Bearing fault diagnosis method and system combined with multi-scale discrete wavelet

Abstract

The invention relates to a bearing fault diagnosis method and system combined with multi-scale discrete wavelets. The method comprises the following steps: firstly, acquiring vibration signal data of a rolling mill bearing under different working conditions and fault types, dividing the vibration signal data into segments with fixed lengths, adding corresponding labels, and constructing four types of fault data sets; thirdly, constructing a discrete wavelet network by using a formula, fusing an improved attention mechanism, and constructing a deep discrete wavelet dense network structure; then, data feature screening and multi-scale feature fusion are carried out on the four types of fault data sets, the network structure is input, and a confusion matrix related to faults is generated; and finally, based on the confusion matrix data, a random neighborhood embedding clustering algorithm is used to obtain fault category determination results of the bearing under different working conditions. By adopting the method, the accuracy and robustness of bearing fault diagnosis can be remarkably improved, and powerful technical support is provided for real-time fault diagnosis of industrial rotating machinery in a scene containing strong noise and multi-working-condition unbalanced data.

Description

Technical Field

[0001] The present invention belongs to mechanical engineering, and in particular relates to a bearing fault diagnosis method and system combined with multi-scale discrete wavelets. Background Art

[0002] As mechanical equipment becomes larger and more complex, bearing failures, a core component, account for as much as 45%-55% of all failures. Industrial scenarios present challenges such as strong noise interference, multiple operating conditions (idling, low-speed steel biting, and accelerated rolling), and data imbalance. Traditional deep learning methods face three major challenges: First, mainstream approaches require preprocessing the one-dimensional vibration signal into a two-dimensional image, which not only loses temporal information but also is cumbersome and inefficient. Second, deep neural networks are not robust to high-noise signals (e.g., SNR = -4), and are prone to overfitting or "black box" problems with poor feature interpretation. Third, existing wavelet networks are only applied to shallow or initial layers, lacking systematic analysis of key parameters such as the number of decomposition layers and vanishing distance, and are difficult to embed into deep networks for multi-scale feature reuse. Although previous studies have attempted to combine wavelet transforms with CNNs (e.g., WaveletKernelNet and MWA-CNN), their feature fusion capabilities within deep networks are limited, and they fail to fully leverage the multi-scale nature of wavelet transforms and the feature reuse mechanism of dense networks. Summary of the Invention

[0003] Based on this, it is necessary to provide a bearing fault diagnosis method and system combining multi-scale discrete wavelets that can significantly improve the accuracy and robustness of bearing fault diagnosis in order to address the above technical problems.

[0004] In a first aspect, the present application provides a bearing fault diagnosis method combined with multi-scale discrete wavelets, comprising:

[0005] Obtain vibration signal data of rolling mill bearings under different operating conditions and fault types.

[0006] The vibration signal data is divided into fixed-length segments and corresponding labels are added to each segment to construct four types of fault data sets.

[0007] The formula is used to construct a discrete wavelet network and an improved attention mechanism is added to construct a deep discrete wavelet dense network structure.

[0008] After screening the data features of the four types of fault data sets, multi-scale feature fusion is performed and input into the deep discrete wavelet dense network structure to generate the confusion matrix related to the fault.

[0009] The confusion matrix data is processed based on the random neighborhood embedding clustering algorithm to obtain the fault category judgment results of the bearing under different working conditions.

[0010] In one embodiment, the vibration signal data is divided into fixed-length segments and corresponding labels are added to each segment to construct four types of fault data sets, including:

[0011] Based on the three-part architecture of the rolling mill experimental platform, the sensor layout plan is clarified and a multi-component collaborative experimental environment is constructed.

[0012] In the experimental environment, the rolling mill bearings are set according to the preset working conditions combined with artificial fault settings to form four types of bearing fault samples including normal conditions.

[0013] According to the preset working conditions and speed conditions, Shannon's theorem is used to perform automatic sliding window non-overlapping segmentation to divide the four types of bearing fault sample data into fixed-length segment sequences.

[0014] Based on the fragment sequence, the insufficient samples are discarded and corresponding labels are added to each fragment sequence according to the fault type and working condition to generate a labeled one-dimensional sample data set.

[0015] Noise interferences with different noise ratios are superimposed on the one-dimensional sample data set to construct four types of fault data sets with different signal-to-noise ratio characteristics.

[0016] In one embodiment, the deep discrete wavelet dense network structure is constructed by the following steps, including:

[0017] The following formula is used to construct the expression of discrete wavelet in neural network to obtain discrete wavelet network.

[0018]

[0019] Among them, x α,g [n] represents the approximate wavelet coefficients processed by the low-pass filter after α-1 layers, x α,h [n] represents the detail wavelet coefficients processed by the high-pass filter after α-1 layers, g[k] and h[k] are the high-frequency and low-frequency filter coefficients respectively, and 2n-k represents the convolution operation with a downsampling value of 2;

[0020] The characteristic length of the discrete wavelet network is verified by experiments, and the length of the filters in each layer is obtained.

[0021] S=N+2*Kα=N+2*(V-1)α

[0022] L α =N / 2 α +2K(1-1 / 2 α )=N / 2 α +2(V-1)(1-1 / 2 α )

[0023] Among them, L αRepresents the length of the αth layer filter, S represents the total length of the wavelet network, N represents the total length of the input features, α represents the number of decomposition layers, V represents the vanishing distance, V=K+1, and K represents an additional parameter item.

[0024] The length of wavelet coefficients in each layer is determined by the derived filter lengths of each layer.

[0025] The frequency attention mechanism is used to process the length of each layer of wavelet coefficients to obtain matrices of different feature scales;

[0026] The different scales in different feature scale matrices are stacked on the feature dimension, and the output value of the network block is made consistent with the input value feature dimension through the α-layer residual network structure, thus obtaining a wavelet residual network block structure for multi-scale feature fusion.

[0027] In one embodiment, after obtaining the wavelet residual network block structure for multi-scale feature fusion, the method further includes:

[0028] The similarities between the discrete wavelet network structure and the convolutional neural network structure are compared and inferred to obtain similarity results.

[0029] The one-dimensional single-channel convolutional neural network is expressed using the following formula:

[0030]

[0031] Among them, y i Represents the i-th sequence of the convolutional neural network output, Y=[y1,y2,…,y M-1 ,y M ] represents a one-dimensional vector of length M, and the sequence length should satisfy x i Represents the y-th sequence of convolutional neural network input, X=[x1,x2,…,x L-1 ,x L ] is a one-dimensional vector of length L, w j Represents the jth convolution kernel, W=[w1,w2,…,w K-1 ,w K ], K represents the total length, s represents the step size, and b represents an independent scalar bias value.

[0032] Based on the similarity results, the four densely connected convolution modules in the dense block are replaced with a multi-scale discrete wavelet residual network block structure, and the improved attention mechanism is alternately inserted as a transition block to obtain a deep discrete wavelet dense network structure.

[0033] In one embodiment, after filtering the data features of four types of fault data sets, multi-scale feature fusion is performed and input into a deep discrete wavelet dense network structure to generate a confusion matrix related to the fault, including:

[0034] The multi-channel time-frequency feature tensors of the four types of fault data sets are obtained; the multi-channel time-frequency feature tensors are generated by discrete wavelet transform of the original vibration signal.

[0035] The multi-channel time-frequency feature tensor is weighted according to the channel dimension to obtain a weighted feature tensor.

[0036] The dynamic aggregation of features is used to perform cross-scale convolution fusion on the segmented features of the weighted feature tensor to generate a multi-scale fused feature tensor.

[0037] Input the multi-scale fusion feature tensor into the deep discrete wavelet dense network to obtain the predicted label;

[0038] Compare the predicted labels with the actual labels, and count the cases where the predicted labels are consistent or inconsistent with the actual labels to obtain the comparison results;

[0039] A confusion matrix is ​​generated based on the comparison results; the confusion matrix includes indicators such as accuracy and recall rate of the deep discrete wavelet dense network structure in classifying different fault types.

[0040] In one embodiment, a random neighborhood embedding clustering algorithm is used based on confusion matrix data to obtain the fault category determination results of the bearing under different working conditions, including:

[0041] The random neighborhood embedding clustering algorithm is used to process the multidimensional feature vector of vibration signal data to obtain the cluster center and the corresponding fault category label;

[0042] Obtain the real-time vibration signal of the bearing to be tested; extract the features of the real-time vibration signal based on the fault characteristics of the confusion matrix and generate the target feature vector;

[0043] Calculate the Euclidean distance between the target feature vector and each cluster center;

[0044] The Euclidean distance is judged based on a preset distance threshold. If the Euclidean distance is less than the distance threshold, the real-time vibration signal is marked as the corresponding fault category;

[0045] Based on the fault category, a fault category determination result with the working condition type and fault category label is obtained.

[0046] In a second aspect, the present application also provides a bearing fault diagnosis system combined with multi-scale discrete wavelets, the system comprising:

[0047] The data processing module is used to obtain vibration signal data under different operating conditions and fault types of rolling mill bearings. It is also used to divide the vibration signal data into fixed-length segments and add corresponding labels to each segment to construct four types of fault data sets.

[0048] The network construction module is used to construct a discrete wavelet network using the formula and add an improved attention mechanism to construct a deep discrete wavelet dense network structure.

[0049] The fault classification module is used to screen the data features of the four types of fault data sets, perform multi-scale feature fusion, and input the deep discrete wavelet dense network structure to generate a confusion matrix related to the fault. It is also used to use the random neighborhood embedding clustering algorithm based on the confusion matrix data to obtain the fault category judgment results of the bearing under different working conditions.

[0050] In a third aspect, the present application further provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above method when executing the computer program.

[0051] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the above method when executed by a processor.

[0052] The bearing fault diagnosis method, system, computer device, and storage medium using multi-scale discrete wavelet analysis first acquires vibration signal data from rolling mill bearings under different operating conditions and fault types, divides the data into fixed-length segments, and assigns corresponding labels to construct four fault datasets. Next, a discrete wavelet network is constructed using a formula and an improved attention mechanism is incorporated to create a deep discrete wavelet dense network structure. Subsequently, data feature screening and multi-scale feature fusion are performed on the four fault datasets, and the data is input into the aforementioned network structure to generate a confusion matrix related to the faults. Finally, based on the confusion matrix data, a random neighborhood embedding clustering algorithm is applied to determine the bearing fault category under different operating conditions. This method significantly improves the efficiency and accuracy of bearing fault diagnosis, effectively extracts fault features, and enables more accurate fault classification. This method enhances the reliability of fault diagnosis and adapts to complex and changing rolling mill operating scenarios, providing strong support for the timely detection and resolution of bearing faults, reducing equipment failure rates, and improving production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0054] Figure 1 A flowchart of a bearing fault diagnosis method combining multi-scale discrete wavelets provided in an embodiment of the present invention;

[0055] Figure 2A structural block diagram of a bearing fault diagnosis system combined with multi-scale discrete wavelets provided in an embodiment of the present invention;

[0056] Figure 3 A schematic diagram of dividing various data sets under the adjusted rolling mill variable operating conditions provided by an embodiment of the present invention;

[0057] Figure 4 A schematic diagram of a histogram comparing different models under different noisy rolling mill bearing data sets provided by an embodiment of the present invention;

[0058] Figure 5 Schematic diagram of the accuracy curves of the rolling mill test set of different models under different noises provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0060] In one embodiment, Figure 1 As shown, the present application provides a bearing fault diagnosis method combined with multi-scale discrete wavelet, which may include the following steps:

[0061] Step S101: Acquire vibration signal data of a rolling mill bearing under different operating conditions and fault types.

[0062] Specifically, axial and radial acceleration sensors were placed on the work roll chocks of the rolling mill test platform. Vibration detection software was used to collect raw vibration signals based on different fault types (inner ring fracture, outer ring crack, rolling element missing, and normal state) and operating conditions (idling, low-speed steel bite, and accelerated rolling, corresponding to loads N0-0 hp, N1-1 hp, and N2-2 hp). The motor drive speed was set to 600 r / min, and the sensors simultaneously collected data and stored it as an .XLSX file, ensuring that the signals covered the time domain characteristics of different fault modes.

[0063] Step S102 : Divide the vibration signal data into fixed-length segments and add corresponding labels to each segment to construct four types of fault data sets.

[0064] Specifically, according to the Shannon sampling theorem, the original vibration signal was segmented into a fixed-length, non-overlapping sliding window of 2048 points, discarding segments shorter than this length to avoid data distortion. Twelve labels (0-11) were generated based on the combination of three operating conditions and four fault types. Labels "0, 4, 8" correspond to normal conditions, "1, 5, 9" to inner race faults, "2, 6, 10" to outer race faults, and "3, 7, 11" to rolling element faults, generating a total of 1129 one-dimensional samples. Furthermore, four levels of Gaussian noise (SNR = -4, -2, 4, and no noise) were added to the samples to construct four fault datasets with different signal-to-noise ratio characteristics, simulating the complex noise interference found in industrial environments.

[0065] Step S103: construct a discrete wavelet network using the formula and add an improved attention mechanism to construct a deep discrete wavelet dense network structure.

[0066] Specifically, a specific formula is first used to construct a representation of discrete wavelets in a neural network, forming a discrete wavelet network. Experiments verify the characteristic length of the discrete wavelet network, and then determine the filter length of each layer, based on which the wavelet coefficient length of each layer is derived. A frequency attention mechanism is used to process the wavelet coefficient length of each layer, resulting in matrices of different feature scales. The different scales in the different feature scale matrices are stacked along the feature dimension, and with the help of an α-layer residual network structure, the output value of the network block is aligned with the input value feature dimension, resulting in a wavelet residual network block structure that integrates multi-scale features. Subsequently, the similarities between the discrete wavelet network and the convolutional neural network structures are compared. Based on this, the four densely connected convolutional modules in the dense block are replaced with a multi-scale discrete wavelet residual network block structure. The improved attention mechanism is also interleaved as a transition block, ultimately resulting in a deep discrete wavelet dense network structure.

[0067] Step S104 , after filtering the data features of the four types of fault data sets, multi-scale feature fusion is performed and input into a deep discrete wavelet dense network structure to generate a confusion matrix related to the fault.

[0068] The preprocessed four-category fault data set is input into the deep discrete wavelet dense network (DDenseNet) in batches. The network realizes multi-layer decomposition of the signal through the multi-scale discrete wavelet residual block (MWRN): the signal is decomposed into α layers using the low-pass filter h[k] and the high-pass filter g[k] to obtain the low-frequency approximation coefficient x α,g [n] and high frequency detail coefficient x α,h[n] The frequency attention mechanism (FAM) is used to weight the coefficients of each layer, fuse features of different scales, and maintain the consistency of feature dimensions through residual connections. Simultaneously, the improved SKAttention transition block uses the Split-Fuse-Select operation to adaptively select multi-scale features, suppressing noise and highlighting fault-related features, ultimately generating a confusion matrix containing time-frequency domain correlation information.

[0069] Step S105 : Using a random neighborhood embedding clustering algorithm based on the confusion matrix data, a fault category determination result of the bearing under different working conditions is obtained.

[0070] Specifically, the confusion matrix data generated by deep discrete wavelet dense network processing is analyzed and processed using a t-distributed random neighborhood embedding clustering algorithm to identify and analyze the fault feature data contained in the confusion matrix. This algorithm identifies underlying patterns and regularities within the data and generates multiple cluster centers. During this process, the real-time vibration signal of the bearing to be tested is first subjected to feature extraction to convert it into a corresponding feature vector. Next, the Euclidean distance between this feature vector and each cluster center is calculated. These Euclidean distances are determined using a preset distance threshold. If the Euclidean distance between a feature vector and a cluster center is less than the threshold, the real-time vibration signal is labeled as belonging to the fault category corresponding to that cluster center. This allows the fault category determination results to be derived, combining the different operating conditions of the bearing with the corresponding operating condition type and fault category label.

[0071] The bearing fault diagnosis method, system, computer device, and storage medium using multi-scale discrete wavelet analysis first acquires vibration signal data from rolling mill bearings under different operating conditions and fault types, divides the data into fixed-length segments, and assigns corresponding labels to construct four fault datasets. Next, a discrete wavelet network is constructed using a formula and an improved attention mechanism is incorporated to create a deep discrete wavelet dense network structure. Subsequently, data feature screening and multi-scale feature fusion are performed on the four fault datasets, and the data is input into the aforementioned network structure to generate a confusion matrix related to the faults. Finally, based on the confusion matrix data, a random neighborhood embedding clustering algorithm is applied to determine the bearing fault category under different operating conditions. This method significantly improves the efficiency and accuracy of bearing fault diagnosis, effectively extracts fault features, and enables more accurate fault classification. This method enhances the reliability of fault diagnosis and adapts to complex and changing rolling mill operating scenarios, providing strong support for the timely detection and resolution of bearing faults, reducing equipment failure rates, and improving production efficiency.

[0072] In one embodiment, dividing the vibration signal data into fixed-length segments and adding corresponding labels to each segment to construct a four-category fault dataset may include the following steps:

[0073] Step S201 : clarify the sensor layout plan based on the three-part structure of the rolling mill experimental platform and build a multi-component collaborative experimental environment.

[0074] Step S202 : setting the rolling mill bearings according to preset working conditions in an experimental environment in combination with artificial fault settings to form four types of bearing fault samples including normal conditions.

[0075] Step S203 : According to the preset working conditions and speed conditions, the Shannon theorem is used to perform automatic sliding window non-overlap segmentation to divide the four types of bearing fault sample data into segment sequences of fixed length.

[0076] In step S204 , insufficient samples are discarded based on the segment sequence and corresponding labels are added to each segment sequence according to the fault type and working condition to generate a labeled one-dimensional sample data set.

[0077] In step S205 , noise interferences of different noise ratios are superimposed on the one-dimensional sample data set to construct four types of fault data sets containing different signal-to-noise ratio features.

[0078] Specifically, based on the three-component architecture of the rolling mill test platform—the rolling mill test equipment, the test console, and the sensor acquisition system—accelerometers were placed in the axial and radial positions of the bearing housing to create a multi-component collaborative signal acquisition environment. Three fault conditions—inner ring fracture, outer ring crack, and rolling element loss—were manually set. Combined with three preset operating conditions (corresponding to loads N0-0 hp, N1-1 hp, and N2-2 hp)—idling, low-speed steel biting, and accelerated rolling—four types of bearing fault samples, encompassing normal conditions, were generated. Based on the Shannon sampling theorem, the signal was automatically segmented using a 2048-point non-overlapping sliding window with a motor speed of 600 r / min. After discarding insufficient samples, the segment sequences were labeled according to the fault type and operating condition combination (a total of 12 categories), generating a dataset of 1,129 labeled one-dimensional samples. Furthermore, the dataset was overlaid with four noise levels: SNR = -4, -2, 4, and no noise. Four fault datasets with varying signal-to-noise ratio characteristics were constructed to ensure data coverage of the complex noise conditions found in industrial scenarios.

[0079] This example significantly improves the standardization and usability of fault data through standardized experimental environments and data processing strategies. A multi-component collaborative sensor layout ensures comprehensive vibration signal acquisition, avoiding single-dimensional data bias. Artificial fault setting and operating condition division enable controllable generation of fault samples, facilitating targeted training of subsequent models. Sliding window segmentation and labeling system construction guided by Shannon's theorem ensure the temporal integrity and class distinguishability of data. Noise enhancement processing strengthens the dataset's ability to simulate high-noise industrial environments. The dataset constructed through these steps provides high-quality training samples for the Deep Discrete Wavelet Dense Network (DDenseNet) across multiple operating conditions and noise levels.

[0080] In one embodiment, the deep discrete wavelet dense network structure can be constructed by the following steps:

[0081] Step S301: Use the following formula to construct the expression of discrete wavelet in the neural network to obtain a discrete wavelet network.

[0082]

[0083] Among them, x α,g [n] represents the approximate wavelet coefficients processed by the low-pass filter after α-1 layers, x α,h [n] represents the detail wavelet coefficients processed by the high-pass filter after α-1 layers, g[k] and h[k] are the high-frequency and low-frequency filter coefficients respectively, and 2n-k represents the convolution operation with a downsampling value of 2;

[0084] Step S302: Verify the characteristic length of the discrete wavelet network through experiments to obtain the length of each layer of filters.

[0085] S=N+2*Kα=N+2*(V-1)α

[0086] L α =N / 2 α +2K(1-1 / 2 α )=N / 2 α +2(V-1)(1-1 / 2 α )

[0087] Among them, L α Represents the length of the αth layer filter, S represents the total length of the wavelet network, N represents the total length of the input features, α represents the number of decomposition layers, V represents the vanishing distance, V=K+1, and K represents an additional parameter item.

[0088] Step S303: Determine the length of each layer of wavelet coefficients by using the derived filter lengths of each layer.

[0089] Step S304: Process the length of each layer of wavelet coefficients using the frequency attention mechanism to obtain different feature scale matrices;

[0090] In step S305, different scales in different feature scale matrices are stacked on the feature dimension, and the output value of the network block is made consistent with the input value feature dimension through the α-layer residual network structure, thereby obtaining a wavelet residual network block structure for multi-scale feature fusion.

[0091] This embodiment uses the hierarchical decomposition of low-pass / high-pass filters to achieve hierarchical extraction of low-frequency trends (such as operating condition characteristics) and high-frequency mutations (such as fault shocks) in vibration signals, avoiding the insufficient representation of complex features by traditional single-scale convolution. Through the feature length formula and residual connection, the dimensional consistency of the wavelet network block in the deep architecture is ensured, the limitation of the existing wavelet layer being only applicable to shallow networks is solved, and seamless integration with deep architectures such as dense networks (DenseNet) is supported. The frequency attention mechanism weights the multi-scale coefficients, enhances the expression strength of fault-related features, suppresses noise interference, and combines feature dimension stacking with residual connection to achieve efficient fusion of cross-scale information, thereby improving the robustness of the model to multi-operating conditions and strong noise data. Experiments show that the structure can still maintain a diagnostic accuracy of more than 83% in data containing SNR=-4 noise, which is a significant improvement over traditional convolution modules.

[0092] In one embodiment, after obtaining the wavelet residual network block structure for multi-scale feature fusion, the following steps may be further included:

[0093] The similarities between the discrete wavelet network structure and the convolutional neural network structure are compared and inferred to obtain similarity results.

[0094] The one-dimensional single-channel convolutional neural network is expressed using the following formula:

[0095]

[0096] Among them, y i Represents the i-th sequence of the convolutional neural network output, Y=[y1,y2,…,y M-1 ,y M ] represents a one-dimensional vector of length M, and the sequence length should satisfy x i Represents the i-th sequence of convolutional neural network input, X=[x1,x2,…,x L-1 ,x L ] is a one-dimensional vector of length L, w j Represents the jth convolution kernel, W=[w1,w2,…,w K-1 ,w K ], K represents the total length, s represents the step size, and b represents an independent scalar bias value.

[0097] Based on the similarity results, the four densely connected convolution modules in the dense block are replaced with a multi-scale discrete wavelet residual network block structure, and the improved attention mechanism is alternately inserted as a transition block to obtain a deep discrete wavelet dense network structure.

[0098] Specifically, based on similarity, the four densely connected convolutional modules in the traditional dense block (DenseBlock) are replaced with a multi-scale discrete wavelet residual network block (MWRN), and an improved SKAttention is introduced as a transition block to alternately embed the network. Specifically, each MWRN block extracts multi-scale features through multi-layer decomposition of discrete wavelet transform, and after weighted fusion using the frequency attention mechanism, the feature dimensions are kept consistent through residual connections; the SKAttention transition block adaptively selects cross-layer features through the Split-Fuse-Select operation, enhancing the model's ability to distinguish high-frequency fault features from low-frequency working condition features. The resulting deep discrete wavelet dense network (DDenseNet) uses DenseNet as its skeleton, and through the synergy of MWRN and SKAttention, it achieves efficient reuse of multi-scale features and noise robustness optimization.

[0099] This embodiment uses the multi-scale decomposition of discrete wavelet transform to replace the single convolution operation, so that the network can capture the time domain details and frequency domain trends of the signal at the same time, solving the limitations of traditional CNN in extracting features of non-stationary signals. Experiments show that on a dataset containing SNR=-4 noise, the diagnosis accuracy of DDenseNet for bearing faults is more than 12% higher than that of traditional DenseNet. Through the residual connection and feature reuse mechanism, the redundancy of network parameters is reduced, while maintaining high diagnostic accuracy, the computational complexity is reduced, and it is suitable for real-time fault diagnosis scenarios. Comparative experiments show that the size of DDenseNet is about 16% larger than that of CNN of the same depth, and the classification accuracy of multiple working conditions is improved by more than 20%.

[0100] In one embodiment, after filtering the data features of four types of fault data sets, performing multi-scale feature fusion and inputting it into a deep discrete wavelet dense network structure to generate a confusion matrix related to the fault may include the following steps:

[0101] Step S401 , obtaining multi-channel time-frequency feature tensors of four types of fault data sets; the multi-channel time-frequency feature tensors are generated by discrete wavelet transform of the original vibration signal.

[0102] Step S402 , performing channel dimension weight assignment on the multi-channel time-frequency feature tensor to obtain a weighted feature tensor.

[0103] Step S403: Perform cross-scale convolution fusion on the segmented features of the weighted feature tensor using dynamic feature aggregation to generate a multi-scale fused feature tensor.

[0104] Step S404: input the multi-scale fusion feature tensor into a deep discrete wavelet dense network to obtain a predicted label;

[0105] Step S405: Compare the predicted labels with the actually added labels, and count the consistency and inconsistency between the predicted labels and the actual labels to obtain a comparison result;

[0106] Step S406: Generate a confusion matrix based on the comparison results; the confusion matrix includes indicators such as accuracy and recall rate of the deep discrete wavelet dense network structure in classifying different fault types.

[0107] Furthermore, the multi-channel time-frequency feature tensors of each batch of data generated by discrete wavelet transform of the original vibration signal are first obtained. Then, the algorithm is used to assign channel dimension weights to the multi-channel time-frequency feature tensors to obtain a weighted feature tensor. Subsequently, the segmented features of the weighted feature tensor are cross-scale convolutionally fused through dynamic feature aggregation to generate a multi-scale fused feature tensor. The multi-scale fused feature tensor is input into the deep discrete wavelet dense network to obtain the predicted label. The predicted label is then compared with the actual added label, and the consistency and inconsistency between the two are counted to obtain the comparison result. Finally, a confusion matrix is ​​generated based on the comparison results. The confusion matrix contains indicators such as the accuracy and recall rate of the deep discrete wavelet dense network structure in the classification of different fault types.

[0108] This embodiment generates multi-channel time-frequency feature tensors through discrete wavelet transforms, fully exploiting the characteristic information of the original vibration signal. A deep discrete wavelet dense network derives predicted labels based on the multi-scale fusion feature tensors. By comparing these predicted labels with the actual labels, a confusion matrix containing key indicators is generated. This helps accurately evaluate the network's diagnostic capabilities for different fault types, providing a strong basis for optimizing diagnostic models and improving the accuracy and reliability of bearing fault diagnosis.

[0109] In one embodiment, a random neighborhood embedding clustering algorithm is used based on confusion matrix data to obtain a fault category determination result of a bearing under different working conditions, which may include the following steps:

[0110] Step S501: Using a random neighborhood embedding clustering algorithm to process the multidimensional feature vector of the vibration signal data, obtain cluster centers and corresponding fault category labels;

[0111] Step S502: obtaining a real-time vibration signal of the bearing to be tested; performing feature extraction on the real-time vibration signal based on the fault features of the confusion matrix to generate a target feature vector;

[0112] Step S503, calculating the Euclidean distance between the target feature vector and each cluster center;

[0113] Step S504: judging the Euclidean distance based on a preset distance threshold, and marking the real-time vibration signal as a corresponding fault category if the Euclidean distance is less than the distance threshold;

[0114] Step S505 : obtaining a fault category determination result with a working condition type and a fault category label based on the fault category.

[0115] Preferably, a clustering algorithm is first used to process the multidimensional feature vector of the vibration signal data, thereby obtaining the cluster center and the corresponding fault category label. Next, the real-time vibration signal of the bearing to be tested is obtained, and a feature extraction operation is performed on it based on the fault characteristics of the confusion matrix to generate a target feature vector. Then, the Euclidean distance between the target feature vector and each cluster center is calculated. The calculated Euclidean distance is judged according to a pre-set distance threshold. Once the Euclidean distance is less than the distance threshold, the real-time vibration signal is marked as the corresponding fault category. Finally, based on the determined fault category and combined with the operating condition type of the bearing, a fault category determination result with the operating condition type and fault category label is obtained.

[0116] This embodiment constructs a complete fault diagnosis system to ensure the standardization and systematic nature of the diagnostic process. In terms of diagnostic accuracy, a clustering algorithm is used to process multidimensional feature vectors to determine the cluster center, providing a reliable benchmark for subsequent judgments. Feature extraction and distance calculation and comparison of real-time vibration signals can accurately identify the fault category of real-time signals. At the same time, the judgment results are given in combination with the operating condition type, making the diagnostic information more comprehensive and accurate, helping staff to quickly understand the operating status of the bearings, take targeted maintenance measures in a timely manner, reduce the risk of equipment failure, ensure the stable operation of the rolling mill, and improve production efficiency.

[0117] In one embodiment, Figure 2 As shown, the present application also provides a bearing fault diagnosis system combined with multi-scale discrete wavelet, which may include:

[0118] The data processing module 601 is used to obtain vibration signal data of rolling mill bearings under different working conditions and fault types; it is also used to divide the vibration signal data into fixed-length segments and add corresponding labels to each segment to construct four types of fault data sets.

[0119] The network construction module 602 is used to construct a discrete wavelet network using a formula and add an improved attention mechanism to construct a deep discrete wavelet dense network structure.

[0120] The fault classification module 603 is used to screen the data features of the four types of fault data sets, perform multi-scale feature fusion, and input the deep discrete wavelet dense network structure to generate a confusion matrix related to the fault; it is also used to use the random neighborhood embedding clustering algorithm based on the confusion matrix data to obtain the fault category judgment results of the bearing under different working conditions.

[0121] In this bearing fault diagnosis system, which incorporates multi-scale discrete wavelets, the data processing module acquires vibration signal data from rolling mill bearings under different operating conditions and fault types, divides it into fixed-length segments, and assigns corresponding labels to construct four fault datasets. The network construction module uses a specific formula to construct a discrete wavelet network, incorporating an improved attention mechanism to build a deep discrete wavelet dense network structure. The fault classification module performs feature screening and multi-scale feature fusion on the four fault datasets. This is then fed into the deep discrete wavelet dense network structure to generate a confusion matrix. Based on this matrix, a random neighborhood embedding clustering algorithm is then used to determine the bearing fault category under different operating conditions. This system significantly improves the efficiency and accuracy of bearing fault diagnosis, effectively extracting fault features and enabling more accurate fault classification. This system enhances the reliability of fault diagnosis and adapts to complex and changing rolling mill operating scenarios, providing strong support for the timely detection and resolution of bearing faults, reducing equipment failure rates, and improving production efficiency.

[0122] In one embodiment, Figure 3 As shown, this application also provides a schematic diagram of the division of each data set under the variable working conditions of the rolling mill. The chart shows the data set division results of the rolling mill bearing under different working conditions (idling, low-speed steel biting, accelerated rolling) and fault types (normal, inner ring fault, outer ring fault, rolling element fault). The data covers 3 working conditions × 4 types of faults, with a total of 12 labels (0-11), among which normal state samples are evenly distributed under each working condition (labels 0, 4, 8), and fault samples are labeled according to different working conditions (such as inner ring fault corresponds to 1, 5, 9). The data set uses Shannon's theorem to split the original signal with a 2048-point sliding window to generate 1129 one-dimensional samples, and adds 4 noise levels of SNR = -4, -2, 4 and no noise to form a multi-signal-to-noise ratio data set for training and verifying the robustness of the model in complex environments.

[0123] In one embodiment, Figure 4 As shown, this application also provides a schematic diagram of a histogram comparing different models under different noisy rolling mill bearing datasets. The chart compares the test set accuracy of DDenseNet and its variants (using db36, coif3 wavelet functions and adding SKAttention), the original DenseNet, MWA-CNN4, MWA-CNN6 and other models under different signal-to-noise ratios (SNR = -4, -2, 0, 4). The results show:

[0124] When there is no noise (SNR=0), the accuracy of the DDenseNet model with SKAttention reaches over 95%, while the original DenseNet is 92.8%. The accuracy of MWA-CNN6 fluctuates greatly due to overfitting;

[0125] When the noise is strong (SNR = -4), the coif3 wavelet + SKAttention model has the highest accuracy (83%), which is significantly higher than MWA-CNN4 (59.3%) and DenseNet (67.3%).

[0126] Overall, DDenseNet and its variants show stronger noise robustness at all noise levels, verifying the effectiveness of multi-scale wavelet features and attention mechanism.

[0127] In one embodiment, Figure 5 As shown, this application also provides a schematic diagram of the mill test set accuracy curve of different models under different noises, which shows the accuracy convergence characteristics of different models during the training process:

[0128] DDenseNet variants: The model using the db36 wavelet converges the fastest, with rapid accuracy gains and good stability under low noise (SNR=4). The model using the coif3 wavelet converges more slowly, but has smaller curve oscillations and the best stability under strong noise (SNR=-4).

[0129] Traditional models: The MWA-CNN series suffers from large oscillation amplitudes and unstable convergence, with significant accuracy fluctuations, especially at low signal-to-noise ratios. The original DenseNet accuracy improvement is slow, and the advantages of multi-scale features are not fully realized.

[0130] Impact of the attention mechanism: The model with SKAttention outperforms the non-attention version at all noise levels, indicating that the attention mechanism effectively enhances the model's ability to selectively extract key features and reduces the impact of noise interference.

[0131] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0132] In one embodiment, a computer device is provided, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the bearing fault diagnosis method and system combined with multi-scale discrete wavelets as described above are implemented.

[0133] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0134] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0135] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. The bearing fault diagnosis method combined with multi-scale discrete wavelet is characterized by: The method comprises: Obtain vibration signal data of rolling mill bearings under different operating conditions and fault types; Dividing the vibration signal data into fixed-length segments and adding corresponding labels to each segment to construct four types of fault data sets; Use the formula to construct a discrete wavelet network and add an improved attention mechanism to construct a deep discrete wavelet dense network structure; After screening the data features of the four types of fault data sets, multi-scale feature fusion is performed and input into the deep discrete wavelet dense network structure to generate a confusion matrix related to the fault; Based on the confusion matrix data, a random neighborhood embedding clustering algorithm is used to obtain the fault category determination results of the bearing under different working conditions.

2. The method according to claim 1, characterized in that The vibration signal data is divided into fixed-length segments and corresponding labels are added to each segment to construct four types of fault data sets, including: Based on the three-part architecture of the rolling mill experimental platform, the sensor layout plan is clarified and a multi-component collaborative experimental environment is constructed; Under the experimental environment, the rolling mill bearings are set according to preset working conditions in combination with artificial fault settings to form four types of bearing fault samples including normal conditions; According to the preset working conditions and speed conditions, the Shannon theorem is used to perform automatic sliding window non-overlap segmentation to divide the four types of bearing fault sample data into fixed-length segment sequences; Based on the fragment sequence, the insufficient point samples are discarded and corresponding labels are added to each fragment sequence according to the fault type and working condition to generate a labeled one-dimensional sample data set; Noise interferences with different noise ratios are superimposed on the one-dimensional sample data set to construct four types of fault data sets containing different signal-to-noise ratio features.

3. The method according to claim 1, characterized in that The deep discrete wavelet dense network structure is constructed by the following steps, including: The following formula is used to construct the expression of discrete wavelet in neural network and obtain discrete wavelet network: Among them, x α,g [n] represents the approximate wavelet coefficients processed by the low-pass filter after α-1 layers, x α,h [n] represents the detail wavelet coefficients processed by the high-pass filter after α-1 layers, g[k] and h[k] are the high-frequency and low-frequency filter coefficients respectively, and 2n-k represents the convolution operation with a downsampling value of 2; The characteristic length of the discrete wavelet network is verified by experiments to obtain the length of each layer filter; S=M+2*Kα=N+2*(V-1)α L α =N / 2 α +2K(1-1 / 2 α )=N / 2 α +2(V-1)(1-1 / 2 α ) Among them, L α Represents the length of the αth layer filter, S represents the total length of the wavelet network, N represents the total length of the input features, α represents the number of decomposition layers, V represents the vanishing distance, V=K+1, K represents the additional parameter term; Determine the length of each layer of wavelet coefficients by the derived filter lengths of each layer; The frequency attention mechanism is used to process the length of the wavelet coefficients of each layer to obtain matrices of different feature scales; The different scales in the different feature scale matrices are stacked on the feature dimension, and the output value of the network block is made consistent with the input value feature dimension through the α-layer residual network structure, thereby obtaining a wavelet residual network block structure for multi-scale feature fusion.

4. The method according to claim 3, characterized in that After obtaining the wavelet residual network block structure of multi-scale feature fusion, the method further includes: Comparing and inferring the similarity between the discrete wavelet network structure and the convolutional neural network structure to obtain a similarity result; The one-dimensional single-channel convolutional neural network is expressed using the following formula: Among them, y i Represents the i-th sequence of the convolutional neural network output, Y=[y1,y2,…,y M-1 ,y M ] represents a one-dimensional vector of length M, and the sequence length should satisfy x i Represents the i-th sequence of convolutional neural network input, X=[x1,x2,…,x L-1 ,x L ] is a one-dimensional vector of length L, w j Represents the jth convolution kernel, W=[w1,w2,…,w K-1 ,w K ], K represents the total length, s represents the step size, and b represents an independent scalar bias value; Based on the similarity results, the four densely connected convolution modules in the dense block are replaced with the multi-scale discrete wavelet residual network block structure, and the improved attention mechanism is alternately inserted as a transition block to obtain the deep discrete wavelet dense network structure.

5. The method according to claim 1, wherein The method of filtering the data features of the four types of fault data sets, performing multi-scale feature fusion and inputting the feature into the deep discrete wavelet dense network structure to generate a confusion matrix related to the fault includes: Obtaining a multi-channel time-frequency feature tensor of the four types of fault data sets; the multi-channel time-frequency feature tensor is generated by discrete wavelet transform of the original vibration signal; Performing channel dimension weight allocation on the multi-channel time-frequency feature tensor to obtain a weighted feature tensor; Performing multi-scale convolution fusion on the segmented features of the weighted feature tensor by utilizing dynamic feature aggregation to generate a multi-scale fused feature tensor; Inputting the multi-scale fusion feature tensor into the deep discrete wavelet dense network to obtain a predicted label; Comparing the predicted labels with the actually added labels, and counting the cases where the predicted labels are consistent or inconsistent with the actual labels to obtain a comparison result; A confusion matrix is ​​generated according to the comparison results; the confusion matrix includes indicators such as accuracy and recall rate of the deep discrete wavelet dense network structure in classifying different fault types.

6. The method according to claim 1, characterized in that The random neighborhood embedding clustering algorithm is used based on the confusion matrix data to obtain the fault category determination results of the bearing under different working conditions, including: A random neighborhood embedding clustering algorithm is used to process the multidimensional feature vector of the vibration signal data to obtain cluster centers and corresponding fault category labels; Acquire a real-time vibration signal of the bearing to be tested; perform feature extraction on the real-time vibration signal based on the fault characteristics of the confusion matrix to generate a target feature vector; Calculating the Euclidean distance between the target feature vector and each cluster center; The Euclidean distance is judged based on a preset distance threshold, and if the Euclidean distance is less than the distance threshold, the real-time vibration signal is marked as a corresponding fault category; A fault category determination result is obtained based on the fault category, which includes an operating condition type and the fault category label.

7. The bearing fault diagnosis system combined with multi-scale discrete wavelet is characterized by: The system comprises: A data processing module is used to obtain vibration signal data of rolling mill bearings under different operating conditions and fault types; it is also used to divide the vibration signal data into fixed-length segments and add corresponding labels to each segment to construct four types of fault data sets; The network construction module is used to construct a discrete wavelet network using the formula and add an improved attention mechanism to construct a deep discrete wavelet dense network structure; The fault classification module is used to screen the data features of the four types of fault data sets, perform multi-scale feature fusion, and input the deep discrete wavelet dense network structure to generate a confusion matrix related to the fault; it is also used to use the random neighborhood embedding clustering algorithm based on the confusion matrix data to obtain the fault category judgment results of the bearing under different working conditions.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images