Rolling bearing fault diagnosis method and device based on multi-scale residual attention

Abstract

The application discloses a rolling bearing fault diagnosis method and device based on multi-scale residual attention, and relates to the field of mechanical equipment fault diagnosis.The method comprises the following steps: S1, obtaining a vibration monitoring signal of a rolling bearing and preprocessing; S2, constructing a rolling bearing fault diagnosis model and training; S3, using the trained model to perform rolling bearing fault diagnosis.The feature extraction layer of the rolling bearing fault diagnosis model extracts local impact features under different convolution scales through a multi-scale feature fusion module, expands the time domain receptive field and captures periodic fault features under different time spans through a multi-scale hollow convolution module, and fuses local convolution features, time sequence dependent features and key channel features based on attention through a dense residual attention module and performs residual fusion, thereby improving the problems of insufficient fault feature extraction, insufficient time sequence dependent modeling and weak diagnosis robustness of traditional methods under strong noise and complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment fault diagnosis technology, and in particular to a method and apparatus for diagnosing rolling bearing faults based on multi-scale residual attention. Background Technology

[0002] Rolling bearings are critical support components in rotating machinery, widely used in motors, pumps, fans, gearboxes, machine tool spindles, and various industrial transmission systems. During long-term operation, rolling bearings are susceptible to damage from alternating loads, insufficient lubrication, installation misalignment, impact vibration, and complex operating conditions, leading to various forms of damage, including inner ring failure, outer ring failure, rolling element failure, and cage failure. When rolling bearing failures are not detected in time, they can further exacerbate equipment vibration, reduce transmission accuracy, increase component wear, and even cause complete machine downtime and production accidents. Therefore, timely and accurate identification of the operating status of rolling bearings is crucial for ensuring the safe operation of rotating machinery, reducing maintenance costs, and improving the reliability of industrial production.

[0003] Currently, rolling bearing fault diagnosis methods mainly include those based on human experience, those based on signal processing, and those based on data. Methods based on human experience typically rely on maintenance personnel to judge operational phenomena such as vibration, noise, and temperature rise. While simple to implement, the diagnostic results are heavily influenced by human experience, making it difficult to meet the needs of continuous monitoring and early fault identification in complex industrial environments. Methods based on signal processing typically extract fault features through Fourier transform, wavelet transform, empirical mode decomposition, and envelope spectrum analysis. These methods can reveal periodic impact components in vibration signals to some extent. However, they usually require manual feature selection, setting of decomposition parameters, or reliance on prior knowledge. Under complex operating conditions such as high noise, variable speed, and variable load, they are prone to problems such as insufficient feature extraction, poor parameter adaptability, and insufficient generalization ability.

[0004] With the development of sensor and artificial intelligence technologies, data-driven fault diagnosis methods based on deep learning have gradually become an important research direction in the field of mechanical equipment health monitoring. Convolutional neural networks (CNNs) can directly extract local fault features from one-dimensional vibration signals, avoiding extensive manual feature design, and have shown good recognition capabilities in rolling bearing fault diagnosis. However, traditional CNNs typically use a fixed kernel scale for feature extraction, making it difficult to simultaneously capture weak impact features, periodic fluctuation features, and long-range dependency features at different time scales. When rolling bearings are in the early stages of faulting or under strong background noise, the impact component of the fault is often weak, and single-scale convolutional structures are prone to missing key local features, thus affecting diagnostic accuracy.

[0005] Furthermore, while existing multi-scale convolutional networks can extract features at different scales through multiple convolutional kernels, some networks only extend at the local convolutional scale, lacking further modeling of periodic fault features under different receptive fields, making it difficult to fully characterize the repetitive impact patterns across time spans in bearing fault signals. Simultaneously, some attention mechanisms simply weight channel features, failing to fully integrate local convolutional features and temporal dependencies between consecutive time steps, resulting in limited robustness of the model under complex operating conditions and noise interference. Summary of the Invention

[0006] To address the above problems, this invention proposes a rolling bearing fault diagnosis method and device based on multi-scale residual attention. The method extracts local impact features at different convolutional scales through a multi-scale feature fusion module, expands the temporal receptive field and captures periodic fault features across different time spans through a multi-scale dilated convolution module, and fuses local convolutional features, temporal dependency features, and attention-based key channel features using a dense residual attention module, thereby improving the shortcomings of traditional methods such as insufficient fault feature extraction, inadequate temporal dependency modeling, and weak diagnostic robustness under strong noise and complex operating conditions.

[0007] On the one hand, the rolling bearing fault diagnosis method based on multi-scale residual attention specifically includes:

[0008] S1, acquire vibration monitoring signals during the operation of the rolling bearing and perform preprocessing;

[0009] S2, construct and train a rolling bearing fault diagnosis model to obtain a trained rolling bearing fault diagnosis model; the rolling bearing fault diagnosis model includes several feature extraction layers, a global average pooling layer and a classification layer connected in sequence; all the feature extraction layers are stacked in ascending order of the number of channels; each feature extraction layer includes a multi-scale feature fusion module, a multi-scale dilated convolution module, a dense residual attention module and a channel adaptation module.

[0010] The preprocessed vibration monitoring signal is processed by a multi-scale feature fusion module to extract local impact features at different convolutional scales, resulting in local impact features. These local impact features are then processed by a multi-scale dilated convolution module to extract periodic fault features under different receptive fields, resulting in periodic fault features. The periodic fault features are further processed by a dense residual attention module to extract local features, temporal-dependent features, and attention-based key channel features, followed by residual fusion to obtain fused and enhanced features. These enhanced features are then processed by a channel adaptation module for channel adjustment and output. Finally, the output of the channel adaptation module in the last feature extraction layer is processed by a global average pooling layer and a classification layer to output a rolling bearing fault classification prediction.

[0011] S3. Use the trained rolling bearing fault diagnosis model to perform rolling bearing fault diagnosis and obtain rolling bearing fault classification prediction.

[0012] Preferably, the vibration monitoring signal during the operation of the rolling bearing includes a one-dimensional vibration acceleration signal during the operation of the rolling bearing; the one-dimensional vibration acceleration signal includes a normal state vibration signal and a fault state vibration signal; the fault state vibration signal includes at least one of the following: inner ring fault vibration signal, outer ring fault vibration signal, rolling element fault vibration signal, and cage fault vibration signal, distinguished according to the fault location.

[0013] Preferably, the fault category labels of the training set of the rolling bearing fault diagnosis model include: setting corresponding fault category labels for vibration signals under normal conditions; and setting corresponding fault category labels for vibration signals under fault conditions at the same fault location according to the degree of fault.

[0014] Preferably, the preprocessing includes abnormal data filtering, missing data completion, fixed-length sample truncation, and normalization.

[0015] Preferably, the multi-scale feature fusion module is represented as follows:

[0016] ;

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] in, This represents the fine-scale features output by the first convolutional branch; Represents a non-linear activation function; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 3. Indicates the first The input features of each feature extraction layer; This represents the mesoscale features output by the second convolutional branch; This represents a one-dimensional convolution operation with a kernel size of 5. This represents the coarse-scale features output by the third convolutional branch; This demonstrates a one-dimensional convolution operation with a kernel size of 7. This represents the pooling characteristics of the pooling branch output; This represents a one-dimensional pointwise convolution operation with a kernel size of 1. This represents the max pooling operation; This indicates the multi-branch feature after channel splicing; This indicates a channel-level concatenation operation; Indicates the first The multi-scale fused features are output by the multi-scale feature fusion module in each feature extraction layer.

[0023] Preferably, the multi-scale dilated convolution module is represented as follows:

[0024] ;

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] This represents the short receptive field feature output by the first dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 1. Represents a non-linear activation function; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 3. Indicates the first The multi-scale fused features output by the multi-scale feature fusion module in each feature extraction layer; This represents the mid-receptive field feature of the output of the second dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 2. This represents the long receptive field feature of the output of the third dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 4. This indicates a channel-level concatenation operation; This indicates the multi-receptor field characteristics after channel splicing; No. Periodic fault features output by the multi-scale dilated convolution module in each feature extraction layer.

[0030] Preferably, the dense residual attention module is represented as follows:

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] in, This represents the local short-scale impact feature of the output of the first convolutional path; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 5. Indicates the characteristics of periodic faults; This represents the local mesoscale impact feature of the output of the second convolutional path; This represents the fusion feature of dual-path convolution; This represents the temporal dependency features output by the temporal modeling unit; This represents a one-dimensional pointwise convolution operation with a kernel size of 1. This represents a bidirectional long short-term memory network; This indicates a feature dimension transformation operation; Indicates channel attention weights; This represents the Sigmoid activation function; This indicates a global average pooling operation; Represents a non-linear activation function; Indicates attention fusion features; This indicates a channel-by-channel multiplication operation; Indicates the first The fused enhanced features are output by the dense residual attention modules in each feature extraction layer.

[0039] Preferably, the loss function of the rolling bearing fault diagnosis model is the cross-entropy loss function.

[0040] On the one hand, the rolling bearing fault diagnosis method based on multi-scale residual attention has the following specific steps:

[0041] On the other hand, the rolling bearing fault diagnosis device based on multi-scale residual attention includes the following:

[0042] The monitoring signal acquisition and preprocessing module is used to acquire and preprocess the vibration monitoring signals during the operation of the rolling bearing.

[0043] The fault diagnosis model construction and training module is used to construct and train a rolling bearing fault diagnosis model to obtain a trained rolling bearing fault diagnosis model. The rolling bearing fault diagnosis model includes several feature extraction layers, a global average pooling layer, and a classification layer connected in sequence. All the feature extraction layers are stacked in ascending order of the number of channels. Each feature extraction layer includes a multi-scale feature fusion module, a multi-scale dilated convolution module, a dense residual attention module, and a channel adaptation module.

[0044] The preprocessed vibration monitoring signal is processed by a multi-scale feature fusion module to extract local impact features at different convolutional scales, resulting in local impact features. These local impact features are then processed by a multi-scale dilated convolution module to extract periodic fault features under different receptive fields, resulting in periodic fault features. The periodic fault features are further processed by a dense residual attention module to extract local features, temporal-dependent features, and attention-based key channel features, followed by residual fusion to obtain fused and enhanced features. These enhanced features are then processed by a channel adaptation module for channel adjustment and output. Finally, the output of the channel adaptation module in the last feature extraction layer is processed by a global average pooling layer and a classification layer to output a rolling bearing fault classification prediction.

[0045] The rolling bearing fault classification and prediction module is used to perform rolling bearing fault diagnosis using a trained rolling bearing fault diagnosis model, and to obtain rolling bearing fault classification and prediction.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] (1) This invention enables the enhanced multi-scale residual attention network to simultaneously extract fine-scale local impact features, mesoscale periodic fluctuation features, coarse-scale trend features and pooling stability features from rolling bearing vibration signals through a multi-scale feature fusion module, thereby enhancing the model’s ability to represent different fault impact scales.

[0048] (2) This invention enables the enhanced multi-scale residual attention network to extract local impact information, periodic repetition information and long-range dependency information in the rolling bearing vibration signal within the receptive field range corresponding to different void ratios through the multi-scale dilated convolution module, thereby improving the model's ability to represent weak fault impacts and complex periodic fault characteristics.

[0049] (3) Through the dense residual attention module, the enhanced multi-scale residual attention network can simultaneously utilize the fault impact features, bidirectional temporal dependency features and channel attention weights under different local convolution scales to adaptively enhance the key fault features in the rolling bearing vibration signal, and retain the original input features through residual connections to alleviate the gradient decay problem in the deep network training process.

[0050] (4) This invention achieves full extraction of local impact features, periodic fault features and time-dependent features in rolling bearing vibration signals through the synergistic effect of multi-scale feature fusion module, multi-scale dilated convolution module and dense residual attention module, and enhances the expression of key fault features through channel attention and residual connection. Compared with traditional manual feature extraction methods and single-scale convolutional network methods, this invention can improve the accuracy, robustness and adaptability of rolling bearing fault diagnosis under complex working conditions and strong noise environment. Attached Figure Description

[0051] The present invention will now be described in further detail with reference to the accompanying drawings;

[0052] Figure 1 This is a flowchart of a rolling bearing fault diagnosis method based on multi-scale residual attention, according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart of vibration signal preprocessing and sample construction for a rolling bearing fault diagnosis method based on multi-scale residual attention, according to an embodiment of the present invention.

[0054] Figure 3 This is a diagram illustrating the overall architecture of the rolling bearing fault diagnosis method based on multi-scale residual attention, as described in this embodiment of the invention.

[0055] Figure 4 This is a schematic diagram of the structure of the multi-scale feature fusion module in the rolling bearing fault diagnosis method based on multi-scale residual attention according to an embodiment of the present invention.

[0056] Figure 5 This is a schematic diagram of the structure of the multi-scale dilated convolution module of the rolling bearing fault diagnosis method based on multi-scale residual attention in an embodiment of the present invention.

[0057] Figure 6 This is a schematic diagram of the dense residual attention module in the rolling bearing fault diagnosis method based on multi-scale residual attention according to an embodiment of the present invention.

[0058] Figure 7 This is a flowchart of the fault diagnosis model testing and fault category output of the rolling bearing fault diagnosis method based on multi-scale residual attention according to an embodiment of the present invention.

[0059] Figure 8 This is a structural block diagram of a rolling bearing fault diagnosis device based on multi-scale residual attention, according to an embodiment of the present invention. Detailed Implementation

[0060] The present invention will be further described below through specific embodiments.

[0061] like Figure 1As shown, the specific steps of the rolling bearing fault diagnosis method based on multi-scale residual attention are as follows:

[0062] S1: Acquire vibration monitoring signals during the operation of the rolling bearing and perform preprocessing.

[0063] Vibration monitoring data of the rolling bearing is acquired and preprocessed. Specifically, the vibration monitoring data consists of one-dimensional vibration signals collected during the operation of the rolling bearing. The preprocessing steps for the vibration monitoring data are the same as those for the test set, including outlier filtering, missing value completion, fixed-length sample truncation, and normalization, thereby obtaining vibration sample data that meets the input requirements of the fault diagnosis model. For the steps of acquiring and preprocessing vibration monitoring data of the rolling bearing, please refer to [link to relevant documentation]. Figure 2 As shown, the specific steps include the following:

[0064] S11, acquire historical vibration signals of the rolling bearing under different health and fault conditions.

[0065] Specifically, one-dimensional vibration signals of the rolling bearing under different operating conditions are collected by vibration sensors installed on the rolling bearing, bearing housing, or rotating machinery housing. Historical vibration signals include vibration signals under normal conditions and vibration signals under fault conditions; the vibration signals under fault conditions include at least one of the following: inner ring fault vibration signals, outer ring fault vibration signals, rolling element fault vibration signals, and cage fault vibration signals.

[0066] S12 preprocesses historical vibration signals to filter out abnormal data and obtain vibration sample data of a fixed length.

[0067] Specifically, such as Figure 3 As shown, the process first determines whether there are missing sampling points, abnormal amplitude points, or abnormal fluctuation segments in the historical vibration signal. If abnormal data is found in the historical vibration signal, the abnormal data is deleted or corrected; if there are data gaps in the historical vibration signal, they are filled in using adjacent valid sampling points.

[0068] After anomaly filtering and missing data completion, the historical vibration signal is truncated according to a preset window length and sliding step size to obtain multiple vibration sample data of fixed length. Let the historical vibration signal be: ;No. Each vibration sample is represented as: Among them, the first The starting sampling point for each vibration sample is: In the formula, This represents the historical vibration signal of the rolling bearing. Indicates the first Vibration amplitude at each sampling point This represents the total number of sampling points for historical vibration signals. Indicates the first A vibration sample of fixed length, Indicates the preset window length. Indicates the sliding step size. Indicates the first The starting sampling point of each vibration sample.

[0069] when At that time, there is overlap between adjacent vibration samples, which is used to increase the number of samples and improve the coverage of fault impact segments; when At this time, there is no overlap between adjacent vibration samples, which is used to sample historical vibration signals at equal intervals to obtain vibration sample data of a fixed length.

[0070] S13, based on the health status, fault location, fault type, or fault severity of the rolling bearing, set corresponding fault category labels for the vibration sample data.

[0071] Specifically, based on the operating state of the rolling bearing, the vibration sample data is divided into normal state samples and fault state samples. The fault category label corresponding to the normal state samples is set to 0, i.e.: In the formula, Indicates the first Fault category label for each vibration sample Indicates the first The vibration sample belongs to the normal state category.

[0072] For fault status samples, set corresponding fault category labels based on fault location, fault type, or fault severity. When the... The vibration sample belongs to the first When a fault condition is detected, its fault category label is set as follows: In the formula, Indicates the fault category number, Indicates the total number of rolling bearing condition categories. This represents the total number of fault condition categories. For example, fault categories can include normal condition, inner ring fault, outer ring fault, rolling element fault, and cage fault. They can also be further subdivided into different fault degrees based on actual diagnostic needs. The set of rolling bearing condition categories combining normal condition, different fault locations, and different fault degrees is represented as: In the formula, This represents a set of labels for rolling bearing failure categories.

[0073] Based on the vibration sample data and the corresponding fault category labels, a labeled sample dataset is constructed: In the formula, This represents a labeled sample dataset. Indicates the first A normalized vibration sample, Indicates the first Fault category label corresponding to each vibration sample This represents the total number of vibration samples.

[0074] S14. The vibration sample data with fault category labels is divided into training set, validation set, and test set. Specifically, the vibration sample data under normal conditions is divided into normal training samples, normal validation samples, and normal test samples according to a preset ratio; the vibration sample data under each fault condition is divided into fault training samples, fault validation samples, and fault test samples according to a preset ratio.

[0075] Then, normal training samples and fault training samples are combined into a training set, normal validation samples and fault validation samples are combined into a validation set, and normal test samples and fault test samples are combined into a test set. The training set is used for parameter learning of the fault diagnosis model, the validation set is used for parameter selection and early stopping judgment during model training, and the test set is used for performance evaluation after model training is completed.

[0076] S15 performs normalization processing on the vibration sample data in the training set, validation set, and test set respectively to eliminate amplitude scale differences between different vibration samples. Specifically, suppose the training set contains... There are 3 vibration samples, each with a length of 1. ,in, Indicates the first training set The vibration sample at the ... Vibration amplitude at each sampling point.

[0077] Based on the vibration amplitudes of all vibration samples in the training set, calculate the global minimum amplitude and global maximum amplitude of the training set: In the formula, This represents the global minimum amplitude of all vibration samples in the training set. This represents the global maximum amplitude value of all vibration samples in the training set.

[0078] Based on the global minimum amplitude and global maximum amplitude, the vibration sample data in the training set, validation set, and test set are subjected to maximum and minimum value normalization. The normalization model is as follows: In the formula, Indicates the first The vibration sample of the first Normalized vibration amplitude values ​​from each sampling point This represents a very small constant used to prevent the denominator from being zero.

[0079] Normalized Each vibration sample is represented as: ;

[0080] And adjust it to the input format of a one-dimensional convolutional network: ;in, Indicates the number of input channels. Indicates the length of the vibration sample.

[0081] S2, construct and train the rolling bearing fault diagnosis model to obtain the trained rolling bearing fault diagnosis model.

[0082] like Figure 3 As shown, a rolling bearing fault diagnosis model based on multi-scale residual attention is presented.

[0083] Construct an enhanced multi-scale residual attention network. Specifically, such as... Figure 3 As shown, the enhanced multi-scale residual attention network comprises multiple sequentially connected feature extraction layers, global average pooling layers, and classification layers. Each feature extraction layer includes a multi-scale feature fusion module (MFF), a multi-scale dilated convolution module (MSDC), a dense residual attention module (DRB), and a channel adaptation module (CA). Let the th... The input features of each feature extraction layer are Then the first The output features of each feature extraction layer Represented as:

[0084] ;

[0085] In the formula, Indicates the first The multi-scale feature fusion module in the feature extraction layer Indicates the first Multi-scale dilated convolutional modules in each feature extraction layer Indicates the first Dense residual attention modules in each feature extraction layer Indicates the first Channel adaptation module in each feature extraction layer.

[0086] In a preferred embodiment, the enhanced multi-scale residual attention network includes three feature extraction layers. The first feature extraction layer has 1 input channel and 16 output channels; the second feature extraction layer has 16 input channels and 32 output channels; and the third feature extraction layer has 32 input channels and 64 output channels. By progressively increasing the number of channels, shallow local impact features and deep fault discrimination features in the rolling bearing vibration signal can be extracted layer by layer.

[0087] The multi-scale feature fusion module extracts local impact features at different convolution scales from the normalized vibration sample data; and the multi-scale dilated convolution module extracts periodic fault features from the local impact features under different receptive fields.

[0088] Specifically, such as Figure 4 As shown, the multi-scale feature fusion module includes a first convolution branch, a second convolution branch, a third convolution branch, and a pooling branch. The first convolution branch uses a one-dimensional convolution with a kernel size of 3, the second convolution branch uses a one-dimensional convolution with a kernel size of 5, the third convolution branch uses a one-dimensional convolution with a kernel size of 7, and the pooling branch uses max pooling and a one-dimensional pointwise convolution with a kernel size of 1.

[0089] The outputs of the first convolution branch, the second convolution branch, the third convolution branch, and the pooling branch are respectively represented as:

[0090] ;

[0091] ;

[0092] ;

[0093] ;

[0094] In the formula, Represents fine-scale features. Indicates mesoscale features. Indicates coarse-scale features, Represents pooling characteristics, This indicates a batch normalization operation. This represents a non-linear activation function.

[0095] The features output from the four branches are concatenated along the channel dimension, and then fused using a one-dimensional pointwise convolution with a kernel size of 1 to obtain multi-scale fused features:

[0096] ;

[0097] ;

[0098] In the formula, This indicates a channel-level concatenation operation. This represents the multi-scale fused features output by the multi-scale feature fusion module.

[0099] A multi-scale feature fusion module is used to replace the basic convolutional module with a single convolutional scale, to avoid the problem that single-scale convolution cannot simultaneously represent the fault impact components at different time scales. Through the multi-scale feature fusion module, the enhanced multi-scale residual attention network can simultaneously extract fine-scale local impact features, mesoscale periodic fluctuation features, coarse-scale trend features, and pooling stability features from rolling bearing vibration signals, thereby enhancing the model's ability to represent different fault impact scales.

[0100] like Figure 5 As shown, the multi-scale dilated convolution module includes multiple parallel dilated convolution branches. In a preferred embodiment, the multi-scale dilated convolution module includes three dilated convolution branches with dilation rates of 1, 2, and 4, respectively, for extracting periodic fault features under short, medium, and long receptive fields.

[0101] Let the multi-scale fused features output by the multi-scale feature fusion module be... ,but:

[0102] ;

[0103] The outputs of the three dilated convolution branches are represented as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] The features output from the three dilated convolution branches are concatenated along the channel dimension, and then batch normalization and non-linear activation are performed sequentially to obtain the output features of the multi-scale dilated convolution module:

[0108] ;

[0109] ;

[0110] In the formula, This represents the periodic fault characteristics output by the multi-scale dilated convolution module. Through the above processing, local impact information and periodic repetition information can be extracted from the rolling bearing vibration signal within different receptive fields.

[0111] By using a multi-scale dilated convolution module, the enhanced multi-scale residual attention network can extract local impact information, periodic repetition information, and long-range dependency information from rolling bearing vibration signals within the receptive field corresponding to different dilation rates, thereby improving the model's ability to represent weak fault impacts and complex periodic fault characteristics.

[0112] The local impact features, periodic fault features, time-dependent features, and key channel features are fused and enhanced using a dense residual attention module. The fault diagnosis model is then trained based on the fused and enhanced features to establish a nonlinear mapping relationship between vibration sample data and fault category labels.

[0113] Specifically, such as Figure 6 As shown, the dense residual attention module includes a dual-path convolutional unit, a temporal modeling unit, a channel attention unit, and a residual connection unit. Let the first... The periodic fault features output by the multi-scale dilated convolutional module in each feature extraction layer are: ,but: ;

[0114] A dual-path convolutional unit includes a first convolutional path and a second convolutional path. The first convolutional path uses a one-dimensional convolution with a kernel size of 3, and the second convolutional path uses a one-dimensional convolution with a kernel size of 5. The outputs of the first and second convolutional paths are represented as follows:

[0115] ;

[0116] ;

[0117] The features output from the first and second convolutional paths are fused to obtain dual-path convolutional fused features:

[0118] ;

[0119] In the formula, This represents the dual-path convolutional fusion feature. The temporal modeling unit uses a bidirectional long short-term memory network to extract the temporal dependency relationship between sampling times, resulting in temporal dependency features:

[0120] ;

[0121] In the formula, Representing time-dependent features, This indicates a feature dimension transformation operation. This represents a bidirectional long short-term memory network.

[0122] The channel attention unit calculates channel weights for the dual-path convolutional fusion features, resulting in channel attention weights:

[0123] ;

[0124] In the formula, Indicates channel attention weights. This indicates a global average pooling operation. Represents a non-linear activation function. This represents the Sigmoid activation function. The dual-channel convolutional fusion features are recalibrated based on the channel attention weights and then fused with the temporal dependency features to obtain the attention-fused features: In the formula, Indicates attention fusion features, This indicates a channel-by-channel multiplication operation.

[0125] By using residual connection units, the input features of the dense residual attention module are added to the attention fusion features to obtain the output features of the dense residual attention module:

[0126] ;

[0127] In the formula, Indicates the first Fusion-enhanced features output by dense residual attention modules in each feature extraction layer;

[0128] The dense residual attention module enables the enhanced multi-scale residual attention network to simultaneously utilize fault impact features, bidirectional temporal dependency features, and channel attention weights at different local convolutional scales to adaptively enhance key fault features in rolling bearing vibration signals. Furthermore, the original input features are preserved through residual connections to alleviate the gradient decay problem during deep network training.

[0129] The fused and enhanced features output from the dense residual attention module are processed by the channel adaptation module and then input into the next feature extraction layer. The fused and enhanced features output from the last feature extraction layer are input into the global average pooling layer to obtain the global feature vector.

[0130] ;

[0131] In the formula, Indicates the first The global feature vector corresponding to each vibration sample This represents the fused and enhanced feature output from the last feature extraction layer. The global feature vector is then input into the classification layer to obtain the... The classification scores for each fault category corresponding to each vibration sample: In the formula, Represents the classification score vector. This represents the weight matrix of the classification layer. This represents the bias vector of the classification layer.

[0132] The classification scores are converted into predicted probabilities for each fault category using the Softmax function.

[0133] ;

[0134] In the formula, Indicates the first The vibration sample belongs to the first Predicted probability of fault category; Indicates the first The vibration sample corresponds to the first Classification scores for fault categories, This indicates the total number of rolling bearing condition categories.

[0135] During model training, the cross-entropy loss function is used to measure the difference between the predicted probability and the true fault category label:

[0136] ;

[0137] In the formula, Indicates training loss, This indicates the number of vibration samples in the training set. Indicates the first The true fault category label for each vibration sample Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise. Indicates the first The vibration sample belongs to the first The predicted probability of each fault category is calculated. The model parameters of the enhanced multi-scale residual attention network are iteratively updated based on the cross-entropy loss function to establish a nonlinear mapping relationship between vibration sample data and fault category labels, thereby obtaining the trained rolling bearing fault diagnosis model.

[0138] The trained fault diagnosis model is tested using a normalized test set until it is deemed satisfactory, resulting in a pre-trained rolling bearing fault diagnosis model based on multi-scale residual attention. Otherwise, the model is retrained. The model's suitability is determined by the accuracy, precision, recall, and overall evaluation metrics of the test results.

[0139] Specifically, such as Figure 7 As shown, the normalized test set is input into the trained fault diagnosis model to obtain the predicted fault category for each vibration sample in the test set. The fault category is determined based on the maximum predicted probability. Predicted fault category for each vibration sample: In the formula, Indicates the first Predicted fault category for each vibration sample This represents the fault category number that maximizes the predicted probability.

[0140] Based on the actual and predicted fault categories for each vibration sample in the test set, accuracy, precision, recall, and a comprehensive evaluation metric are calculated. The formula for calculating accuracy is:

[0141] ;

[0142] In the formula, Indicates accuracy rate. This indicates the number of vibration samples that were correctly predicted in the test set. This indicates the total number of vibration samples in the test set;

[0143] The formula for calculating accuracy is:

[0144] ;

[0145] The formula for calculating recall rate is:

[0146] ;

[0147] The formula for calculating the comprehensive evaluation index is as follows:

[0148] ;

[0149] In the formula, Indicates accuracy. Indicates recall rate, This represents the comprehensive evaluation indicators. Indicates the actual number of cases. Indicates the number of false positives. Indicates the number of false negatives. This represents a very small constant used to prevent the denominator from being zero.

[0150] When the accuracy, precision, recall, and comprehensive evaluation index all meet the preset thresholds, the trained fault diagnosis model is deemed qualified, and a pre-trained rolling bearing fault diagnosis model based on multi-scale residual attention is obtained; otherwise, the model training parameters, network structure parameters, or training rounds are adjusted, and the model is retrained until it meets the fault diagnosis requirements.

[0151] S3. Use the trained rolling bearing fault diagnosis model to perform rolling bearing fault diagnosis and obtain rolling bearing fault classification prediction.

[0152] The preprocessed vibration monitoring data is input into a pre-trained rolling bearing fault diagnosis model based on multi-scale residual attention to obtain the fault category of the rolling bearing.

[0153] Specifically, the pre-trained rolling bearing fault diagnosis model has established a nonlinear mapping relationship between vibration sample data and fault category labels. After inputting the preprocessed vibration monitoring data into the model, the corresponding fault category can be output. The fault category includes at least one of the following: normal state, inner ring fault, outer ring fault, rolling element fault, and cage fault.

[0154] To verify the effectiveness of the proposed fault diagnosis model, the PU rolling bearing dataset was selected as the experimental object. A 14-category rolling bearing fault classification task was constructed, with the input vibration sample length set to 2048, the number of input channels to 1, and the number of fault categories to 14. The feature channels of the enhanced multi-scale residual attention network were configured as [1, 16, 32, 64]. The multi-scale feature fusion module used one-dimensional convolutional branches with kernel sizes of 3, 5, and 7, as well as max pooling branches. The multi-scale dilated convolutional module used parallel dilated convolutional branches with dilation rates of 1, 2, and 4. The dense residual attention module used a dual-path convolutional structure with kernel sizes of 3 and 5, and combined bidirectional long short-term memory networks and channel attention mechanisms for feature fusion. The model output used a global average pooling layer and a fully connected classification layer. During training, the cross-entropy loss function was used to optimize the model parameters. In the experiment, the test samples were evaluated under different signal-to-noise ratio conditions, and accuracy, precision, recall, and F1 score were used as model performance evaluation indicators. The experimental parameter settings are shown in Table 1.

[0155] Table 1: Experimental parameter settings.

[0156]

[0157] As shown in Table 2, in the 14-category classification task of the PU rolling bearing dataset, the rolling bearing fault diagnosis model based on multi-scale residual attention constructed in this invention achieved high classification performance under different signal-to-noise ratio (SNR) conditions. When the SNR was 0 dB, the model's accuracy, precision, recall, and F1 score all reached 99.91%; when the SNR decreased to -8 dB, the model's accuracy still reached 96.95%, and the F1 score reached 96.96%. Experimental results show that the model of this method can maintain good fault recognition ability under noise interference conditions, indicating that the multi-scale feature fusion module, multi-scale dilated convolution module, and dense residual attention module can effectively enhance the expression ability of key fault features in rolling bearing vibration signals, and improve the diagnostic accuracy and robustness of the model in complex noise environments.

[0158] Table 2: Classification results of fault diagnosis models under different signal-to-noise ratios.

[0159]

[0160] like Figure 8 As shown, the present invention also discloses a rolling bearing fault diagnosis device based on multi-scale residual attention, comprising:

[0161] The monitoring signal acquisition and preprocessing module 801 is used to acquire and preprocess the vibration monitoring signals during the operation of the rolling bearing.

[0162] The fault diagnosis model construction and training module 802 is used to construct and train a rolling bearing fault diagnosis model to obtain a trained rolling bearing fault diagnosis model. The rolling bearing fault diagnosis model includes several feature extraction layers, a global average pooling layer, and a classification layer connected in sequence. All the feature extraction layers are stacked in ascending order of the number of channels. Each feature extraction layer includes a multi-scale feature fusion module, a multi-scale dilated convolution module, a dense residual attention module, and a channel adaptation module.

[0163] The preprocessed vibration monitoring signal is processed by a multi-scale feature fusion module to extract local impact features at different convolutional scales, resulting in local impact features. These local impact features are then processed by a multi-scale dilated convolution module to extract periodic fault features under different receptive fields, resulting in periodic fault features. The periodic fault features are further processed by a dense residual attention module to extract local features, temporal-dependent features, and attention-based key channel features, followed by residual fusion to obtain fused and enhanced features. These enhanced features are then processed by a channel adaptation module for channel adjustment and output. Finally, the output of the channel adaptation module in the last feature extraction layer is processed by a global average pooling layer and a classification layer to output a rolling bearing fault classification prediction.

[0164] The rolling bearing fault classification and prediction module 803 is used to perform rolling bearing fault diagnosis using a trained rolling bearing fault diagnosis model and obtain rolling bearing fault classification and prediction.

[0165] The specific implementation of the rolling bearing fault diagnosis device based on multi-scale residual attention is the same as that of the rolling bearing fault diagnosis method based on multi-scale residual attention, and will not be described again in this embodiment.

[0166] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A method for fault diagnosis of rolling bearings based on multi-scale residual attention, characterized in that, Includes the following steps: S1, acquire vibration monitoring signals during the operation of the rolling bearing and perform preprocessing; S2, construct and train a rolling bearing fault diagnosis model to obtain a trained rolling bearing fault diagnosis model; the rolling bearing fault diagnosis model includes several feature extraction layers, a global average pooling layer and a classification layer connected in sequence; all the feature extraction layers are stacked in ascending order of the number of channels; each feature extraction layer includes a multi-scale feature fusion module, a multi-scale dilated convolution module, a dense residual attention module and a channel adaptation module. The preprocessed vibration monitoring signal is processed by a multi-scale feature fusion module to extract local impact features at different convolutional scales, resulting in local impact features; the local impact features are then processed by a multi-scale dilated convolution module to extract periodic fault features at different receptive fields, resulting in periodic fault features. Periodic fault features are processed by a dense residual attention module to extract local features, temporal dependent features, and attention-based key channel features, and then residual fusion is performed to obtain fused enhanced features. The fused and enhanced features are adjusted by the channel adaptation module and then output; the output of the channel adaptation module of the last feature extraction layer is passed through the global average pooling layer and the classification layer to output the rolling bearing fault classification prediction. S3. Use the trained rolling bearing fault diagnosis model to perform rolling bearing fault diagnosis and obtain rolling bearing fault classification prediction.

2. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The vibration monitoring signal during the operation of the rolling bearing includes a one-dimensional vibration acceleration signal during the operation of the rolling bearing; the one-dimensional vibration acceleration signal includes a normal state vibration signal and a fault state vibration signal; the fault state vibration signal includes at least one of the following: inner ring fault vibration signal, outer ring fault vibration signal, rolling element fault vibration signal and cage fault vibration signal, distinguished according to the fault location.

3. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 2, characterized in that, The fault category labels for the training set of the rolling bearing fault diagnosis model include: setting corresponding fault category labels for vibration signals under normal conditions; and setting corresponding fault category labels for vibration signals under fault conditions at the same fault location according to the degree of fault.

4. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The preprocessing includes abnormal data filtering, missing data completion, fixed-length sample truncation, and normalization.

5. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The multi-scale feature fusion module is represented as follows: ； ； ； ； ； ； in, This represents the fine-scale features output by the first convolutional branch; Represents a non-linear activation function; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 3. Indicates the first The input features of each feature extraction layer; This represents the mesoscale features output by the second convolutional branch; This represents a one-dimensional convolution operation with a kernel size of 5. This represents the coarse-scale features output by the third convolutional branch; This demonstrates a one-dimensional convolution operation with a kernel size of 7. This represents the pooling characteristics of the pooling branch output; This represents a one-dimensional pointwise convolution operation with a kernel size of 1. This represents the max pooling operation; This indicates the multi-branch feature after channel splicing; This indicates a channel-level concatenation operation; Indicates the first The multi-scale fused features are output by the multi-scale feature fusion module in each feature extraction layer.

6. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The multi-scale dilated convolution module is represented as follows: ； ； ； ； ； This represents the short receptive field feature output by the first dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 1. Represents a non-linear activation function; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 3. Indicates the first The multi-scale fused features output by the multi-scale feature fusion module in each feature extraction layer; This represents the mid-receptive field feature of the output of the second dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 2. This represents the long receptive field feature of the output of the third dilated convolution branch; This represents a one-dimensional dilated convolution operation with a kernel size of 5 and a dilation rate of 4. This indicates a channel-level concatenation operation; This indicates the multi-receptor field characteristics after channel splicing; No. Periodic fault features output by the multi-scale dilated convolution module in each feature extraction layer.

7. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The dense residual attention module is represented as follows: ； ； ； ； ； ； ； in, This represents the local short-scale impact feature of the output of the first convolutional path; This indicates a batch normalization operation; This represents a one-dimensional convolution operation with a kernel size of 5. Indicates the characteristics of periodic faults; This represents the local mesoscale impact feature of the output of the second convolutional path; This represents the fusion feature of dual-path convolution; This represents the temporal dependency features output by the temporal modeling unit; This represents a one-dimensional pointwise convolution operation with a kernel size of 1. This represents a bidirectional long short-term memory network; This indicates a feature dimension transformation operation; Indicates channel attention weights; This represents the Sigmoid activation function; This indicates a global average pooling operation; Represents a non-linear activation function; Indicates attention fusion features; This indicates a channel-by-channel multiplication operation; Indicates the first The fused enhanced features are output by the dense residual attention modules in each feature extraction layer.

8. The rolling bearing fault diagnosis method based on multi-scale residual attention according to claim 1, characterized in that, The loss function of the rolling bearing fault diagnosis model is the cross-entropy loss function.

9. A rolling bearing fault diagnosis device based on multi-scale residual attention, characterized in that, Including the following: The monitoring signal acquisition and preprocessing module is used to acquire and preprocess the vibration monitoring signals during the operation of the rolling bearing. The fault diagnosis model construction and training module is used to construct and train a rolling bearing fault diagnosis model to obtain a trained rolling bearing fault diagnosis model. The rolling bearing fault diagnosis model includes several feature extraction layers, a global average pooling layer, and a classification layer connected in sequence. All the feature extraction layers are stacked in ascending order of the number of channels. Each feature extraction layer includes a multi-scale feature fusion module, a multi-scale dilated convolution module, a dense residual attention module, and a channel adaptation module. The preprocessed vibration monitoring signal is processed by a multi-scale feature fusion module to extract local impact features at different convolutional scales, resulting in local impact features; the local impact features are then processed by a multi-scale dilated convolution module to extract periodic fault features at different receptive fields, resulting in periodic fault features. Periodic fault features are processed by a dense residual attention module to extract local features, temporal dependent features, and attention-based key channel features, and then residual fusion is performed to obtain fused enhanced features. The fused and enhanced features are adjusted by the channel adaptation module and then output; the output of the channel adaptation module of the last feature extraction layer is passed through the global average pooling layer and the classification layer to output the rolling bearing fault classification prediction. The rolling bearing fault classification and prediction module is used to perform rolling bearing fault diagnosis using a trained rolling bearing fault diagnosis model, and to obtain rolling bearing fault classification and prediction.

Images