Mining ventilator bearing fault diagnosis method based on CEEMDAN and multi-scale spatio-temporal information fusion graph neural network

Abstract

The invention discloses a mining ventilator bearing fault diagnosis method based on a CEEMDAN and a multi-scale spatio-temporal information fusion map neural network. The method comprises the steps of collecting a mining ventilator bearing vibration signal; the vibration signals of the mining ventilator bearing are preprocessed; cEEMDAN is carried out on the preprocessed mining ventilator bearing vibration signal, and the signal is decomposed into a plurality of IMF components through adaptive multi-scale decomposition; constructing a multi-scale space-time convolution module, and adopting convolution kernels of different scales to extract time sequence characteristics of each IMF component in parallel; introducing a channel attention mechanism to perform adaptive weighting on the multi-scale spatial-temporal features, and dynamically strengthening the characterization intensity of the key fault mode by calculating the importance weight of the channel features; and aggregating the multi-scale spatial-temporal features through GCN to generate global feature representation, and inputting the global feature representation into a classifier to output a fault diagnosis result. Through combination of the CEEMDAN and the multi-scale spatio-temporal information fusion graph neural network, the technical bottlenecks of noise covering, spatio-temporal correlation deficiency and insufficient generalization ability are broken through.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing fault diagnosis, and in particular to a bearing fault diagnosis method for a mining ventilator based on a CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network. Background Art

[0002] Mine ventilators, as core equipment in underground coal mine ventilation systems, have a high degree of operational reliability that is directly related to mine safety and the safety of workers. Bearings, critical transmission components in ventilators, are subjected to high-frequency vibration, high temperatures, high pressures, and mechanical stresses under complex operating conditions. These bearings are highly susceptible to failure due to fatigue wear, lubrication failure, or material defects. According to statistics, bearing failures account for over 35% of the total ventilator failure rate. Early fault signatures are weak and easily masked by background noise, making it difficult to achieve high-precision diagnosis using traditional detection methods. This severely restricts the level of intelligent coal mine operation and maintenance.

[0003] Currently, vibration signal analysis is the primary technical approach for bearing fault diagnosis, but traditional methods have significant limitations. First, the strong noise in vibration signals and the complex multi-band coupled fault characteristics in mining environments make it difficult for conventional time-frequency analysis methods to effectively isolate key information. Second, the spatial topology of bearing components is strongly correlated with the time series characteristics of vibration signals. However, traditional methods focus solely on single-dimensional information, making it difficult to establish a complete fault evolution model. For example, 1) wavelet transforms rely on the manual selection of basis functions and the number of decomposition levels, making them less adaptable to non-stationary signals and prone to misjudging modulo maxima due to high-frequency noise. 2) While empirical mode decomposition (EMD) can decompose signals, it suffers from modal aliasing, making it difficult for the decomposed intrinsic mode functions (IMFs) to accurately correspond to fault characteristic frequencies. Decomposition accuracy decreases significantly, especially in strong noise environments. 3) Existing deep learning models, such as convolutional neural networks, typically use fixed-scale convolution kernels, are unable to adaptively extract multi-scale features, and lack the ability to model spatial correlation information, resulting in poor generalization in noisy environments.

[0004] In addition, traditional methods mostly use a single-stage "feature extraction-classification" process, which does not fully explore the spatiotemporal correlation of vibration signals: 1) In the time dimension, the timing 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 is often regarded as an independent sequence, resulting in the loss of spatial correlation information; 3) In terms of scale effect, fault characteristics present a multi-scale distribution in the time-frequency domain, and a single-scale analysis is difficult to cover the full frequency band information, resulting in a high diagnostic missed detection rate.

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

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

[0007] Aiming at the problems of insufficient feature extraction and low spatial-temporal information correlation of traditional methods in complex noise environments, the present invention provides a mining fan bearing fault diagnosis method based on CEEMDAN and multi-scale spatial-temporal 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, multi-level intelligent diagnosis framework was proposed. This approach provides a highly accurate and robust intelligent solution for fault diagnosis of mining fan bearings. This approach overcomes technical bottlenecks such as noise masking, lack of spatiotemporal correlation, and insufficient generalization.

[0008] The technical solution adopted by the present invention is: The bearing fault diagnosis method for mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network includes the following steps: Step 1: Collect the vibration signal of the mining fan bearing; Step 2: Preprocess the mining fan bearing vibration signal collected in step 1, including outlier processing and data normalization; Step 3: Perform complete ensemble empirical mode decomposition (CEEMDAN) on the mining fan bearing vibration signal preprocessed 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. Step 4: Construct a multi-scale spatiotemporal convolution module and use convolution kernels of different scales to extract the temporal characteristics of each IMF component in parallel, covering the multi-band distribution of fault characteristics; Step 5: Introduce a channel attention mechanism to adaptively weight multi-scale spatiotemporal features. By calculating the importance weights of channel features, the representation strength of key fault modes is dynamically enhanced to suppress noise interference. Step 6: Aggregate multi-scale spatiotemporal features through GCN to generate global feature representation, which is then input into the classifier to output the fault diagnosis result.

[0009] In step 2, outliers are first processed by using the local mean method. The calculation formula is shown in formula (1): (1); In formula (1): is the missing data value; is the number of existing data values ​​around the missing data; Indicates the index where the current missing value is located; Represents an index The corresponding value; Indicates the index corresponding to the data.

[0010] The data normalization method uses Min-Max normalization, and the calculation is shown in formula (2): (2); In formula (2): is the current data value; and They are The minimum and maximum values ​​in the vector; is the normalized value.

[0011] In step 3, CEEMDAN is an improved signal decomposition algorithm based on EMD, which can effectively eliminate the white noise residue in the IMF component. The decomposition steps are as follows: S3.1: Determine the addition of Gaussian white noise Number of times and standard deviation , , adding it to the original signal Get new signal , then perform EMD decomposition and take the average value of the first-order eigenmode function to obtain the component and removal The residual after ,in The calculation formula is as follows: (3); In formula (3): Indicates the number of times Gaussian white noise is added, Indicates in The first-order IMF components are obtained from the integrated decomposition.

[0012] S3.2: In Add white noise Get a new signal , perform EMD decomposition and average the first-order intrinsic mode function to obtain the component and removal The residual after ,in The calculation formula is as follows: (4); In formula (4): Indicates in The second-order IMF components are obtained from the integrated decomposition.

[0013] S3.3: Continue to repeat S3.2 until a monotonic residual signal that is no longer suitable for decomposition appears. At this time, the original signal is decomposed into IMF components and residual components ,in: is the residual component, Indicates the A portion.

[0014] In step 4, convolution kernels of different scales are used to extract the temporal characteristics of each IMF component obtained in step 3 in parallel. The structure of the multi-scale convolution module is shown in FIG. Figure 2 The core operation of the convolutional neural network is the local receptive field mechanism provided by the convolution operation. The convolution operation is shown in the following formula: (5); In formula (5), Indicates that Features extracted by convolution kernels; Indicates the corresponding weight; represents the bias term; Represents one-dimensional input; express Enter the number of data points in .

[0015] In formula (5), , Is a positive integer.

[0016] After the convolution operation, the nonlinear transformation is performed through the ReLU activation function, which is expressed as follows: (6); In formula (6) Indicates that Features extracted by convolution kernels; , Is a positive integer.

[0017] The method used in the pooling layer is the maximum pooling layer, and the maximum pooling definition is as follows: (7); In formula (7), Indicates The pooling area related to the feature map The maximum pooling layer output value; Represents the pooling area Located in the middle The elements at Represents the coordinates of a vector in the pooling area, represents the row coordinates, Represents column coordinates.

[0018] 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: (8); In formula (8), represents the feature matrix; represents the eigenvector; is the number of channels, that is, the number of features extracted by convolution kernels of different scales.

[0019] 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; In step 4, the time series characteristics of each IMF component are extracted. The IMF components here are obtained in step 3. The time series characteristics of each IMF component are described in S3.3.

[0020] The step 5 comprises the following steps: S5.1: The convolution kernels of different scales in step 4 are obtained with sizes of The input feature map is subjected to global maximum pooling and global average pooling in the spatial dimension, and each feature map obtains two The feature map of 、 and Represent the height, width and number of channels of the input feature map respectively.

[0021] S5.2: The results of global maximum pooling and global average pooling are fed into a shared multi-layer perceptron to learn and obtain two The feature map of MLP is , Indicates the preset reduction ratio used to control the compression degree of the multilayer perceptron.

[0022] The activation function is Relu, and the calculation formula is as follows: (9); In formula (9), Represents the output of the activation function ReLU; Represents input; If you enter If it is greater than 0, the output is By itself, if you enter If it is less than or equal to 0, the output is 0.

[0023] The number of neurons in the second layer of MLP is , reduction ratio It is a preset integer that controls the compression level of the multilayer perceptron.

[0024] S5.3: The MLP outputs are summed and then mapped through the Sigmoid activation function. The calculation formula is as follows: (10); In formula (10), Represents the output of the Sigmoid function, is the input of the neuron, is the base of natural logarithms.

[0025] The output range of the Sigmoid function is , any real number Mapped to the probability interval, the channel attention weight matrix is ​​obtained To express the importance weight of channel features, Indicates the The weight of each channel.

[0026] S5.4: Apply the attention weight matrix to the multi-scale convolution output obtained in step 3 to obtain a weighted feature matrix to dynamically enhance the representation strength of key failure modes: (11); In formula (11), represents the weighted feature matrix, Channel attention weight matrix, The characteristic matrix obtained by expression (8) is: Indicates multiplication.

[0027] In step 5, the importance weights of channel features are calculated to dynamically enhance the representation strength of key fault modes. Specifically, the channel attention weight matrix calculated in S5.3 is: That is, it represents the importance of channel features; S5.4 dynamically enhances the representation strength of critical failure modes by multiplying the weight matrix with the multi-scale convolution output obtained in step 3.

[0028] In step 6, a two-layer graph convolutional structure is used in GCN. The first layer CNN is used for basic feature aggregation, and the second layer CNN is used to deeply capture the global relationship between multi-scale features. Figure 3 As shown; Each node is represented by spatiotemporal features extracted under convolution kernels of different scales, and edges are constructed based on the similarity between features. Figure 3 In the figure, X1, X2, X3, and X4 represent the spatiotemporal features extracted from the raw data by the first convolutional layer, and the solid black lines indicate the connections between these features. Z1, Z2, Z3, and Z4 are the outputs of the first GCN layer. Z1, Z2, Z3, and Z4 correspond to X1, X2, X3, and X4, respectively. The Z vector contains richer structural information and connectivity than the X vector. Y1 and Y4 represent the final outputs.

[0029] The similarity is measured using Euclidean distance, and the calculation formula is as follows (12); In formula (12), represents the Euclidean distance between two points, 、 Represent the x-axis coordinates of the first and second points respectively, 、 Represent the y-axis coordinates of the first and second points respectively, 、 Represents the z-axis coordinates of the first point and the second point respectively.

[0030] The connection relationship between feature nodes is the 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 formula (14): Indicates the Feature representation of the layer; Indicates the Feature representation of the layer; is the adjacency matrix, ; is the degree matrix; It is the weight matrix that GCN needs to learn; is the activation function; The initial input of GCN is the weighted feature matrix calculated by formula (11), which makes the features learned by GCN focus more on key channels; The classifier uses the softmax layer: (15); In formula (15): Indicates the The probability of an element, is the first elements, It is The exponential of the input value, is the sum of all element exponents.

[0031] The loss function uses the cross entropy loss function, which is defined as follows: (16); In formula (16), represents the overall average loss, , represents the number of all samples; Indicates the number of all sample categories; Indicates the samples; Indicates the The predicted probability of each category; represents an indicator variable; Indicates the The samples belong to The predicted probability of each category.

[0032] Finally, combined with formula (15), the fault type corresponding to the sample is calculated by the following formula: (17); In formula (17), Represents the output category predicted by the model; Indicates taking the index with the highest probability, that is, choosing The category corresponding to the maximum probability in the output is taken as the final classification result.

[0033] The present invention provides a mining fan bearing fault diagnosis method based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network, and the technical effects are as follows: 1) Feature enhancement and improved fault identifiability in strong noise environments. CEEMDAN performs adaptive multi-scale decomposition on the original vibration signal, effectively separating strong background noise from weak fault features. This effectively improves the identifiability of fault features.

[0034] 2) Combining multi-scale convolutional networks with GCN, the spatiotemporal features are extracted in parallel through convolution kernels of different scales, which solves the problem of key feature loss caused by single-scale analysis. The multi-scale spatiotemporal features of GCN are aggregated to generate a global feature representation, thereby improving the performance of fault diagnosis.

[0035] 3) The channel attention mechanism dynamically strengthens the weight of key fault channels to suppress the interference of noise channels, thereby improving the diagnostic accuracy of the model under changing working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The present invention will be further described below with reference to the accompanying drawings and examples: Figure 1 This is a flow chart of bearing fault diagnosis according to the present invention.

[0037] Figure 2 This is a schematic diagram of the multi-scale convolution module structure of the present invention.

[0038] Figure 3 Schematic diagram of the GCN structure.

[0039] Figure 4 Classification confusion matrix for the model based on the test set. DETAILED DESCRIPTION

[0040] A mining fan bearing fault diagnosis method based on CEEMDAN and a multi-scale spatiotemporal information fusion graph neural network is proposed. 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 diagnosis framework is proposed. This provides a high-precision and robust intelligent solution for mining fan bearing fault diagnosis, which has important engineering value for promoting the intelligent operation and maintenance of coal mine equipment.

[0041] First, CEEMDAN is used to perform adaptive multi-scale decomposition of vibration signals, effectively separating strong noise interference and enhancing the identifiability of fault features. Secondly, a multi-scale spatiotemporal convolution module is constructed to extract spatiotemporal features through parallel convolution kernels, and a channel attention mechanism is introduced to achieve dynamic weight distribution and learn to cover fault information in the entire frequency band. Finally, a graph convolutional network is used to aggregate multi-scale spatiotemporal features and establish a unified representation model for temporal vibration and spatial correlation. This method breaks through the technical bottlenecks of noise masking, lack of spatiotemporal correlation, and insufficient generalization ability. The flowchart is as follows: Figure 1 shown.

[0042] Example: Experimental data was collected under four operating conditions: 0 HP, 1 HP, 2 HP, and 3 HP, corresponding to speeds of 1797 rpm, 1772 rpm, 1750 rpm, and 1730 rpm. Vibration sensors were placed at the drive and fan ends, with a frequency of 12 kHz at the fan end and 48 kHz at the drive end. Defects of 0.007 inch, 0.014 inch, and 0.021 inch in size were created on the bearing inner and outer rings and rolling elements using electrical discharge machining (EDM). This generated nine types of fault data and one type of normal data.

[0043] This experiment used data collected from the drive end, with a sampling frequency of 48kHz, a workload of 3HP, and a rotational speed of 1730 rpm. To improve model training efficiency, 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. See Table 1 for details.

[0044] Table 1 Bearing data set

[0045] As shown in Table 2, the CEEMDAN+ multi-scale spatiotemporal information fusion graph neural network in the method of the present invention can correctly identify most of the bearing fault samples. F 1 scores are all greater than 0.95, among which labels 0, 1, 2, 4, 6, 8, and 9 F 1 score greater than 0.98, labels 1, 2, 4, 6, 8, 9 F 1The score is as high as 1.00, indicating that the model’s ability to identify various types of faults is at a high level.

[0046] Table 2 Evaluation indicators of the test set

[0047] Figure 4This is the confusion matrix between the model prediction value and the actual value in one of the experimental results. The CEEMDAN+ multi-scale spatiotemporal information fusion graph neural network in the method of the present invention can well distinguish various fault modes.

[0048] In summary, the present invention uses CEEMDAN to perform adaptive multi-scale decomposition of the original vibration signal, effectively separating strong noise interference and enhancing the identifiability of weak fault features. Subsequently, the multi-scale convolutional network extracts temporal features through parallel convolution kernels and combines the channel attention mechanism to achieve dynamic weighted enhancement of key fault modes. Next, the multi-scale spatiotemporal features of GCN are aggregated to generate a global feature representation, thereby improving the performance of fault diagnosis. Experimental verification shows that the accuracy of the model on the test set reaches 99.0%, which means that the model can correctly identify most bearing fault samples, providing a highly reliable technical solution for the intelligent operation and maintenance of coal mine equipment.

Claims

1. A mining fan bearing fault diagnosis method based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network is characterized by The following steps are involved: Step 1: Collect the vibration signal of the mining fan bearing; Step 2: Preprocess the mining fan bearing vibration signal collected in step 1, including outlier processing and data normalization; Step 3: Perform complete ensemble empirical mode decomposition (CEEMDAN) on the vibration signal of the mining fan bearing preprocessed 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 use convolution kernels of different scales to extract the temporal features of each IMF component in parallel; Step 5: Introduce a channel attention mechanism to adaptively weight multi-scale spatiotemporal features, and dynamically enhance the representation of key fault modes by calculating the importance weights of channel features; Step 6: Aggregate multi-scale spatiotemporal features through GCN to generate global feature representation, which is then input into the classifier to output the fault diagnosis result.

2. The mining fan bearing fault diagnosis method based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 1 is characterized by: In step 2, outliers are first processed by using the local mean method. The calculation formula is shown in formula (1): （1）； In formula (1): is the missing data value; is the number of existing data values ​​around the missing data; Indicates the index where the current missing value is located; Represents an index The corresponding value; Indicates the index corresponding to the data.

3. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 2 is characterized by: The data normalization method uses Min-Max normalization, and the calculation is shown in formula (2): （2）； In formula (2): is the current data value; and They are The minimum and maximum values ​​in the vector; is the normalized value.

4. The mining fan bearing fault diagnosis method based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 1 is characterized by: In step 3, CEEMDAN can effectively eliminate the white noise residue in the IMF component. The decomposition steps are as follows: S3.1: Determine the addition of Gaussian white noise Number of times and standard deviation , , adding it to the original signal Get new signal , then perform EMD decomposition and take the average value of the first-order eigenmode function to obtain the component and removal The residual after ,in The calculation formula is as follows: （3）； In formula (3): Indicates the number of times Gaussian white noise is added, Indicates in The first-order IMF components obtained from the integrated decomposition; S3.2: In Add white noise Get a new signal , perform EMD decomposition and average the first-order intrinsic mode function to obtain the component and removal The residual after ,in The calculation formula is as follows: （4）； In formula (4): Indicates in The second-order IMF components obtained from the integrated decomposition; S3.3: Continue to repeat S3.2 until a monotonic residual signal that is no longer suitable for decomposition appears. At this time, the original signal is decomposed into IMF components and residual components ,in: is the residual component, Indicates the A portion.

5. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 4 is characterized by: In step 4, convolution 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 formula (5), Indicates that Features extracted by convolution kernels; Indicates the corresponding weight; represents the bias term; Represents one-dimensional input; express Enter the number of data points in ; , is a positive integer; After the convolution operation, the nonlinear transformation is performed through the ReLU activation function, which is expressed as follows: （6）； In formula (6) Indicates that Features extracted by convolution kernels; , is a positive integer; The method used in the pooling layer is the maximum pooling layer, and the maximum pooling definition is as follows: （7）； In formula (7), Indicates The pooling area related to the feature map The maximum pooling layer output value; Represents the pooling area Located in the middle The elements at Represents the coordinates of a vector in the pooling area, represents the row coordinates, Represents column coordinates.

6. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 5 is characterized by: Dropout is introduced to avoid overfitting. By configuring an appropriate ratio, hidden neurons in the convolutional neural network are excluded to improve the robustness and generalization performance of the model, and finally the feature matrix is ​​obtained: （8）； In formula (8), represents the feature matrix; represents the eigenvector; is the number of channels, that is, the number of features extracted by convolution kernels of different scales.

7. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 6 is characterized by: The step 5 comprises the following steps: S5.1: The convolution kernels of different scales in step 4 are obtained with sizes of The input feature map is subjected to global maximum pooling and global average pooling in the spatial dimension, and each feature map obtains two The feature map of 、 and Represent the height, width and number of channels of the input feature map respectively; S5.2: The results of global maximum pooling and global average pooling are fed into a shared multi-layer perceptron to learn and obtain two The feature map of MLP is , Indicates that the preset reduction ratio is used to control the compression degree of the multi-layer perceptron; The activation function is Relu, and the calculation formula is as follows: （9）； In formula (9), Represents the output of the activation function ReLU; Represents input; If you enter If it is greater than 0, the output is By itself, if you enter If it is less than or equal to 0, the output is 0; The number of neurons in the second layer of MLP is , reduction ratio is a preset integer used to control the compression level of the multilayer perceptron; S5.3: The MLP outputs are summed and then mapped through the Sigmoid activation function. The calculation formula is as follows: （10）； In formula (10), Represents the output of the Sigmoid function, is the input of the neuron, is the base of natural logarithms; The output range of the Sigmoid function is , any real number Mapped to the probability interval, the channel attention weight matrix is ​​obtained To express the importance weight of channel features, Indicates the 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 to dynamically enhance the representation strength of key failure modes: （11）； In formula (11), represents the weighted feature matrix, Channel attention weight matrix, The characteristic matrix obtained by expression (8) is: Indicates multiplication.

8. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 7 is characterized by: In step 5, the importance weights of channel features are calculated to dynamically enhance the representation strength of key failure modes, as follows: S5.3 Calculated channel attention weight matrix That is, it represents the importance of channel features; S5.4 dynamically enhances the representation strength of critical failure modes by multiplying the weight matrix with the multi-scale convolution output obtained in step 3.

9. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 8, characterized in that: In step 6, a two-layer graph convolutional structure is used in GCN. The first layer CNN is used for basic feature aggregation, and the second layer CNN is used to deeply capture the global relationship between multi-scale features. The spatiotemporal features extracted under different scale convolution kernels represent each node, 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 convolution layer, and the black solid line represents the connection relationship between these features. Z1, Z2, Z3, and Z4 are the outputs of the first layer of GCN, respectively. 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. The similarity is measured using Euclidean distance, and the calculation formula is as follows （12）； In formula (12), represents the Euclidean distance between two points, 、 Represent the x-axis coordinates of the first and second points respectively, 、 Represent the y-axis coordinates of the first and second points respectively, 、 Represent the z-axis coordinates of the first point and the second point respectively; The connection relationship between feature nodes is the 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 formula (14): Indicates the Feature representation of the layer; Indicates the Feature representation of the layer; is the adjacency matrix, ; is the degree matrix; It is the weight matrix that GCN needs to learn; is the activation function.

10. The method for fault diagnosis of bearings of mining ventilators based on CEEMDAN and multi-scale spatiotemporal information fusion graph neural network according to claim 9, characterized in that: The initial input of GCN is the weighted feature matrix calculated by formula (11), which makes the features learned by GCN focus more on key channels; The classifier uses the softmax layer: （15）； In formula (15): Indicates the The probability of an element, is the first elements, It is The exponential of the input value, is the sum of all element indices; The loss function uses the cross entropy loss function, which is defined as follows: （16）； In formula (16), represents the overall average loss, , represents the number of all samples; Indicates the number of all sample categories; Indicates the samples; Indicates the The predicted probability of each category; represents an indicator variable; Indicates the The samples belong to The predicted probability of each category; Finally, combined with formula (15), the fault type corresponding to the sample is calculated by the following formula: （17）； In formula (17), Represents the output category predicted by the model; Indicates taking the index with the highest probability, that is, choosing The category corresponding to the maximum probability in the output is taken as the final classification result.

Images