Rolling bearing fault diagnosis method and system

Abstract

The application provides a rolling bearing fault diagnosis method and system, and relates to the technical field of fault diagnosis, and the specific scheme comprises the following steps: according to different sampling frequencies of vibration signals, multi-scale convolutional neural networks are used to extract multi-scale features from the vibration signals to be diagnosed; based on a hierarchical attention mechanism, the multi-scale features are adaptively assigned weights to obtain weighted features; the time sequence features of the weighted features are extracted through a long short-term memory network; the time sequence features are input into a classification layer to obtain a fault diagnosis result; the application is based on a multi-scale CNN-LSTM network, learns features of different time scales, and accurately identifies fault features; considering that the hierarchical attention mechanism can distinguish the weight ratio of each feature, some key information is selected for processing, and the accuracy and efficiency of the training model are improved.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular relates to a method and system for diagnosing rolling bearing faults. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Rolling bearings are common mechanical components widely used in industrial production. In actual production, the failure of a single rolling bearing can cause the entire conveyor belt to stop, requiring a lot of manpower and resources for troubleshooting and repair. Therefore, using a reliable fault diagnosis method to achieve predictive maintenance of automated conveyor equipment can reduce enterprise costs and improve work efficiency, and has extremely high research value.

[0004] Traditional fault diagnosis relies on expert experience and accurate mathematical models. When faced with complex fault data with multiple features, the classification effect is poor and it lacks adaptability. Compared with traditional mathematical model methods, deep learning models focus on learning from historical and empirical data, which can extract deeper features from vibration fault data and obtain more accurate prediction results. Deep learning models show great promise in fault diagnosis, do not require researchers to have a professional educational background, and are more suitable for data-driven research.

[0005] CNNs have excellent feature extraction capabilities and can capture sensitive fault features without requiring expert knowledge, making them increasingly popular in various fault diagnosis methods. LSTM is a time recurrent neural network improved from recurrent neural networks with long-term memory. Due to its excellent ability to process time-series signals and its ability to predict development trends based on historical information, it is widely used in time series processing.

[0006] Although both of the above models have been widely used in the field of fault diagnosis, there are still limitations to using them alone. Because the vibration signal of rolling bearings has complex time-scale characteristics, if only a single CNN network with a fixed kernel size is used, the features obtained are limited and it is not easy to obtain features of other time scales. If only a single LSTM network is used, high-standard preprocessing of the original vibration signal is required to build a sufficiently complex diagnostic model in order to obtain ideal results.

[0007] Therefore, existing rolling bearing fault diagnosis methods require prior signal processing knowledge and rely on manual extraction of fault features to achieve fault diagnosis. Such methods are difficult to capture key fault features, have poor universality, and are inefficient. Summary of the Invention

[0008] To overcome the shortcomings of the prior art, this invention provides a rolling bearing fault diagnosis method and system. Based on a multi-scale CNN-LSTM network, it learns features at different time scales to accurately identify fault features. Considering that the hierarchical attention mechanism can determine the weight ratio of each feature, some key information is selected for processing to improve the accuracy and efficiency of the training model.

[0009] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0010] The first aspect of this invention provides a method for diagnosing rolling bearing faults.

[0011] A method for diagnosing rolling bearing faults, comprising:

[0012] Based on different sampling frequencies of the vibration signal, a multi-scale convolutional neural network is used to extract multi-scale features from the vibration signal to be diagnosed.

[0013] Based on the hierarchical attention mechanism, weights are adaptively assigned to multi-scale features to obtain weighted features;

[0014] Temporal features of weighted features are extracted using a long short-term memory network;

[0015] The temporal features are input into the classification layer to obtain the fault diagnosis results.

[0016] Furthermore, the multi-scale convolutional neural network employs multiple one-dimensional convolutional neural networks with different kernel sizes and pooling window sizes to extract features at different scales, thereby obtaining multi-scale features.

[0017] Furthermore, the one-dimensional convolutional neural network consists of a convolutional layer, a dropout layer, a pooling layer, and a flattening layer connected in sequence.

[0018] Furthermore, it also includes: after obtaining multi-scale features, connecting the features at different scales to obtain a multi-scale feature matrix.

[0019] Furthermore, the adaptive weight allocation for multi-scale features specifically involves:

[0020] The multi-scale features are processed using tanh to obtain the adaptive weight coefficient matrix;

[0021] The weighted features are obtained by multiplying the multi-scale features with the weight coefficients at the corresponding positions in the adaptive weight coefficient matrix.

[0022] Furthermore, the Long Short-Term Memory Network consists of multiple units, and the data transmission between different units is controlled by three thresholds: the input gate, the forget gate, and the output gate.

[0023] The input gate controls the input of current information. When information passes through the input unit, it is multiplied by the input gate to determine whether to write the current information.

[0024] The forget gate determines whether to reset the previous information and multiplies it with the previous memory information to determine whether the information is retained;

[0025] The output gate controls the output of the currently memorized information, and multiplies the current memorized information to determine whether to output the information.

[0026] Furthermore, the classification layer includes a fully connected layer and a softmax function; finally, the softmax function converts the output of the fully connected layer into a probability distribution of the rolling bearing health status, thus obtaining a label for the rolling bearing failure type.

[0027] A second aspect of the present invention provides a rolling bearing fault diagnosis system.

[0028] A rolling bearing fault diagnosis system includes a feature extraction module, a feature weighting module, a time series extraction module, and a fault classification module.

[0029] The feature extraction module is configured to extract multi-scale features from the vibration signal to be diagnosed using a multi-scale convolutional neural network based on different sampling frequencies of the vibration signal.

[0030] The feature weighting module is configured to: adaptively assign weights to multi-scale features based on a hierarchical attention mechanism to obtain weighted features;

[0031] The temporal extraction module is configured to extract time-series features of weighted features through a long short-term memory network;

[0032] The fault classification module is configured to input time-series features into the classification layer to obtain fault diagnosis results.

[0033] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a rolling bearing fault diagnosis method as described in the first aspect of the present invention.

[0034] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of a rolling bearing fault diagnosis method as described in the first aspect of the present invention.

[0035] The above one or more technical solutions have the following beneficial effects:

[0036] This invention employs a multi-scale CNN network to extract features at different scales from the original vibration signal. It uses a hierarchical attention mechanism to adaptively assign weights to the feature information in order to extract key features. Then, it uses an LSTM layer to further extract the temporal features of the input signal. Finally, the softmax function of the fully connected layer is used to achieve fault classification.

[0037] This invention is based on a multi-scale CNN-LSTM network to learn features at different time scales and accurately identify fault features. Considering that the hierarchical attention mechanism can determine the weight ratio of each feature, some key information is selected for processing to improve the accuracy and efficiency of training the model.

[0038] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0040] Figure 1 This is a flowchart of the method in the first embodiment.

[0041] Figure 2 This is a structural diagram of the LSTM network in the first embodiment.

[0042] Figure 3 The image shows the original vibration acceleration signal of the 0hp load bearing in the first embodiment.

[0043] Figure 4 The image shows the original vibration acceleration signal of the 3hp load bearing in the first embodiment.

[0044] Figure 5 The training curves show the model accuracy and loss function for the first embodiment with 0hp and 3hp loads.

[0045] Figure 6 The fault classification confusion matrix diagram for the 0hp load and 3hp load models of the first embodiment.

[0046] Figure 7 This is a system structure diagram of the second embodiment. Detailed Implementation

[0047] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0048] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0049] Example 1

[0050] In one or more embodiments, a rolling bearing fault diagnosis method is disclosed. By extracting features from the vibration signal to be diagnosed, the health status of the rolling bearing can be detected without preprocessing. That is, the rolling bearing fault is diagnosed by two major steps: feature extraction and fault classification. Feature extraction is further divided into three steps: multi-scale feature extraction, feature weighting, and temporal feature extraction. Therefore, the method of this embodiment will be described according to the steps of feature extraction, feature weighting, temporal extraction, and fault classification.

[0051] A method for diagnosing rolling bearing faults, such as Figure 1 As shown, it includes the following steps:

[0052] Step S1: Based on the different sampling frequencies of the vibration signal, use a multi-scale convolutional neural network to extract multi-scale features from the vibration signal to be diagnosed.

[0053] Since the vibration signals of bearing failures have different frequencies, in order to extract features at each frequency more accurately, a multi-scale convolutional neural network uses three one-dimensional convolutional neural networks with different kernel sizes and pooling window sizes. The vibration signal to be diagnosed is simultaneously transmitted to the three convolutional networks to extract features at different scales.

[0054] Convolutional network 1 uses a 200*1 convolutional kernel and a 20*1 pooling kernel; convolutional network 2 uses a 100*1 convolutional kernel and a 10*1 pooling kernel; and convolutional network 3 uses a 50*1 convolutional kernel and a 5*1 pooling kernel. Dropout layers (with coefficients set to 0.5) and Flatten layers are used in all three convolutional networks to further optimize model performance. The Dropout layer randomly discards a certain percentage of neurons to avoid overfitting; the Flatten layer flattens multidimensional tensor data into a one-dimensional vector, facilitating processing by subsequent fully connected layers.

[0055] A Convolutional Neural Network (CNN) is a feedforward neural network with convolutional computation and a deep structure. It is widely used in image processing and natural language processing and consists of sequentially connected convolutional layers, dropout layers, pooling layers, and flattening layers.

[0056] Convolutional layers utilize filters with shared weights to extract features from the input signal, mapping local signals from the previous layer to the next. Shared weights offer significant advantages, greatly reducing the computational burden of complex nonlinear transformations. Since the vibration signal is a one-dimensional time series, this embodiment employs one-dimensional convolution, defined as follows:

[0057]

[0058] Where k and b are the weights and biases of the i-th and l-th layers, respectively, and x is the j-th local input of the l-th layer.

[0059] A Dropout layer is connected after the convolutional layer to avoid overfitting, followed by a pooling layer to select and filter the extracted features; max pooling is the most common pooling operation, which selects the largest statistic to obtain the local features of the region and reduces the number of parameters.

[0060] The result of the (l+1)th layer of the i-th channel after pooling is:

[0061]

[0062] Where q represents the t-th neuron in the l-th layer of the i-th channel, and w is the width of the pooling nucleus.

[0063] After obtaining the multi-scale features, the features at different scales need to be input into the Concatenate layer for merging. The Concatenate layer connects the three output features along the last dimension (feature dimension) to obtain the merged output features. This output feature matrix contains the features extracted by all three convolutional layers, ensuring that the model can extract features at different scales and improve classification performance.

[0064] Step S2: Based on the hierarchical attention mechanism, adaptively assign weights to multi-scale features to obtain weighted features.

[0065] Considering that the hierarchical attention mechanism can determine the weight ratio of each feature, this embodiment adds a hierarchical attention layer to weight the features, selects some key features for processing, and improves the accuracy and efficiency of training the model.

[0066] The hierarchical attention mechanism was first applied to natural language processing technology. It can give weighted attention to different levels of input and output in order to capture the rich structural and semantic information of text. By giving weighted attention to different levels, such as words, sentences, and paragraphs, this technology enables the model to understand the semantic and structural information of text more accurately.

[0067] This embodiment designs a similar hierarchical attention network, that is, adding a hierarchical attention layer between the original CNN and LSTM networks. The specific operation of the hierarchical attention layer is as follows:

[0068] (1) The multi-scale features output by the multi-scale convolutional neural network are copied into two copies. One copy is used to obtain the adaptive weight coefficient matrix, and the other copy directly retains the features learned by the CNN.

[0069] (2) Perform tanh processing on the multi-scale features to obtain the adaptive weight coefficient matrix.

[0070] Multi-scale features are mapped to a set of weight vectors that reflect the importance of each feature in the feature map.

[0071] (3) Multiply the multiscale features with the corresponding weight coefficients in the adaptive weight coefficient matrix (processed through the Multiply layer) to obtain weighted features. The features multiplied with high weight coefficients are effective features, while the features multiplied with low weight coefficients are inefficient or ineffective features. Effective features usually have a greater impact on the classification results, while inefficient or ineffective features have a smaller impact on the classification results. Weighted features can enhance effective features and suppress the importance of inefficient or ineffective features.

[0072] Layered attention, expressed by the formula:

[0073] X tan =tanh(Y) (3)

[0074] X = Y·X tan (4)

[0075] Here, Y is the output of the three-scale CNN network, i.e., the multi-scale feature matrix, and tanh represents the hyperbolic tangent activation function. After passing Y through the hyperbolic tangent activation function tanh, the adaptive weight coefficient matrix X is obtained. tan Multiply by the matrix Y output by the CNN network to obtain the weighted feature X.

[0076] Layered attention assigns different weights to each output vector of a CNN, enabling the model to focus its attention on key features and reduce the influence of other features. The use of layered attention mechanism can make the results expressed by CNN more reasonable in the current task.

[0077] Step S3: Extract the temporal features of the weighted features through a long short-term memory network.

[0078] After multi-scale convolutional neural networks and hierarchical attention layers, weighted features are sent to a long short-term memory (LSTM) network to further learn the temporal characteristics of the extracted features, thus obtaining temporal features.

[0079] Long Short-Term Memory (LSTM) is an improvement on traditional Recurrent Neural Networks (RNNs). Both are used to process sequential information. However, RNNs suffer from vanishing gradients during backpropagation, making them ineffective at capturing long-term dependencies in data. Therefore, to address the vanishing and exploding gradient problems during training with long sequences, a special type of RNN (LSTM) was proposed. LSTM outperforms ordinary RNNs on longer sequences. Its structure is as follows: Figure 2 As shown.

[0080] LSTM networks introduce state values ​​and "gate" control structures on top of recurrent neural networks. To selectively filter information, data transmission between different units in the hidden layer is controlled by three thresholds: the input gate, the forget gate, and the output gate. The forget gate determines whether to reset previous information and multiplies it with previous memory information to decide whether to retain the information. The input gate controls the input of current information; when information passes through the input unit, it is multiplied by the input gate to determine whether to write it into the current information. The output gate controls the output of currently remembered information; it is multiplied by the current memory information to determine whether to output the information.

[0081] The specific definition of LSTM is as follows:

[0082] First, information forgetting is processed, which is done through an S-shaped unit called the forgetting gate, as shown in formula (5):

[0083] f t =σ(W f [h t-1 ,x t ]+b f (5)

[0084] Among them, h t-1 and x t Used to determine which information should be updated through the operation of the input gate.

[0085] After this, the LSTM network adds new information through the input gate; h t-1 and x t New candidate cell location information is obtained through a tanh layer. The update process is shown in formulas (6) and (7):

[0086] it =σ(W i [h t-1 ,x t ]+b i (6)

[0087]

[0088] Update cell location information C t-1 To the new location information C t It can be represented as:

[0089]

[0090] The output gate determines the output characteristics of the final unit, as shown in formulas (7) and (8):

[0091] o t =σ(W o [h t-1 ,x t ]+b o (9)

[0092] h t =o t ·tanh(C t (10)

[0093] In the above formulas, i t f t o t These are the input gate, forget gate, and output gate; σ is the sigmoid activation function; W i W f W o W c b corresponds to the weight coefficients of neurons in different gate hidden layers. i b f b o b c These correspond to the deviation coefficients of different gates; C t These represent the short-term and long-term states, respectively; tanh is the hyperbolic tangent activation function; x and h are the current input and output information, respectively.

[0094] Step S4: Input the temporal features into the classification layer to obtain the fault diagnosis results.

[0095] The classification layer consists of a fully connected layer and a softmax function. The fully connected layer connects each node to all nodes in the previous layer to combine the extracted features. In other words, the fully connected layer maps the input features to fault scores. The softmax function converts the fault scores output by the fully connected layer into a probability distribution of the rolling bearing health status, thus obtaining a label for the rolling bearing fault type and diagnosing the health status of the rolling bearing.

[0096] Experimental verification and result analysis

[0097] The simulation environment in this embodiment is Windows 10, using the Tensorflow 2.1 learning framework, and the computer configuration is Intel(R) Core(TM) i7-8700 CPU, NVIDIA GeForce GTX 1660 GPU, and 16GB of memory.

[0098] Dataset Introduction

[0099] The data in this embodiment comes from the rolling bearing test bench at Case Western Reserve University. The CWRU bearing dataset provided by the Case Western Reserve University Bearing Data Center is widely used in bearing fault diagnosis research. It mainly consists of a motor, torque sensor, power meter and electronic control equipment. The motor is set to run at speeds of 1797, 1772, 1750 and 1730 rpm, respectively, with each speed corresponding to loads of 0, 1, 2 and 3 HP. Bearing wear faults of 7, 14, 21 and 28 mils are manually set, and the sampling frequencies of the vibration signals are 12K and 48K.

[0100] This model uses drive-end signals with a sampling frequency of 48KHz, a bearing speed of 1797rpm, and bearing loads of 0HP and 3HP. Bearing conditions include normal, inner raceway fault, outer raceway fault, and ball fault, with wear levels of normal (7mils), light (14mils), moderate (14mils), and severe (21mils), respectively. These can be further categorized into 10 bearing conditions: normal (N), light rolling element fault (7_BA), light inner ring fault (7_IR), light outer ring fault (7_OR), moderate rolling element fault (14_BA), moderate inner ring fault (14_IR), moderate outer ring fault (14_OR), severe rolling element fault (21_BA), severe inner ring fault (21_IR), and severe outer ring fault (21_OR).

[0101] To better prepare and evaluate the method, the experimental dataset contained no outliers, and the data for each sample conformed to normal conditions, requiring no additional preprocessing. A total of 487,963 samples were used, with the training and test sets divided at 70% and 30% respectively. The original vibration signals of 10 bearing states under 0hp and 3hp loads are shown below. Figure 3 and Figure 4 As shown.

[0102] Experimental parameter settings

[0103] The experiment employed a mini-batch training method, setting the batch size to 100 and the training iterations to 10. The Adam algorithm learning rate was set to 0.001. To improve the model's training accuracy, each dataset was trained 10 times, and the average value was used as the final result. The parameters of each layer of this model are shown in Table 1.

[0104] Table 1. Parameters of each layer in the multi-scale CNN-LSTM network based on hierarchical attention mechanism.

[0105]

[0106]

[0107] Experimental Results Analysis

[0108] The model achieved an average accuracy of 99.96% during training with a 0hp load, and an average F1 score and average recall of 99.82% and 99.82% respectively during validation on the test set. During training with a 3hp load, the model achieved an average accuracy of 99.91%, and an average F1 score and average recall of 99.69% and 99.69% respectively during validation on the test set. The training curves for the model's accuracy and loss function under 0hp and 3hp loads are shown below. Figure 5 As shown, the model fault classification confusion matrix for 0hp load and 3hp load is as follows: Figure 6 As shown.

[0109] Table 2 shows the results of comparing the classification performance of the proposed method with that of CNN, CNN-LSTN, and CNN-Attention-LSTM. The evaluation metrics are fault detection accuracy, recall, and F1 score. As can be seen from Table 2, the average accuracy, average F1 score, and average recall of the proposed method are all higher than those of the original CNN, CNN-LSTN, and CNN-Attention-LSTM methods, which proves that the proposed method has better fault classification ability and can better adapt to various working conditions.

[0110] Table 2 Comparison of Fault Diagnosis Performance of Four Methods

[0111]

[0112] Based on the above experimental results, this embodiment proposes a novel method for rolling bearing fault diagnosis. This method does not require complex data preprocessing; it only requires adding fault labels to the original vibration signals before directly inputting them into the network for learning. It improves upon the original CNN-LSTM method by replacing the CNN network for feature extraction with a multi-scale convolutional neural network, thus enhancing feature extraction capabilities. Furthermore, it adds a hierarchical attention mechanism layer to adaptively assign weights to the features of each sample learned by the neural network. These improvements enhance the model's classification performance, achieving an average accuracy of 99.94%, and an evaluation index of 99.76% using F1 score and recall. The multi-scale CNN-LSTM network based on the hierarchical attention mechanism has the advantages of fast convergence speed during training and high recognition accuracy, showing great potential in the field of motor bearing fault diagnosis.

[0113] Example 2

[0114] In one or more embodiments, a rolling bearing fault diagnosis system is disclosed, such as Figure 7 As shown, it includes a feature extraction module, a feature weighting module, a time series extraction module, and a fault classification module:

[0115] The feature extraction module is configured to extract multi-scale features from the vibration signal to be diagnosed using a multi-scale convolutional neural network based on different sampling frequencies of the vibration signal.

[0116] The feature weighting module is configured to: adaptively assign weights to multi-scale features based on a hierarchical attention mechanism to obtain weighted features;

[0117] The temporal extraction module is configured to extract time-series features of weighted features through a long short-term memory network;

[0118] The fault classification module is configured to input time-series features into the classification layer to obtain fault diagnosis results.

[0119] Example 3

[0120] The purpose of this embodiment is to provide a computer-readable storage medium.

[0121] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a rolling bearing fault diagnosis method as described in Embodiment 1 of this disclosure.

[0122] Example 4

[0123] The purpose of this embodiment is to provide an electronic device.

[0124] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in a rolling bearing fault diagnosis method as described in Embodiment 1 of this disclosure.

[0125] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for diagnosing rolling bearing faults, characterized in that, include: Based on different sampling frequencies of the vibration signal, a multi-scale convolutional neural network is used to extract multi-scale features from the vibration signal to be diagnosed. The multi-scale convolutional neural network employs multiple one-dimensional convolutional neural networks with different kernel sizes and pooling window sizes to extract features at different scales, thus obtaining multi-scale features. The one-dimensional convolutional neural network consists of a convolutional layer, a dropout layer, a pooling layer, and a flattening layer connected in sequence. After obtaining the multi-scale features, the features at different scales are connected to obtain a multi-scale feature matrix. Based on the hierarchical attention mechanism, weights are adaptively assigned to multi-scale features to obtain weighted features. Specifically, the multi-scale features are processed by the hyperbolic tangent activation function to obtain an adaptive weight coefficient matrix. The multi-scale features are multiplied by the weight coefficients at the corresponding positions in the adaptive weight coefficient matrix to obtain the weighted features. Temporal features of weighted features are extracted using a long short-term memory network; The temporal features are input into the classification layer to obtain the fault diagnosis results.

2. The rolling bearing fault diagnosis method as described in claim 1, characterized in that, The Long Short-Term Memory Network consists of multiple units, and the data transmission between different units is controlled by three thresholds: the input gate, the forget gate, and the output gate. The input gate controls the input of current information. When information passes through the input unit, it is multiplied by the input gate to determine whether to write the current information. The forget gate determines whether to reset the previous information and multiplies it with the previous memory information to determine whether the information is retained; The output gate controls the output of the currently memorized information, and multiplies the current memorized information to determine whether to output the information.

3. The rolling bearing fault diagnosis method as described in claim 1, characterized in that, The classification layer includes a fully connected layer and a softmax function; finally, the softmax function converts the output of the fully connected layer into a probability distribution of the rolling bearing health status, thus obtaining a label for the rolling bearing failure type.

4. A rolling bearing fault diagnosis system, characterized in that, It includes a feature extraction module, a feature weighting module, a time series extraction module, and a fault classification module: The feature extraction module is configured to: extract multi-scale features from the vibration signal to be diagnosed using a multi-scale convolutional neural network based on different sampling frequencies of the vibration signal; the multi-scale convolutional neural network employs multiple one-dimensional convolutional neural networks with different kernel sizes and pooling window sizes to extract features at different scales, thereby obtaining multi-scale features; wherein, the one-dimensional convolutional neural network consists of a convolutional layer, a dropout layer, a pooling layer, and a flattening layer connected in sequence; after obtaining the multi-scale features, the features at different scales are connected to obtain a multi-scale feature matrix; The feature weighting module is configured to: adaptively assign weights to multi-scale features based on a hierarchical attention mechanism to obtain weighted features; specifically: process the multi-scale features with a hyperbolic tangent activation function to obtain an adaptive weight coefficient matrix; multiply the multi-scale features with the weight coefficients at the corresponding positions in the adaptive weight coefficient matrix to obtain the weighted features; The temporal extraction module is configured to extract time-series features of weighted features through a long short-term memory network; The fault classification module is configured to input time-series features into the classification layer to obtain fault diagnosis results.

5. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-3.

6. A storage medium characterized in that it non-transitory stores computer-readable instructions, wherein, When the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-3 are executed.

Images