Belt conveying equipment bearing fault diagnosis method based on multiple modes

Abstract

The invention discloses a multi-modal-based belt conveying equipment bearing fault diagnosis method, which comprises the following steps of: preprocessing a multi-modal data set based on continuous wavelet transform, extracting time-frequency characteristics based on RepLKNet, extracting time-sequence characteristics of a BiGRU network based on GlobalAttention optimization, and fusing the multi-modal characteristics of the time-frequency characteristics and the time-sequence characteristics. On one hand, local features in a larger range can be obtained, more and higher-level fault feature information can be captured, on the other hand, feature splicing fusion is carried out on time-frequency image features extracted by RepLKNet and vibration one-dimensional signal time sequence features processed by BiGRU-GlobalAttention, efficient and accurate classification performance can be kept, robustness and generalization of the model are improved, and the method is suitable for large-scale popularization and application. Therefore, the fault type of the bearing of the belt type conveying equipment can be diagnosed more accurately, and the method is of great significance to the application of online fault diagnosis of the rolling bearing of the large-scale belt type conveying equipment.

Description

Technical Field

[0001] The invention relates to the technical field of mechanical fault diagnosis, and in particular to a belt conveyor equipment bearing fault diagnosis method based on multi-mode. Background Art

[0002] Rolling bearings play a key role in industrial production and are widely used in belt conveyor equipment such as mines and coal mines. However, under harsh conditions, rolling bearings are prone to failure, and failures may cause damage to the entire belt conveyor equipment. Therefore, accurate diagnosis of rolling bearing failures is crucial to ensure the safe, stable and efficient operation of belt conveyor equipment.

[0003] Most of the existing fault feature extraction is based on vibration measurement signals, based on one of the three modal signals in the time domain, frequency domain and wavelet domain. When using frequency domain analysis, the energy of the bearing will be transferred to the mid-frequency and high-frequency bands; when using time domain analysis, although the time domain signal is more sensitive to the defects of the rolling bearing, its perception ability in terms of amplitude and frequency is limited; using wavelet domain analysis, multi-resolution and multi-scale features of the rolling bearing can be extracted, but it is too sensitive to micro variables, and it is easy to make misjudgments and missed judgments in the case of large interference, and sometimes the conclusion lacks uniqueness, and it is difficult to select an appropriate wavelet basis.

[0004] It can be seen from the above that signal analysis that relies only on a single domain is often unable to fully extract the information of other modal signals. This is inadequate when dealing with complex fault diagnosis, especially rolling bearing faults under variable operating conditions, and may lead to a decrease in the timeliness and accuracy of diagnosis. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects in the prior art and provide a belt conveyor bearing fault diagnosis method based on multi-mode.

[0006] To achieve the above object, a multi-modal belt conveyor bearing fault diagnosis method of the present invention utilizes a neural network learning model, and the fault diagnosis method comprises: S1. Preprocessing of multimodal data sets: converting the one-dimensional time series signal of vibration into a two-dimensional time-frequency signal (also called time-frequency image) based on continuous wavelet transform (CWT), while removing outliers or performing filtering, smoothing and other processing; S2. Extract time-frequency image features based on RepLKNet (large kernel re-parameterized convolutional neural network); Introducing RepLKNet into time-frequency images can obtain local features in a larger range and capture higher-level information of the signal; S3. Optimizing a BiGRU (Bidirectional Gated Recurrent Unit) network based on GlobalAttention (global attention mechanism) and extracting temporal features of the multimodal data. The BiGRU network optimized based on GlobalAttention is recorded as BiGRU-GlobalAttention. S4, multimodal feature fusion, the time-frequency image features extracted by RepLKNet and the one-dimensional vibration signal time series features processed by BiGRU-GlobalAttention are spliced ​​and fused, which can more comprehensively reflect the fault characteristics of the signal, maintain efficient and accurate classification performance, and improve the robustness and generalization ability of the model in practical applications; the splicing and fusion includes processing the output of RepLKNet with the flatten() function, and then processing it with the output of BiGRU-GlobalAttention with the concat() function, and then entering the fully connected layer to obtain the probability of occurrence of different types of faults (i.e., the weight coefficient diagnoses the fault type). Among them, flatten() and concat() are PyTorch library functions. PyTorch is the python version of torch, which is Facebook's open source neural network framework and is a prior art.

[0007] There is no order of precedence for steps S2 and S3.

[0008] Furthermore, the multimodal data includes vibration measurement data of the belt conveyor equipment bearing.

[0009] Furthermore, the wavelet basis of the continuous wavelet transform includes:

[0010] in, is the angular frequency, t is the time variable; the length of the wavelet basis scale sequence is less than or equal to the length of the signal.

[0011] Furthermore, the data set preprocessing also includes: transforming the vibration one-dimensional time series signal through continuous wavelet transform to generate detailed information of the signal at different times and scales, and visualize the frequency and time characteristics of the signal.

[0012] Furthermore, the RepLKNet network uses a large 31x31 convolution kernel instead of the traditional 3x3 convolution kernel, which innovatively expands the receptive field of the model and improves the feature capture capability.

[0013] Furthermore, the GlobalAttention-based BiGRU optimization includes calculating a weight vector at each time step and using it as a feature representation of the BiGRU output. The weight vector represents the degree of attention the model pays to each part of the input sequence. By weighting the features at all positions, the model can focus on important time domain features in a more targeted manner.

[0014] Furthermore, the different types of faults include bearing inner ring faults, outer ring faults, rolling element damage and other faults in different parts, which are classified here according to the fault parts.

[0015] Furthermore, the different types of faults also include different fault sizes and / or combinations of different load conditions and fault locations.

[0016] Furthermore, the fault diagnosis method also includes an iterative process: In each iteration, the input data is forward propagated through steps S2, S3, S4 and the fully connected layer to generate a predicted category probability distribution; the loss value between the predicted result and the true label is calculated using the cross entropy loss function; then the backpropagation algorithm is used to calculate the gradient layer by layer starting from the output layer; finally, based on the Adam optimization algorithm, the network weights are updated according to the calculated gradient, and the cycle is repeated back to step S2; The iterative process terminates when the maximum change in the loss function value in K consecutive iterations is less than or equal to the loss function threshold or reaches the maximum number of iterations, where K is the threshold number of consecutive iterations.

[0017] The advantages and beneficial effects of the present invention are as follows: In order to solve the problems of low accuracy in fault diagnosis of rolling bearings of belt conveyor equipment under complex working conditions and great difficulty in online fault diagnosis application, the present invention adopts multi-domain feature fusion technology, and the fault identification accuracy can reach 99.81%. It can accurately judge whether the bearing has faults and the type of fault in the actual working environment, providing an important reference for the repair and maintenance of mechanical equipment, and is of great significance for ensuring production safety, improving product quality and preventing accidents, especially the application of online fault diagnosis of rolling bearings of large belt conveyor equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a principle block diagram of a belt conveyor bearing fault diagnosis method based on multi-mode in the present invention; Figure 2 It is an experimental trend diagram of the fault identification accuracy of the method of the present invention as the number of iterations increases. DETAILED DESCRIPTION

[0019] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0020] Embodiment 1: The present invention provides a belt conveyor bearing fault diagnosis method based on multi-mode, such as Figure 1 As shown, using a neural network learning model, the fault diagnosis method includes: S1. Preprocessing of multimodal data sets: converting the one-dimensional time series signal of vibration into a two-dimensional time-frequency signal (also called time-frequency image) based on continuous wavelet transform (CWT), while removing outliers or performing filtering, smoothing and other processing; S2. Extract time-frequency image features based on RepLKNet (large kernel re-parameterized convolutional neural network); Since wavelet transform is originally conducive to extracting different frequency domain information, introducing RepLKNet into time-frequency images can obtain local features in a larger range and capture higher-level information of the signal; S3. Optimizing a BiGRU (Bidirectional Gated Recurrent Unit) network based on GlobalAttention (global attention mechanism) and extracting temporal features of the multimodal data. The BiGRU network optimized based on GlobalAttention is recorded as BiGRU-GlobalAttention. The BiGRU structure output consists of a forward network structure and a reverse network structure. The two GRU outputs in different time directions are integrated into the final output in the form of matrix splicing. The forward output and reverse output are directly added as the model hidden state output. ,in, For positive output, For reverse output.

[0021] Since the vibration signals before and after the motor bearing fault and at time t contain rich feature information, the BiGRU rolling bearing fault recognition network is obtained by improving the GRU neural network model. The BiGRU rolling bearing fault recognition network adopts a bidirectional gated recurrent unit to extract historical information and future information at the same time, which can achieve better feature extraction effect. Although BiGRU can capture the bidirectional dependencies in the signal by modeling the time series signal from both the forward and reverse directions, it pays equal attention to different parts of the input signal, that is, there is no mechanism to highlight those parts that are more critical to the prediction results.

[0022] This paper introduces the GlobalAttention mechanism, which is a mechanism for strengthening the model's attention to different parts of the input sequence. It calculates a weight vector at each time step to represent the model's attention to each part of the input sequence. These weights are then applied to the feature representation of the BiGRU output, and by weighting the features at all positions, the model can focus more specifically on important time domain features.

[0023] The main codes for initialization and forward propagation implemented by the software in this embodiment are as follows: ① Initialization phase code # RepLKNet31Small feature extraction self.RepLKNet31Small = create_RepLKNet31Small(num_classes=output_dim) # Adaptive average pooling self.adaptive_pool = nn.AdaptiveAvgPool2d(1) # BiGRU parameters self.num_layers = len(hidden_layer_sizes) # Number of bigru layers self.bigru_layers = nn.ModuleList() # List used to save BiGRU layers # Define the first layer of BiGRU self.bigru_layers.append(nn.GRU(time_input_dim, hidden_layer_sizes[0], batch_first=True, bidirectional=True)) # Define subsequent BiGRU layers for i in range(1, self.num_layers): self.bigru_layers.append( nn.GRU(hidden_layer_sizes[i - 1] * 2, hidden_layer_sizes[i], batch_first=True, bidirectional=True)) # Define global attention layer # The default dimension of the attention layer is the BiGRU output layer dimension self.globalAttention = GlobalAttention(hidden_layer_sizes[-1]) # Define fully connected layer self.classifier = nn.Linear(256 + hidden_layer_sizes[-1] * 2, output_dim) ②Forward propagation code # Time domain features are fed into BiGRU # Data preprocessing # Note: Here the data is stacked. A 1*1024 matrix is ​​divided and stacked into a shape of 32*32, which reduces the length of the input sequence. bigru_out = x_seq.view(batch_size, 16, 64) # torch.Size([32, 32, 32]) # Send to BiGRU layer # Change the input shape to adapt to the network input [batch, seq_length, H_in] # hidden Get hidden layer data hidden = [] for bigru in self.bigru_layers: bigru_out, hidden = bigru(bigru_out) ## Perform a forward propagation of the BiGRU layer (b,l,w) # Send to the global attention layer gatt_features = self.globalAttention(hidden[-1], bigru_out) #torch.Size([32, 128]) S4, multimodal feature fusion, the time-frequency image features extracted by RepLKNet and the one-dimensional vibration signal time series features processed by BiGRU-GlobalAttention are spliced ​​and fused, which can more comprehensively reflect the fault characteristics of the signal, maintain efficient and accurate classification performance, and improve the robustness and generalization ability of the model in practical applications; the splicing and fusion includes processing the output of RepLKNet with the flatten() function, and then processing it with the output of BiGRU-GlobalAttention with the concat() function, and then entering the fully connected layer to obtain the probability of occurrence of different types of faults (i.e., the weight coefficient diagnoses the fault type). Among them, flatten() and concat() are PyTorch library functions. PyTorch is the python version of torch, which is Facebook's open source neural network framework and is a prior art.

[0024] There is no order of precedence for steps S2 and S3.

[0025] Currently, bearing faults are mainly divided into inner ring faults, outer ring faults, and rolling element damage according to the fault locations. The time-frequency characteristics and time series characteristics of the vibration data generated by different faults are different. At the same time, the fault data (or fault phenomena) under different fault sizes or different load conditions are also different. Different corresponding characteristic values ​​can be obtained by analyzing the typical fault data of different fault types and their combinations.

[0026] The difficulty in processing measured data lies in how to accurately extract these characteristic values, especially the extraction and identification of fault characteristic values ​​of different fault locations, different fault sizes, and different damage types under different load conditions.

[0027] Based on the traditional neural network learning model, the present invention creatively uses convolutional neural network to extract features from the time-frequency signal of the original vibration signal after wavelet transform. At the same time, BiGRU-GlobalAttention is used to extract time series features from the original time series data, and the time-frequency features are spliced ​​and fused with the time series features to obtain more accurate bearing fault diagnosis information.

[0028] Preferably, the multimodal data includes vibration measurement data of a bearing of the belt conveyor equipment.

[0029] Preferably, the wavelet basis of the continuous wavelet transform comprises:

[0030] in, is the angular frequency, t is the time variable; the length of the wavelet basis scale sequence is less than or equal to the length of the signal, the longer the scale sequence is, the higher the scale resolution (frequency resolution), and the more subtle frequency changes can be detected. However, too long a scale sequence length will also increase the computational complexity, so this embodiment selects a scale of 128, a signal length of 512, and a wavelet function cmor100-1 (a complex Morlet wavelet with high bandwidth and low center frequency, suitable for transient signal analysis requiring high time resolution. The bandwidth is 100 and the center frequency is 1).

[0031] Preferably, the data set preprocessing further includes: transforming the vibration one-dimensional time series signal through continuous wavelet transform to generate detailed information of the signal at different times and scales, and visualize the frequency and time characteristics of the signal.

[0032] Preferably, the RepLKNet network described in this embodiment uses a large 31x31 convolution kernel instead of a traditional 3x3 convolution kernel, innovatively expanding the receptive field of the model and improving the feature capture capability.

[0033] Preferably, the GlobalAttention-based BiGRU optimization includes calculating a weight vector at each time step (specifically implemented as the forward propagation code described above) and using it as the feature representation of the BiGRU output. The weight vector represents the degree of attention paid by the model to each part of the input sequence. By weighting the features at all positions, the model can focus on important time domain features in a more targeted manner.

[0034] Preferably, the different types of faults include bearing inner ring fault, outer ring fault or rolling element damage and other faults in different parts, which are classified here according to the fault location.

[0035] Preferably, the different types of faults further include different fault sizes and / or combinations of different load conditions and fault locations.

[0036] Preferably, the fault diagnosis method further comprises an iterative process: In each iteration, the input data is forward propagated through steps S2, S3, S4 and the fully connected layer to generate a predicted category probability distribution; the loss value between the predicted result and the true label is calculated using the cross entropy loss function; then the backpropagation algorithm is used to calculate the gradient layer by layer starting from the output layer; finally, based on the Adam optimization algorithm, the network weights are updated according to the calculated gradient, and the cycle is repeated back to step S2; The iterative process terminates when the maximum change in the loss function value in K consecutive iterations is less than or equal to the loss function threshold or reaches the maximum number of iterations, where K is the threshold number of consecutive iterations.

[0037] In this embodiment, the continuous iteration threshold K is set to 50, the loss function threshold is set to 0.001, the maximum number of iterations is set to 5000, and the cross entropy loss function is set to

[0038] Where Y is the true label vector, P is the positive category probability vector predicted by the model; for the i-th category, the total number of categories is N, in this embodiment N=10, y i is the true label of the sample. If the sample belongs to this category, the value is 1, otherwise it is 0. i is the probability that the model predicts that the sample belongs to the positive category.

[0039] The experiment in this embodiment uses the CWRU bearing dataset of the University of Western Reserve University, which is publicly available in academic research. The dataset collects data from acceleration sensors installed near the motor bearings and simulates a variety of bearing fault conditions. The dataset records the normal operating state and different types of bearing fault conditions, including bearing inner ring, outer ring and rolling element damage faults. Each fault condition also distinguishes different fault sizes and different load conditions.

[0040] This embodiment judges the advantages and disadvantages of the method of the present invention by processing and analyzing the known data set, and establishes several sample data sets, each with a length of 512 points and an overlap rate of adjacent sample data points of 0.5; for the case of single load, three fault locations (inner ring fault, outer ring fault, rolling element damage), and three fault sizes (0.007 inches, 0.014 inches, and 0.021 inches), S2 outputs a vector of size [32, 256], and S3 outputs a vector of size [32, 128]. After fusion, it becomes a vector of size [32, 256+128], which belongs to 10 classification tasks (plus the normal state, that is, 3*3+1=10). After passing through the fully connected layer of the neural network model, a vector of size [32, 10] is output, thereby determining the probability that the input sample belongs to each category of fault type.

[0041] The experimental results are as follows Figure 2 As shown in the figure, after 50 iterations, the fault identification accuracy rate reached 99.81%. This is significantly better than the fault identification accuracy rate of about 95% in the traditional method for belt conveyor bearing fault diagnosis.

[0042] Embodiment 2: The difference from Example 1 is that this embodiment targets 2 types of loads (1 horsepower, 2 horsepower), 3 types of fault locations (inner ring fault, outer ring fault, rolling element damage), and 3 types of fault sizes (0.007 inches, 0.014 inches, and 0.021 inches), which are 19 classification tasks (plus the normal state, i.e., N=2*3*3+1=19). After passing through the fully connected layer of the neural network model, a vector of size [32, 19] is output, thereby determining the probability that the input sample belongs to each type of fault.

[0043] The basic principle of the present invention is: the present invention provides a bearing fault diagnosis method for belt conveyor equipment based on multimodality. First, the time-frequency of the vibration one-dimensional time series signal is visualized based on the continuous wavelet transform CWT; then, the time-frequency image is feature extracted based on RepLKNet to capture higher-level information of the signal; then, BiGRU is used to extract time series features, and the GlobalAttention mechanism is used to dynamically assign importance weights of different time steps, so that the model pays more attention to critical moment information; finally, the time-frequency image features extracted by RepLKNet and the one-dimensional vibration signal features processed by BiGRU-GlobalAttention are feature spliced ​​and fused to improve the generalization of the model. The present invention is of great significance to the research on bearing fault diagnosis of belt conveyor equipment and the application of online fault diagnosis.

[0044] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the technical principles of the present invention, several improvements and modifications can be made, such as the selection of wavelet basis, the selection of RepLKNet network convolution kernel, different fault classification or combination methods, iterative condition setting, etc. These improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A multi-modal belt conveyor bearing fault diagnosis method, characterized in that: Using a neural network learning model, the fault diagnosis method includes: S1, multimodal data set preprocessing, based on continuous wavelet transform, the vibration one-dimensional time series signal is converted into a two-dimensional time-frequency signal; S2, extracting time-frequency image features based on RepLKNet; S3. Optimizing the BiGRU network based on GlobalAttention and extracting the temporal features of the multimodal data. The BiGRU network optimized based on GlobalAttention is recorded as BiGRU-GlobalAttention. S4, multimodal feature fusion, feature splicing and fusing the time-frequency image features extracted by RepLKNet and the time series features processed by BiGRU-GlobalAttention; the splicing and fusion includes processing the output of RepLKNet by the flatten() function, and then processing it together with the output of BiGRU-GlobalAttention by the concat() function, and then entering the fully connected layer to obtain the probability of occurrence of different types of faults; There is no order of precedence for steps S2 and S3.

2. A belt conveyor bearing fault diagnosis method based on multi-mode according to claim 1, characterized in that: The multimodal data includes vibration measurement data of bearings of the belt conveyor equipment.

3. A belt conveyor bearing fault diagnosis method based on multi-mode according to claim 1, characterized in that: The wavelet basis of the continuous wavelet transform comprises: ， The length of the wavelet basis scale sequence is less than or equal to the length of the signal.

4. A belt conveyor bearing fault diagnosis method based on multi-mode according to claim 1, characterized in that: The data set preprocessing also includes: transforming the vibration one-dimensional time series signal through continuous wavelet transform to generate detailed information of the signal at different times and scales, and visualize the frequency and time characteristics of the signal.

5. The multi-modal belt conveyor bearing fault diagnosis method according to claim 1, characterized in that: The RepLKNet network uses a large convolution kernel of 31x31.

6. The multi-modal belt conveyor bearing fault diagnosis method according to claim 1, characterized in that: The GlobalAttention-based BiGRU optimization includes calculating a weight vector at each time step and using it as a feature representation of the BiGRU output, wherein the weight vector represents the degree of attention paid by the model to each part of the input sequence.

7. The multi-modal belt conveyor bearing fault diagnosis method according to claim 1, characterized in that: The different types of faults include bearing inner ring fault, outer ring fault or rolling element damage.

8. A belt conveyor bearing fault diagnosis method based on multi-mode according to claim 7, characterized in that: The different types of faults also include different fault sizes and / or combinations of different load conditions and fault locations.

9. The multi-modal belt conveyor bearing fault diagnosis method according to claim 1, characterized in that: The fault diagnosis method also includes an iterative process: In each iteration, the input data is forward propagated through steps S2, S3, S4 and the fully connected layer to generate a predicted category probability distribution; the loss value between the predicted result and the true label is calculated using the cross entropy loss function; then the backpropagation algorithm is used to calculate the gradient layer by layer starting from the output layer; finally, based on the Adam optimization algorithm, the network weights are updated according to the calculated gradient, and the cycle is repeated back to step S2; The iterative process terminates when the maximum change in the loss function value in K consecutive iterations is less than or equal to the loss function threshold or reaches the maximum number of iterations, where K is the threshold number of consecutive iterations.

Images