A Fault Diagnosis Method for Mine Ventilation Fan Bearings Based on CEEMDAN and Multi-Scale Spatiotemporal Information Fusion Graph Neural Network

Abstract

A fault diagnosis method for mine ventilation fan bearings based on CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network includes: acquiring vibration signals of the mine ventilation fan bearings; preprocessing the vibration signals; applying CEEMDAN to the preprocessed vibration signals, decomposing the signals into multiple IMF components through adaptive multi-scale decomposition; constructing a multi-scale spatiotemporal convolution module, using convolution kernels of different scales to extract the temporal features of each IMF component in parallel; introducing a channel attention mechanism to adaptively weight the multi-scale spatiotemporal features, dynamically strengthening the representation strength of key fault modes by calculating the importance weights of channel features; aggregating the multi-scale spatiotemporal features through GCN to generate a global feature representation, which is then input into a classifier to output the fault diagnosis result. This invention overcomes the technical bottlenecks of noise masking, lack of spatiotemporal correlation, and insufficient generalization ability by combining CEEMDAN with a multi-scale spatiotemporal information fusion graph neural network.

Description

Technical Field

[0001] This invention relates to the field of bearing fault diagnosis technology, specifically to a method for diagnosing bearing faults in mine ventilation fans based on a graph neural network that fuses CEEMDAN with multi-scale spatiotemporal information. Background Technology

[0002] As the core equipment of underground ventilation systems in coal mines, the operational reliability of mine ventilation fans directly affects the safety of mine production and the lives of workers. Bearings, as key transmission components of ventilation fans, endure high-frequency vibration, high temperature and high pressure, and mechanical stress under complex operating conditions for extended periods, making them highly susceptible to failure due to fatigue wear, lubrication failure, or material defects. Statistics show that bearing failures account for more than 35% of the total failure rate of ventilation fans. Furthermore, early fault characteristics are weak and easily masked by background noise, making it difficult for traditional detection methods to achieve high-precision diagnosis, severely restricting the level of intelligent operation and maintenance in coal mines.

[0003] Currently, vibration signal analysis is the main technical means for bearing fault diagnosis, but traditional methods have significant limitations: on the one hand, the strong noise and complex multi-frequency coupled fault characteristics of vibration signals in mining environments make it difficult for conventional time-frequency analysis methods to effectively separate key information; on the other hand, there is a strong correlation between the spatial topology of bearing components and the time series characteristics of vibration signals, but traditional methods only focus on single-dimensional information and cannot establish a complete fault evolution model. For example: 1) Wavelet transform relies on manual selection of basis functions and decomposition levels, which is not adaptable to non-stationary signals, and high-frequency noise can easily lead to misjudgment of modulus maxima; 2) Although Empirical Mode Decomposition (EMD) can decompose signals, it suffers from mode aliasing, and the decomposed Intrinsic Mode Function (IMF) is difficult to accurately correspond to fault characteristic frequencies, especially in strong noise environments where the decomposition accuracy drops significantly; 3) Existing deep learning models, such as convolutional neural networks, usually use fixed-scale convolution kernels, which cannot adaptively extract multi-scale features and do not adequately model spatial correlation information, resulting in poor generalization ability of the model in noisy environments.

[0004] Furthermore, traditional methods often employ a single-stage process of "feature extraction-classification," failing to fully explore the spatiotemporal correlation of vibration signals: 1) In the time dimension, the temporal characteristics of bearing faults are easily masked by noise, and conventional time-domain statistical indicators cannot characterize the nonlinear fault evolution process; 2) In the spatial dimension, the dynamic contact relationship between the vibration transmission path and the bearing assembly is not effectively modeled, and sensor data are often treated as independent sequences, resulting in the loss of spatial correlation information; 3) In terms of scale effects, fault features exhibit multi-scale distribution in the time-frequency domain, and single-scale analysis is difficult to cover information across the entire frequency band, leading to a high rate of missed diagnoses.

[0005] In recent years, scholars have attempted to improve diagnostic performance by combining multi-scale analysis with deep learning, but there are technical bottlenecks: 1) Multi-scale feature fusion and multi-dependency cascaded multi-layer convolutional networks lack a dynamic weight allocation mechanism for features at different scales, resulting in insufficient representation strength of key fault modes; 2) Some studies have introduced graph convolutional networks (GCNs) to model spatial topology, but have not deeply coupled spatiotemporal information, and spatial correlation remains only at the static graph structure level, failing to reflect the dynamic vibration transmission process; 3) The contradiction between noise resistance and feature integrity remains unresolved. Although existing attention mechanisms can enhance channel feature weights, they are not combined with multi-scale decomposition techniques, making it difficult to retain weak fault features during noise suppression; 4) The model's generalization ability is limited by the dependence on labeled data. The complex operating conditions of mining equipment and the sparse fault samples lead to a significant performance decline across equipment or operating conditions.

[0006] Therefore, there is an urgent need for a diagnostic technology that can overcome the limitations of traditional methods and achieve adaptive extraction of multi-scale features, deep coupling of spatiotemporal correlation information, and improved model generalization ability in noisy environments. Summary of the Invention

[0007] To address the shortcomings of traditional methods, such as insufficient feature extraction and low spatiotemporal information correlation in complex noisy environments, this invention provides a fault diagnosis method for bearings of mine ventilation fans based on CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network.

[0008] By combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) with a multi-scale spatiotemporal information fusion graph neural network, a phased, multi-level intelligent diagnostic framework is proposed, providing a high-precision and robust intelligent solution for fault diagnosis of mine ventilation fan bearings. This method overcomes the technical bottlenecks of noise masking, lack of spatiotemporal correlation, and insufficient generalization ability.

[0009] The technical solution adopted in this invention is as follows:

[0010] A fault diagnosis method for bearings of mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network includes the following steps:

[0011] Step 1: Collect vibration signals from the bearings of the mine ventilation fan;

[0012] Step 2: Preprocess the vibration signal of the mine ventilation fan bearing collected in Step 1, including outlier handling and data normalization;

[0013] Step 3: Perform complete ensemble empirical mode decomposition (CEEMDAN) on the vibration signal of the mine ventilation fan bearing after preprocessing in Step 2. The signal is decomposed into multiple IMF components through adaptive multi-scale decomposition to separate strong noise interference and enhance the identifiability of weak fault features.

[0014] Step 4: Construct a multi-scale spatiotemporal convolution module, using convolution kernels of different scales to extract the temporal features of each IMF component in parallel, covering the multi-frequency distribution of fault features;

[0015] Step 5: Introduce a channel attention mechanism to adaptively weight multi-scale spatiotemporal features, and dynamically enhance the representation strength of key fault modes by calculating the importance weight of channel features, thereby suppressing noise interference;

[0016] Step 6: Aggregate multi-scale spatiotemporal features using GCN to generate a global feature representation, and input it into the classifier to output the fault diagnosis result.

[0017] In step 2, outliers are first processed using the local mean method, and the calculation formula is shown in equation (1):

[0018] (1);

[0019] In formula (1): For missing data values; The number of existing data values ​​surrounding the missing data; Indicates the index where the current missing value is located; Indicates index The corresponding value; This indicates the index corresponding to the data.

[0020] The data normalization method uses Min-Max normalization, and the calculation is shown in equation (2):

[0021] (2);

[0022] In formula (2): This is the current data value; and They are respectively The minimum and maximum values ​​in the vector; This is the normalized value.

[0023] In step 3, CEEMDAN is an improved signal decomposition algorithm based on EMD, which can effectively eliminate white noise residue in IMF components. The decomposition steps are as follows:

[0024] S3.1: Determine to add Gaussian white noise Number of times and standard deviation , Add it to the original signal New signals were obtained Then, EMD decomposition is performed, and the average value of the first-order intrinsic mode functions is taken to obtain the components. and removal Residual after ,in The calculation formula is as follows:

[0025] (3);

[0026] In formula (3): This indicates the number of times Gaussian white noise is added. Indicates the first The first-order IMF components are obtained from the decomposition of the integrated system.

[0027] S3.2: In Add white noise Receive new signals EMD decomposition was performed, and the components were obtained by averaging the first-order intrinsic mode functions. and removal Residual after ,in The calculation formula is shown below:

[0028] (4);

[0029] In equation (4): Indicates the first The second-order IMF component is obtained from the decomposition of the integrated system.

[0030] S3.3: Continue repeating S3.2 until a monotonic residual signal appears that is no longer suitable for decomposition. At this point, the residual signal obtained from the original signal is... IMF components and residual components ,in: For residual components, Indicates the first Each component.

[0031] In step 4, convolutional kernels of different scales are used to extract the temporal features of each IMF component obtained in step 3 in parallel. The structure diagram of the multi-scale convolutional module is shown below. Figure 2 As shown. The core operation of a convolutional neural network is the local receptive field mechanism provided by the convolution operation, and its convolution operation is shown in the following formula:

[0032] (5);

[0033] In equation (5), Indicates by the first Features extracted by each convolutional kernel; Indicates the corresponding weight; Indicates the bias term; Indicates one-dimensional input; express Enter the number of data points.

[0034] In equation (5), , It is a positive integer.

[0035] After the convolution operation, a nonlinear transformation is performed using the ReLU activation function, expressed as follows:

[0036] (6);

[0037] In formula (6) Indicates by the first Features extracted by each convolutional kernel; , It is a positive integer.

[0038] The method used in the pooling layer is max pooling, which is defined as follows:

[0039] (7);

[0040] In equation (7), Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling layer output value; Represents the pooling region Central Elements at the location; Represents the coordinates of a vector within the pooling region. Represents row coordinates, Represents column coordinates.

[0041] Dropout is introduced to avoid overfitting. By configuring an appropriate ratio to exclude hidden neurons in the convolutional neural network, the robustness and generalization performance of the model are improved. Finally, the feature matrix is ​​obtained:

[0042] (8);

[0043] In equation (8), Represents the characteristic matrix; Represents the eigenvector; The number of channels is the number of features extracted by convolutional kernels of different scales.

[0044] In step 4, the multi-scale spatiotemporal convolution module specifically includes: the technical solution in S3.3, and the contents recorded in equations (5) to (8), which are the relevant formulas of the multi-scale spatiotemporal convolution module;

[0045] In step 4, the temporal features of each IMF component are extracted. The IMF components here are those obtained in step 3. The temporal features of each IMF component are described in S3.3.

[0046] Step 5 includes the following steps:

[0047] S5.1: Obtain the size of the convolution kernels with different scales in step 4 respectively. The input feature maps are subjected to global max pooling and global average pooling in the spatial dimension, resulting in two values ​​for each feature map. The feature map, where, , and These represent the height, width, and number of channels of the input feature map, respectively.

[0048] S5.2: The results of global max pooling and global average pooling are fed into a shared multilayer perceptron for learning, resulting in two... Feature map; the number of neurons in the first layer of the MLP is , This indicates the preset reduction ratio used to control the degree of compression of the multilayer perceptron.

[0049] The activation function is ReLU, and the calculation formula is shown below:

[0050] (9);

[0051] In equation (9), This represents the output of the ReLU activation function; Indicates input; Indicates if input If it is greater than 0, then the output is If input If the value is less than or equal to 0, the output will be 0.

[0052] The number of neurons in the second layer of the MLP is Reduction ratio This is a preset integer used to control the compression level of the multilayer perceptron.

[0053] S5.3: The results of the MLP output are summed, and then processed by the Sigmoid activation function. The calculation formula is as follows:

[0054] (10);

[0055] In equation (10), This represents the output of the Sigmoid function. It is the input of the neuron. It is the base of the natural logarithm.

[0056] The output range of the Sigmoid function is: , for any real number Mapping to the probability interval yields the channel attention weight matrix. To represent the importance weight of channel features, where, Indicates the first The weight of each channel.

[0057] S5.4: Apply the attention weight matrix to the multi-scale convolution output obtained in step 3 to obtain a weighted feature matrix, which dynamically enhances the representation strength of key fault modes.

[0058] (11);

[0059] In equation (11), This represents the weighted eigenma matrix. Channel attention weight matrix, The characteristic matrix obtained by expression (8) Indicates multiplication.

[0060] In step 5, the representation strength of key failure modes is dynamically enhanced by calculating the importance weights of channel features. Specifically, this involves: S5.3 calculating the channel attention weight matrix. This represents the importance of channel features;

[0061] S5.4 The representation strength of critical fault modes is dynamically enhanced by multiplying the weight matrix with the multi-scale convolution output obtained in step 3.

[0062] In step 6, the GCN employs a two-layer graph convolutional structure. The first CNN layer is used for basic feature aggregation, and the second CNN layer is used to capture the global relationships between multi-scale features. Figure 3 As shown;

[0063] Each node is represented by spatiotemporal features extracted under convolutional kernels of different scales, and edges are constructed based on the similarity between features. Figure 3In the diagram, X1, X2, X3, and X4 represent the spatiotemporal features extracted from the original data by the first convolutional layer, and the solid black lines represent the connections between these features. Z1, Z2, Z3, and Z4 are the outputs of the first GCN layer, corresponding to X1, X2, X3, and X4 respectively. The Z vector contains richer structural information and connections than the X vector. Y1 and Y4 represent the final outputs.

[0064] Similarity is measured using Euclidean distance, and the calculation formula is as follows:

[0065] (12);

[0066] In equation (12), This represents the Euclidean distance between two points. , These represent the x-coordinates of the first and second points, respectively. , These represent the y-coordinates of the first and second points, respectively. , These represent the z-axis coordinates of the first and second points, respectively.

[0067] The connection relationships between feature nodes are an adjacency matrix. The calculation formula is as follows:

[0068] (13);

[0069] Based on this, GCN is used for feature aggregation. The basic propagation formula of GCN is as follows:

[0070] (14);

[0071] In equation (14): Indicates the first Layer feature representation; Indicates the first Layer feature representation; It is an adjacency matrix. ; It is a degree matrix; This is the weight matrix that GCN needs to learn; It is an activation function;

[0072] The initial input of GCN is the weighted feature matrix calculated by equation (11), which makes the features learned by GCN focus more on key channels;

[0073] The classifier uses a softmax layer:

[0074] (15);

[0075] In equation (15): Indicates the first The probability of each element. It is the first of the input vectors One element, It is the first The index of each input value, It is the sum of the indices of all elements.

[0076] The loss function used is the cross-entropy loss function, defined as follows:

[0077] (16);

[0078] In equation (16), This represents the average loss of the entire population. , Indicates the total number of samples; Indicates the number of all sample categories; Indicates the first One sample; Indicates the first The predicted probabilities of each category; Indicates the indicator variable; Indicates the first The sample belongs to the first The predicted probability of each category.

[0079] Finally, combining equation (15), the fault type corresponding to the sample is calculated using the following formula:

[0080] (17);

[0081] In equation (17), Indicates the output category predicted by the model; This indicates selecting the index with the highest probability, i.e., choosing... The category with the highest probability in the output is taken as the final classification result.

[0082] This invention discloses a fault diagnosis method for bearings of mine ventilation fans based on CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network. The technical effects are as follows:

[0083] 1) Feature enhancement and improved fault identifiability in high-noise environments. CEEMDAN is used to adaptively decompose the original vibration signal into multiple scales, effectively separating strong background noise from weak fault features. This significantly improves the identifiability of fault features.

[0084] 2) By combining multi-scale convolutional networks with GCN, spatiotemporal features are extracted in parallel using convolutional kernels of different scales, which solves the problem of key feature loss caused by single-scale analysis. Furthermore, the spatiotemporal features of GCN are aggregated to generate global feature representations, thereby improving the performance of fault diagnosis.

[0085] 3) By dynamically strengthening the weight of key fault channels through the channel attention mechanism, the interference of noise channels is suppressed, thereby improving the diagnostic accuracy of the model under changing operating conditions. Attached Figure Description

[0086] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0087] Figure 1 This is a flowchart of the bearing fault diagnosis process of the present invention.

[0088] Figure 2 This is a schematic diagram of the multi-scale convolution module structure of the present invention.

[0089] Figure 3 This is a schematic diagram of the GCN structure.

[0090] Figure 4 This is the confusion matrix for model classification based on the test set. Detailed Implementation

[0091] This invention proposes a fault diagnosis method for mine ventilation fan bearings based on CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network. By combining Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) with a multi-scale spatiotemporal information fusion graph neural network, a phased and multi-level intelligent diagnostic framework is proposed. This provides a high-precision and robust intelligent solution for fault diagnosis of mine ventilation fan bearings, and has significant engineering value for promoting the intelligent operation and maintenance of coal mine equipment.

[0092] First, CEEMDAN is used to adaptively decompose the vibration signal at multiple scales, effectively separating strong noise interference and enhancing the identifiability of fault features. Second, a multi-scale spatiotemporal convolutional module is constructed, extracting spatiotemporal features through parallel convolutional kernels and introducing a channel attention mechanism to achieve dynamic weight allocation, learning to cover fault information across the entire frequency band. Finally, a graph convolutional network is used to aggregate multi-scale spatiotemporal features, establishing a unified representation model of temporal vibration and spatial correlation. This method overcomes the technical bottlenecks of noise masking, lack of spatiotemporal correlation, and insufficient generalization ability. The flowchart is as follows: Figure 1 As shown.

[0093] Example:

[0094] Experimental data were collected under four operating conditions: 0HP, 1HP, 2HP, and 3HP, corresponding to speeds of 1797 r / min, 1772 r / min, 1750 r / min, and 1730 r / min, respectively. Vibration sensors were placed at the drive end and the fan end, with a frequency of 12 kHz at the fan end and 48 kHz at the drive end. Defects of 0.007 inches, 0.014 inches, and 0.021 inches in size were manufactured on the inner ring, outer ring, and rolling elements of the bearing using electrical discharge machining (EDM), generating nine types of fault data and one type of normal data.

[0095] The experiment used data collected at the drive end, with a sampling frequency of 48kHz, a workload of 3HP, and a rotational speed of 1730r / min. To improve the efficiency of model training, the original vibration signal was downsampled with a step size of 8. To verify the effectiveness of the model, the data was divided into training and test sets in an 8:2 ratio, as detailed in Table 1.

[0096] Table 1 Bearing Dataset

[0097]

[0098] As shown in Table 2, the CEEMDAN+ multi-scale spatiotemporal information fusion graph neural network in the method of this invention can correctly identify most bearing fault samples. All labeled... F All scores were greater than 0.95, with the scores for labels 0, 1, 2, 4, 6, 8, and 9 being particularly high. F A score greater than 0.98, with labels 1, 2, 4, 6, 8, and 9. F A score of 1.00 indicates that the model has a high level of ability to identify various types of faults.

[0099] Table 2 Evaluation metrics for the test set

[0100]

[0101] Figure 4 The confusion matrix between the model prediction and the actual value is one of the experimental results. The CEEMDAN+ multi-scale spatiotemporal information fusion graph neural network in the method of this invention can distinguish various fault modes very well.

[0102] In summary, this invention uses CEEMDAN to adaptively decompose the original vibration signal into multiple scales, effectively separating strong noise interference and enhancing the identifiability of weak fault features. Subsequently, a multi-scale convolutional network extracts temporal features through parallel convolutional kernels and combines a channel attention mechanism to dynamically weight and enhance key fault modes. Next, GCN multi-scale spatiotemporal features are aggregated to generate a global feature representation, improving fault diagnosis performance. Experimental results show that the model achieves 99.0% accuracy on the test set, meaning it can correctly identify most bearing fault samples, providing a highly reliable technical solution for intelligent operation and maintenance of coal mine equipment.

Claims

1. A fault diagnosis method for bearings of mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network, characterized in that... Includes the following steps: Step 1: Collect vibration signals from the bearings of the mine ventilation fan; Step 2: Preprocess the vibration signal of the mine ventilation fan bearing collected in Step 1, including outlier handling and data normalization; Step 3: Perform complete ensemble empirical mode decomposition (CEEMDAN) on the vibration signal of the mine ventilation fan bearing after preprocessing in Step 2, and decompose the signal into multiple IMF components through adaptive multi-scale decomposition. Step 4: Construct a multi-scale spatiotemporal convolution module and extract the temporal features of each IMF component in parallel using convolution kernels of different scales; Step 5: Introduce a channel attention mechanism to adaptively weight multi-scale spatiotemporal features, and dynamically enhance the representation strength of key failure modes by calculating the importance weight of channel features; Step 6: Aggregate multi-scale spatiotemporal features using GCN to generate a global feature representation, and input it into the classifier to output the fault diagnosis result; Step 5 includes the following steps: S5.1: Obtain the size of the convolution kernels with different scales in step 4 respectively. The input feature maps are subjected to global max pooling and global average pooling in the spatial dimension, resulting in two values ​​for each feature map. The feature map, where, , and These represent the height, width, and number of channels of the input feature map, respectively. S5.2: The results of global max pooling and global average pooling are fed into a shared multilayer perceptron for learning, resulting in two... Feature map; the number of neurons in the first layer of the MLP is , This indicates the preset reduction ratio used to control the degree of compression of the multilayer perceptron; The activation function is ReLU, and the calculation formula is shown below: （9）； In equation (9), This represents the output of the ReLU activation function; Indicates input; Indicates if input If it is greater than 0, then the output is If input If the value is less than or equal to 0, the output is 0; The number of neurons in the second layer of the MLP is Reduction ratio This is a preset integer used to control the compression level of the multilayer perceptron; S5.3: The results of the MLP output are summed, and then processed by the Sigmoid activation function. The calculation formula is as follows: （10）； In equation (10), This represents the output of the Sigmoid function. It is the input of the neuron. It is the base of the natural logarithm; The output range of the Sigmoid function is: , for any real number Mapping to the probability interval yields the channel attention weight matrix. To represent the importance weight of channel features, where, Indicates the first The weight of each channel; S5.4: Apply the attention weight matrix to the multi-scale convolution output obtained in step 3 to obtain a weighted feature matrix, which dynamically enhances the representation strength of key fault modes. （11）； In equation (11), This represents the weighted eigenma matrix. Channel attention weight matrix, The characteristic matrix obtained by expression (8) Indicates multiplication; In step 5, the representation strength of key failure modes is dynamically enhanced by calculating the importance weights of channel features, as detailed below: The channel attention weight matrix calculated in S5.3 This represents the importance of channel features; S5.4 By multiplying the weight matrix with the multi-scale convolution output obtained in step 3, the representation strength of key failure modes is dynamically enhanced. In step 6, the GCN employs a two-layer graph convolutional structure. The first CNN layer is used for basic feature aggregation, and the second CNN layer is used to capture the global relationships between multi-scale features. Each node is represented by the spatiotemporal features extracted by convolutional kernels at different scales, and edges are constructed based on the similarity between features. X1, X2, X3, and X4 represent the spatiotemporal features extracted from the original data by the first convolutional layer, and the black solid lines represent the connection relationships between these features. Z1, Z2, Z3, and Z4 are the outputs of the first GCN layer, and Z1, Z2, Z3, and Z4 correspond to X1, X2, X3, and X4, respectively. The Z vector contains richer structural information and connection relationships than the X vector. Y1 and Y4 represent the final outputs. Similarity is measured using Euclidean distance, and the calculation formula is as follows: （12）； In equation (12), This represents the Euclidean distance between two points. , These represent the x-coordinates of the first and second points, respectively. , These represent the y-coordinates of the first and second points, respectively. , These represent the z-axis coordinates of the first and second points, respectively. The connection relationships between feature nodes are an adjacency matrix. The calculation formula is as follows: （13）； Based on this, GCN is used for feature aggregation. The basic propagation formula of GCN is as follows: （14）； In equation (14): Indicates the first Layer feature representation; Indicates the first Layer feature representation; It is an adjacency matrix. ; It is a degree matrix; This is the weight matrix that GCN needs to learn; It is an activation function.

2. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 1, characterized in that: In step 2, outliers are first processed using the local mean method, and the calculation formula is shown in equation (1): （1）； In formula (1): For missing data values; The number of existing data values ​​surrounding the missing data; Indicates the index where the current missing value is located; Indicates index The corresponding value; This indicates the index corresponding to the data.

3. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 2, characterized in that: The data normalization method uses Min-Max normalization, and the calculation is shown in equation (2): （2）； In formula (2): This is the current data value; and They are respectively The minimum and maximum values ​​in the vector; This is the normalized value.

4. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 1, characterized in that: In step 3, CEEMDAN can effectively eliminate residual white noise in the IMF component. The decomposition steps are as follows: S3.1: Determine to add Gaussian white noise Number of times and standard deviation , Add it to the original signal New signals were obtained Then, EMD decomposition is performed, and the average value of the first-order intrinsic mode functions is taken to obtain the components. and removal Residual after ,in The calculation formula is as follows: （3）； In formula (3): This indicates the number of times Gaussian white noise is added. Indicates the first The first-order IMF components obtained from the integrated decomposition; S3.2: In Add white noise Receive new signals EMD decomposition was performed, and the components were obtained by averaging the first-order intrinsic mode functions. and removal Residual after ,in The calculation formula is shown below: （4）； In equation (4): Indicates the first The second-order IMF component obtained from the integrated decomposition; 3.3: Continue repeating S3.2 until a monotonic residual signal appears that is no longer suitable for decomposition. At this point, the residual signal obtained from the original signal is... IMF components and residual components ,in: For residual components, Indicates the first Each component.

5. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 4, characterized in that: In step 4, convolutional kernels of different scales are used to extract the temporal features of each IMF component obtained in step 3 in parallel. The convolution operation of the multi-scale convolution module is shown in the following formula: （5）； In equation (5), Indicates by the first Features extracted by each convolutional kernel; Indicates the corresponding weight; Indicates the bias term; This represents a one-dimensional input; express Enter the number of data points in the input field; , It is a positive integer; After the convolution operation, a nonlinear transformation is performed using the ReLU activation function, expressed as follows: （6）； In formula (6) Indicates by the first Features extracted by each convolutional kernel; , It is a positive integer; The method used in the pooling layer is max pooling, which is defined as follows: （7）； In equation (7), Indicates the relationship with the first Pooling regions related to each feature map The maximum pooling layer output value; Represents the pooling region Central Elements at the location; Represents the coordinates of a vector within the pooling region. Represents row coordinates, Represents column coordinates.

6. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 5, characterized in that: Dropout is introduced to avoid overfitting. By configuring an appropriate ratio to exclude hidden neurons in the convolutional neural network, the robustness and generalization performance of the model are improved, ultimately yielding the feature matrix: （8）； In equation (8), Represents the characteristic matrix; Represents the eigenvector; The number of channels is the number of features extracted by convolutional kernels of different scales.

7. The method for fault diagnosis of bearings in mine ventilation fans based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network as described in claim 1, characterized in that: The initial input of GCN is the weighted feature matrix calculated by equation (11), which makes the features learned by GCN focus more on key channels; The classifier uses a softmax layer: （15）； In equation (15): Indicates the first The probability of each element. It is the first of the input vectors One element, It is the first The index of each input value, It is the sum of the indices of all elements; The loss function used is the cross-entropy loss function, defined as follows: （16）； In equation (16), This represents the average loss of the entire population. , Indicates the total number of samples; Indicates the number of all sample categories; Indicates the first One sample; Indicates the first The predicted probabilities of each category; Indicates the indicator variable; Indicates the first The sample belongs to the first The predicted probabilities of each category; Finally, combining equation (15), the fault type corresponding to the sample is calculated using the following formula: （17）； In equation (17), Indicates the output category predicted by the model; This indicates selecting the index with the highest probability, i.e., choosing... The category with the highest probability in the output is taken as the final classification result.

Images