved classification filter of the dual LSTM network includes: a first LSTM network, an attention mechanism, a second LSTM network, a regularization and a fully connected layer;

[0059] The improved dual LSTM network is used to build a complex nonlinear model between input and output; the hidden state of the first LSTM network (LSTM_1) provides input for the second LSTM network (LSTM_2), and the output of LSTM_2 is fed into a fully connected layer for classification; the input is divided into several steps to enter the LSTM network; the output of the next step will be affected by the output of the previous step; therefore, the LSTM network makes good use of the characteristics of the timing signal and fully extracts the internal features of the vibration signal; finally, the softmax function converts the neuron output into a probability distribution of more than 10 rolling bearing health states; the construction of the CNN-LSTM model is completed;

[0060] The LSTM network includes: input, forget, output, cell and hidden gate updates can be expressed as follows:

[0061] i t =σ(W i x t +V i h t-1 +b i ) (7)

[0062] f t =σ(W f x t +V i h t-1 +b f ) (8)

[0063] o t =σ(W o x t +V o h t-1 +b o ) (9)

[0064] c t =f t ⊙c t-1 +i t ⊙tanh(W c x t +V c h t-1 +b c ) (10)

[0065] h t =o t ⊙tanh(c t ) (11)

[0066] Where b represents the bias; W and V represent the input and hidden state weights respectively; in the tth update step, the input gate i, forget gate f, input gate o and unit state c are updated by the input and hidden states of step n-1.

[0067] Taking into account the generalization processing problem of the hidden layer output of the LSTM network, an attention model is added between the hidden layer and the output layer to calculate the attention weight value of the LSTM hidden layer output, and then the weighted average calculation is performed to obtain the output; the addition of the attention mechanism allows the model to pay more attention to the time points and features that are most critical for fault diagnosis, thereby improving the performance and generalization ability of the model.

[0068] Train and test the CNN-LSTM model;

[0069] Dataset selection: The Case Western Reserve University (CWRU) rolling bearing dataset is used to verify the proposed CNN-LSTM model; data collected from an accelerometer located at the drive end (DE) is selected with a sampling frequency of 48kHz; the motor power and speed are fixed at 3hp and 1730rpm respectively; three bearing fault types are considered: ball, inner ring and outer ring faults; each fault type is induced in the rolling bearing by spark discharge and has different fault diameters (0.007, 0.014 and 0.021 inches); therefore, there are a total of 10 faults: 3 fault types × 3 defect sizes + 1 normal signal; each fault has a different and unique vibration signal.

[0070] Training set and test set division: In order to ensure the effectiveness and robustness of the model of the present invention, each bearing state (including healthy state) contains 240 samples obtained from the accelerometer in the time domain, and a function is used to divide the data into 70% for training and 30% for testing; therefore, there are a total of 1680 training samples and 720 test samples, each of which contains 2000 continuous vibration signals; the model uses accuracy, precision, recall and F1-score to verify the model.

[0071] Parameter initialization:

[0072] Initialize the best accuracy and model parameters, use dropout in training to avoid overfitting after the two-layer LSTM network, and use mean square error (MSE) as the loss function to evaluate the network performance. The mathematical expression is as follows:

[0073]

[0074] Where n is the number of samples, x i is the true value of the i-th sample, y i is the model's predicted value for the i-th sample.

[0075] The model parameters are then updated using the Adam optimizer, with the learning rate set to 0.0006. Backpropagation is used to update the weights, and the cycle is repeated for 400 epochs. The best model parameters recorded during the training process are loaded to complete the bearing fault classification.

[0076] According to the experimental records, the test accuracy rate reached up to 99.72%. Figure 3 The confusion matrix is ​​shown in Figure 2.

[0077] Table 1 and Table 2 are the accuracy comparison between the proposed model and the existing method, and the ablation experiment respectively;

[0078] Table 1 Comparative experimental results of CWRU dataset

[0079]

[0080] Table 2 Model ablation experiment results

[0081]

[0082] The present invention utilizes a dual convolutional neural network to extract time domain features, combined with an improved LSTM network to capture temporal dependencies; normalization processing is added during the feature extraction process to improve training stability and convergence speed, and a squeeze excitation module weights features to enhance the expressiveness of key features; the present invention has the advantage of providing a network model that is compact in structure, occupies less resources, is easy to integrate and deploy, has efficient data processing capabilities, provides high-precision fault diagnosis results, and can maintain high accuracy in noisy environments.

[0083] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A bearing fault diagnosis method based on double convolution of squeeze excitation, characterized in that: The following steps are involved: Step 1: Collect the original signal of the rolling bearing fault and pre-process the original signal; Step 2: Build a CNN-LSTM model, extract local features and global features of the time domain signal of the original signal through a dual-channel convolutional neural network, and fuse the local and global features; use the improved dual LSTM network to classify the fused features and output the fault type.

2. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 1 is characterized in that: The dual-channel convolutional neural network includes: a first channel and a second channel. The first channel is a stacked convolution module. The stacked convolution module consists of two improved convolution layers and a pooling layer. The two improved convolution layers are composed of 1D convolution, a normalization layer and a squeeze excitation module. The convolution kernel sizes of the two 1D convolutions are 20×20 and 10×10 respectively. The second channel includes: 2 stacked convolution modules. The convolution kernel sizes of the two 1D convolutions in the stacked convolution module are both 6×6.

3. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 2 is characterized in that: The improved dual LSTM network includes: the first LSTM network, the attention mechanism, the second LSTM network, the regularization and the fully connected layer.

4. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 1 is characterized in that: The CNN-LSTM model was validated using accuracy, precision, recall and F1-score.

5. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 1 is characterized in that: The mean square error is used as the loss function of the CNN-LSTM model.

6. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 1 is characterized in that: Preprocessing includes Wiener filtering to remove noise.

7. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 1 is characterized in that: The original signal includes: fault type, fault defect and normal signal.

8. The bearing fault diagnosis method based on double convolution of squeeze excitation according to claim 7 is characterized in that: Fault types include: ball, inner race and outer race faults.

9. The bearing fault diagnosis system based on double convolution of squeeze excitation is characterized by: include: a memory for storing instructions executable by a processor; A processor, used to execute instructions to implement the bearing fault diagnosis method based on double convolution of squeeze excitation as described in any one of claims 1-8.

10. A computer readable medium storing computer program code, characterized in that: When the computer program code is executed by a processor, the computer program code implements the bearing fault diagnosis method based on double convolution of squeeze excitation as described in any one of claims 1 to 8.

Images