Intelligent fault diagnosis method for rolling bearing-rotor system based on VMTF-RBCBAM-Net

Abstract

A rolling bearing-rotor system fault intelligent diagnosis method based on VMTF-RBCBAM-Net comprises the steps that vibration acceleration signals of a rolling bearing-rotor system in different states are integrated into speed signals, the speed signals are coded into two-dimensional images through a Markov transition field, then the generated two-dimensional images are marked, and the marked two-dimensional images are stored in a database; dividing into a training set, a verification set and a test set in proportion; then, an RBCBAM-Net model fusing a residual error structure and a channel-space attention mechanism is constructed, and the model is trained by using training set data; and finally, evaluating the model performance through test set data, and performing visual analysis on the classification effect of the RBCBAM-Net model by adopting a confusion matrix and a t-SNE method. According to the method, the multi-scale key features can be adaptively extracted from the VMTF image of the equipment vibration speed signal, the gradient disappearance phenomenon in the training process is effectively avoided, the method has high feature extraction capacity and generalization performance, and a good effect is shown in intelligent recognition of rolling bearing-rotor system faults.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical equipment fault diagnosis, and particularly relates to an intelligent fault diagnosis method for a rolling bearing-rotor system based on VMTF-RBCBAM-Net (Velocity Markov Transition Field-Residual Block Convolutional Block Attention Module-Net). Background Technique

[0002] Industrial auxiliaries such as pumps, motors, and fans are widely used in national economic pillar industries such as petroleum, chemical industry, metallurgy, power generation, and energy. Their main structure belongs to a typical rolling bearing-rotor system, which has the characteristics of compact structure and dense deployment. Due to the often harsh operating environment of the equipment and the long-term action of alternating loads and impact loads, mechanical faults are extremely likely to occur during the operation of the equipment, seriously affecting the safe and stable operation of the equipment. Due to the large number of auxiliary equipment, the traditional diagnosis method based on manual analysis has problems such as diagnosis lag and low efficiency, and the diagnosis accuracy highly depends on the experience and level of diagnostic engineers, which cannot meet the needs of modern industrial enterprises for equipment operation and maintenance. Therefore, carrying out research on intelligent fault diagnosis for rolling bearing-rotor systems is of great significance for ensuring the safe operation of equipment and improving the production efficiency of enterprises.

[0003] In recent years, with the rapid development of deep learning technology, many researchers have combined feature extraction with deep learning theory and applied it to the field of mechanical equipment fault diagnosis. The specific method (Jiang Jiaguo, Guo Manli. Rolling bearing fault diagnosis method based on MTF and DenseNet [J]. Industry and Mine Automation, 2022, 48(09): 63-8.) is as follows: encode the collected vibration acceleration signal into an MTF image, use it as the input of DenseNet, extract the fault features of the rolling bearing vibration signal through DenseNet, and then realize the fault identification of the rolling bearing. In addition, some scholars perform short-time Fourier transform on the collected vibration signal to generate a time-frequency diagram, and then send it into a CNN for motor fault diagnosis and prediction. (Ribeiro Junior RF, dos Santos Areias IA, Campos MM, et al. Fault detection and diagnosis in electric motors using convolution neural network and short-time fourier transform [J]. Journal of Vibration Engineering & Technologies, 2022, 10(7): 2531-2542). However, most of the existing methods directly encode the acceleration signal into a Markov transfer field (MTF) image for processing. In the rolling bearing-rotor system, the high-frequency vibration interference in the acceleration signal often masks the characteristics of low-frequency faults, making it difficult to effectively identify its typical fault characteristics. And due to the characteristics of fixed hierarchical receptive fields and single convolution kernel sizes in fixed-structure deep networks (such as standard DenseNet / CNN, etc.), there are limitations in the adaptive extraction of multi-scale fault features, resulting in the difficulty of achieving an ideal level of model accuracy and generalization ability when dealing with different degrees of rolling bearing-rotor system faults.

[0004] Therefore, for different types and degrees of rolling bearing-rotor system faults, how to effectively extract the vibration characteristics of typical faults and combine them with an efficient deep learning model is the key to improving the accuracy of intelligent fault diagnosis. Summary of the Invention

[0005] To overcome the shortcomings of the above-mentioned existing technologies, the purpose of the present invention is to provide an intelligent fault diagnosis method for a rolling bearing-rotor system based on VMTF-RBCBAM-Net. By using the RBCBAM-Net model with a residual structure and a channel-spatial attention mechanism, it can adaptively extract key features at multiple scales from the VMTF images of the device vibration velocity signals, possess good feature extraction capabilities and generalization performance, and effectively achieve the accurate identification of typical faults in the rolling bearing-rotor system.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] An intelligent fault diagnosis method for a rolling bearing-rotor system based on VMTF-RBCBAM-Net includes the following steps:

[0008] 1) Use an acceleration sensor to collect the vibration acceleration signal S ij (n) of typical faults in the rolling bearing-rotor system. The typical faults include: imbalance, misalignment, looseness, and dynamic-static rubbing faults, where: i represents the fault type, j represents the jth acceleration signal sample of the ith type of fault, and n is the index position of the sampling point of the vibration acceleration signal sample, n = 1, 2,... N, and N represents the sample length;

[0009] 2) Adopt the integral formula: Integrate the vibration acceleration signal S ij (n) into the velocity signal V ij (n), where: Δt represents the sampling time interval;

[0010] 3) Use the formula: V' ij (n) = (V ij (n) - min[V ij (n)]) / (max[V ij (n)] - min[V ij (n)]) to normalize the velocity signal V ij (n) to obtain V' ij (n);

[0011] 4) Use the Markov transform to encode V' ij (n) into a VMTF image, and then divide it into a training set, a validation set, and a test set according to a ratio. Among them: the training set and the validation set contain VMTF images of the same degree of faults, and the test set contains VMTF images of different degrees of faults;

[0012] 5) Construct an RBCBAM-Net network model. The RBCBAM-Net network model consists of a convolutional layer, a pooling layer, an RBCBAM module, a global average pooling layer, and a classification module;

[0013] 6) Input the training set data prepared in step 4) into the RBCBAM-Net network model constructed in step 5) for training, and save the trained RBCBAM-Net network model;

[0014] 7) Import the VMTF images of the test set in step 4) into the trained RBCBAM-Net network model in step 6) for fault diagnosis, and use the confusion matrix and t-SNE method to visually display the classification effect.

[0015] In step 4), the specific operation of encoding V' ij (n) into VMTF images is as follows:

[0016] First, divide V' ij (n) into Q partitions, and map it to the corresponding state q, q = 1, 2,..., Q according to the value range of each sampling point V' ij (n);

[0017] Then, construct a Q×Q Markov transition matrix to represent the probability of transitioning from one state to another state. The specific formula is:

[0018]

[0019] where: w pq represents the probability of transitioning from partition p to partition q, N pq represents the number of times of transitioning from state p to state q, and N p represents the total number of times state p appears;

[0020] Next, construct a T×T transition field matrix M to represent the state transition probability within different time windows. The specific formula is:

[0021]

[0022] where: m pq represents the transition probability from time window t p to t q , is the number of times of transitioning from state p to state q within time window t p , represents the total number of times state p appears within time window t p .

[0023] In step 5), the convolutional layer and pooling layer of the RBCBAM-Net network model first perform convolutional and pooling operations on the VMTF images to obtain the feature vector x;

[0024] The RBCBAM-Net network model contains multiple RBCBAM modules. Each RBCBAM module includes two convolutional layers, two BN layers, the CBAM attention mechanism (channel attention module CAM and spatial attention module SAM), and a residual connection. After convolutional operations and feature fusion, Y is output. out , and the specific steps are as follows:

[0025] The first step: The feature vector x undergoes convolutional operations and batch normalization processing to generate the feature map Y 1 and Y 2 ;

[0026] Y 1 = BN(Conv m×m (x)) (3)

[0027] Y 2 = ReLU(BN(Conv m×m (Y 1 ))) (4)

[0028] In the formula: Conv m×m is the convolutional operation, where m = 3, 4... 11; BN is the batch normalization operation, and ReLU is the activation function;

[0029] The second step: The channel attention module CAM first performs average pooling and max pooling operations on Y 2 to obtain and Then, two channel attention vectors are obtained through MLP; and then normalized by the Sigmoid function to obtain the channel attention weight M C (Y 2 ); Finally, it is applied to Y 2 to obtain the feature map Y C , and the specific formula is as follows:

[0030]

[0031] Y C = Y 2 × M C (Y 2 ) (6)

[0032] In the formula: σ is the Sigmoid activation function, MLP represents the multi-layer perceptron for processing the pooling results, AvgPool(Y 2 ) represents the average pooling operation on Y 2 , and MaxPool(Y 2 ) represents the max pooling operation on Y 2 ;

[0033] Step 3: The spatial attention module SAM first operates on Y c to perform average pooling and max pooling in the channel dimension to obtain two spatial attention feature maps; then they are concatenated and a convolution operation is performed on them to obtain the spatial attention weight M S (Y c ); finally, M S (Y C ) is used for Y C , and the feature map Y S after spatial attention adjustment is obtained;

[0034] M S (Y C ) = σ(Conv t×t (Concat[AvgPool spatial (Y C ),MaxPool spatial (Y C ))) (7)

[0035] Y S = Y C ×M S (Y C ) (8)

[0036] where: Conv t×t is the convolution operation, t = 3, 4…11; Concat represents the operation of concatenating feature maps in the channel dimension, combining different pooling results, AvgPool spatial (Y C ) and maxPool spatial (Y C ) respectively represent global average pooling and global max pooling of the feature map Y C in the channel dimension;

[0037] Step 4: After the feature map Y S is processed by the ReLU function, it is added to the residual Shortcut(x) to obtain the output Y out , and the specific formula is:

[0038] Y out = ReLU(Y S ) + Shortcut(x) (9)

[0039] where: Shortcut(x) = x.

[0040] In step 5), the classification module consists of a fully connected layer, an activation function, a Dropout layer, and a Softmax function, and is responsible for mapping the extracted high-dimensional feature vectors to specific fault categories.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] The present invention fully combines the vibration characteristics of typical faults of equipment. First, the velocity signal obtained by integrating the vibration acceleration signal is used, and the velocity signal is encoded into a VMTF image by using a Markov transition field, effectively suppressing high-frequency signal interference and improving the processing effect of fault feature extraction. Then, the RBCBAM-Net model that fuses the residual structure and the channel-spatial attention mechanism is proposed, which can adaptively extract key features of the image at multiple scales, significantly improving the feature extraction ability and generalization performance of the model, and providing an effective solution for the intelligent diagnosis of typical faults of rolling bearing-rotor systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 is a flowchart of an embodiment of the present invention.

[0044] Figure 2 is a layout diagram of the INV1216 type test bench and acceleration sensors in an embodiment of the present invention.

[0045] Figure 3 In (a) is the process of encoding the velocity signal V ij (n) into a VMTF image in an embodiment of the present invention, and (b), (c), (d), (e), (f) are the VMTF images corresponding to each state, and the order is normal, unbalance, misalignment, pedestal looseness, and rubbing between the moving and static parts in sequence.

[0046] Figure 4 is a structural diagram of the RBCBAM-Net model in an embodiment of the present invention.

[0047] Figure 5 is a training curve diagram comparing the VMTF-RBCBAM-Net model with traditional methods in an embodiment of the present invention, where: Fig. (a) is the accuracy curve of the training set, and Fig. (b) is the loss value curve of the training set.

[0048] Figure 6 is a comparison diagram of the verification set accuracy between the VMTF-RBCBAM-Net model and traditional methods in an embodiment of the present invention.

[0049] Figure 7 is a confusion matrix of the test set of the VMTF-RBCBAM-Net model in an embodiment of the present invention.

[0050] Figure 8 is a t-SNE distribution diagram of the test set of the VMTF-RBCBAM-Net model in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0051] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments and the accompanying drawings.

[0052] Refer to Figure 1 , an intelligent fault diagnosis method for a rolling bearing-rotor system based on VMTF-RBCBAM-Net, comprising the following steps:

[0053] 1) Use an acceleration sensor to collect the vibration acceleration signal S ij (n) of typical faults of the rolling bearing-rotor system. The typical faults include: imbalance, misalignment, looseness, and dynamic-static rubbing faults, where: i represents the fault type, j represents the jth acceleration signal sample of the ith type of fault, n is the index position of the sampling point of the vibration acceleration signal sample, n = 1, 2, … N, and N represents the sample length;

[0054] In this embodiment, a rolling bearing-rotor system fault is simulated through a rotor test bench of model INV1216. Refer to Figure 2 , this test bench is composed of components such as a motor, a coupling, a main shaft, a rotor disc, and a bearing housing. The rotor disc is placed in the middle position of the main shaft. The main shaft is connected to the motor through a coupling. The rolling bearing is installed inside the bearing housing; vibration acceleration sensors are respectively installed in the vertical direction of bearing housing 1 and bearing housing 2, with the model CTC-AC192 and a sensitivity of 100 mV / g; the working speed of the motor is stabilized at 1500 rpm, and the MachVIEW analysis system is used in the experiment to collect the vibration acceleration signal S ij (n). The sampling frequency f s is set to 2048 Hz, and the sample length N is 4096; the fault settings of the rolling bearing-rotor system are shown in Table 1, simulating imbalance, misalignment, support looseness, and dynamic-static rubbing faults respectively. The vibration acceleration signals collected at bearing housing 1 in each fault state are used for model training and testing. The specific method is as follows:

[0055] Imbalance fault: Install screws of different masses on the mass disc of the rotor disc to simulate different degrees of imbalance states;

[0056] Misalignment fault: Add gaskets on both sides of bearing housing 1 to change the elevation of the bearing housing and simulate the misalignment state of the shafting;

[0057] Support looseness fault: Adjust the tightness of the bolts on both sides of bearing housing 1 to simulate the support looseness state;

[0058] Dynamic-static rubbing fault: Install a rubbing screw above the rubbing bracket and adjust its position to make it contact the rotor surface to simulate the dynamic-static rubbing fault.

[0059] Table 1 Fault information table

[0060] Fault type Fault degree 1 Fault degree 2 Normal state / / Unbalance Counterweight 4.7g Counterweight 6.4g Misalignment Shim up 0.2mm Shim up 0.4mm Loose support One-side bolt loose Both-side bolts loose Static-dynamic rubbing Intermittent touch Continuous touch

[0061] 2) Adopt the integral formula: Integrate the vibration acceleration signal S ij (n) into the velocity signal V ij (n), where: Δt represents the sampling time interval, and Δt = 1 / f s ;

[0062] 3) Use the formula: V' ij (n) = (V ij (n) - min[V ij (n)]) / (max[V ij (n)] - min[V ij (n)]), to normalize the velocity signal V ij (n) to obtain V' ij (n), to ensure that the scales of the input data are relatively consistent;

[0063] 4) Use the Markov transform to encode the velocity signal V' ij (n) into a two-dimensional image VMTF, with the image size of 224×224. The conversion process and the generated image are as shown in Figure 3 (a);

[0064] Divide the VMTF image into a training set, a validation set, and a test set. The number of samples and labels corresponding to each fault type are shown in Table 2, where: the training set and the validation set are VMTF images of the same degree of fault, and the test set is VMTF images of different degrees of fault; Figure 3 In (b), (c), (d), (e), (f) are the VMTF images corresponding to each state, and the order is normal, unbalance, misalignment, bearing looseness, and rotor-stator rub respectively.

[0065] Table 2 Dataset Division

[0066]

[0067] In step 4), the specific operation of encoding V' ij (n) into the VMTF image is as follows:

[0068] First, divide V' ij (n) into Q partitions, take Q = 16, and map it to the corresponding state q according to the value range of each sampling point V' ij (n), where q = 1, 2,..., Q;

[0069] Then, construct a Q×Q Markov transition matrix, representing the probability of transitioning from one state to another state. The specific formula is:

[0070]

[0071] where: w pq represents the probability of transfer from partition p to partition q, and N pq represents the number of times of transfer from state p to state q, and N p represents the total number of times state p appears;

[0072] Then construct a T×T transfer field matrix M, which represents the state transfer probability within different time windows. The specific formula is:

[0073]

[0074] where: m pq represents the transfer probability from time window t p to t q , is the number of times of transfer from state p to state q within time window t p , represents the total number of times state p appears within time window t p ;

[0075] 5) Network construction: Construct the RBCBAM-Net network model, as Figure 4 shown, whose structure includes multiple convolutional layers, pooling layers, RBCBAM modules, fully connected layers, Dropout layers and Softmax functions; specifically as follows:

[0076] Convolutional layer and pooling layer: First, perform convolution and pooling operations on the VMTF image to extract preliminary features and obtain the feature vector x;

[0077] RBCBAM module: Each RBCBAM module enhances the feature representation through residual connection and CBAM attention mechanism to alleviate the problem of gradient disappearance; each module includes two 3×3 convolutional layers, two BN layers, CBAM attention mechanism and residual connection;

[0078] Global average pooling layer: Perform global average pooling on the feature map processed by multiple RBCBAM modules to further compress the feature dimension;

[0079] The specific steps are:

[0080] The first step: The feature vector x undergoes convolution operation and batch normalization processing to generate the feature maps Y 1 and Y 2 ;

[0081] Y 1 = BN(Conv m×m (x)) (3)

[0082] Y 2 = ReLU(BN(Convm×m (Y 1 ))) (4)

[0083] Where: Conv m×m is the convolution operation, m = 3, 4... 11, and in this embodiment, m = 3, that is, 3×3 convolution is adopted; BN is the batch normalization operation, and ReLU is the activation function;

[0084] Step 2: The channel attention module CAM first performs average pooling and max pooling operations on Y 2 to obtain and Then, two channel attention vectors are obtained through MLP; and then they are normalized by the Sigmoid function to obtain the channel attention weight M C (Y 2 ); Finally, it is applied to Y 2 to obtain the feature map Y C , and the specific formula is as follows:

[0085]

[0086] Y C = Y 2 × M C (Y 2 ) (6)

[0087] Where: σ is the Sigmoid activation function, MLP represents the multi-layer perceptron to process the pooling result, AvgPool(Y 2 ) represents the average pooling operation on Y 2 , and MaxPool(Y 2 ) represents the max pooling operation on Y 2 ;

[0088] Step 3: The spatial attention module SAM first performs average pooling and max pooling on Y C in the channel dimension to obtain two spatial attention feature maps; then they are concatenated and a convolution operation is performed on them to obtain the spatial attention weight M S (Y C ); Finally, M S (Y C ) is used for Y C to obtain the spatially attention-adjusted feature map Y S ;

[0089] M S (Y C ) = σ(Conv t×t (Concat[AvgPool spatial (Y C ),MaxPoolspatial (Y C )])) (7)

[0090] Y S = Y C × M S (Y C ) (8)

[0091] In the formula: Conv t×t is the convolution operation, where t = 3, 4... 11, and in this embodiment, t = 7, that is, a 7×7 convolution kernel is used; Concat represents the operation of splicing feature maps in the channel dimension, merging different pooling results, AvgPool spatial (Y C ) and maxPool spatial (Y C ) respectively represent global average pooling and global max pooling of the feature map Y C in the channel dimension;

[0092] Fourth step: After the feature map Y S is processed by the ReLU function, it is added to the residual Shortcut(x) to obtain the output Y out , and the specific formula is:

[0093] Y out = ReLU(Y S ) + Shortcut(x) (9)

[0094] where: Shortcut(x) = x;

[0095] Classification module: It includes multiple fully connected layers, activation functions, Dropout layers and Softmax activation functions, and is responsible for mapping high-dimensional feature vectors to specific fault categories;

[0096] Figure 4 The RBCBAM module shown realizes dynamic enhancement and optimization of features through residual connection and CBAM attention mechanism; each module includes two 3×3 convolutional layers, two BN layers, CBAM attention mechanism (channel attention module CAM and spatial attention module SAM) and residual connection; after the feature vector x enters the RBCBAM module, it outputs Y out after a series of convolutional calculations and feature fusions;

[0097] 6) Input the training set data prepared in step 4) into the RBCBAM-Net network model constructed in step 5) for training, and save the trained RBCBAM-Net network model;

[0098] In this embodiment, the Adam optimizer is used, the learning rate is set to 0.00001, the batch size is 32, the number of iterations is 50, and the cross-entropy loss function is used to optimize the model. The training curves of the method proposed in this embodiment are compared with those of four traditional methods, namely AMTF-BP, AMTF-LeNet-5, AMTF-AlexNet, and AMTF-RBCBAM-Net. The results are as Figure 5 shown: The AMTF-BP method requires 35 iterations to reach the basic convergence state, and the final accuracy is only 85%, and the loss value remains at about 0.4; The AMTF-LeNet-5 network converges after 23 iterations, the accuracy reaches about 96%, and the loss value drops to 0.05; The AMTF-AlexNet method converges after 30 iterations, the accuracy is about 95%, and the loss value remains at 0.04; The AMTF-RBCBAM-Net method converges after 15 training cycles, and the final accuracy is increased to 97%, and the loss value drops to 0.03; The VMTF-RBCBAM-Net method proposed in this embodiment converges after 5 iterations, and the accuracy is close to 100%; The loss curve decreases rapidly during the iteration process and completely converges after 5 iterations, the loss value is close to 0, and the curve is stable and smooth. Compared with the above four methods, the present invention is superior in terms of accuracy, convergence, and stability.

[0099] In this embodiment, the recognition effects of the above four traditional methods and the VMTF-RBCBAM-Net method on the validation set are compared. The results are as Figure 6 shown: The classification accuracy of the VMTF-RBCBAM-Net method reaches 97%, and its effect is significantly better than the above four traditional methods.

[0100] 7) Import the VMTF images of the test set in step 4) into the RBCBAM-Net network model trained in step 6) for fault diagnosis, and use the confusion matrix and t-SNE method to visually display the classification effect.

[0101] In this embodiment, the confusion matrix and t-SNE are used to visually display the results of the VMTF-RBCBAM-Net method on the test set, as Figure 7 、 Figure 8 shown: Figure 7 The confusion matrix in Figure 8 shows that the average accuracy of the model for various types of faults reaches 98%; The t-SNE distribution in shows that different types of faults have good separability in the feature space and there is almost no obvious overlap between them, indicating that the model proposed by the present invention has a good classification effect on the rolling bearing-rotor system faults.

Claims

1. An intelligent fault diagnosis method for rolling bearing-rotor system based on VMTF-RBCBAM-Net, characterized in that: The following steps are involved: 1) Use acceleration sensors to collect the vibration acceleration signal S of typical faults in rolling bearing-rotor systems ij (n), typical faults include: imbalance, misalignment, looseness and dynamic-static friction faults, where: i represents the fault type, j represents the jth acceleration signal sample of the i-th fault, n is the index position of the vibration acceleration signal sample sampling point, n = 1, 2, ... N, N represents the sample length; 2) Using the integral formula: The vibration acceleration signal S ij (n) The integral is the speed signal V ij (n), where: Δt represents the sampling time interval; 3) Using the formula: V' ij (n)=(V ij (n)-min[V ij (n)]) / (max[V ij (n)]-min[V ij (n)]), for the speed signal V ij (n) is normalized to obtain V' ij (n); 4) Use Markov transformation to transform V' ij (n) Encode into VMTF images, and then divide them into training set, validation set and test set according to the ratio, where: the training set and validation set contain VMTF images with the same degree of fault, and the test set contains VMTF images with different degrees of fault; 5) Construct the RBCBAM-Net network model, which consists of a convolutional layer, a pooling layer, an RBCBAM module, a global average pooling layer, and a classification module; 6) Input the training set data prepared in step 4) into the RBCBAM-Net network model constructed in step 5) for training, and save the trained RBCBAM-Net network model; 7) Import the VMTF images of the test set in step 4) into the RBCBAM-Net network model trained in step 6) for fault diagnosis, and use the confusion matrix and t-SNE method to visualize the classification effect.

2. The method according to claim 1, characterized in that Step 4) V' ij (n) The specific operation of encoding into VMTF image is: First, V' ij (n) is divided into Q partitions, and each sampling point V' ij The value range of (n) is mapped to the corresponding state q, q = 1, 2, ..., Q; Then, a Q×Q Markov transition matrix is ​​constructed to represent the probability of transitioning from one state to another. The specific formula is: Where: w pq represents the probability of transferring from partition p to partition q, N pq represents the number of transitions from state p to state q, N p Indicates the total number of times state p appears; Then construct a T×T transfer field matrix M to represent the state transition probability in different time windows. The specific formula is: Where: m pq represents the time window t p to q The transition probability, From the time window t p The number of transitions from state p to state q, represents the time window t p The total number of times internal state p occurs.

3. The method according to claim 1, characterized in that In step 5), the convolution layer and pooling layer of the RBCBAM-Net network model first perform convolution and pooling operations on the VMTF image to obtain a feature vector x; The RBCBAM-Net network model contains multiple RBCBAM modules. Each RBCBAM module contains two convolutional layers, two BN layers, CBAM attention mechanism and residual connection. After convolution operation and feature fusion, the output is Y out .

4. The method according to claim 3, characterized in that The specific steps are as follows: Step 1: The feature vector x is processed by convolution and batch normalization to generate feature maps Y1 and Y2; Y1=BN(Conv m×m (x)) (3)Y2=ReLU(BN(Conv m×m (Y1))) (4) Where: Conv m×m is a convolution operation, m=3,4…11; BN is a batch normalization operation, and ReLU is an activation function; Step 2: The channel attention module CAM first performs average pooling and maximum pooling operations on Y2 to obtain and Then, two channel attention vectors are obtained through MLP; and then normalized through the Sigmoid function to obtain the channel attention weight M C (Y2); finally, apply it to Y2 to obtain the feature map Y C , the specific formula is as follows: Y C =Y2×M C (Y2) (6) Where: σ is the Sigmoid activation function, MLP represents the pooling result of multi-layer perceptron processing, AvgPool(Y2) represents the average pooling operation on Y2, and MaxPool(Y2) represents the maximum pooling operation on Y2; Step 3: The spatial attention module SAM first C Average pooling and maximum pooling are performed on the channel dimension to obtain two spatial attention feature maps; then they are concatenated and convolved to obtain the spatial attention weight M. S (Y C ); Finally, M S (Y C ) for Y C , get the feature map Y after spatial attention adjustment S ; M S (AND C )=σ(Conv t×t (Concat[AvgPool spatial (AND C ),MaxPool spatial (AND C )])) (7) AND S =And C ×M S (AND C ) (8) Where: Convt ×t is a convolution operation, t=3,4…11; Concat represents the concatenation operation of the feature map of the channel dimension, merging different pooling results, AvgPool spatial (Y C ) and maxPool spatial (Y C ) represent the feature map Y C Perform global average pooling and global maximum pooling in the channel dimension; Step 4: Feature map Y S After being processed by the ReLU function, it is added to the residual Shortcut(x) to get the output Y out , the specific formula is: Y out =ReLU(Y S )+Shortcut(x) (9)wherein: Shortcut(x)=x.

5. The method according to claim 1, characterized in that In step 5), the classification module consists of a fully connected layer, an activation function, a Dropout layer, and a Softmax function, which is responsible for mapping the extracted high-dimensional feature vector to a specific fault category.

Images