Rolling bearing fault diagnosis method based on GAF-MMCNN

Abstract

The invention discloses a rolling bearing fault diagnosis method based on GAF-MMCNN, and the method comprises the steps: obtaining a vibration signal, and obtaining a bearing data set; data preprocessing: sampling the one-dimensional original data of various bearing fault types to obtain a data sample; a GAF coding method is adopted to obtain two-dimensional images of various faults, one-dimensional vibration signals of the bearing are converted into two-dimensional feature images, and the two-dimensional feature images are input into a neural network for feature extraction; constructing a convolutional neural network (CNN) model; based on the GAF image and the CNN structure, constructing a GAF-MMCNN fault diagnosis model; and inputting the image into the MMCNN fault diagnosis model to output a classification result. The MMCNN module designed by the invention integrates a multi-scale branch and a multi-scale cascade structure, obtains information of different scales by using convolution kernels of different scales, fuses shallow spatial information and deep semantic information, and extracts more features from vibration signals; the CFAM channel attention module can adaptively allocate the weight of each channel in the network model training process, and improves the accuracy of bearing fault diagnosis.

Description

Technical Field

[0001] The present invention relates to a rolling bearing fault diagnosis method, in particular to a rolling bearing fault diagnosis method based on GAF-MMCNN. Background Art

[0002] With the continuous advancement of "smart manufacturing," modern mechanical equipment is developing toward automation, efficiency, and intelligence. Rolling bearings are a crucial component of mechanical systems and play an indispensable role in applications such as automobiles. Their performance impacts the efficiency of mechanical equipment. Due to harsh operating environments, rolling bearings are prone to failures such as pitting corrosion, which can lead to cracks and spalling on the inner and outer rings and rolling elements during operation. Bearing vibration signals contain bearing fault information, which can be analyzed and processed to identify bearing health. Because bearings typically operate under complex and variable operating conditions such as variable loads and speeds, extracting bearing fault signal characteristics and diagnosing fault types under these conditions and with insufficient sample size is crucial.

[0003] Traditional rolling bearing fault diagnosis methods require manual signal processing and rely on signal processing techniques and expert experience. While these methods can effectively predict bearing faults, they are subject to significant human intervention, often resulting in poor robustness, weak generalization, and difficulty for technicians to perform fault diagnosis tasks. Hinton et al. first proposed the concept of deep learning, which uses machine learning models with multiple hidden layers to mine valuable feature information from data. Deep learning is currently used in various fields, including recognition, natural language processing, semantic segmentation, and speech processing, and is now also involved in the field of fault diagnosis. With the advancement of computer hardware, current research on bearing fault diagnosis is largely based on deep learning, with convolutional neural networks (CNNs) being a popular approach. Their powerful data mining and feature-adaptive learning capabilities have attracted significant attention, but researchers are increasingly dissatisfied with their practical applications. While improvements to CNN architectures can improve training efficiency, the problem of fault information loss in raw signals persists in the field of bearing fault diagnosis. Effectively utilizing this information has become a key issue in the field.

[0004] Research has shown that superimposing one-dimensional vibration signals to generate two-dimensional fault feature images and extracting image features using a multidimensional convolutional neural network eliminates the manual effort required to extract fault information from raw bearing signals. The Gramian angular field (GAF) algorithm can fully leverage the relevant information in time series to convert one-dimensional time series into two-dimensional images, creating a one-to-one mapping between the image and the one-dimensional time series. Each element in the matrix represents the temporal correlation strength of a pair of signal data points by superimposing them in the time interval direction, facilitating the extraction of rich information from the time series. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a rolling bearing fault diagnosis method based on GAF-MMCNN, which effectively improves the generalization performance of the model and solves the problem of low fault classification accuracy when the actual working conditions of rolling bearings vary greatly and the sample size is insufficient.

[0006] Technical solution: The present invention includes step S1, acquiring a vibration signal to obtain a bearing data set;

[0007] Step S2: data preprocessing, sampling one-dimensional raw data of various bearing fault types to obtain data samples;

[0008] In step S3, the GAF encoding method is used to obtain two-dimensional images of various faults. The one-dimensional vibration signal of the bearing is converted into a two-dimensional feature image and input into the neural network for feature extraction. First, the data samples are normalized. Second, the polar coordinate system is used instead of the typical Cartesian coordinate system to represent the time series. Finally, the trigonometric sum / difference between each point is considered, and the angular perspective is used to identify the time correlation within different time intervals, thus preserving the time correlation of the original vibration signal.

[0009] Step S4, constructing a convolutional neural network (CNN) model;

[0010] Step S5: Based on the GAF image and CNN structure, a GAF-MMCNN fault diagnosis model is constructed. The model integrates the multi-scale branch module and the multi-scale cascade module, constructs a convolutional fusion attention module (CFAM), enhances image features, and assigns attention weights to each channel.

[0011] Step S6: input the image into the MMCNN fault diagnosis model to output the classification result.

[0012] The step S3 specifically includes:

[0013] Step S31, time series X={x1,x2,...,x n} contains n observations, rescale X so that all values fall within the interval [-1, 1] or [0, 1] using the following formula:

[0014]

[0015] Step S32, by encoding the value as the cosine of the angle and the timestamp as the radius, the rescaled time series in the polar coordinate system is expressed by formula

[0016]

[0017] Among them, t i is the timestamp, N is a constant factor;

[0018] In step S33, after converting the rescaled time series into a polar coordinate system, angular perspective is used to identify temporal correlations within different time intervals by considering the trigonometric sums / differences between each point. GAF is defined as follows:

[0019]

[0020] Where I is the unit row vector, and Represents different row vectors.

[0021] In step S22, when converted to the polar coordinate system, the scaled data of the two normalization operations correspond to different angle ranges respectively. The data in the range [-1, 1] corresponds to the arc cosine function angle range of [0, π], and the data in the range [0, 1] corresponds to the arc cosine value range of [0, π / 2].

[0022] In step S33, samples of each fault type are divided into a training set and a test set at a ratio of 3:1.

[0023] The step S4 specifically includes:

[0024] In step S41, the expression of the convolution operation is as follows:

[0025]

[0026] in, is the output feature map of layer l, f(·) is the activation function, M is the number of output feature maps, is the weight of the convolution kernel, is the bias of layer l;

[0027] Step S42: Pooling operation, the formula is described as:

[0028]

[0029] in, is the output value of the t-th neuron in the ith channel of the l-th layer, W is the size of the pooling kernel, and P i l+1 (j) is the output value of the l+1th channel of the i-th layer;

[0030] Step S43: Select the ReLU activation function, the expression is as follows:

[0031]

[0032] Step S44: The fully connected layer expression is as follows:

[0033] y l =f{(w l ) T x l-1 +b l}

[0034] Among them, x l-1 is the output value of layer l-1, w l is the weight, y l is the output of layer l, b l is the bias term;

[0035] Step S45: The Softmax function formula is as follows:

[0036]

[0037] Where i = 1, 1, ..., k ..., k, K is the number of output nodes, P i represents the probability of each output,

[0038] The step S5 specifically includes:

[0039] Step S51, constructing a convolutional fusion attention module CFAM;

[0040] Step S52: CFAM includes feature compression and weight generation. In the feature compression stage, GAP and GMP are used to compress the input features along the height and width dimensions, and the outputs are PGAP and PGMP respectively. In the weight generation stage, the compressed features are flattened into a one-dimensional tensor and stacked along the channel dimension to output the double-pooled features. The entire CFAM calculation process is as follows:

[0041]

[0042] P GMP =maxx(i,j)

[0043]

[0044] Y T =CovT(Cov(Y P ))

[0045] z=reshape(Y T )

[0046] Among them, P GAP and P GMP are the one-dimensional data after GAP and GMP compression, i and j are the positions of the input value x in the two-dimensional data, flatten(·) represents the stacking along the channel dimension, Y P is the fused pooling information, Cov(·) smoothing and CovT(·) smoothing represent convolution and transposed convolution operations, Y T is the middle layer of weight information, reshape(·) represents the shape of 1×1×C, and z is the final generated channel attention weight;

[0047] Step S53: In order to construct MMCNN, the multi-scale branch structure and the multi-scale cascade structure are integrated. Meanwhile, the BN layer and the CFAM attention mechanism are added.

[0048] The CFAM utilizes global average pooling (GAP) and global maximum pooling (GMP) to compress features.

[0049] The CFAM uses one-dimensional convolution to reduce the size and number of ascending channels, selects transposed convolution to restore the size, and the number of channels becomes 1; finally, the Sigmoid activation function is selected to generate weights and multiply the weights with the input features.

[0050] The GAF-MMCNN structure includes: feature screening network; deep feature extraction network; auxiliary classifier; main classifier.

[0051] The step S6 uses the accuracy ACC to evaluate the experimental results, and the formula is as follows:

[0052]

[0053] Among them, TP represents the frequency of predicting positive samples as positive, TN represents the frequency of predicting negative samples as negative, FP represents the frequency of predicting negative samples as positive, and FN represents the frequency of predicting positive samples as negative.

[0054] Beneficial effects: The MMCNN module designed in the present invention integrates multi-scale branches and multi-scale cascade structures, can use convolution kernels of different scales to obtain information of different scales, fuse shallow spatial information and deep semantic information, and extract more features from vibration signals; the CFAM channel attention module can adaptively allocate the weights of each channel during the network model training process, can more accurately locate valuable features, and extract more effective feature information with fewer samples, further improving the accuracy of bearing fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 is a flow chart of the present invention;

[0056] Figure 2 1 is a structural diagram of a convolutional fusion attention module (CFAM) in an embodiment of the present invention;

[0057] Figure 3 is a multi-scale branch structure in an embodiment of the present invention;

[0058] Figure 4 It is a multi-scale cascade structure in the example of the present invention. DETAILED DESCRIPTION

[0059] The present invention will be further described below with reference to the accompanying drawings.

[0060] like Figure 1 As shown, the rolling bearing fault diagnosis method based on GAF-MMCNN of this embodiment includes the following steps:

[0061] Step S1: Acquire a vibration signal to obtain a rolling bearing data set.

[0062] The rolling bearing signals used in the experiment are from the Bearing Data Center of the Electrical Laboratory of Case Western Reserve University in the United States. The experimental platform consists of a 1.5kW (2hp) motor, a torque sensor / converter, and a power tester. The speed is 1772r / min, the sampling frequency is 12kHz, and the sampling time is 10s. Single-point faults are introduced into the test bearings using electrospark machining, and the fault diameters are 7mils, 14mils, 21mils, 28mils, and 40mils (1mil-0.001 inches). All experimental data use drive end acceleration data, which is called vibration acceleration signal. The present invention uses 6 types of fault data and normal types, namely: (a) Normal. (b) 7mils, slight inner rail fault. (c) 28mils, severe inner rail fault. (d) 7mils, slight outer rail fault. (e) 28mils, severe outer rail fault. (f) 7mils, slight ball fault. (g) 28mils, severe ball fault. The data is segmented according to the length of the set sample. In the present invention, the length of the segmented samples is 2048, with a total of 2800 samples and signals in 7 different states.

[0063] Step S2, data preprocessing, samples the one-dimensional original data of the seven bearing fault types in the CWRU dataset to obtain data samples.

[0064] In step S3, the GAF encoding method is used to obtain two-dimensional images of seven types of faults. The one-dimensional vibration signal of the bearing is converted into a two-dimensional feature image and input into a neural network for feature extraction. This method is based on the GAF image that considers the time-position relationship. The two-dimensional feature image after GAF transformation has richer fault information. First, the data samples are normalized. Then, the polar coordinate system is used instead of the typical Cartesian coordinate system to represent the time series. Finally, the trigonometric sum / difference between each point is considered to use angular perspective to identify the time correlation within different time intervals, thereby retaining the time correlation of the original vibration signal. Specifically, it includes:

[0065] Step S31, time series X={x1,x2,...,x n} contains n observations, rescale X so that all values fall in the interval [-1, 1] or [0, 1] using the following formula:

[0066]

[0067]

[0068] Step S32, by encoding the value as the cosine of the angle and the timestamp as the radius, the rescaled time series in the polar coordinate system is expressed by formula

[0069]

[0070] Among them, t i It is the timestamp and N constant factor, which is used to adjust the span of the polar coordinate system. When converted to the polar coordinate system, the scaled data of the two normalization operations correspond to different angle ranges. The data in the range [-1, 1] corresponds to the arc cosine function angle range of [0, π], and the data in the range [0, 1] corresponds to the arc cosine value range of [0, π / 2];

[0071] Step S33, after converting the rescaled time series into a polar coordinate system, angular perspective is used to identify temporal correlations within different time intervals by considering the trigonometric sum / difference between each point. GAF is defined as follows:

[0072]

[0073] Where I is the unit row vector, and Representing different row vectors, they are converted into 2800 initial feature maps through GAF. The 400 samples of each fault type are divided into training set and test set at a ratio of 3:1.

[0074] Step S4: Construct a convolutional neural network (CNN) model. Specifically, it includes:

[0075] In step S41, the convolution layer and pooling layer are the main components of the convolutional neural network. The convolution layer is used to extract the hidden features of the image, and the convolution kernel is used to scan different locations of the feature image. The expression of the convolution operation is as follows:

[0076]

[0077] in, is the output feature map of layer l, f(·) is the activation function, M is the number of output feature maps, is the weight of the convolution kernel, is the bias of layer l.

[0078] Step S42: The pooling layer is usually connected to the convolution layer. Its main function is to reduce the dimension of the feature map output by the convolution layer and the computational complexity of the network parameters. Commonly used pooling layers include maximum pooling and average pooling, which replace the entire area information of a certain area of the image with the maximum value or average value. Therefore, the pooling operation can be understood as a blurring process of the image. The formula is described as:

[0079]

[0080] in, is the output value of the t-th neuron in the ith channel of the l-th layer, W is the size of the pooling kernel, and P il+1 (j) is the output value of the l+1th channel of the i-th layer.

[0081] Step S43: The activation function is mainly used to provide nonlinear modeling capabilities for the network model. The nonlinear activation function can improve the stability of neural network model training and accelerate model convergence. The present invention selects the ReLU activation function, which is expressed as follows:

[0082]

[0083] Step S44: The connection layer is usually connected to the last layer of the network structure and is used to map the fault features learned by the model to the sample space for classification. The fully connected layer converts the convolution or pooling output of the previous layer into a one-dimensional feature vector, which is then fully connected to the next layer. The fully connected layer expression is as follows:

[0084] y l =f{(w l ) T x l-1 +b l}

[0085] Among them, x l-1 is the output value of layer l-1, w l is the weight, y l is the output of layer l, b l is the bias term.

[0086] Step S45: The Softmax classifier is an activation function that ensures the non-negativity of the predicted probability. Normalization can use the features output by the fully connected layer as input and convert them into a probability distribution with a sum of 1. The fault category with the highest probability is the predicted result. The Softmax function formula is as follows:

[0087]

[0088] Where i = 1, 1, ..., k, K is the number of output nodes, P i represents the probability of each output,

[0089] Step S5: Based on the GAF image and CNN structure, a GAF-MMCNN fault diagnosis model is constructed. This model integrates a multi-scale branch module and a multi-scale cascade module to fully extract image features. It then constructs a convolutional fusion attention module (CFAM) to enhance image features and assign attention weights to each channel, thereby improving the classification accuracy and robustness of bearing fault diagnosis. Specifically, it includes:

[0090] In step S51, the attention mechanism aims to assign different attention weights to different parts of the model input, adjust the model's learning focus on key information, and thus improve the model's performance. SENet is the most common channel attention mechanism, but it uses a pooling method to compress input features into one dimension, which is prone to losing fault information. Therefore, the present invention proposes a convolutional fusion attention module (CFAM). Unlike SENet, CFAM uses global average pooling (GAP) and global maximum pooling (GMP) to compress features to prevent partial information loss. The CFAM attention structure is as follows: Figure 2 shown.

[0091] Step S52, CFAM includes feature compression and weight generation. In the feature compression stage, GAP and GMP are used to compress the input features along the height and width dimensions, and the outputs are PGAP and PGMP respectively. In the weight generation stage, the compressed features are flattened into a one-dimensional tensor and stacked along the channel dimension to facilitate the integration of one-dimensional convolution operations in the attention mechanism module and output double-pooled features. One-dimensional convolution is used to reduce the size and number of ascending channels, and then transposed convolution is selected to restore the size, and the number of channels becomes 1. Finally, the Sigmoid activation function is selected to generate weights and multiply them with the input features. The entire CFAM calculation process is as follows:

[0092]

[0093] P GMP =max x(i,j)

[0094]

[0095] Y T =CovT(Cov(Y P ))

[0096] z=reshape(Y T )

[0097] Among them, P GAP and P GMP are the one-dimensional data after GAP and GMP compression, i and j are the positions of the input value x in the two-dimensional data, flatten(·) represents the stacking along the channel dimension, Y P is the fused pooling information, Cov(·) smoothing and CovT(·) smoothing represent convolution and transposed convolution operations, Y T is the middle layer of weight information, reshape(·) represents the shape of 1×1×C, and z is the final generated channel attention weight.

[0098] In step S53, in the convolutional neural network, layers with different convolution kernel sizes will form different receptive fields. The multi-scale convolutional neural network fuses the results of a series of convolution operations with various convolution kernel sizes. There are two main types of traditional multi-scale feature fusion networks: one is the multi-scale feature fusion network, and the other is the multi-scale cascade fusion network. The first is the multi-scale branch structure, such as Figure 3 As shown, AP represents average pooling. The characteristic of the multi-scale structure is that each branch in the same layer of the network has a different convolution kernel and uses 1×1 convolution to reduce the number of channels to achieve simple operations. The second is the multi-scale cascade structure, such as Figure 4 As shown in Figure 1. There is only one backbone, and the entire network chain gradually deepens from the shallow layer, while the size of the feature map decreases, and the image obtained by upsampling the small-scale features is fused with the large-scale feature map.

[0099] Step S54, the multi-scale branch structure model is relatively small, and high-precision fault classification cannot be achieved when a small number of modules are stacked. For this reason, the present invention proposes an MMCNN that integrates a multi-scale branch structure and a multi-scale cascade structure to avoid model overfitting and long training time caused by a complex network structure. At the same time, a BN layer is added to speed up the convergence of the model to prevent gradient disappearance, and a CFAM attention mechanism is introduced to improve the model's attention to different channel features. The GAF-MMCNN structure consists of four parts: (1) feature screening network; (2) deep feature extraction network; (3) auxiliary classifier; (4) main classifier.

[0100] In step S6, the image is input into the MMCNN fault diagnosis model to output the classification results. To find the optimal model parameters, the training batch, learning rate, auxiliary classification weights, and initial convolution kernel size of the network model are adjusted. The training results are observed and the parameters with the highest prediction accuracy are selected. The accuracy (ACC) is used to evaluate the experimental results. The formula is as follows:

[0101]

[0102] Among them, TP represents the frequency of predicting positive samples as positive, TN represents the frequency of predicting negative samples as negative, FP represents the frequency of predicting negative samples as positive, and FN represents the frequency of predicting positive samples as negative.

Claims

1. A rolling bearing fault diagnosis method based on GAF-MMCNN, characterized in that: The following steps are involved: Step S1, acquiring a vibration signal to obtain a bearing data set; Step S2: data preprocessing, sampling one-dimensional raw data of various bearing fault types to obtain data samples; In step S3, the GAF encoding method is used to obtain two-dimensional images of various faults. The one-dimensional vibration signal of the bearing is converted into a two-dimensional feature image and input into the neural network for feature extraction. First, the data samples are normalized. Second, the polar coordinate system is used instead of the typical Cartesian coordinate system to represent the time series. Finally, the trigonometric sum / difference between each point is considered, and the angular perspective is used to identify the time correlation within different time intervals, thus preserving the time correlation of the original vibration signal. Step S4, constructing a convolutional neural network (CNN) model; Step S5: Based on the GAF image and CNN structure, a GAF-MMCNN fault diagnosis model is constructed. The model integrates the multi-scale branch module and the multi-scale cascade module, constructs a convolutional fusion attention module (CFAM), enhances image features, and assigns attention weights to each channel. Step S6: input the image into the MMCNN fault diagnosis model to output the classification result.

2. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 1, characterized in that: The step S3 specifically includes: Step S31, time series X={x1,x2,...,x n } contains n observations, rescale X so that all values fall within the interval [-1, 1] or [0, 1] using the following formula: Step S32, by encoding the value as the cosine of the angle and the timestamp as the radius, the rescaled time series in the polar coordinate system is expressed by formula Among them, t i is the timestamp, N is a constant factor; In step S33, after converting the rescaled time series into a polar coordinate system, angular perspective is used to identify temporal correlations within different time intervals by considering the trigonometric sums / differences between each point. GAF is defined as follows: Where I is the unit row vector, and Represents different row vectors.

3. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 2 is characterized in that: In step S32, when converted to the polar coordinate system, the scaled data of the two normalization operations correspond to different angle ranges respectively. The data in the range [-1, 1] corresponds to the arc cosine function angle range of [0, π], and the data in the range [0, 1] corresponds to the arc cosine value range of [0, π / 2].

4. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 2, characterized in that: In step S33, samples of each fault type are divided into a training set and a test set at a ratio of 3:

1.

5. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 1, characterized in that: The step S4 specifically includes: In step S41, the expression of the convolution operation is as follows: in, is the output feature map of layer l, f(·) is the activation function, M is the number of output feature maps, is the weight of the convolution kernel, is the bias of layer l; Step S42: Pooling operation, the formula is described as: in, is the output value of the t-th neuron in the ith channel of the l-th layer, W is the size of the pooling kernel, and is the output value of the l+1th channel of the i-th layer; Step S43: Select the ReLU activation function, the expression is as follows: Step S44: The fully connected layer expression is as follows: y l =f{(w l ) T x l-1 +b l } Among them, x l-1 is the output value of layer l-1, w l is the weight, y l is the output of layer l, b l is the bias term; Step S45: The Softmax function formula is as follows: Where i = 1, 1, ..., k, K is the number of output nodes, P i represents the probability of each output, 6. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 1, characterized in that: The step S5 specifically includes: Step S51, constructing a convolutional fusion attention module CFAM; Step S52: CFAM includes feature compression and weight generation. In the feature compression stage, GAP and GMP are used to compress the input features along the height and width dimensions, and the outputs are PGAP and PGMP respectively. In the weight generation stage, the compressed features are flattened into a one-dimensional tensor and stacked along the channel dimension to output the double-pooled features. The entire CFAM calculation process is as follows: P GMP =max x(i,j) Yes T =CovT(Cov(Y P )) z=reshape(Y T ) Among them, P GAP and P GMP are the one-dimensional data after GAP and GMP compression, i and j are the positions of the input value x in the two-dimensional data, flatten(·) represents the stacking along the channel dimension, Y P is the fused pooling information, Cov(·) smoothing and CovT(·) smoothing represent convolution and transposed convolution operations, Y T is the middle layer of weight information, reshape(·) represents the shape of 1×1×C, and z is the final generated channel attention weight; Step S53: In order to construct MMCNN, the multi-scale branch structure and the multi-scale cascade structure are integrated. Meanwhile, the BN layer and the CFAM attention mechanism are added.

7. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 6, characterized in that: The CFAM utilizes global average pooling (GAP) and global maximum pooling (GMP) to compress features.

8. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 6, characterized in that: The CFAM uses one-dimensional convolution to reduce the size and number of ascending channels, selects transposed convolution to restore the size, and the number of channels becomes 1; finally, the Sigmoid activation function is selected to generate weights and multiply the weights with the input features.

9. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 6, characterized in that: The GAF-MMCNN structure includes: feature screening network; deep feature extraction network; auxiliary classifier; main classifier.

10. The rolling bearing fault diagnosis method based on GAF-MMCNN according to claim 1, characterized in that: The step S6 uses the accuracy ACC to evaluate the experimental results, and the formula is as follows: Among them, TP represents the frequency of predicting positive samples as positive, TN represents the frequency of predicting negative samples as negative, FP represents the frequency of predicting negative samples as positive, and FN represents the frequency of predicting positive samples as negative.

Images