A Fault Diagnosis Method for Spindle Bearings of Mining Drilling Rigs Based on Improved Transformer

Abstract

This invention discloses a fault diagnosis method for main shaft bearings of mining drilling rigs based on an improved Transformer. A gated gain lightweight Transformer network is constructed to perform sliding window framing and standardization on the acquired one-dimensional vibration signal, mapping it to a high-dimensional feature space through one-dimensional convolution. A gated gain denoising module extracts feature statistics through a dual-branch approach of global average pooling and global max pooling, adaptively generating feature thresholds. Sigmoid-gated gain soft thresholding is then used to suppress noise and preserve weak fault impact features. The denoised features are input into a dimensionality-reducing Former layer, where deep fault features are extracted layer by layer through layer normalization, multi-head self-attention, channel attention-guided convolutional dimensionality reduction, and a feedforward network. Finally, the bearing fault category is output after global average pooling and a fully connected layer. This invention achieves high diagnostic accuracy under strong noise conditions, with low model parameter count and inference overhead, making it suitable for on-site condition monitoring and predictive maintenance of main shaft bearings in mining drilling rigs.

Description

Technical Field

[0001] This invention belongs to the field of mining equipment condition monitoring, vibration signal processing and deep learning intelligent fault diagnosis technology, specifically involving a fault diagnosis method for the spindle bearing of a mining drilling rig based on an improved Transformer. Background Technology

[0002] Mining drilling rigs are core equipment in coal mining, mining engineering, and underground drilling operations. Their spindle bearings operate under conditions of low speed, heavy load, frequent load fluctuations, strong impact loads, and complex environmental noise coupling, making them highly susceptible to localized damage to the inner ring, outer ring, rolling elements, or cage. Failure to detect these faults in a timely manner can lead to decreased drilling efficiency, unplanned equipment downtime, and even safety accidents. Therefore, accurate and real-time fault diagnosis of the operating status of mining drilling rig spindle bearings is of significant engineering importance for improving equipment reliability, reducing maintenance costs, and enabling predictive maintenance.

[0003] Currently, bearing fault diagnosis based on vibration signal analysis is the most widely used technical approach in industry. Traditional diagnostic methods rely on manually extracting time-domain statistical indicators, frequency-domain spectral features, and time-frequency domain features, and then combining them with classifiers such as support vector machines and random forests to complete the identification. This type of method has certain effectiveness in stable operating conditions and low noise scenarios. However, in the field operation of mining drilling rigs, vibration signals are mixed with various interference components such as coal and rock crushing impact, mechanical structure coupled vibration, and motor electromagnetic noise. The weak impact characteristics of early faults are easily masked by strong background noise, and manually designed features are difficult to stably represent the fault state. The diagnostic accuracy and robustness cannot meet the requirements of the field.

[0004] With the development of deep learning technology, models such as convolutional neural networks, residual networks, and Transformers are gradually being applied to the field of bearing fault diagnosis. Among them, convolutional neural networks can automatically extract local features, but their receptive field is limited, resulting in insufficient ability to model long-term temporal dependencies. Ordinary Transformer models rely on self-attention mechanisms to capture global temporal dependencies, but their large number of parameters and computational overhead lead to insufficient real-time performance and deployment difficulties in scenarios where the computing power of edge monitoring terminals for mining drilling rigs is limited. Meanwhile, most existing deep learning fault diagnosis models lack dedicated denoising mechanisms for strong background noise and have insufficient ability to enhance weak fault impact features, resulting in a significant decrease in diagnostic accuracy when directly applied to the complex working conditions of mining drilling rigs.

[0005] In summary, existing technologies cannot simultaneously meet the multiple requirements of strong noise resistance, high diagnostic accuracy, and lightweight edge deployment for fault diagnosis of spindle bearings in mining drilling rigs, and there is an urgent need to propose targeted improvement solutions. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problems faced by existing technologies in the fault diagnosis of main shaft bearings of mining drilling rigs: strong background noise masking early weak fault impacts leading to low diagnostic accuracy, large number of parameters and high computational cost of ordinary Transformer models making edge deployment difficult, and lack of scene adaptability and robustness of existing models. The invention provides a fault diagnosis method for main shaft bearings of mining drilling rigs based on an improved Transformer.

[0007] To achieve the above objectives, the technical solution adopted in this invention is as follows: a fault diagnosis method for mining drilling rig spindle bearings based on an improved Transformer. This method constructs a gated gain lightweight Transformer (GGLT) network, adaptively suppressing strong background noise and retaining weak fault features through a front-end gated gain denoising module. A lightweight dimensionality reduction Former feature extraction layer extracts long-range temporal features while progressively compressing dimensions and reducing computational overhead, ultimately achieving high-precision and lightweight fault diagnosis for mining drilling rig spindle bearings. The specific steps are as follows: S1. Data Acquisition and Dataset Construction: Collect one-dimensional vibration signals of the main shaft bearing of the mining drilling rig under different operating conditions, label the signals according to the bearing operating conditions, and construct a dataset including normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault.

[0008] S2. Data Preprocessing and Dataset Partitioning: The one-dimensional vibration signal is framed using a sliding window to obtain multiple fixed-length vibration signal segments; all segments are Z-score standardized to eliminate the influence of amplitude scale differences; the labeled dataset is divided into training set, validation set and test set according to a preset ratio, which are used for model training and updating, performance verification and effect testing, respectively.

[0009] S3 and GGLT Network Architecture Construction: A gated gain lightweight Transformer network is built, consisting of an input layer, a gated gain denoising module (GGDM), a three-layer stacked dimensionality reduction Former feature extraction layer, and an output layer connected in series.

[0010] S4. High-dimensional feature mapping: The training set is input into the input layer, and the vibration signal is subjected to local feature extraction and high-dimensional mapping through a one-dimensional convolutional layer. Combined with batch normalization and nonlinear activation, the initial feature map is obtained.

[0011] S5. Adaptive Denoising Processing: The initial feature map is input into the gated gain denoising module. The overall energy and local peak statistics of each channel are extracted through global average pooling and global max pooling dual branches. After encoding by a fully connected layer and fusion with learnable weights, a channel-level adaptive feature threshold is generated. Based on the threshold, the Sigmoid gated gain coefficient is calculated. The noise features are attenuated and the fault impact features are preserved through soft thresholding, resulting in a denoised feature map.

[0012] S6. Deep Temporal Feature Extraction: The denoised feature map is input into a three-layer stacked dimensionality reduction Former feature extraction layer. Each layer is processed sequentially through layer normalization, multi-head self-attention, residual connection, channel attention recalibration, convolutional dimensionality reduction, feedforward network and residual connection. Multi-head self-attention captures long-range temporal dependencies, channel attention enhances the weights of fault-sensitive channels, and convolutional dimensionality reduction compresses redundant feature dimensions step by step, finally extracting deep fault features.

[0013] S7. Network Model Training and Update: Input deep fault features into the output layer, compress the sequence dimension through global average pooling, and then output the probability of each category through a fully connected layer to obtain the diagnosis result; based on the true label and prediction result in step S2, calculate the prediction error as the classification loss using the cross-entropy loss function, and use the Adam optimizer to backpropagate and update the network parameters. Set the initial learning rate to 0.0001 and the batch size to 64, and iterate until the accuracy of the validation set converges to prevent the network model from overfitting, thus completing the network model training and update.

[0014] S8. Fault Diagnosis Application: After the vibration signal of the main shaft bearing of the mining drilling rig to be diagnosed is preprocessed in the same way, it is input into the updated GGLT network and the corresponding bearing operating status category is output to realize end-to-end fault diagnosis.

[0015] Furthermore, in step S2, each fixed-length vibration signal segment contains 2048 sampling points, and the sliding window has a sliding step size of 1600. Overlapping sampling reduces information loss caused by fault impacts located at the window boundary. Z-score normalization is calculated based on the mean and standard deviation of the training set, mapping the signal to a standard distribution with a mean of 0 and a standard deviation of 1.

[0016] Further, in step S3, the input layer consists of a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation layer in sequence; wherein the one-dimensional convolutional layer has 64 output channels, a kernel size of 64, a stride of 16, and uses Same padding to convert the one-dimensional vibration signal of length 2048 into an initial feature map with 64 channels and a sequence length of 128.

[0017] Furthermore, in step S5, the adaptive threshold of the gated gain denoising module simultaneously integrates the overall channel energy and local peak information, and can automatically adjust the threshold size according to the noise intensity of the input signal; the sigmoid gated gain achieves soft threshold denoising, avoiding feature abrupt changes and fault information loss caused by hard thresholding. The specific processing steps are as follows: S5.1 Perform absolute value transformation on the input initial feature map to obtain an amplitude feature map that only represents the magnitude of the feature amplitude; S5.2 Input the amplitude feature map into the global average pooling branch and the global max pooling branch respectively, and extract the overall energy statistical features and peak intensity statistical features of each channel respectively; S5.3 Input the two types of statistical features into the corresponding fully connected layers for nonlinear encoding to obtain the average response encoding feature and the peak response encoding feature; S5.4. By using learnable weighted parameters that satisfy normalization constraints, the two types of encoded features are weighted and fused to generate channel-level adaptive feature thresholds. S5.5 Based on the adaptive feature threshold and the learnable scale adjustment factor, the Sigmoid gated gain coefficient is calculated. The gain coefficient is multiplied element by element with the initial feature map to complete soft thresholding denoising and obtain the denoised feature map.

[0018] Further, in step S6, the output channel dimensions of the three stacked dimensionality-reduced Form feature extraction layers are reduced to 64, 32, and 16 respectively. Channel attention guides the dimensionality reduction, ensuring that feature information of fault-sensitive channels is preferentially preserved during the reduction process, minimizing accuracy loss while reducing parameter and computational load. The dimensionality-reduced Form feature extraction layer is a lightweight Transformer encoder structure, and the processing order is as follows: S6.1 After performing the first layer normalization on the input feature map, input it into the multi-head self-attention unit, extract long-range temporal dependencies, and perform the first residual connection between the output and the input. S6.2 The residual output is recalibrated by the channel attention unit to enhance the weight of fault-sensitive channels and suppress redundant channels; S6.3. The recalibrated features are reduced in dimensionality through one-dimensional convolution to compress the feature dimension. S6.4. After dimensionality reduction, the features are normalized in a second layer. The input feedforward network is nonlinearly mapped, and the output is connected to the dimensionality-reduced features in a second residual connection to obtain the output features of the current layer.

[0019] Compared with the prior art, the present invention has the following advantages: 1. The gated gain denoising module proposed in this invention extracts multi-dimensional statistical information through dual-branch pooling to generate an adaptive threshold. Combined with a sigmoid-gated soft modulation mechanism, it can effectively suppress interference components under strong background noise while accurately preserving early, weak fault impact characteristics. Experimental verification shows that even under extremely strong noise conditions with a signal-to-noise ratio (SNR) of 0, the diagnostic accuracy of this method remains above 80%, and the advantage becomes more significant with stronger noise, effectively adapting to complex mining conditions.

[0020] 2. The dimensionality reduction Former feature extraction layer designed in this invention innovatively embeds channel attention and convolutional dimensionality reduction into the Transformer encoder structure. Through a collaborative mechanism of "first enhancing sensitive channels and then compressing redundant dimensions," it progressively reduces feature dimensions while preserving core fault features. The final model has only 50K parameters and a single-sample inference time of only 10ms. The number of parameters and computational cost are only 1 / 5 of that of the traditional Transformer model, which can be directly deployed on the edge monitoring terminal of mining drilling rigs to meet real-time diagnostic requirements.

[0021] 3. This invention is specifically optimized for the characteristics of low-speed heavy load, strong impact load, and large fluctuations in operating conditions of the spindle bearing of mining drilling rigs. On a real mining drilling rig operating condition dataset, the average diagnostic accuracy reaches over 94%, which is significantly better than mainstream models such as 1DCNN, VGG16, ResNet18, and ordinary Transformer. It can still maintain stable diagnostic performance under complex operating conditions with load fluctuations, and has strong scene adaptability and high diagnostic robustness. Attached Figure Description

[0022] Figure 1 This is a structural block diagram of the Gated Gain Denoising Module (GGDM) in this invention.

[0023] Figure 2 This is a diagram showing the overall architecture of the Gated Gain Lightweight Transformer (GGLT) network in this invention.

[0024] Figure 3 This is a comparison chart showing the diagnostic accuracy of various fault diagnosis models on the Jiangnan University bearing dataset under different signal-to-noise ratio conditions.

[0025] Figure 4 This is a comparison chart of the diagnostic accuracy of each model on the Z1 working condition dataset of the main shaft bearing of a mining drilling rig. Detailed Implementation

[0026] The present invention will be further described below.

[0027] Example 1: This example uses the fault diagnosis of the spindle bearing of a certain type of mining drilling rig as an application scenario to fully illustrate the specific implementation process of the method of the present invention. Those skilled in the art can reproduce the present invention and complete the fault diagnosis based on the content of this example.

[0028] In this embodiment, the operating status of the mining drilling rig spindle bearing is divided into five categories: normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault. Vibration signals are collected by an accelerometer mounted on the spindle bearing housing, with a sampling frequency of 10kHz.

[0029] The specific implementation steps are as follows: S1. Data Acquisition and Dataset Construction: Collect raw one-dimensional vibration signals of the main shaft bearing of the mining drilling rig under five operating conditions, with a continuous signal length of no less than 100 seconds for each condition. Label the signals according to the bearing operating conditions to construct the raw sample dataset.

[0030] S2. Data Preprocessing and Dataset Partitioning: S2.1 Sliding Window Framing: The original long vibration signal is segmented using a sliding window, with a window length L = 2048 sampling points and a sliding step size p = 1600 sampling points. The segmentation formula is as follows:

[0031]

[0032] in, Let m be the m-th fixed-length vibration signal segment, N be the total length of the original signal, and M be the total number of segments obtained from the segmentation. Segmentation using an overlapping sliding window reduces information loss caused by fault impacts located at the window edges.

[0033] S2.2 Z-score Standardization: All fixed-length vibration signal segments are standardized to eliminate the impact of amplitude scale differences between samples on model training. The calculation formula is as follows:

[0034] Where x is the amplitude of the original vibration signal, μ is the mean of all samples in the training set, and σ is the standard deviation of all samples in the training set. This is the standardized signal amplitude.

[0035] S2.3 Dataset Partitioning: All labeled samples are randomly divided into training, validation, and test sets in a 6:2:2 ratio. The training set is used to update network parameters, the validation set is used to monitor model performance and trigger early stopping during training, and the test set is used to finally evaluate the model's diagnostic effectiveness.

[0036] S3 and GGLT Network Architecture Construction: Construct a gated gain lightweight Transformer (GGLT) network. The overall architecture is as follows: Figure 2 As shown, the network consists of the following layers from input to output: input layer, gated gain denoising module (GGDM), first-layer dimensionality reduction Form, second-layer dimensionality reduction Form, third-layer dimensionality reduction Form, and output layer.

[0037] The specific configurations for each layer are as follows: (1) Input layer: It consists of a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation layer in sequence. The kernel size of the one-dimensional convolutional layer is 64, the stride is 16, the number of output channels is 64, and Same padding is used; the input is a one-dimensional vibration signal with a length of 2048, and the output is an initial feature map with 64 channels and a sequence length of 128.

[0038] (2) Gated Gain Denoising Module (GGDM): The structure is as follows Figure 1 As shown, the input is the initial feature map, and the output is the denoised feature map of the same dimension.

[0039] (3) Three stacked Former feature extraction layers for dimensionality reduction: the first layer has 128 input channels and 64 output channels; the second layer has 64 input channels and 32 output channels; and the third layer has 32 input channels and 16 output channels. Each layer has the same internal structure and includes layer normalization, multi-head self-attention, residual connection, channel attention, convolutional dimensionality reduction, feedforward network, and residual connection unit.

[0040] (4) Output layer: It consists of a global average pooling layer and a fully connected layer. The global average pooling layer compresses the sequence dimension by average, compressing a sequence of length 128 into a single vector with an output dimension of 16. The fully connected layer maps the 16-dimensional features into a 5-dimensional output, corresponding to 5 types of bearing operating states. Finally, the probability of each category is obtained through the Softmax function.

[0041] S4. High-dimensional feature mapping: The input is a one-dimensional vibration signal of length 2048. The vibration signal is subjected to local feature extraction and high-dimensional mapping through a one-dimensional convolutional layer. Combined with batch normalization and nonlinear activation, the output is an initial feature map with 64 channels and a sequence length of 128.

[0042] S5. Adaptive denoising (GGDM module): This module processes the initial feature map output from the input layer. Input the GGDM module to complete adaptive noise reduction. The specific process is as follows: S5.1 Absolute value transformation: for the initial feature map Taking the absolute value of each element yields an amplitude feature map that only represents the magnitude of the amplitude. :

[0043] S5.2, Dual-branch pooling statistics: The amplitude characteristic map... Input the Global Average Pooling (GAP) branch and the Global Max Pooling (GMP) branch respectively to obtain the statistical vectors of the channel dimension:

[0044]

[0045] in, This is the average pooling statistics vector with a dimension of 64, representing the overall energy level of each channel; This is a max-pooling statistical vector with a dimension of 64, representing the peak impact intensity of each channel.

[0046] S5.3 Fully Connected Layer Encoding: The two statistical vectors are input into their respective fully connected layers for non-linear encoding to obtain the encoded features:

[0047]

[0048] The fully connected layer consists of two linear transformation layers and one ReLU activation layer, which first reduces the dimensionality and then increases it to enhance the nonlinear expressive power, while the output encoding feature dimension remains 64.

[0049] S5.4 Adaptive Threshold Generation: Two encoded features are weighted and fused using learnable weighting parameters α and β to generate a channel-level adaptive feature threshold λ.

[0050]

[0051] Here, α and β are parameters automatically learned during network training, both initially set to 0.5. This threshold simultaneously integrates the overall channel energy and local peak information, and can be automatically adjusted according to the noise intensity of the input signal.

[0052] S5.5, Sigmoid Gated Gain and Denoising: The gating gain coefficient is calculated based on an adaptive threshold, and soft thresholding is applied to the initial feature map.

[0053]

[0054] Where G is the gated gain coefficient matrix, and its dimension is the same as the initial feature. Figure 1τ is a learnable scaling factor used to control the steepness of the sigmoid function; ε is a minimal constant with a value of To prevent the denominator from being zero; This indicates element-wise multiplication; This is the feature map after denoising.

[0055] When the feature amplitude is greater than the threshold, the gain coefficient approaches 1, and the fault impact feature is preserved; when the feature amplitude is less than the threshold, the gain coefficient approaches 0, and the noise component is attenuated. This soft thresholding mechanism avoids the feature abrupt changes caused by hard thresholding and effectively preserves the detailed information of the fault features.

[0056] S6. Deep Temporal Feature Extraction (Dimensionality Reduction Former Feature Extraction Layer): This involves extracting the denoised feature map... Input a three-layer stacked dimensionality-reduced Former module and extract deep fault features layer by layer. Taking the first layer of the dimensionality-reduced Former as an example, the specific processing procedure is as follows: S6.1, First Layer Normalization and Multi-Head Self-Attention: For input features Perform layer normalization to obtain normalized features. :

[0057] Will Input a multi-head self-attention unit, obtain the query matrix Q, key matrix K, and value matrix V through linear mapping, and calculate the self-attention output:

[0058]

[0059] in, Here, h represents the dimension of single-head attention, and h is the number of attention heads. In this embodiment, h=4. This is the output mapping matrix.

[0060] Compare the self-attention output with the original input Perform residual connection to obtain the first residual feature. :

[0061] S6.2 Channel Attention Recalibration: For the first residual feature Perform channel attention weighting to enhance fault-sensitive channels and suppress redundant channels: First, the channel statistics vector s is obtained through global average pooling:

[0062] Then, the channel weight vector w is obtained by passing two fully connected layers and sigmoid activation:

[0063] in, These are the parameters for the first layer fully connected layer. δ represents the parameters of the second fully connected layer, and δ is the ReLU activation function.

[0064] Finally, the weights are multiplied channel by channel by channel with the original features to obtain the recalibrated features. :

[0065] S6.3 Convolution Dimensionality Reduction: One-dimensional convolution with a kernel size of 1 is used to perform channel dimensionality reduction on the recalibrated features, compressing the number of channels from 128 to 64, resulting in dimensionality-reduced features. :

[0066] This step performs dimensionality reduction under the guidance of channel attention, prioritizing the retention of fault-sensitive channels with high weights to minimize information loss caused by dimensionality reduction.

[0067] S6.4 Second Layer Normalization and Feedforward Network: Dimensionality reduction features A second layer normalization is performed, and the input to the feedforward network undergoes a nonlinear mapping:

[0068]

[0069] in, The parameters are for the feedforward network, which consists of two linear transformation layers, with the hidden layer having a dimension twice that of the input.

[0070] Feedforward network output and dimensionality reduction features Perform residual connections to obtain the final output of the first-layer dimensionality-reduced Former. :

[0071] The processing logic of the second and third layers of the dimensionality reduction Form is the same as that of the first layer, only the input and output channel dimensions are different: the second layer compresses the number of channels from 64 to 32, and the third layer compresses the number of channels from 32 to 16, further extracting deep abstract features. Finally, after three layers of processing, deep fault features are obtained. The dimensions are 16 channels and 128 sequences.

[0072] S7, Network Model Training Update: S7.1 Classification Output: This classifies deep fault characteristics. The input and output layers first compress the sequence dimension using global average pooling to obtain a 16-dimensional global feature vector f:

[0073] The feature vectors are then mapped to a 5-dimensional output vector o through a fully connected layer, and the probability distributions for each category are obtained through the Softmax function:

[0074]

[0075] in, These are the parameters for the fully connected layer. This represents the probability that the input sample belongs to class c.

[0076] The category with the highest probability is selected as the final diagnosis result:

[0077] S7.2 Loss Calculation and Parameter Update: The error between the predicted result and the true label is calculated using the cross-entropy loss function:

[0078] in, This is a one-hot encoding for the real label.

[0079] The Adam optimizer is used for backpropagation to update all learnable parameters of the network. The initial learning rate is set to 0.0001, and the batch size is 64. During training, the validation set accuracy is calculated in each round, and an early stopping strategy is adopted: if the validation set accuracy does not improve for 10 consecutive rounds, training is stopped, and the optimal model weights are saved.

[0080] S8. Fault Diagnosis Application: The vibration signal of the main shaft bearing of the mining drilling rig to be diagnosed is processed by dividing it into frames using the same sliding window as the training set and standardizing it with Z-score. Then, it is input into the trained GGLT network. The network automatically outputs the corresponding bearing operating status category, realizing end-to-end intelligent fault diagnosis.

[0081] Experimental verification: To verify the effectiveness of the method of the present invention, a comparative experiment was conducted using the publicly available Jiangnan University bearing dataset and the self-developed mining drill rig spindle bearing dataset. The method of the present invention (GGLT) was compared with four mainstream deep learning fault diagnosis methods: 1DCNN, VGG16, ResNet18, and ordinary 1DTransformer.

[0082] (1) Introduction to the dataset ① Jiangnan University Bearing Dataset: This dataset includes three operating speeds: 600 rpm, 800 rpm, and 1000 rpm. Each operating condition contains four states: normal, inner race fault, outer race fault, and rolling element fault. Each state contains 300 samples, with a total sample length of 2048 points. This dataset is used to verify the noise resistance of the model. In the experiment, Gaussian white noise of different intensities was added to the original data. The signal-to-noise ratio (SNR) is used to characterize the noise intensity; the smaller the SNR, the stronger the noise. See Table 1 for details.

[0083] Table 1. Parameters of the Jiangnan University Bearing Dataset

[0084] ② Mining Drill Rig Spindle Bearing Dataset: This dataset includes three load conditions: Z1, Z2, and Z3. Each condition contains five states: normal, inner ring fault, outer ring fault, rolling element fault, and cage fault. Each state has 150 samples, with a total sample length of 2048 points. This dataset is used to verify the diagnostic performance of the model under real mining conditions, as shown in Table 2.

[0085] Table 2 Drilling Rig Spindle Bearing Dataset Parameters

[0086] (2) Experimental setup: All models used the same data preprocessing and training parameters: sample length 2048 points, Z-score normalization, Adam optimizer, initial learning rate 0.0001, batch size 64, training epochs 100, and early stopping strategy. Each experiment was repeated 5 times, and the average accuracy was taken as the final result.

[0087] (3) Experimental results and analysis: ① Noise resistance performance comparison: Diagnostic accuracy of each model under different signal-to-noise ratio conditions, such as... Figure 3 As shown in Table 3.

[0088] Table 3. Diagnostic accuracy (%) of each model under different signal-to-noise ratios in J2 and J3 operating conditions.

[0089] The results show that the diagnostic accuracy of all models decreases with increasing noise intensity, but the method of this invention (GGLT) maintains the highest accuracy under all noise conditions. Even under extreme strong noise conditions with SNR=0, the accuracy of GGLT remains above 80%, more than 5 percentage points higher than the second-best method, with a small standard deviation and high stability, demonstrating that the GGDM module can effectively suppress strong background noise and preserve weak fault characteristics.

[0090] ② Comparison of diagnostic performance under real working conditions: Diagnostic results of each model on the mining drill rig spindle bearing dataset are as follows: Figure 4 As shown in Table 4.

[0091] Table 4. Average diagnostic accuracy (%) of each model under different load conditions

[0092] Depend on Figure 4 As can be seen, in the Z1 dataset experiments, the method of this invention achieved the best results in all five experiments, with accuracies of 96.0%, 96.0%, 98.0%, 97.3%, and 95.3%, respectively, representing improvements of 3.3%, 1.3%, 4.0%, 5.3%, and 4.1% compared to the second-best method. This indicates that the method of this invention can effectively extract deep fault features from the vibration signals of the spindle bearings of mining drilling rigs and has strong anti-interference capabilities under real-world working conditions. The experimental results also show that GGLT's diagnostic accuracy on the Z1 dataset is superior to the other four comparative models.

[0093] As shown in Table 4, the method of this invention achieved the highest average diagnostic accuracy on both the Z2 and Z3 datasets, at 96.80% and 91.77% respectively, with an average of 94.29%, significantly higher than 1DCNN, VGG16, ResNet18, and 1DTransformer. Especially under the Z3 working condition with the largest load fluctuations, the accuracy of other models decreased significantly, while GGLT maintained an accuracy of over 91%, demonstrating that this method has extremely strong adaptability and robustness to the complex working conditions of mining drilling rigs.

[0094] ③ Lightweight Performance Comparison: To verify the effectiveness of the Gated Gain Denoising Module (GGDM module) and the dimensionality reduction Former feature extraction layer in this invention, an ablation comparison experiment was designed. Three models were constructed: a complete GGLT model, a GGLT model with the GGDM module removed, and a GGLT model using a traditional Transformer to replace the dimensionality reduction Former feature extraction layer, denoted as Model 1, Model 2, and Model 3, respectively. The Z1 dataset was used as an example for the experiment. This ablation experiment was used to verify the denoising capability of the GGDM module and the effectiveness of the lightweight design of the dimensionality reduction Former feature extraction layer; the comparison of the number of parameters, computational cost, and inference time for each model is shown in Table 5.

[0095] Table 5 Comparison of Lightweight Performance of Each Model

[0096] The results show that the complete GGLT model of this invention achieves the highest diagnostic accuracy while having lower parameter count, computational cost, and inference time. Compared with the complete GGLT model, the diagnostic accuracy of Model 2 decreased by 4 percentage points after removing the GGDM module, indicating that the GGDM module can effectively suppress environmental noise and retain weak fault characteristics. Model 3, after replacing the dimensionality-reduced Former feature extraction layer with the traditional Transformer, achieved a diagnostic accuracy close to that of this invention, but the number of parameters increased from 50K to 250K, the computational cost increased from 10M to 30M, and the inference time increased from 10ms to 50ms. This indicates that the dimensionality-reduced Former feature extraction layer can significantly reduce model complexity and improve inference efficiency while maintaining high diagnostic accuracy. It reduces the number of parameters and computational cost by 80%, achieving the optimal balance between accuracy and efficiency, and fully meeting the deployment requirements of mining edge terminals.

[0097] In summary, the fault diagnosis method for spindle bearings of mining drilling rigs proposed in this invention has the advantages of strong noise resistance, high diagnostic accuracy and extreme lightweight, and can effectively adapt to the complex working conditions and edge deployment requirements of mining drilling rigs, thus having high engineering application value.

Claims

1. A fault diagnosis method for spindle bearings of mining drilling rigs based on an improved Transformer, characterized in that, Includes the following steps: S1. Collect one-dimensional vibration signals of the main shaft bearing of the mining drilling rig, and form a dataset after labeling the bearing's operating status. S2. After performing sliding window framing and standardization on the one-dimensional vibration signal, the dataset is divided into training set, validation set and test set. S3. Construct a gated gain lightweight Transformer network, including an input layer, a gated gain denoising module, a multi-layer dimensionality reduction Former feature extraction layer, and an output layer; S4. Input the training set into the input layer, and use one-dimensional convolution to map the one-dimensional vibration signal to a high-dimensional feature space to obtain the initial feature map. S5. Input the initial feature map into the gated gain denoising module, adaptively generate channel-level feature thresholds based on the global statistical information of the features, and obtain the denoised feature map through Sigmoid gated gain soft thresholding. S6. Input the denoised feature map into the dimensionality reduction Former feature extraction layer to extract deep fault features step by step. S7. Input the deep fault features into the output layer to obtain the diagnosis results. Calculate the classification loss based on the validation set and diagnosis results from step S2, and update the network parameters through the optimizer. S8. Input the vibration signal of the main shaft bearing of the mining drilling rig to be diagnosed into the updated network, and output the corresponding bearing operating status category.

2. The method according to claim 1, characterized in that, In step S2, the one-dimensional vibration signal is divided into multiple fixed-length vibration signal segments by sliding window segmentation; the standardization process adopts Z-score standardization, which maps vibration signals of different scales to a distribution with a mean of 0 and a standard deviation of 1.

3. The method according to claim 1, characterized in that, In step S3, the input layer sequentially includes a one-dimensional convolutional layer, a batch normalization layer, and a ReLU activation layer; the output layer sequentially includes a global average pooling layer and a fully connected classification layer, and the classification layer outputs the category probability corresponding to the bearing's operating state.

4. The method according to claim 3, characterized in that, The one-dimensional convolutional layer uses the Same padding method to convert the one-dimensional vibration signal into an initial feature map corresponding to the number of channels in the one-dimensional convolutional layer.

5. The method according to claim 1, characterized in that, In step S5, the specific processing steps of the gated gain denoising module are as follows: S5.1 Perform absolute value transformation on the input initial feature map to obtain an amplitude feature map that only represents the magnitude of the feature amplitude; S5.2 Input the amplitude feature map into the global average pooling branch and the global max pooling branch respectively, and extract the overall energy statistical features and peak intensity statistical features of each channel respectively; S5.3 Input the two types of statistical features into the corresponding fully connected layers for nonlinear encoding to obtain the average response encoding feature and the peak response encoding feature; S5.

4. By using learnable weighted parameters that satisfy normalization constraints, the two types of encoded features are weighted and fused to generate channel-level adaptive feature thresholds. S5.5 Based on the adaptive feature threshold and the learnable scale adjustment factor, the Sigmoid gated gain coefficient is calculated. The gain coefficient is multiplied element by element with the initial feature map to complete soft thresholding denoising and obtain the denoised feature map.

6. The method according to claim 1, characterized in that, In step S6, the dimensionality reduction Former feature extraction layer is a lightweight Transformer encoder structure, and the processing order is as follows: S6.1 After performing the first layer normalization on the input feature map, input it into the multi-head self-attention unit, extract long-range temporal dependencies, and perform the first residual connection between the output and the input. S6.2 The residual output is recalibrated by the channel attention unit to enhance the weight of fault-sensitive channels and suppress redundant channels; S6.

3. The recalibrated features are reduced in dimensionality through one-dimensional convolution to compress the feature dimension. S6.

4. After dimensionality reduction, the features are normalized in a second layer. The input feedforward network is nonlinearly mapped, and the output is connected to the dimensionality-reduced features in a second residual connection to obtain the output features of the current layer.

7. The method according to claim 6, characterized in that, The output channel dimension of the multi-layer dimensionality reduction Former feature extraction layer decreases sequentially, thereby achieving step-by-step compression of the feature dimension.

8. The method according to claim 1, characterized in that, In step S7, the prediction error is calculated using the cross-entropy loss function as the classification loss, the network parameters are updated using the Adam optimizer, and an early stopping strategy is adopted based on the validation set accuracy during the training process.

9. The method according to claim 1, characterized in that, The bearing operating states include five categories: normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault.

Images