A Fault Diagnosis Method Based on the Fusion of Residual Learning and Attention Mechanism

Abstract

This invention discloses a fault diagnosis method based on the fusion of residual learning and attention mechanisms. The steps are as follows: (1) Fuse the raw signal data from multiple sensors, then normalize the data and convert it into a two-dimensional grayscale dataset as input to the model; (2) Build a model consisting of an inverted bottleneck structure residual block, a self-attention and convolutional combined residual block (CA), a squeeze and excitation network residual block (SE), and a multilayer perceptron classifier (MLP); (3) The inverted bottleneck structure residual block extracts sample features, and CA enriches the extracted fault features; (4) SE uses channel squeezing and excitation operations to focus on useful features and ignore invalid features; (5) The MLP outputs each fault category; (6) The loss obtained from each iteration's cross-entropy is backpropagated to update the model parameters. When the final iteration number is reached, the network model is determined for fault type identification. This invention extracts more effective information, has a faster computation speed, and a higher recognition rate.

Description

Technical Field

[0001] This invention addresses the diversity and complexity of equipment failures in industrial applications by employing a fault diagnosis method based on a deep learning model and a multi-residual block neural network, thus belonging to the field of fault diagnosis. Background Technology

[0002] With the continuous development of modern industries, industrial equipment is becoming increasingly precise and complex. However, this increased complexity also significantly increases the likelihood of equipment malfunctions, leading to substantial losses for enterprises and even personal injury. Therefore, fault diagnosis methods are of paramount importance. When rolling bearings experience fatigue, wear, corrosion, or other damage, they generate additional vibrations during the operation of machinery. Therefore, it is possible to collect and analyze the vibration signals of the bearings during operation to determine if a fault has occurred. Research on rolling bearing fault diagnosis based on vibration signals can be divided into two categories: traditional fault diagnosis methods and deep learning-based methods. Traditional fault diagnosis methods, which rely on expert knowledge, often face a knowledge acquisition bottleneck, lacking effective ways to express diagnostic knowledge and struggling to achieve high-precision fault diagnosis. Therefore, how to reduce the involvement of expert knowledge and achieve high-precision intelligent fault diagnosis based on the inherent feature information of the monitoring data itself has become a current research hotspot.

[0003] With the continuous advancement of information technology, deep learning-based fault diagnosis has gradually become a widely used method in the field of fault diagnosis. Deep learning can automatically learn data features from raw signals, mapping the complex relationship between vibration signals and bearing faults. It can quickly and effectively monitor the bearing's operating status using vibration signals without manual feature extraction, which is of significant research value in improving the efficiency of fault diagnosis. In early deep learning fault diagnosis methods, some traditional models, such as Support Vector Machines (SVM), Naive Bayes, and decision trees, played a certain role. However, due to limitations in the computing power and iteration methods of computer hardware at the time, they could not accurately identify complex fault identification problems. Later, convolutional neural networks achieved good research results in computer vision, autonomous driving, medical imaging, and other fields, and have been widely applied in deep learning-based fault diagnosis. This demonstrates the potential of deep learning in the field of fault diagnosis. However, at present, in order to improve the accuracy of fault identification, the proposed convolutional neural network models are relatively complex, with a large number of parameters, an ever-deepening number of network layers, and a large storage space required. As a result, phenomena such as network degradation and overfitting have gradually emerged. Therefore, some scholars have proposed to use the classic convolutional neural network model residual network to effectively suppress the phenomenon of network degradation. Moreover, this model has good flexibility and can be combined with various modules to achieve the desired effect.

[0004] In real-world industrial environments, equipment parts are often exposed to harsh conditions, and the data collected by sensors often contains a great deal of interference. To suppress this interference and extract effective features from faults, an attention mechanism combined with residual blocks is used to effectively identify the fault type. Compared to ordinary convolutional neural networks, this approach can extract more effective information, has a faster processing speed, and a higher recognition rate. Summary of the Invention

[0005] The purpose of this invention is to provide a fault diagnosis method that combines multiple residual modules. This method fuses one-dimensional raw signals collected by multiple sensors and transforms them into a two-dimensional grayscale dataset with distinct features. This greatly preserves feature information, eliminates the need for cumbersome parameter adjustments, has a high conversion rate, and provides richer features compared to vibration signals collected by a single sensor, while effectively resisting interference features. The method employs inverted bottleneck residual blocks to extract fault features and residual blocks combining self-attention and convolution to extract a large number of effective features, enriching the fault feature set. The obtained information is then integrated and sent to squeeze and excite (SE) residual blocks for centralized processing. Through channel feature transformation operations, effective fault features are prioritized while ineffective features are ignored, ultimately accurately outputting complex fault types.

[0006] Collection of raw data for S1 fault

[0007] This invention aims to monitor the different operating states of industrial rotating equipment under complex working conditions. Multiple acceleration sensors are installed near the location of equipment failure. Signals under various rotational speeds are collected using a fixed sampling frequency and sampling time. The vibration signals from multiple sensors are fused and transmitted to a computer. MATLAB software is used on the computer to preprocess the collected signals.

[0008] S2 Data Processing and Sample Construction

[0009] To address the limitations of raw signals acquired by a single sensor, a correlation fusion approach using signals from multiple sensors is employed. Assume x1(n), x2(n), ..., x... t (n) represents the signals collected by t sensors, and the correlation between sensor i and sensor j can be expressed as follows:

[0010]

[0011] Where N represents the total number of signals, n represents the signal length, and m is the time coordinate shift value of the discrete signals from different channels in the correlation calculation. Since the signals are acquired simultaneously by multiple sensors, and a single signal point is acquired very quickly, k is marked as 0 for simplicity. Therefore, the fused signal from multiple sensors can be represented as:

[0012]

[0013] To address the issue of individual abnormal data values ​​and mitigate their impact, the collected raw data was normalized. Finally, MATLAB software was used to convert the normalized data sequence into 224×224 grayscale images as input samples for the model. Each fault type contained 680 grayscale images. The raw data was normalized using formula (3) to normalize the data to the range [0,1] d, as shown in the formula below:

[0014]

[0015] Where Y represents the original fused data, and max(Y) and min(Y) represent the maximum and minimum values ​​in the original data, respectively.

[0016] S3 performs classification using an improved multi-residual neural network model.

[0017] The established multi-residual neural network model structure mainly consists of four parts: inverted bottleneck structure residual blocks, self-attention and convolution combined residual blocks (CA), squeezed and activated residual blocks (SE), and a multilayer perceptron classifier (MLP). The diagnostic process for this model structure is as follows:

[0018] (1) First, the original signals from multiple sensors at different speeds are fused. Then, the fused signals are processed into grayscale data as input data. Different categories of grayscale images are labeled with 0, 1, 2, 3, ..., C using Python software, where C is the number of categories.

[0019] (2) The model is built and the network model parameters are initialized in the PyTorch framework environment of Python software. It mainly consists of four parts: inverted bottleneck structure residual blocks, self-attention and convolution combined residual blocks (CA), squeezed and activated residual blocks (SE), and a multilayer perceptron classifier (MLP). Combining these four parts in the order described above allows for the full processing of fault feature information. Each part is described in detail below:

[0020] In the inverted bottleneck residual block, the structure is composed of a deep convolutional layer with a 7×7 kernel, a batch normalization layer, a point convolutional layer with a 1×1 kernel, an activation function ReLU, and another point convolutional layer with a 1×1 kernel, connected sequentially. The shape of this structure resembles an inverted bottleneck. First, a deep convolutional layer with a large kernel is used to extract features, which has good robustness. Then, point convolutional layers are used to summarize the features from each channel of the deep convolution, effectively preventing information loss. In the field of deep learning, the combination of deep convolutional layers and point convolutional layers is also called grouped convolution, which can reduce the amount of computation and reduce computational loss.

[0021] In the self-attention and convolution combined residual block (CA), the structural sequence is mainly a convolutional layer with a 1×1 kernel, a batch normalization layer, a ReLU activation function, a two-branch structure of self-attention mechanism and convolution, a batch normalization layer, a ReLU activation function, a convolutional layer with a 1×1 kernel, a batch normalization layer, and a ReLU activation function connected sequentially. The two-branch structure of self-attention mechanism and convolution consists of two parallel branches: a convolution operation and a self-attention operation. In this module, the initial 1×1 convolutional layer is used to process features, then the process is divided into two branches: one for convolution operations and one for self-attention operations. Finally, a convolutional layer with a 1×1 kernel aggregates the features from both branches. This module mainly utilizes the convolution and self-attention dual-branch feature processing, combining the feature extraction capability of convolution with the localized feature processing capability of self-attention to obtain rich fault features, facilitating subsequent processing of the most relevant fault features.

[0022] Squeezing and Activating Residual Blocks (SE) is a commonly used channel attention mechanism. It consists of a basic residual block, a squeezing process, and an activation process. The basic residual block is formed by sequentially connecting a 1×1 convolutional layer, a batch normalization layer, a 3×3 convolutional layer, a batch normalization layer, another 1×1 convolutional layer, and another batch normalization layer to extract input features. Then, in the squeezing process, a global pooling layer compresses the two-dimensional features of each channel of the input features into a single real number. This real number has a global receptive field, while the number of channels remains unchanged. Finally, in the activation process, the first fully connected layer compresses C channels into C / r channels, reducing computational cost, where r is a scaling factor. After comparing various values ​​of r, it was found that r=16 achieves the best balance between overall performance and computational efficiency. Then, after passing through the ReLU activation function, a second fully connected layer operation is performed to restore it to the original C channels. Finally, after passing through the Sigmoid function, a vector is obtained, which is multiplied with the previous basic residual block result to complete the channel attention mechanism process. This allows the rich features obtained from the convolution and self-attention mechanism residual blocks to be filtered, important fault features to be selected and features irrelevant to faults to be suppressed, thereby obtaining important features in an automatic learning manner.

[0023] The MLP classifier mainly consists of an input layer, hidden layers, and an output layer. To reduce the number of parameters, each of the input, hidden, and output layers uses only one fully connected layer. A ReLU activation function is used for non-linear transformation between every two fully connected layers. The output results are compared with the actual labeled fault types, and the results are sorted according to the probability values ​​of each label to obtain the final classification result. The MLP classifier is a high-performing classification algorithm in deep learning classification tasks. When combined with the aforementioned residual blocks, it can improve diagnostic performance.

[0024] The network is trained using sample labels assigned to fault types as its output, with 30 training iterations. The output of the MLP classifier is compared with the actual labeled fault types, and backpropagation is performed based on the obtained loss values ​​to update the parameters, ultimately yielding the final classification result.

[0025] Adjustment of S4 model parameters

[0026] After each round of minimum batch training, the network structure calculates the loss value based on the results, automatically performs backpropagation for feedback, and adjusts the parameters of each part of the model in a timely manner. When the final training reaches the final set value, the model completes training and the fault diagnosis process ends.

[0027] Compared with the prior art, the inventiveness of this invention is mainly reflected in:

[0028] This invention primarily addresses a fault diagnosis method for suppressing interference in collected data in industrial environments and utilizing a neural network model combined with multiple residual blocks. The original signal undergoes information fusion and data normalization to suppress interference from abnormal data values. The processed signal data is then converted into a two-dimensional grayscale image. This conversion process makes the differences between different fault characteristics more apparent, amplifying the feature differences between different faults. This allows the multi-residual network model to fully leverage its advantages in processing image data, automatically extracting features from the input data without requiring prior knowledge in related fields. This method combines various modules to focus on capturing key features while ignoring invalid features. Each residual block uses as few activation functions as possible, reducing the computational load and preventing the loss of useful information. Furthermore, the combination of different modules lays the groundwork for subsequent intelligent fault diagnosis methods targeting different problems. Attached Figure Description

[0029] Figure 1 This is a flowchart of the intelligent diagnostic method for the multi-residual block structure neural network of the present invention.

[0030] Figure 2 This is a simplified diagram of the wind power equipment test bench and the sensor installation locations.

[0031] Figure 3 This is a diagram of the multi-residual block structure neural network of the present invention.

[0032] Figure 4 It is a diagram of a two-branch structure combining convolution and self-attention.

[0033] Figure 5 The confusion matrix of the proposed method on the wind power monitoring test bench

[0034] Figure 6 The proposed method classifies T-SNE visualization.

[0035] Figure 7 The graph shows the classification accuracy versus loss curve of the proposed method. Detailed Implementation

[0036] The following is a detailed explanation based on the accompanying diagrams and specific procedures.

[0037] according to Figure 1 The flowchart of the intelligent diagnosis method of a multi-residual block structure neural network is as follows: (1) First, signals are collected by multiple sensors installed on the wind power condition monitoring test bench; (2) The collected sensor signals are fused and then normalized to convert the normalized data into a two-dimensional grayscale dataset; (3) The grayscale dataset is input into the set network model, such as... Figure 1Within the dashed frame is the proposed network model structure, which sequentially passes through inverted bottleneck structure residual blocks, convolutional and self-attention residual blocks (CA), and squeezed and activated residual blocks (SE); (4) finally, the loss value is calculated based on the model results, and backpropagation is performed in a timely manner to update the parameters of the network model and continuously train it. The following is a detailed explanation of the multi-residual block structure neural network structure diagram in conjunction with the diagnostic method flowchart.

[0038] The wind power monitoring equipment test bench has six operating states for the test gears: inner bearing ring failure (gear tooth breakage), inner bearing ring failure (gear normal), inner bearing ring failure (gear wear), outer bearing ring failure (gear tooth breakage), outer bearing ring failure (gear normal), and outer bearing ring failure (gear wear). Each state is assigned a status label from 0 to 5. Three accelerometers are positioned above and to the side of the gearbox, and above the bearing housing (CH1 to CH3), respectively, to ensure optimal acquisition of status information for each operating component. The signal speed is 1500 r / min, and the sampling frequency is 10 kHz.

[0039] In step 2), correlation data fusion is performed on signals collected from multiple sensors. The specific process is as follows:

[0040] Assume x1(n), x2(n), ..., x t (n) represents the signals collected by t sensors, and the correlation between sensor i and sensor j can be expressed as follows:

[0041]

[0042] Where N represents the total number of signals, n represents the signal length, and m is the time coordinate shift value of the discrete signals from different channels in the correlation calculation. Since the signals are acquired simultaneously by multiple sensors, and a single signal point is acquired very quickly, k is marked as 0 for simplicity. Therefore, the fused signal from multiple sensors can be represented as:

[0043]

[0044] The merged data is then normalized to remove the influence of outliers, and the data is normalized to the range d of [0,1]. The formula is as follows:

[0045]

[0046] Where Y represents the original fused data, and max(Y) and min(Y) represent the maximum and minimum values ​​in the original data, respectively. The normalized data d is converted into a 224×224 two-dimensional grayscale image sample using MATLAB software as the input data x.

[0047] In step 3), a multi-residual block neural network is established to achieve fault identification. The detailed structure of the entire model is as follows: Figure 3 As shown.

[0048] (1) First, the input data is preprocessed by a convolutional layer with a 7×7 kernel, a batch normalization layer, and a max pooling layer. The preprocessing process is shown in formula (3). The formulas below are labeled with different subscripts to distinguish the results of each process.

[0049] y i =max{H(f(x)),s} (3)

[0050] Where x represents the input data, f(·) represents the convolution operation, H(·) represents the batch normalization operation, and y i The output features of the preprocessing are represented by s, and the pooling step size is represented by s. After comparing various step size values, s=2 is used to keep the input and output channels consistent.

[0051] (2) Then, features are extracted through the inverted bottleneck residual block structure. This inverted bottleneck residual block contains grouped convolutions, which improves the computation speed of the model while extracting features. The residual block structure first uses a large 7×7 convolutional kernel and then two 1×1 convolutional kernels to form a structure similar to an inverted bottleneck, effectively preventing information loss. Its main calculation process is shown in formula (4):

[0052] y o =P(d(y) i (4)

[0053] In the formula y i Let d(·) represent the input features, P(·) represent the depthwise convolution operation, and y represent the pointwise convolution operation. o This represents the output of the entire inverted bottleneck structure.

[0054] (3) Secondly, a convolutional and self-attention residual block (CA) is used to process the input features of the previous residual block. First, the residual block shares some operations, and then the features are processed by the self-attention and convolutional branches before being aggregated together to further enrich the fault features and obtain more fault feature information. The parallel branch structure of convolution and self-attention is shown in the attached figure. Figure 3 As shown, the specific process is as follows: First, the input features are processed by a convolutional layer with a 1×1 kernel. The expression is as follows:

[0055] y c =f1(y o (5)

[0056] In the formula, yo represents the input feature, f1(·) represents the convolution operation with a 1×1 convolution kernel, and y cThis represents the output of a convolutional layer with a 1×1 convolutional kernel.

[0057] Then, the input features are processed by a convolution process in one branch, performing a linear mapping, the expression of which is:

[0058] g c =Ky c (6)

[0059] Where K represents the convolution kernel weight, y c This represents the features extracted by a 1×1 convolution kernel. Next, the resulting feature maps are aggregated together, and their mathematical expression is shown below:

[0060] g c '=∑g c (7)

[0061] In another branch, the self-attention operation first takes the feature y as input. c The projection is a query q, a key k, and a value v, and the formula is as follows:

[0062] q = W q y c ,k=W k y c v = W v y c (8)

[0063] Among them W q W k and W v This is the weight matrix for query q, key k, and value v. Next, the obtained attention weights and values ​​are integrated, and their mathematical expression is shown below:

[0064] g s =∑A(q,k)v (9)

[0065] In the formula, A(·) represents the attention calculation weight, and g s This represents the output of the self-attention operation.

[0066] Finally, the results from the two branches are added together to obtain the output g. w Its mathematical expression is shown in (10). Thus, a large number of rich features were obtained, containing more fault information.

[0067] g w =g c ′+g s (10)

[0068] (4) The features obtained in the CA residual block are then passed to the squeezed and excitation residual block (SE). This structure adds an attention mechanism to the feature channel dimension in the basic residual block, including two key operations: squeeze and excitation. The process is as follows: The squeeze operation is a global flat pooling layer that performs a global flat pooling operation on the input feature g. w The c channels are compressed into c real numbers, resulting in 64 channels. The mathematical expression for this is:

[0069] S = Avg(g) w (11)

[0070] In the formula, Avg(·) represents the averaging operation, and S represents the output of the extrusion process. Then, channel recovery is performed through excitation. In the first step, the first fully connected layer reduces the dimensionality of channel c by a factor of r, then performs nonlinearization using the ReLU activation function. Next, the second fully connected layer increases the dimensionality by a factor of r. After comparing various values ​​of c and r, it is finally determined that c=64 and r=16 achieve the best balance between overall performance and computational efficiency. The second step of the excitation operation first uses the Sigmoid activation function to estimate the channel weights, obtaining a vector value. The mathematical expression for the excitation process is shown below.

[0071] α=F ex (S,W) (12)

[0072] In the formula, S represents the output result of the extrusion process, W represents the fully connected operation in the excitation process, and F... ex (·) represents the excitation function, and α represents the vector value output by the excitation process.

[0073] Since the squeezing and excitation processes operate on the features of the basic residual block, the vector value obtained from the excitation process is multiplied by the feature value obtained from the basic residual block to complete the entire channel attention process. The calculation is as follows:

[0074] m=α·T(g w (13)

[0075] Where m represents the output of the squeezed and excited residual block (SE), T(·) represents the basic residual block function, and g w It represents the output feature of convolution and self-attention residual block (CA), where α represents the vector value output by the activation process.

[0076] (5) Finally, the classification task is completed by the MLP classifier, and the results of each fault category are output. The mathematical expression is as follows:

[0077] y l =M(∑w l m) (12)

[0078] In the formula, m represents the input feature, and y represents the input feature. l For the output result, w l Here, M represents the connection weights of the intermediate layer neurons, and M(·) represents the activation function. After the training phase, backpropagation is performed based on the obtained cross-entropy loss value to continuously adjust the training parameters. The process of calculating the cross-entropy loss value is shown below:

[0079]

[0080] In the formula, C represents the number of categories. There are 6 states, so C = 6, and the labels corresponding to the 6 states are 0, 1, 2, 3, 4, and 5. β represents the 16 random label values ​​corresponding to the actual faults, such as 4, 2, 1, 2, 3, 5, 3, 6, 1, 7, 2, 4, 7, 2, 4, 6; γ is the predicted value, which is determined by the result y. l The model generates 16 predicted fault label values. These 16 label values ​​are determined by the minimum batch size of 16 and are formed by combining labels corresponding to 6 different states. The model performs calculations based on the batch size each time. Finally, backpropagation is continuously performed based on the obtained loss values ​​to update the parameters. When the model iterates to the set final iteration number of 30, the minimum loss value is obtained, and the final output result is obtained.

[0081] These six state signals are used as follows: 480 samples for each type are selected as training samples, and 320 samples for each type are selected as test samples, for a total of 4800 samples. The output layer sets sample labels based on the fault types recorded during signal acquisition as the network output. In the parameter settings, the minimum batch size is 16, and the learning rate is set to 0.001. In step 4), the network is trained 30 times, using backpropagation and stochastic gradient descent during training. The optimizer is Adam, as referenced in Adam: AMethod for Stochastic Optimization.

[0082] The classification performance of the proposed model is represented by a confusion matrix, such as... Figure 5 As shown, the horizontal axis of the confusion matrix represents the predicted class value of the sample, and the vertical axis represents the actual state of the sample. It can be seen that the predicted class corresponds one-to-one with the actual class, and the accuracy is approximately 100%.

[0083] To verify the classification performance of the proposed method, the results were visualized using T-SNE, as shown below. Figure 6 As shown in the results, each category can form a distinct cluster without overlap with other categories, effectively classifying the various fault categories. Figure 7 As shown in the accuracy vs. loss curve, during the continuous iteration process, the fault identification accuracy gradually approaches 1, and the loss value gradually approaches 0, proving the effectiveness of the invention.

Claims

1. A fault diagnosis method based on multi-residual block combination, characterized by: The original signals of multiple sensors under different working conditions are fused, and then the fused data is normalized to suppress the influence of abnormal values in the signals, and finally the normalized data is converted into a two-dimensional gray image as the input of the model to enhance the characteristics between different faults and fully exert the performance of the model in processing picture data; a network model combining multiple residual blocks is proposed, which enriches fault features by combining multiple residual blocks, extracts fault features related to faults, and performs local processing on prominent fault features in the obtained rich features to suppress irrelevant feature interference, and finally the processed results are recognized by a multi-layer perception classifier MLP to identify faults; The loss value is calculated through the training process to update the model parameters by back propagation, and when the final iteration number is reached, the final output result is obtained, so as to perform fault diagnosis; the structure of the network model combining multiple residual blocks is as follows: the model is built under the Pytorch framework, mainly including four parts of inverted bottle neck structure residual block, self-attention and convolution combined residual block CA, squeeze and excitation residual block SE and multi-layer perception classifier MLP; the four parts are combined in turn according to the introduction order, as follows: In the inverted bottle neck structure residual block, the structure sequence is a 7x7 convolution kernel depth convolution layer, a batch normalization layer, a 1x1 convolution kernel point convolution layer, a ReLU activation function and a 1x1 convolution kernel point convolution layer connected in turn; In the self-attention and convolution combined residual block CA, the structure sequence mainly includes a 1x1 convolution kernel convolution layer, a batch normalization layer, a ReLU activation function, a self-attention mechanism and convolution double branch structure, a batch normalization layer, a ReLU activation function, a 1x1 convolution kernel convolution layer, a batch normalization layer and a ReLU activation function connected in turn, wherein the self-attention mechanism and convolution double branch structure are composed of two parallel branches of convolution operation and self-attention operation; in this module, the features are first processed by the preceding 1x1 convolution layer, then divided into two convolution operation and self-attention operation branches, and finally a 1x1 convolution kernel convolution layer is used to aggregate the features of the two branches; The squeeze and excitation residual block SE is composed of a basic residual block, a squeeze process and an excitation process. The basic residual block is sequentially connected by a convolution layer with a 1x1 convolution kernel, a batch normalization layer, a convolution layer with a 3x3 convolution kernel, a batch normalization layer, a convolution layer with a 1x1 convolution kernel and a batch normalization layer, and is used to extract input features. Then, the squeeze process is performed by a global pooling layer to compress the two-dimensional features of each channel of the input features into one real number, i.e., 64 channels. The real number has a global receptive field, and the number of channels remains unchanged. In the excitation process, the first fully connected layer compresses C channels into C / r channels, i.e., 4 channels, to reduce the computational load, where r is a scaling coefficient. After comparing various values of C and r, the overall performance and computational efficiency are balanced when the number of channels C is 64 and the scaling coefficient r is 16. Then, the ReLU activation function is used, and the second fully connected layer is operated to restore the number of channels to the original 64. Finally, the Sigmoid function is used to obtain a vector, which is multiplied by the result of the basic residual block to complete the channel attention mechanism process. The MLP classifier is mainly composed of an input layer, a hidden layer and an output layer. In order to reduce the number of parameters, each of the input layer, the hidden layer and the output layer only uses one fully connected layer. The ReLU activation function is used for non-linear transformation between each two fully connected layers. The output result is compared with the actual labeled fault type. The final classification result is obtained by sorting the probability values of each label. 2.The method according to claim 1, characterized in that The signal fusion method of multiple sensors is as follows: The multiple sensors are used for signal acquisition, and the multiple sensor signal fusion is used to avoid the limitation of using single sensor information, and the correlation between the sensors is represented as follows: is the signal collected by the sensors, and the correlation between the sensors and the sensors is represented as follows: （1） wherein, represents the total number of signals, represents the length of a signal, is the time coordinate shift value of the different channel discrete signals; since the signals are collected simultaneously by multiple sensors, and a single signal point is quickly collected by the sensor, is marked as 0; therefore, the multiple sensor fusion signal is represented as: （2）。 3.The method according to claim 1, characterized in that The training process of the model is as follows: (1) The input data is preprocessed by a convolution layer with a 7x7 convolution kernel, a batch normalization layer and a max pooling layer. The preprocessing process is shown in formula (3), where different subscripts are used to distinguish the results of each process. (3) where denotes the input feature, f() represents the convolution operation, H() denotes the batch normalization operation, denotes the pre-processed output feature, denotes the pooling step, after comparing various step values, s = 2 is adopted to keep the input and output channels consistent; (2) Then, the feature is extracted by the inverted bottleneck residual block structure. The inverted bottleneck residual block structure contains grouped convolution, which can extract features while improving the operation speed of the model. The residual block structure first uses a large convolution kernel of 7x7, and then uses two convolution layers with a 1x1 convolution kernel to form an inverted bottleneck structure. The main calculation process is shown in formula (4): (4) In the formula represents the input feature, d() represents the depth convolution operation, P() represents the point convolution operation, represents the output of the entire inverted bottle-neck structure; (3) The input feature is processed by the residual block CA with convolution and self-attention. The specific process is as follows: the input feature is first processed by a convolution layer with a 1x1 convolution kernel. The expression is as follows: (5) In the formula represents the input feature, f1() represents the convolution operation of the 1x1 convolution kernel, represents the output of the convolution layer of the 1x1 convolution kernel; Then, the input feature is processed by a convolution process in a branch, which is a linear mapping of the input feature. （6） wherein represent the convolution kernel weights, values randomly generated by the network model, represents the feature extracted by the 1x1 convolution kernel; secondly, the mapping feature map is aggregated together, and the mathematical expression is as follows: （7） In another branch, self-attention operation, first input features Projection as query q, key k and value v, formula as follows: （8） where , and are the weight matrices for query q, key k and value v, which are randomly generated by the system; secondly, the resulting attention weights and values are integrated, and the mathematical expression is as follows: （9） wherein A() represents an attention computation weight, an output representing a self-attention operation; Finally, the results of the two branches are added to obtain the output The mathematical expression is shown as (10); thus, a large number of rich features are obtained, and more fault information is contained. （10） (4) The features obtained in the CA residual block are transmitted to the squeeze and excitation residual block SE. The structure increases the attention mechanism for the feature channel dimension in the basic residual block, including two key operations of squeeze and excitation. The process is as follows: the squeeze operation is a global average pooling layer, which compresses c channels of input features into c real numbers, i.e. the number of channels is 64, and the mathematical expression is as follows: ​ (11) wherein, denotes an averaging operation, denotes the output of the squeezing process; then the channel restoration is performed through the excitation operation, the first full connection layer reduces the dimension of c channels by r times in the first step, and then the ReLU activation function is used for nonlinearization, followed by the second full connection layer for increasing the dimension by r times, and through comparison of various values of c and r, it is finally concluded that the overall performance and computational efficiency are most balanced when c=64 and r=16; the excitation operation in the second step first uses the Sigmoid activation function to estimate the weight of the channel to obtain a vector value; the mathematical expression of the excitation process is as follows: (12) wherein denotes the output of the squeezing process, W denotes the fully connected operation in the excitation process, F ex denotes the activation function, denotes the vector value of the output of the excitation process; Since the squeeze and excitation process operates on the features of the basic residual block, the vector value obtained by the excitation process is multiplied by the features obtained by the basic residual block to complete the entire channel attention process. The calculation is as follows: (13) wherein denotes the output of the squeeze-and-excitation residual block SE, T( ) denotes a base residual block function, is the output feature of the convolution-and-self-attention residual block CA; (5) Finally, the classifier MLP completes the classification task, and the results of each fault category are output. The mathematical expression is as follows: (14) In the formula, is an input feature, is an output result, is a connection weight of an intermediate layer neuron, M( ) represents an activation function; after the training session ends, the obtained cross-entropy loss value is used for back propagation, and the training parameters are constantly adjusted; the cross-entropy calculation loss value process is as follows: (12) In the formula, C represents the number of categories, a total of 6 states, so C=6, and the labels corresponding to the 6 states are 0, 1, 2, 3, 4, and 5; The 16 random label values corresponding to the real fault are 4, 2, 1, 2, 3, 5, 3, 6, 1, 7, 2, 4, 7, 2, 4, and 6; The predicted value is the result The 16 predicted fault label values correspond; the 16 label values here are determined by the minimum batch value 16, and are formed by the label combinations corresponding to the 6 different states, and the model operates in batches each time; finally, the parameters are updated by continuously performing back propagation according to the obtained loss value; Until the iteration reaches the final iteration number 30, the minimum loss value is obtained, and the final output result is obtained.

Images