A bearing weak fault diagnosis method based on ACGP network

Abstract

This invention discloses a bearing weak fault diagnosis method based on the ACGP network. A bearing weak fault diagnosis model based on the ACGP network is constructed, including a convolutional network, an ACGP network, and a classifier. The convolutional network extracts features from the input vibration signal, and the ACGP network processes the extracted features using a channel attention mechanism, cross-channel interaction, and global average pooling. The resulting features are then fed into the classifier to obtain the predicted probability that the input vibration signal belongs to each fault type. This invention constructs a bearing weak fault diagnosis model based on the ACGP (Attention Mechanism-Cross-Channel-Global Pooling) network, achieving high-accuracy fault diagnosis in weak fault environments.

Description

Technical Field

[0001] This invention belongs to the field of bearing fault technology, and more specifically, relates to a bearing weak fault diagnosis method based on ACGP network. Background Technology

[0002] Bearings, as core supporting components of rotating machinery, are widely used in many key fields such as industrial production, rail transportation, aerospace, and new energy equipment. The health of bearings directly determines the operational reliability and service life of rotating machinery. Once a bearing fails, especially an early potential failure, if it is not detected and dealt with in time, it can easily lead to equipment downtime and production interruption. In severe cases, it may even cause equipment damage and safety accidents, resulting in huge economic losses and safety hazards. Therefore, accurate diagnosis of bearing failures through effective means, especially the identification and early warning of early failures, is crucial for ensuring the safe operation of rotating machinery.

[0003] In actual industrial production scenarios, the diagnosis of early bearing faults faces numerous technical challenges, particularly the weak signal characteristics of early faults, which are easily masked by noise. Traditional bearing fault diagnosis methods mainly rely on classical signal processing techniques such as envelope analysis and Fourier transform. These techniques extract fault features by performing spectral and time-domain analysis on bearing vibration signals to achieve fault diagnosis. However, these methods have high requirements for signal stability. In scenarios with weak signals and strong noise in early faults, they suffer from insufficient fault feature extraction, low diagnostic accuracy, and high false positive and false negative rates, failing to meet the needs of accurate diagnosis of early bearing faults in industrial settings.

[0004] In recent years, neural network technology has developed rapidly. With its powerful self-learning, feature mining, and data classification capabilities, it has been widely explored and applied in fields such as mechanical fault diagnosis and data analysis, providing a new technical path for bearing fault diagnosis. However, traditional neural network models still have significant limitations in bearing fault diagnosis applications: on the one hand, model training requires massive amounts of high-quality labeled fault data as support, while while there is ample data on normal bearing operation in industrial settings, early fault data is scarce and labeling costs are high, making it difficult to meet model training needs; on the other hand, model operation requires high computing power, increasing deployment costs in industrial settings; furthermore, traditional neural networks have weak resistance to noise interference, and in the complex interference environment of industrial settings, models are prone to overfitting, leading to decreased diagnostic accuracy and an inability to stably adapt to the complex scenarios of actual industrial production. Therefore, a bearing fault diagnosis solution that can solve the above problems is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a bearing weak fault diagnosis method based on ACGP network. A bearing weak fault diagnosis model based on ACGP (Attention Mechanism-Cross Channel-Global Pooling) network is constructed, and high-accuracy fault diagnosis is achieved in weak fault environments.

[0006] To achieve the above-mentioned objectives, the bearing weak fault diagnosis method based on ACGP network of the present invention includes the following steps:

[0007] S1: Collect fault samples of bearings in early weak fault states of different fault types according to actual needs, and construct a fault sample set. Each fault sample includes a length of [length missing]. Vibration signals and their corresponding fault type labels, length The value should be set according to actual needs;

[0008] S2: Construct a bearing weak fault diagnosis model based on the ACGP network, including a convolutional network, an ACGP network, and a classifier, where:

[0009] Convolutional networks are used to perform convolution operations on input vibration signals, and the resulting convolutional features Send to the ACGP network, where Indicates the number of channels;

[0010] ACGP networks are used to process received convolutional features. Perform channel feature extraction and obtain the channel features The data is sent to the classifier; the ACGP network includes a channel attention weighting module, a cross-channel interaction module, and a global average pooling module, where:

[0011] The channel attention weighting module is used to apply channel attention mechanisms to convolutional features. Extracting Channel Attention Weights And use channel attention weights convolutional features Weighted processing is performed to obtain channel attention features. ,in Represent the Hadamard product, then use channel attention features. Send to the cross-channel interaction module;

[0012] The cross-channel interaction module is used to apply channel attention features using 1×1 cross-channel convolution operations. Perform cross-channel interaction to obtain channel interaction characteristics. And send it to the global average pooling module;

[0013] The global average pooling module is used for channel interaction features. Perform global average pooling to obtain channel features. ;

[0014] A classifier is used to classify based on channel features The predicted probability that the input vibration signal belongs to each fault type is obtained;

[0015] S3: Use the fault sample set from step S1 to train the bearing weak fault diagnosis model, and obtain the trained bearing weak fault diagnosis model.

[0016] S4: When weak fault diagnosis of bearings is required, the data acquisition length is... The vibration signal is input into the trained bearing weak fault diagnosis model to obtain the fault diagnosis result.

[0017] This invention relates to a bearing weak fault diagnosis method based on ACGP network. It constructs a bearing weak fault diagnosis model based on ACGP network, including a convolutional network, an ACGP network, and a classifier. The convolutional network extracts features from the input vibration signal, and the ACGP network processes the extracted features using channel attention mechanism, cross-channel interaction, and global average pooling. The obtained features are then fed into the classifier to obtain the predicted probability that the input vibration signal belongs to each fault type.

[0018] This invention addresses weak fault diagnosis in bearings. By introducing attention mechanisms and cross-channel interaction, it achieves high-accuracy fault diagnosis in weak fault environments. The network has a small number of parameters, strong noise resistance, and is less prone to overfitting. Compared with traditional fault detection models, the accuracy is improved by approximately 5% to 15%, with a particularly significant improvement in accuracy under weak fault environments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a specific implementation of the bearing weak fault diagnosis method based on ACGP network of the present invention.

[0020] Figure 2 This is a structural diagram of the bearing weak fault diagnosis model based on ACGP network in this embodiment;

[0021] Figure 3 This is the confusion matrix diagram of the present invention on four datasets in this embodiment;

[0022] Figure 4 This is an accuracy curve of the training and verification of the present invention in this embodiment;

[0023] Figure 5 This is a comparison chart of the detection accuracy of the ablation experiment in this embodiment;

[0024] Figure 6This is a visualization of the t-SNE dimensionality reduction of the features extracted by the ACGP network of the present invention in this embodiment. Detailed Implementation

[0025] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0026] Example

[0027] Figure 1 This is a flowchart illustrating a specific implementation of the bearing weak fault diagnosis method based on ACGP networks according to the present invention. Figure 1 As shown, the bearing weak fault diagnosis method based on ACGP network of the present invention includes the following steps:

[0028] S101: Collect samples of minor bearing faults:

[0029] Based on actual needs, fault samples of bearings under early weak fault states of different fault types are collected to construct a fault sample set. Each fault sample includes a length of [length missing]. Vibration signals and their corresponding fault type labels, length The value should be set according to actual needs.

[0030] S102: Constructing a bearing weak fault diagnosis model based on ACGP network:

[0031] To achieve high-accuracy bearing fault diagnosis under weak fault conditions, this invention constructs a bearing weak fault diagnosis model based on ACGP network. Figure 2 This is a structural diagram of the bearing weak fault diagnosis model based on the ACGP network in this embodiment. Figure 2 As shown, the bearing weak fault diagnosis model based on the ACGP network in this embodiment includes a convolutional network, an ACGP network, and a classifier. Each module will be described in detail below.

[0032] Convolutional networks are used to perform convolution operations on input vibration signals, and the resulting convolutional features Send to the ACGP network, where Indicates the number of channels.

[0033] Convolutional networks, through local connectivity and weight sharing mechanisms, can automatically learn local features of input data and progressively abstract high-level feature representations through multi-layer stacking. In image processing, small-sized convolutional kernels (such as 3×3) are widely used to maintain the precision of feature extraction while reducing the number of parameters. However, for one-dimensional vibration signals, the correlation between adjacent sampling points is strong, and fault features often manifest as enhancement of specific frequency components or periodic impacts, which may span a long time interval. If the small convolutional kernel design used in image processing is adopted, a large number of convolutional layers are required to obtain a sufficiently large receptive field to capture complete fault features. This not only increases the network depth and computational burden but may also lead to gradient vanishing or overfitting problems due to excessive layers. To address these issues, this embodiment employs a layered progressive convolutional network, aiming to efficiently compress signal length, highlight key feature regions, and maintain feature discriminativeness. Figure 2 As shown, the convolutional network in this embodiment adopts a two-layer convolutional structure, including a cascaded first convolutional layer and a second convolutional layer. The kernel size of the second convolutional layer is larger than that of the first convolutional layer. The output of each convolutional layer is processed using batch normalization (BN) and ReLU activation function.

[0034] In this embodiment, the first convolutional layer uses a kernel size of 3 and a stride of 1. Its main purpose is to perform preliminary feature transformation and local detail enhancement on the original vibration signal. The first convolutional layer captures the feature patterns between adjacent sampling points through lightweight convolution operations, such as the rising edge or subsequent attenuation information of transient impacts generated by faults and contact between mating surfaces, while maintaining the temporal resolution of the signal.

[0035] The second convolutional layer uses a kernel size of 50 and a stride of 20. This is because the impact caused by the contact between the fault and the mating surface decays very rapidly, while the interval between impacts is relatively long compared to the duration of the impact. Larger kernel sizes can achieve wider temporal window coverage, extract feature patterns of signals with a larger span, and avoid the information transmission efficiency loss caused by stacking multiple layers of small kernels. In addition, the larger stride (20) has a significant downsampling effect. On the one hand, it greatly compresses the temporal length of the feature map and reduces the subsequent computation. On the other hand, it enhances the macroscopic representation of the features through interval sampling and suppresses the interference of high-frequency noise and redundant details.

[0036] ACGP networks are used to process received convolutional features. Perform channel feature extraction and obtain the channel features Send to the classifier.

[0037] After convolution processing, the original one-channel vibration signal is transformed into a multi-channel multivariate feature map, where each channel can be considered as a signal representation extracted from the perspective of different convolutional filters. Theoretically, these channels contain rich complementary information, but their contribution to the final fault classification is not equal: some channels may focus on capturing resonant frequency bands or impact modes strongly correlated with the fault, while others may be more sensitive to background noise or irrelevant vibrations. If all channels are processed equally, key features may be diluted by redundant information, reducing the discriminative power of the features. To address these issues, the ACGP network proposed in this invention includes a channel attention weighting module, a cross-channel interaction module, and a global average pooling module, aiming to adaptively strengthen important channels, suppress secondary channels, and fuse cross-channel information to form a more robust feature representation.

[0038] The channel attention weighting module is used to apply channel attention mechanisms to convolutional features. Extracting Channel Attention Weights And use channel attention weights convolutional features Weighted processing is performed to obtain channel attention features. ,in Represent the Hadamard product, then use channel attention features. Send to the cross-channel interaction module.

[0039] The core idea of ​​the channel attention weighting module is to recalibrate the feature map along the channel dimensions by learning the attention weights for each channel. For example... Figure 2 As shown, the channel attention weighting module in this embodiment includes a global average pooling layer, a global max pooling layer, a dimensionality-reduced fully connected layer, a dimensionality-upgraded fully connected layer, an attention weighting layer, and a weighting layer, wherein:

[0040] Global average pooling layers are used for convolutional features Calculate the global average pooling feature for each channel along the sampling time dimension. And send it to the dimension-reduced fully connected layer.

[0041] Global max pooling layers are used to process convolutional features. Calculate the global max pooling feature for each channel along the sampling time dimension. And send it to the dimension-reduced fully connected layer.

[0042] Global average pooling features It reflects the overall average level of the channel response and is quite sensitive to the overall existence of features; while global max pooling features This captures the most salient response in the channel, paying more attention to the local salience of features. Combining the two allows for a more comprehensive description of the importance of the channel.

[0043] Dimensionality reduction fully connected layers are used to process global average pooling features separately. With global max pooling features Perform dimensionality reduction operations to obtain the global average pooling features. With global max pooling features The data is sent to the upgraded fully connected layer. By compressing the channel descriptors to a lower dimension, computational complexity can be reduced and generalization ability enhanced. In practical applications, the dimensionality reduction rate can be set according to actual needs.

[0044] The dimensional fully connected layer is used to process the global average pooling features separately. With global max pooling features Perform a dimensionality upscaling operation to restore the original dimensionality to its original state, resulting in the globally average pooled feature. With global max pooling features And send it to the attention weight layer.

[0045] Attention weight layers are used to apply globally average pooled features and global max pooling features The features are added together, and then the added features are processed using the sigmoid function to obtain the channel attention weights. And send it to the weighted layer, This represents the sigmoid function.

[0046] Weighted layer channel attention weights convolutional features Weighted processing is performed to obtain channel attention features. .

[0047] By employing channel attention weights to weight convolutional features, we can adaptively enhance the responses of channels containing discriminative fault features while weakening channels dominated by noise or irrelevant features, thereby improving the representation quality of features.

[0048] The cross-channel interaction module is used to apply channel attention features using 1×1 cross-channel convolution operations. Perform cross-channel interaction to obtain channel interaction characteristics. And send it to the global average pooling module.

[0049] While channel attention weighting has assessed the importance of individual channels, fault characteristics in practice often manifest as collaborative changes across multiple channels. For example, a resonant frequency component may simultaneously activate filter channels in multiple adjacent frequency bands. To capture this cross-channel correlation pattern, this invention further introduces a 1×1 cross-channel convolution operation. 1×1 convolution essentially performs a linear combination along the channel dimension, enabling information exchange and fusion between channels without altering the time dimension.

[0050] The global average pooling module is used for channel interaction features. Global average pooling is performed on each channel to obtain the channel features. Global average pooling can transform time-series feature maps into fixed-length feature vectors, maintaining robustness to changes in the time dimension. The resulting channel features... Each channel feature vector integrates information from multiple channels and multiple time steps, preserving the differentiated attention weights between channels while incorporating cross-channel interactive features, thus forming a highly condensed and discriminative fault feature representation.

[0051] A classifier is used to classify based on channel features The predicted probability of the input vibration signal belonging to each fault type is obtained. In this embodiment, the classifier includes three stacked fully connected layers. The outputs of the first and second fully connected layers are processed using batch normalization and the ReLU activation function, and the output of the third fully connected layer is processed using the Softmax function to obtain the predicted probability of each fault type.

[0052] S103: Training the bearing weak fault diagnosis model:

[0053] The bearing weak fault diagnosis model is trained using the fault sample set obtained in step S101, resulting in a trained bearing weak fault diagnosis model. Specific parameters for training the bearing weak fault diagnosis model, such as the loss function and parameter update method, can be set according to actual needs. In this embodiment, cross-entropy loss is used as the loss function during bearing weak fault diagnosis model training.

[0054] S104: Fault Diagnosis:

[0055] When weak fault diagnosis of bearings is required, the data acquisition length is [length to be specified]. The vibration signal is input into the trained bearing weak fault diagnosis model to obtain the fault diagnosis result.

[0056] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the present invention.

[0057] In this embodiment, four datasets were used to verify the invention: Harbin Institute of Technology (HIT), Jiangnan University (JNU), Xi'an Jiaotong University (XJTU), and Western Reserve University (CWRU). The sample sizes of the four datasets were 5858, 4755, 48240, and 1440, respectively. Among the four datasets, the HIT dataset was used as a weak fault dataset because of its low signal-to-noise ratio of fault features, high confusion between fault and normal sample features, complex operating conditions, and unbalanced distribution of fault types and samples, making it difficult to effectively extract and distinguish fault modes using conventional methods. In all datasets except the HIT dataset, the model training time did not exceed 500 seconds. Figure 3 This is a confusion matrix diagram of the present invention on four datasets in this embodiment. For example... Figure 3 As shown, the present invention achieves an accuracy of over 96% on all four datasets, with an average accuracy of 98.41%. The test accuracy on the HIT and XJTU datasets exceeds 99%. Since the four datasets are of different sizes, this indicates that the present invention can converge well for classification tasks of varying numbers and dataset sizes, demonstrating the universality of the model's learning ability.

[0058] Figure 4 This is an accuracy curve graph of the training and verification of the present invention in this embodiment. For example... Figure 4 As shown, in this embodiment, the bearing weak fault diagnosis model of the present invention converges quickly, and the validation accuracy curve is basically consistent with the training accuracy curve. There are no cases where the validation accuracy is much lower than the training accuracy or the accuracy curve fluctuates greatly. This indicates that the model effectively suppresses overfitting, has good convergence, and performs stably during training.

[0059] To illustrate the technical effectiveness of the ACGP network in the bearing weak fault diagnosis model of this invention, an ablation experiment was designed. The detection accuracy of the model with the ACGP network removed from the model of this invention, as well as the conventional CNN+FC model, was compared with the complete model of this invention on four datasets. Figure 5 This is a comparison chart of the detection accuracy of the ablation experiment in this embodiment. For example... Figure 5 As shown, the complete model of this invention has the highest detection accuracy on the four datasets.

[0060] Figure 6 This is a visualization of the t-SNE dimensionality reduction of the features extracted by the ACGP network of this invention in this embodiment. For example... Figure 6As shown, in this embodiment, most samples of the same category cluster into independent clusters with relatively clear boundaries, indicating significant differences in the features extracted by the ACGP network. Furthermore, the different clusters are widely distributed on the graph, covering the area from the lower left corner (red cluster) to the upper right corner (gray cluster), demonstrating the spatial extensibility of the features extracted by the ACGP network after dimensionality reduction. In summary, the features extracted by the ACGP network of this invention exhibit good separability, thereby improving the detection accuracy for early-stage weak bearing faults.

[0061] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A bearing weak fault diagnosis method based on ACGP network, characterized in that, Includes the following steps: S1: Collect fault samples of bearings in early weak fault states of different fault types according to actual needs, and construct a fault sample set. Each fault sample includes a length of [length missing]. Vibration signals and their corresponding fault type labels, length The value should be set according to actual needs; S2: Construct a bearing weak fault diagnosis model based on the ACGP network, including a convolutional network, an ACGP network, and a classifier, where: Convolutional networks are used to perform convolution operations on input vibration signals, and the resulting convolutional features Send to the ACGP network, where Indicates the number of channels; ACGP networks are used to process received convolutional features. Perform channel feature extraction and obtain the channel features The data is sent to the classifier; the ACGP network includes a channel attention weighting module, a cross-channel interaction module, and a global average pooling module, where: The channel attention weighting module is used to apply channel attention mechanisms to convolutional features. Extracting Channel Attention Weights And use channel attention weights convolutional features Weighted processing is performed to obtain channel attention features. ,in Represent the Hadamard product, then use channel attention features. Send to the cross-channel interaction module; The cross-channel interaction module is used to apply channel attention features using 1×1 cross-channel convolution operations. Perform cross-channel interaction to obtain channel interaction characteristics. And send it to the global average pooling module; The global average pooling module is used for channel interaction features. Perform global average pooling to obtain channel features. ; A classifier is used to classify based on channel features The predicted probability that the input vibration signal belongs to each fault type is obtained; S3: Use the fault sample set from step S1 to train the bearing weak fault diagnosis model, and obtain the trained bearing weak fault diagnosis model. S4: When weak fault diagnosis of bearings is required, the data acquisition length is... The vibration signal is input into the trained bearing weak fault diagnosis model to obtain the fault diagnosis result.

2. The bearing weak fault diagnosis method according to claim 1, characterized in that, The convolutional network includes a first convolutional layer and a second convolutional layer, wherein the kernel size of the second convolutional layer is larger than that of the first convolutional layer, and the output of each convolutional layer is processed using batch normalization and the ReLU activation function.

3. The bearing weak fault diagnosis method according to claim 1, characterized in that, The channel attention weighting module includes a global average pooling layer, a global max pooling layer, a dimensionality-reduced fully connected layer, a dimensionality-increasing fully connected layer, an attention weight layer, and a weighting layer, wherein: Global average pooling layers are used for convolutional features Calculate the global average pooling feature for each channel along the sampling time dimension. And send it to the dimension-reduced fully connected layer; Global max pooling layers are used to process convolutional features. Calculate the global max pooling feature for each channel along the sampling time dimension. And send it to the dimension-reduced fully connected layer; Dimensionality reduction fully connected layers are used to process global average pooling features separately. With global max pooling features Perform dimensionality reduction operations to obtain the global average pooling features. With global max pooling features Send to the upgraded fully connected layer; The dimensional fully connected layer is used to process the global average pooling features separately. With global max pooling features Perform a dimensionality upscaling operation to restore the original dimensionality to its original state, resulting in the globally average pooled feature. With global max pooling features And send it to the attention weight layer; Attention weight layers are used to apply globally average pooled features and global max pooling features The features are added together, and then the added features are processed using the sigmoid function to obtain the channel attention weights. And send it to the weighted layer, Represents the sigmoid function; Weighted layer channel attention weights convolutional features Weighted processing is performed to obtain channel attention features. .

4. The bearing weak fault diagnosis method according to claim 1, characterized in that, The classifier comprises three stacked fully connected layers. The outputs of the first and second fully connected layers are processed using batch normalization and the ReLU activation function, and the output of the third fully connected layer is processed by the Softmax function to obtain the predicted probability of each fault type.

Images