Rolling bearing fault diagnosis method and system based on wavelet residual shrinkage network

Abstract

The application discloses a rolling bearing fault diagnosis method and system based on a wavelet residual shrinkage network, and the method comprises the following steps: S1, acquiring an original data file of a rolling bearing, parsing a file name, cleaning data, and constructing a standardized vibration signal sample library; S2, constructing a WRSN model; reading a vibration signal from the sample library and performing STFT to obtain a two-dimensional time-frequency spectrum, which is used as the input of the WRSN model; S3, dividing the data set of the two-dimensional time-frequency spectrum into a training set, a verification set and a test set, training the WRSN model through the training set, evaluating the generalization performance of the model after the end of each round, and saving the WRSN model parameters with the optimal performance as a judgment condition; S4, loading the saved optimal WRSN model weight, inputting the test set into the model for forward propagation reasoning; the model performs feature extraction and classification on the time-frequency spectrum, and finally outputs the fault type diagnosis result corresponding to the test sample.

Description

Technical Field

[0001] This invention belongs to the field of industrial artificial intelligence and fault diagnosis technology, and in particular relates to a method and system for fault diagnosis of rolling bearings based on wavelet residual shrinkage network. Background Technology

[0002] Rolling bearings are core components in rotating machinery, and their health directly affects the safe and stable operation of the entire production system. In actual industrial production, bearings often operate under high speed, high load, and complex, high-noise environments for extended periods. If a bearing failure goes undetected, it can lead to equipment downtime and significant economic losses, or even catastrophic safety accidents. Therefore, accurate diagnosis of rolling bearing failure types and severity under high-noise conditions is crucial for industrial safety and predictive maintenance.

[0003] Existing fault diagnosis methods mainly include traditional signal processing methods based on physical models and data-driven deep learning methods. Although early physical modeling methods and traditional signal processing techniques (such as Fourier transform and Hilbert transform) perform reasonably well in low-noise environments, they heavily rely on expert experience for manual feature extraction, and their generalization ability and robustness are poor when processing non-stationary and strong noise signals.

[0004] With the development of deep learning technology, Convolutional Neural Networks (CNNs) and Residual Networks (ResNets) have been widely used due to their powerful end-to-end feature extraction capabilities. To address the noise interference problem, methods such as Deep Residual Shrink Networks (DRSNs) have introduced soft thresholding mechanisms, which have enhanced noise resistance to some extent. However, existing deep learning diagnostic techniques still have the following significant drawbacks in practical applications:

[0005] (1) The first layer of convolutional kernels in traditional convolutional neural networks is usually randomly initialized and trained. This "black box" feature extraction lacks clear physical meaning. When processing broadband vibration signals, it is impossible to effectively distinguish between low-frequency background running signals and high-frequency fault impact signals. This can easily lead to the mixing of signals of different frequency bands with random noise, resulting in frequency aliasing, which can mask weak fault features.

[0006] (2) Most existing denoising mechanisms operate only in the time domain and do not make full use of the physical interpretability of the frequency domain. When faced with harsh working conditions in industrial sites with extremely low signal-to-noise ratios (SNR) or complex noise types (such as mixed noise), their adaptive capabilities and diagnostic accuracy will drop significantly.

[0007] (3) Most existing studies only focus on identifying the location type of the fault (such as inner ring, outer ring, sphere fault), while neglecting the refined assessment of the severity of the fault (such as mild, moderate, and severe damage). In actual operation and maintenance, accurate classification of the severity of the fault is the key basis for formulating maintenance strategies and predicting the remaining service life. Especially under strong noise interference, the characteristics of moderate faults are easily confused with severe faults or background noise, leading to misjudgment.

[0008] Therefore, there is an urgent need in this field for a technical solution that can combine prior physical knowledge with the adaptive capabilities of deep learning to achieve high-precision fault type identification and fault severity determination in a noisy environment. Summary of the Invention

[0009] The purpose of this invention is to provide a rolling bearing fault diagnosis method and system based on wavelet residual shrinkage network, so as to solve the technical problems of frequency aliasing, difficulty in weak feature extraction, and inability to accurately determine the severity of faults in the existing technology under strong noise environment.

[0010] The technical solution of the present invention is as follows:

[0011] A method for fault diagnosis of rolling bearings based on wavelet residual shrinkage networks includes the following steps:

[0012] S1. Obtain publicly available vibration signal datasets for rolling bearings, parse data file names (such as operating condition information), clean data (such as redundant data), and construct a standardized vibration signal sample library.

[0013] S2. Construct a wavelet residual shrinking network (WRSN) model; read vibration signals from the sample library, add noise, and perform short-time Fourier transform (STFT) to obtain a two-dimensional time spectrum, which is used as the input of the wavelet residual shrinking network model;

[0014] S3. Divide the time-spectrum dataset into a training set, a validation set, and a test set. Train the wavelet residual shrinking network model using the training set. Evaluate the generalization performance of the model after each round using the validation set. Save the WRSN model parameters with the highest validation set accuracy as the criterion.

[0015] S4. Load the saved optimal WRSN model weights and input the test set into the model for forward propagation inference; the model extracts and classifies the features of the input time-spectrum graph and finally outputs the fault type diagnosis result corresponding to the test sample.

[0016] Preferably, step S1 includes acquiring data, parsing filenames, and classifying and storing them to construct a standardized vibration signal sample library. The specific process includes:

[0017] S11. Obtain data

[0018] Obtain the raw data file of the rolling bearing; preferably, obtain the HUSTBearing dataset, which contains 9 health states and 11 speed conditions.

[0019] S12. Identify characteristic characters in the original data file name;

[0020] The characteristic characters include a coefficient identifier representing the degree of damage (such as "0.5X") and an abbreviation representing the location of the fault (I, O, B, C).

[0021] If the file name contains a coefficient identifier representing moderate damage (such as "0.5X"), it is determined to be a moderate fault.

[0022] If the file name does not contain a damage coefficient identifier, it is determined to be a severe fault; at the same time, the letter abbreviations (I, O, B, C) representing the faulty part are identified.

[0023] S13. Tag system construction: Based on the aforementioned feature characters, the data is divided into 9 states, including: normal state, moderate inner circle fault, severe inner circle fault, moderate outer circle fault, severe outer circle fault, moderate sphere fault, severe sphere fault, moderate composite fault, and severe composite fault.

[0024] S14. Read the contents of the original data file, retain the vibration signal channels (X, Y, Z channels) collected by the accelerometer, and remove the redundant speed channels (Speed ​​channel) under constant speed conditions.

[0025] S15. Convert the cleaned data into a unified streaming data format (CSV) and automatically categorize it into subfolders named according to health status and body part.

[0026] Preferably, in step S2, constructing the wavelet residual shrinking network (WRSN) model includes:

[0027] S21. Construct a wavelet decomposition convolutional layer (WDConv) at the network input. The wavelet decomposition convolutional layer (WDConv) is a feature extraction layer based on physical priors. The construction method is as follows: select the Daubechies 4 (db4) wavelet as the basis function and obtain its one-dimensional low-pass filter vector G. L and a one-dimensional high-pass filter vector G H The one-dimensional filter vector is expanded into a two-dimensional separable convolution kernel through the outer product operation:

[0028]

[0029] in, Represents the set of real numbers. G is a vector containing 8 real numbers. L Represents a low-pass core, G H Represents the high-pass kernel, V represents the outer product operator, T represents the transpose operator, i and j represent the index positions of the filter coefficient vector, and G is the two-dimensional matrix generated by the outer product operation. L The element in the i-th row and j-th column is This method extends a one-dimensional filter into a two-dimensional filter, enabling it to capture low-frequency information of a signal simultaneously in both the horizontal and vertical directions.

[0030] S22. During the forward propagation of the WRSN model, the WDConv layer adopts a grouped convolution method, setting the number of convolution groups to be equal to the number of channels of the input feature map, and using the two-dimensional convolution kernel constructed in step S21 to independently perform convolution operation on each channel of the input two-dimensional time-spectrum map.

[0031] Preferably, in step S2, channel-wise residual shrinking units (RSBU-CW) are stacked deep within the network to adaptively learn feature thresholds and perform soft thresholding denoising.

[0032] After performing an absolute value operation on the input feature map x, global average pooling is used to compress the two-dimensional feature map into a one-dimensional statistical feature vector. This one-dimensional statistical feature vector is then input into a threshold generation subnetwork of a channel-wise residual shrinking unit, and a scaling parameter α is output through a sigmoid activation function. c The formula is as follows:

[0033]

[0034] Where, α c This represents the scaling factor for the c-th channel of the output learned by the subnetwork, with a value ranging from (0,1). c This represents the feature value of the c-th neuron.

[0035] Then, the scaling parameters are used to calculate the independent threshold for each channel. The formula is as follows:

[0036]

[0037] in, α represents the adaptive threshold for the c-th channel. c This represents the scaling factor of the c-th channel of the subnetwork's learned output, where i, j, and c are the width, height, and channel indices of the feature map x.

[0038] Finally, the calculated threshold is used. The corresponding channels of the original feature map are subjected to soft thresholding to obtain the output feature map, as shown in the following formula:

[0039]

[0040] Where y i,j,c For output features, x i,j,c The original input features are represented by sign(·), which is the sign function.

[0041] Preferably, in step S2, the channel-by-channel residual shrinking unit (RSBU-CW) is used to adaptively reduce noise on the features after WDConv decomposition, and the processing includes:

[0042] After performing an absolute value operation on the input feature map x, the two-dimensional feature map is compressed into a one-dimensional statistical feature vector through global average pooling (GAP).

[0043] The vector is input into a threshold generation subnetwork, and the scaling parameter α is output through the Sigmoid activation function. c Then, the independent threshold for each channel is calculated:

[0044]

[0045] Soft thresholding is applied to the corresponding channels of the original feature map using the threshold:

[0046]

[0047] in, α represents the adaptive threshold for the c-th channel. c This represents the scaling factor for the c-th channel of the output learned by the subnetwork, with a value ranging from (0,1). c is the feature value at the c-th neuron, and i,j,c are the width, height, and channel indices of the feature map x.

[0048] Preferably, in step S2, the vibration signals are labeled according to different fault types and fault severity; and the labeled vibration signals are subjected to STFT.

[0049] In at least one embodiment of the present invention, step S2, which involves labeling the vibration signal according to different fault types and fault severity, specifically includes:

[0050] Label 0 represents the normal state (H);

[0051] Label 1 represents a moderate inner race fault (0.5_I);

[0052] Label 2 represents severe inner race fault (I);

[0053] Label 3 represents a moderate outer ring fault (0.5X_O);

[0054] Label 4 indicates a severe outer ring fault (O);

[0055] Label 5 represents a moderate sphere failure (0.5X_B);

[0056] Label 6 represents severe sphere failure (B);

[0057] Label 7 represents a moderate complex fault (0.5X_C);

[0058] Label 8 represents a severe complex fault (C).

[0059] Each health condition corresponds to eleven different operating conditions, including 20, 25, 30, 35, 40, 60, 65, 70, 75 and a variable speed condition of 0-40-0Hz, where the variable speed condition is a uniform speed change within two seconds.

[0060] All signals were acquired at a sampling frequency of 25600Hz, with 262144 data points recorded each time (approximately 10.2 seconds) to ensure signal integrity and authenticity. The long signal duration also makes the simulated faults more consistent with real-world applications.

[0061] To verify the robustness of the model in complex industrial environments, Gaussian white noise, Laplace noise, pink noise, and mixed noise were introduced into the original dataset to construct multiple different noise datasets.

[0062] The noise intensity is controlled by the signal-to-noise ratio (SNR), and the noise power added is based on a preset SNR value. This represents the signal power. The calculation formula is as follows:

[0063]

[0064] Where Psignal is the signal power;

[0065] Preferably, in step S2, performing STFT on the signal in the noise dataset includes:

[0066] The sampling frequency was set to 25600Hz, the Hann window function was used, the window size was set to 1024, and the number of overlap points was set to 256.

[0067] The transformed two-dimensional time-spectrum image is uniformly scaled to 64×64 pixels to adapt to the network input dimension, and the time-spectrum image is used as the input of the wavelet residual shrinkage network model.

[0068] Preferably, in step 3, training the wavelet residual shrinkage network model includes:

[0069] To improve the model's robustness in complex industrial environments, different types of noise interference were artificially injected into the original acquired vibration signals, including Gaussian white noise, Laplace noise, pink noise, and mixtures of these noises. Subsequently, the noisy one-dimensional signal was converted into a two-dimensional time-spectrum graph using Short Time Fourier Transform (STFT), generating a noisy time-spectrum graph dataset.

[0070] The noisy time-spectrum dataset constructed above was randomly divided into training, validation, and test sets in a 6:2:2 ratio. For model training hyperparameter settings, the batch size was set to 32, and the total number of iterations was set to 100.

[0071] The WRSN model is iteratively trained using the training set. In each iteration, data is input into the model to obtain prediction results, and the difference between the predicted distribution and the true label is calculated using the cross-entropy loss function. The calculation process is as follows: First, the nine outputs of the fully connected layer of the model are converted into a probability distribution using the softmax function, and the calculation formula is as follows:

[0072]

[0073] Among them, z i p is the output value for the i-th category; i Let be the predicted probability of the i-th type of fault. Then, the difference between the predicted probability distribution and the true label is calculated using the cross-entropy loss function, as shown in the formula:

[0074] Among them, y i The true label for the sample. When the sample belongs to the i-th type of fault, y... i =1, otherwise y i =0. The smaller this loss value, the closer the model's prediction is to the actual fault type. The SGD optimization algorithm, combined with the L2 regularization mechanism, updates the network parameters based on the calculated gradients until the loss function converges.

[0075] After each training round, forward inference is performed on the current model using validation set data to calculate the validation set accuracy. The system monitors model performance in real time and automatically saves the WRSN model weight file with the highest validation set accuracy as the criterion for subsequent testing and diagnostics.

[0076] This invention also discloses a rolling bearing fault diagnosis system based on a wavelet residual shrinkage network, used to perform the above method, including:

[0077] The data processing and time-frequency conversion module acquires the raw data file of the rolling bearing, parses the file name and cleans the data, and constructs a standardized vibration signal sample library; it reads the vibration signal from the sample library and performs STFT to obtain a two-dimensional time-frequency spectrum, which is used as the input of the WRSN model;

[0078] The model building module is used to build the wavelet residual shrinking network WRSN model;

[0079] The model training module is used to divide the two-dimensional time-spectrum graph into a training set and a test set, train the WRSN model using the training set, and save the trained WRSN model when the end-of-training condition is met.

[0080] The fault diagnosis module is used to input the test set into the trained WRSN model and output diagnostic results including the fault type and severity.

[0081] Compared with the prior art, the present invention has the following beneficial technical effects:

[0082] This invention uses a fixed wavelet decomposition convolutional layer (WDConv) to replace the traditional first-layer convolution. By using a wavelet basis with clear physical meaning, it achieves frequency band separation between low-frequency background and high-frequency impulse, effectively avoiding the frequency aliasing problem caused by traditional convolution.

[0083] This invention combines the adaptive denoising capabilities of wavelet physical frequency division and residual shrinking unit (RSBU-CW), and the synergistic effect of the two results in higher diagnostic accuracy in complex noise environments.

[0084] This invention can simultaneously identify fault types and make precise judgments on fault severity (mild, medium, and severe). In particular, it has a much higher accuracy rate than existing models in identifying easily confused medium-severity faults, providing a more accurate basis for predictive maintenance of equipment. Attached Figure Description

[0085] Figure 1 This is a flowchart of a bearing fault diagnosis method based on a wavelet residual shrinking network according to a preferred embodiment of the present invention. The flowchart also illustrates the internal data processing logic of the fault diagnosis technology of the present invention.

[0086] Figure 2 This is a schematic diagram of the front-end interactive interface and diagnostic result display provided in a preferred embodiment of the present invention;

[0087] Figure 3 This is a topology diagram of a wavelet residual shrinking network (WRSN) provided in a preferred embodiment of the present invention; the model integrates a wavelet decomposition convolutional layer (WDConv) for physical frequency division and a residual shrinking unit (RSBU-CW) for adaptive denoising.

[0088] Figure 4 This is a schematic diagram of the frequency band separation of the wavelet decomposition convolutional layer (WDConv) in the preprocessing stage provided by a preferred embodiment of the present invention; wherein, (a) is the original STFT time spectrum diagram, (b) is the frequency aliasing effect diagram of the first layer output of the existing DRSN model, (c) is the low-frequency energy (representative channel) diagram extracted by the WRSN of the present invention, (d) is the high-frequency energy (representative channel) diagram extracted by the WRSN of the present invention, (e) is a local magnified diagram of the high-frequency energy, and (f) is the true value of the high-frequency energy;

[0089] Figure 5 This is a visual comparison of low-pass and high-pass kernels extracted from wavelet decomposition convolutional layers according to a preferred embodiment of the present invention.

[0090] Figure 6 This is a training loss versus validation loss curve provided in a preferred embodiment of the present invention;

[0091] Figure 7 This is a comparison chart of the average accuracy and stability error bars of the method of this invention and existing mainstream methods under different noise levels; the comparison models include: 1. VIHGNN, 2. ViT, 3. Swin-T, 4. WDCNN, 5. 1DMCICNN, 6. DRSN, and 7. WRSN, which represents the method of this invention.

[0092] Figure 8 The provided diagram shows a comparison of the classification confusion matrices of different models under 0dB Gaussian noise; where (a)-(g) correspond to VIHGNN, ViT, Swin-T, WDCNN, 1DMCICNN, DRSN, and WRSN (the method of this invention) respectively.

[0093] Figure 9 The provided high-level feature t-SNE visualization distribution diagrams of different models under 0dB Gaussian noise are shown; where (a)-(g) correspond to VIHGNN, ViT, Swin-T, WDCNN, 1DMCICNN, DRSN and WRSN (the method of this invention) respectively.

[0094] Figure 10 This is a block diagram of a rolling bearing fault diagnosis system based on a wavelet residual shrinkage network, according to a preferred embodiment of the present invention. Detailed Implementation

[0095] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this invention. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0096] Example 1

[0097] like Figure 1 As shown, this embodiment provides a rolling bearing fault diagnosis method based on wavelet residual shrinkage network, including the following steps:

[0098] S1. Obtain publicly available vibration signal datasets for rolling bearings, parse the operating condition information in the data file names, clean up redundant data, and construct a standardized vibration signal sample library.

[0099] S2. Construct a wavelet residual shrinking network (WRSN) model; label the vibration signals according to different fault types and fault severity; perform STFT on the labeled vibration signals to obtain a two-dimensional time spectrum, and use the time spectrum as the input of the wavelet residual shrinking network model;

[0100] S3. Divide the dataset of the two-dimensional time-spectrum graph into a training set, a validation set and a test set. Train the wavelet residual shrinkage network model through the training set. Evaluate the generalization performance of the model after each round of the validation set. Save the parameters of the WRSN model with the highest performance based on the highest accuracy of the validation set.

[0101] S4. Load the saved optimal WRSN model weights, input the test set into the model for forward propagation inference; the model extracts deep fault features and finally outputs the fault type diagnosis result corresponding to the test sample, as shown in the output. Figure 2 As shown.

[0102] The following is a more detailed description of each step.

[0103] Step S1: In this embodiment, a standardized vibration signal sample library is first constructed.

[0104] In step S1 of this embodiment, the HUSTBearing dataset is acquired. This dataset was collected using the Spectra-Quest mechanical failure simulator and contains vibration data of the ER-16K rolling bearing under different health conditions. To address the issue that the original data is in Excel format and contains redundant channels, this embodiment designs an automated parsing and cleaning process:

[0105] Filename parsing: Identify characteristic characters in the filename. If the filename contains a coefficient identifier representing moderate damage (0.5X in this example), it is determined to be a moderate fault; if the filename does not contain a damage coefficient identifier, it is determined to be a severe fault; at the same time, identify letter abbreviations (I, O, B, C) representing the faulty part; for example, extract "0.5X" as the moderate damage identifier, extract "B" as the sphere fault identifier, and extract "65Hz" as the speed condition identifier from the filename "0.5X_B_65Hz.xls".

[0106] Based on characteristic characters, the data is divided into 9 states, including: normal state, moderate inner circle fault, severe inner circle fault, moderate outer circle fault, severe outer circle fault, moderate sphere fault, severe sphere fault, moderate combined fault, and severe combined fault.

[0107] Data cleaning: Read the original file content, remove redundant speed recording channels (Speed ​​column) under constant speed conditions, and retain only the X, Y, and Z axis vibration signal channels collected by the acceleration sensor.

[0108] Standardized storage: The cleaned data is converted into a unified CSV streaming format and automatically categorized into the corresponding folders based on the parsed tags.

[0109] In step S2 of this embodiment, the overall structure of the constructed wavelet residual shrinking network (WRSN) model is as follows: Figure 3 As shown. This model combines the physical frequency division characteristics of wavelet transform with the adaptive noise reduction capability of residual shrinking networks. Specifically, constructing the WRSN model includes the following steps:

[0110] S21. Construct a wavelet decomposition convolutional layer (WDConv) at the network input:

[0111] Existing deep learning models typically use randomly initialized convolutional kernels, which can easily lead to the mixing of low-frequency operating signals and high-frequency fault impact signals when processing strong noise signals, i.e., frequency aliasing. To solve this problem, this embodiment designs a WDConv layer. Figure 4 The experimental results shown verify its effectiveness.

[0112] The specific construction method is as follows: In this embodiment, the Daubechies 4 (db4) wavelet is selected as the basis function, and its one-dimensional low-pass filter vector is used. and one-dimensional high-pass filter vector It is extended into a two-dimensional separable convolution kernel through the outer product operation, and the calculation formula is as follows:

[0113]

[0114] Where V represents the outer product operation, T represents the vector transpose; G L G H This represents the generated two-dimensional separable convolution kernel. During the network forward propagation, a grouped convolution method is used, with the number of convolution groups equal to the number of channels in the input feature map. The aforementioned two-dimensional filter kernel is used to independently perform convolution operations on each channel of the input two-dimensional time-spectrum map to achieve physical frequency band separation.

[0115] To intuitively illustrate the essential difference between the WDConv layer and traditional convolutional layers, please refer to [link / reference]. Figure 5 .

[0116] Figure 5 In the image, (a) shows the visualization of the generated two-dimensional low-pass convolution kernel. It can be seen that the convolution kernel exhibits a smooth color transition (uniform weight distribution). This structure enables it to "smooth" the input signal, thereby effectively extracting the low-frequency background trend and steady-state components in the signal.

[0117] Figure 5 In the image, (b) shows a visualization of the generated two-dimensional high-pass convolution kernel. (Compared to...) Figure 5 In stark contrast to (a) in this example, the convolution kernel exhibits a distinct checkerboard pattern of alternating red and blue (positive and negative weights). This specific texture structure endows it with extremely high sensitivity to high-frequency abrupt signals, enabling it to accurately capture transient impacts and edge details in signals, much like an edge detection operator.

[0118] Compared to the randomly initialized "black box" convolutional kernels in traditional CNNs, this invention... Figure 5 The convolutional kernels shown have clear physical texture features, ensuring that the network can process signals according to physical laws in the initial stage.

[0119] S22. During the forward propagation of the WRSN model, the WDConv layer adopts a grouped convolution method, setting the number of convolution groups to be equal to the number of channels of the input feature map, and using the two-dimensional convolution kernel constructed in step S21 to independently perform convolution operation on each channel of the input two-dimensional time-spectrum map.

[0120] In deep network stacking of channel-by-channel residual shrinkage units (RSBU-CW): such as Figure 3As shown in BasicBlock, the RSBU-CW unit is used for adaptively learning feature thresholds and performing soft thresholding denoising. Its construction process includes:

[0121] 1) Feature compression: After performing absolute value operation on the input feature map x, the two-dimensional feature map is compressed into a one-dimensional statistical feature vector through global average pooling (GAP);

[0122] 2) Threshold calculation: The one-dimensional statistical feature vector is input into a threshold generation subnetwork (containing two fully connected layers), and the scaling parameter α is output through the Sigmoid activation function. c The formula is as follows:

[0123]

[0124] Where, α c This represents the scaling factor for the c-th channel of the output learned by the subnetwork, with a value ranging from (0,1). c This represents the feature value of the c-th neuron.

[0125] Then, the scaling parameters are used to calculate the independent threshold for each channel. The formula is as follows:

[0126]

[0127] in, α represents the adaptive threshold for the c-th channel. c This represents the scaling factor of the c-th channel of the subnetwork's learned output, where i, j, and c are the width, height, and channel indices of the feature map x.

[0128] 3) Soft thresholding: using the calculated threshold The corresponding channels of the original feature map are subjected to soft thresholding to obtain the feature map, as shown in the following formula:

[0129]

[0130] in, For output features, The original input features are represented by sign(·), which is the sign function.

[0131] In this embodiment, the specific parameters of the WRSN model are shown in Table 1.

[0132] Table 1

[0133]

[0134] To achieve fault classification, a classification module is constructed at the output of the WRSN model (i.e., the Head section in Table 1). This module first utilizes a Global Average Pooling (GAP) layer to perform average pooling on the 3D feature map output by the last convolutional module (512×7×7 in this embodiment), compressing it into a 512×1×1 one-dimensional feature vector. Compared to the traditional method of directly flattening the feature map and connecting it to a fully connected layer, GAP significantly reduces the number of parameters, not only reducing the risk of model overfitting but also enhancing robustness to spatial translation. Subsequently, this feature vector is input to a fully connected layer (FC), mapping the 512-dimensional feature vector output by GAP to 9 output nodes corresponding to the number of fault categories. Finally, the output result is processed by the Softmax function to obtain the probability distribution of the sample belonging to each category.

[0135] In step S2 of this embodiment, the vibration signal is labeled according to different fault types and severity, and is divided into 9 health states. The category labels determined in this embodiment are shown in Table 2.

[0136] Table 2

[0137]

[0138] In step S2 of this embodiment, the labeled vibration signal is subjected to STFT to obtain a two-dimensional time-frequency spectrum. Specifically, this includes:

[0139] The sampling frequency was set to 25600Hz, the Hann window function was used, the window size was set to 1024, and the number of overlap points was set to 256. This parameter setting can balance time resolution and frequency resolution, effectively capturing the transient impact characteristics of bearing failure.

[0140] To adapt to network input and control computational costs, the transformed two-dimensional time-spectrum image is uniformly scaled (resized) to 64×64 pixels and used as the input tensor of the WRSN model.

[0141] In step S3 of this embodiment, the time-spectrum graph is divided and the WRSN model is trained, specifically including:

[0142] Dataset partitioning: The sample library is randomly divided into training set, validation set and test set in a ratio of 6:2:2.

[0143] Data augmentation: To improve the robustness of the model in noisy environments, Gaussian white noise, Laplace noise, pink noise, and mixed noise (Gaussian + pink) are introduced into the dataset to construct a multi-signal-to-noise ratio (SNR) training environment.

[0144] Network Training: The WRSN model was trained using the partitioned training set. During training, the batch size was set to 32 samples, and the number of epochs was set to 100. The SGD optimizer was used, with an initial learning rate of 0.1, which decreased with each training epoch (0.1 for the first 40 epochs, 0.01 for the middle 40 epochs, and 0.001 for the last 20 epochs). L2 regularization (coefficient 0.0001) was introduced to prevent overfitting. The cross-entropy loss function was used.

[0145] The network training process is as follows: The program executes 100 iterations, each consisting of two phases: training and validation. First, the model enters training mode and loads the training set data. Through forward propagation, the data is sequentially fed into wavelet convolutional layers, residual modules, and the classification head structure to obtain the predicted output. The loss is calculated using the cross-entropy loss function, specifically by first converting the output into a probability distribution using Softmax, and then calculating the log-likelihood loss between this distribution and the true label. Subsequently, the program performs backpropagation on this loss value to calculate the gradients of the parameters in each layer and calls the SGD optimizer to update the network weights. The optimizer is configured with an initial learning rate of 0.1, a momentum of 0.9, and an L2 regularization coefficient to prevent overfitting.

[0146] Immediately following in the same round, the model switches to evaluation mode and loads validation set data. The validation set is used at this stage to monitor model performance and model selection. The program performs forward propagation to calculate the current accuracy and loss, but gradient calculation is disabled, so validation set data does not trigger backpropagation or weight updates. The current validation set accuracy is compared to the historical best accuracy; if the current performance is better, the current model parameters are saved as the optimal model weights. This mechanism ensures that the final retained model is the version with the strongest generalization ability on the validation set, rather than simply the version that best fits the training set. The test set is completely independent of the above process and is only used for final inference evaluation after the entire training process is completed and the optimal model weights are loaded.

[0147] In step S4 of this embodiment, the time-spectrum graph of the test set is input into the trained WRSN model, and the diagnostic results are output.

[0148] Figure 6 The curve showing the change in loss value during model training is presented. It can be seen that as the number of iterations increases, the model loss value decreases rapidly and converges, indicating that the model training is stable. However, due to the presence of high noise, the loss value does not decrease to 0.

[0149] To verify the effectiveness of the method of the present invention, an ablation experiment was conducted below.

[0150] Table 3 shows a comparison of the ablation experimental results of the method of this invention (WRSN) with existing mainstream methods (such as ResNet18 benchmark, DRSN, etc.) in a 0dB mixed noise environment. The results show that introducing WDConv or RSBU-CW alone can improve accuracy, while the WRSN model combining both achieves an accuracy of 92.31%, an improvement of 2.90% compared to the benchmark model. This further confirms... Figure 4 and Figure 5 Conclusion: WDConv ( Figure 5 The clear frequency band separation provided () Figure 4 This provides higher-quality feature input for RSBU-CW's adaptive denoising, and the two produce a synergistic effect.

[0151] Table 3

[0152]

[0153] To fully verify the technical effectiveness of the proposed WRSN model in complex industrial environments, various noise environments were constructed, and detailed comparative experiments were conducted with existing mainstream advanced models. The comparative models included: VIHGNN, ViT, Swin-T, WDCNN, 1DMCICNN, and the Deep Residual Shrinking Network (DRSN) as the benchmark.

[0154] The types of noise in industrial environments are complex and varied. In this embodiment, Gaussian white noise, Laplace noise, pink noise, and mixed noise were superimposed on the original signal, and tests were conducted at different signal-to-noise ratios (SNR, from -10dB to 30dB).

[0155] Gaussian noise environment experiment (corresponding to Table 4). Table 4 is a comparison table of the diagnostic accuracy of the method of the present invention and the existing mainstream methods in the presence of Gaussian white noise.

[0156] Table 4

[0157]

[0158] Table 4 shows a comparison of the diagnostic accuracy of each model under Gaussian white noise conditions (warmer colors indicate higher accuracy).

[0159] Experimental data show that all models perform well when the SNR is greater than 0 dB; however, in environments with strong noise (such as -4 dB and below), the performance of Transformer-type models such as Swin-T drops sharply or even fails. In contrast, the WRSN model of this invention (rightmost column) maintains optimal performance across the entire SNR range. Especially in the 0 dB environment, the accuracy of WRSN reaches 87.72%, significantly better than DRSN's 84.56%.

[0160] Laplace noise environment experiment (corresponding to Table 5). Table 5 is a comparison table of the diagnostic accuracy of the method of the present invention and the existing mainstream methods in the environment with added Laplace noise.

[0161] Table 5

[0162]

[0163] Table 5 shows a comparison of the diagnostic accuracy of each model under Laplace noise conditions.

[0164] Laplace noise exhibits significant impulse characteristics, simulating transient impact interference in industrial environments. Data results show that, thanks to the physical capture capability of the WDConv layer for high-frequency transient components, the WRSN model demonstrates extremely strong robustness against impulse noise, maintaining an accuracy of 77.67% even at -4dB, outperforming all comparative models.

[0165] Pink noise environment experiment (corresponding to Table 6). Table 6 is a comparison table of the diagnostic accuracy of the method of the present invention and the existing mainstream methods in the presence of added pink noise.

[0166] Table 6

[0167]

[0168] Table 6 shows a comparison of the diagnostic accuracy of each model under an environment with added pink noise (1 / f noise).

[0169] Pink noise is mainly concentrated in the low-frequency band, which can easily mask the bearing's rotational frequency characteristics. The low-pass filter of the WDConv layer plays a crucial role here, effectively extracting the low-frequency operating trend. Experiments show that at an extremely low signal-to-noise ratio of -10dB, WRSN can still achieve an accuracy of 53.81%, while most of the comparison models are no longer effective.

[0170] Experiments in mixed noise environment (corresponding to Table 7). Table 7 is a comparison table of the diagnostic accuracy of the method of the present invention and the existing mainstream methods in the environment with added mixed noise (Gaussian + pink).

[0171] Table 7

[0172]

[0173] Table 7 shows a comparison of the diagnostic accuracy of each model under mixed noise (Gaussian + pink) conditions.

[0174] This is the most realistic complex noise environment in industrial settings. The results show that the WRSN model demonstrates the advantage of "cooperative noise reduction": WDConv performs physical separation in the frequency domain, while RSBU-CW performs adaptive soft-threshold denoising in the feature domain. Under 0dB mixed noise, the WRSN accuracy reaches 73.54%, an improvement of 2.19 percentage points compared to the benchmark DRSN model (71.35%), proving the practical value of the proposed solution under complex working conditions.

[0175] like Figure 7 As shown, the average accuracy and error bars of each model under all noise test conditions are displayed. In the figure, the height of the bar represents the average accuracy, and the error bar represents the half-range. It can be seen that WRSN (the pink bar on the far right) has the highest average accuracy in all four noise environments, reaching 77.02% (Gaussian), 78.47% (Laplace), 76.67% (pink), and 66.69% (mixed). It also exhibits the strongest stability; the error bar length of WRSN is the shortest among all models. This indicates that the method of this invention has extremely strong adaptability to fluctuations of different noise intensities and types, is not easily affected by random factors, and has strong generalization ability.

[0176] To verify the ability of this invention to distinguish the severity of faults, this embodiment plotted a confusion matrix under 0dB Gaussian noise.

[0177] like Figure 8 As shown, the classification details of each model are compared:

[0178] Figure 8 In the diagram, (f) represents the confusion matrix of the baseline model DRSN. It can be seen that it performs extremely poorly in identifying moderate inner-ring faults (0.5X_I) and moderate outer-ring faults (0.5X_O), with accuracies of only 25% and 22%, respectively. A large number of samples are misclassified as mixed faults or spherical faults. This indicates that traditional convolution cannot extract weak, moderate damage features.

[0179] Figure 8 In the diagram, (g) represents the confusion matrix of the WRSN of this invention. Under the same noise conditions, the WRSN improves the recognition rate of moderate inner ring faults to 94% and the recognition rate of moderate outer ring faults to 93%.

[0180] This striking contrast ( Figure 8 (f) vs Figure 8 (g) in the paper fully demonstrates that by introducing wavelet decomposition convolution with physical meaning, the present invention successfully solves the technical problem that weak fault features (especially moderate faults) are easily submerged and confused under strong noise, and achieves truly refined diagnosis.

[0181] at last, Figure 9The t-SNE plot further confirms the above conclusion from the perspective of feature distribution, showing that the features extracted by WRSN are significantly better than other models in terms of inter-class separation.

[0182] In summary, the rolling bearing fault diagnosis method based on wavelet residual shrinking networks of this invention, after automated cleaning and STFT transformation of the original vibration signal, directly uses the time-spectrum diagram containing physical frequency band information as model input. This allows the model to achieve refined diagnosis of rolling bearing fault types and severity without relying on complex manual feature engineering. This invention has the advantages of strong noise resistance, high diagnostic accuracy, and good generalization ability, and is particularly suitable for high-noise environments in industrial settings.

[0183] Example 2

[0184] like Figure 10 As shown, based on the same inventive concept, this embodiment of the invention also provides a rolling bearing fault diagnosis system based on a wavelet residual shrinkage network, which is used to perform the fault diagnosis method described in Embodiment 1 above.

[0185] like Figure 1 (Flowchart) and Figure 2 As shown in the (front-end interactive interface), the system in this embodiment mainly includes the following four core functional modules in terms of logical architecture:

[0186] The data processing and time-frequency conversion module's main function is to establish a standard input channel and format the data to be measured. This module first acquires the raw vibration signal data file of the rolling bearing, automatically parses characteristic characters in the filename to identify the operating condition label, and simultaneously removes redundant channels such as those for constant speed, retaining only the vibration signal channel, thereby constructing a standardized vibration signal sample library. It then reads the vibration signals from the sample library for STFT (Simultaneous Time-Frequency Transformation) to generate a two-dimensional time-frequency spectrum for the model input.

[0187] Function Description: This module reads vibration signals from a sample library and uses a sliding window algorithm (window size 1024, overlap 256) to divide long-sequence signals into short-time segments. Then, a STFT is performed on each segment, with a sampling frequency of 25600Hz and a Hann window function. The generated two-dimensional time-spectrum image is normalized and then uniformly scaled to 64×64 pixels using image interpolation. The generated image is saved as a standardized image dataset, serving as the standard input tensor for subsequent deep learning models.

[0188] Model building module: Its main function is to build wavelet residual shrinking network (WRSN) model instances based on the deep learning framework (model topology as follows). Figure 3 (As shown). The network structure constructed by this module includes the following units:

[0189] Wavelet decomposition convolutional unit: Utilizing fixed Daubechies (db4) wavelet basis functions, the one-dimensional low-pass / high-pass filter vector is expanded into a two-dimensional separable convolutional kernel (8×8) through an outer product operation. During the forward propagation of the convolutional layer, grouped convolution is used to perform physical band separation on the input temporal spectrum (e.g., ...). Figure 4 and Figure 5 (As shown), to solve the frequency aliasing problem.

[0190] Adaptive Denoising Unit: Using the channel-wise residual shrinking unit (RSBU-CW), and taking advantage of global average pooling and the statistical distribution of the perceptual feature map of the fully connected layer, a soft threshold is adaptively generated to remove background noise.

[0191] Model training module: Its main function is to manage the iterative optimization process of the model. This module is responsible for randomly dividing the two-dimensional time-spectrum dataset into training, validation, and test sets according to a preset ratio (e.g., 6:2:2). Simultaneously, to enhance robustness, this module also injects Gaussian white noise, Laplace noise, or pink noise into the training data to construct a noisy dataset.

[0192] Function Description: During training, this module inputs data into the model in batches and calculates the loss between the model's predicted values ​​and the true labels using the cross-entropy loss function. Subsequently, the loss values ​​are backpropagated to calculate the gradients of the parameters at each layer, and the network weights of the WRSN model are updated based on the gradients using the SGD optimizer (configured with momentum 0.9 and L2 regularization). This module also evaluates the validation set accuracy after each training round and automatically saves the optimal model weight file.

[0193] Fault Diagnosis Module: Its main function is to perform online or offline fault reasoning. This module loads the optimal WRSN model weights saved during training and inputs the two-dimensional time-spectrum graph to be tested into the model. The specific interface is shown below. Figure 2 As shown.

[0194] Function Description: In evaluation mode, the model performs forward propagation inference and outputs a predicted probability distribution containing nine fault types. This module ultimately outputs the fault type diagnosis result for the rolling bearing (including the specific fault location and severity) based on the model output (e.g., taking the category corresponding to the highest probability). Figure 2 As shown, the system's front-end interface displays the final generated structured diagnostic report, highlighting the final fault conclusions to assist maintenance personnel in developing accurate repair plans.

[0195] Other aspects of this embodiment can be found in the above method embodiments.

[0196] In summary, this invention discloses a method and system for rolling bearing fault diagnosis based on wavelet residual shrinkage network, belonging to the field of industrial artificial intelligence and fault diagnosis technology. The method mainly includes: S1, establishing a standardized sample library: obtaining the original data files (e.g., Excel format) of publicly available rolling bearing datasets, parsing the operating conditions and fault information contained in the file names, batch converting them to a general data format (e.g., CSV), and storing them according to labels; S2, constructing a wavelet residual shrinkage network (WRSN) model. The first layer of this model uses a fixed convolution kernel based on Daubechies wavelets to achieve physical frequency band separation, and the deep layer integrates channel-by-channel residual shrinkage units to achieve adaptive denoising; classifying and labeling the vibration signals in the dataset according to fault type and severity; performing STFT on the labeled signals to generate a two-dimensional time-spectrum graph, which serves as the network input; S3, dividing the dataset into training, validation, and test sets, and training the WRSN model; S4, implementing fault diagnosis using the trained model. This invention utilizes the physical interpretability of wavelet transform to solve the frequency aliasing problem and uses the residual shrinkage mechanism to achieve adaptive denoising under strong noise. It eliminates the need for manual feature extraction and can accurately diagnose the severity of faults. It features high computational accuracy and strong noise resistance.

[0197] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for fault diagnosis of rolling bearings based on wavelet residual shrinking networks, characterized in that, Includes the following steps: S1. Obtain the raw data file of the rolling bearing, parse the file name and clean the data, and build a standardized vibration signal sample library; S2. Construct a wavelet residual shrinking network (WRSN) model; read vibration signals from the sample library and perform short-time Fourier transform (STFT) to obtain a two-dimensional time spectrum, which is used as the input of the WRSN model. S3. Divide the dataset of the two-dimensional time-spectrum graph into a training set, a validation set and a test set. Train the WRSN model using the training set. Evaluate the generalization performance of the model after each round using the validation set. Save the WRSN model parameters with the highest performance based on the highest accuracy of the validation set. S4. Load the saved optimal WRSN model weights and input the test set into the model for forward propagation inference; The model extracts and classifies features from the time-frequency spectrum, and finally outputs the fault type diagnosis result corresponding to the test sample.

2. The rolling bearing fault diagnosis method according to claim 1, characterized in that, Step S1 is as follows: S11. Obtain the original data file of the rolling bearing; S12. Identify characteristic characters in the original data file name; S13. Based on the aforementioned characteristic characters, the data is divided into 9 states, including: normal state, moderate inner ring fault, severe inner ring fault, moderate outer ring fault, severe outer ring fault, moderate sphere fault, severe sphere fault, moderate combined fault, and severe combined fault. S14. Read the contents of the original data file, retain the vibration signal channel, and remove the redundant speed channel under constant speed conditions. S15. Convert the cleaned data into CSV format and categorize and store it.

3. The rolling bearing fault diagnosis method according to claim 1 or 2, characterized in that, In step S2, the wavelet residual shrinking network WRSN model is constructed as follows: S21. Construct the wavelet decomposition convolutional layer WDConv: Select the Daubechies 4 wavelet as the basis function and obtain its corresponding one-dimensional low-pass filter vector. and one-dimensional high-pass filter vector The one-dimensional filter vector is expanded into a two-dimensional separable convolution kernel through the outer product operation, as shown in the following formula: in, Represents the set of real numbers. G is a vector containing 8 real numbers. L Represents a low-pass core, G H Represents the high-pass kernel, V represents the outer product operator, T represents the transpose operator, i and j represent the index positions of the filter coefficient vector, and G is the two-dimensional matrix generated by the outer product operation. L The element in the i-th row and j-th column is ; S22. During the forward propagation of the WRSN model, the WDConv layer adopts a grouped convolution method, setting the number of convolution groups to be equal to the number of channels of the input feature map, and using the two-dimensional convolution kernel constructed in step S21 to independently perform convolution operation on each channel of the input two-dimensional time-spectrum map.

4. The rolling bearing fault diagnosis method according to claim 3, characterized in that, In step S2, the WRSN model further includes a channel-wise residual shrinking unit RSBU-CW for adaptive noise reduction of the features after WDConv decomposition, specifically: After performing an absolute value operation on the input feature map x, global average pooling is used to compress the two-dimensional feature map into a one-dimensional statistical feature vector. The one-dimensional statistical feature vector is input into a threshold generation subnetwork of a channel-wise residual shrinking unit, and a scaling parameter α is output through the Sigmoid activation function. c The formula is as follows: Where, α c This represents the scaling factor for the c-th channel of the output learned by the subnetwork, with a value ranging from (0,1). c The feature value representing the c-th neuron; Then, the scaling parameters are used to calculate the independent threshold for each channel. The formula is as follows: in, This represents the adaptive threshold for the c-th channel, where i, j, and c are the width, height, and channel indices of the feature map x. Finally, the calculated threshold is used. The corresponding channels of the original feature map are subjected to soft thresholding to obtain the output feature map, as shown in the following formula: Among them, y i,j,c For output features, x i,j,c The original input features are represented by sign(·), which is the sign function.

5. The rolling bearing fault diagnosis method according to claim 4, characterized in that, In step S2, the vibration signal is subjected to STFT, and the specific parameters are set as follows: Based on the sampling frequency of the original data, the Hann window function was used, with the window size set to 1024 and the number of overlapping points to 256. The generated two-dimensional time-frequency spectrograms after transformation are uniformly scaled to 64×64 pixels.

6. The rolling bearing fault diagnosis method according to claim 1, characterized in that, In step S3, the training of the WRSN model is performed as follows: Gaussian white noise, Laplace noise, pink noise, and a mixture of the above noises were injected into the original vibration dataset. The noisy signal was converted into a two-dimensional time-spectrum using STFT to construct a noisy dataset for evaluating the model's noise robustness. The noisy dataset is input into the WRSN model, and the loss value between the predicted classification value output by the model and the true label of the sample is calculated using the cross-entropy loss function. The calculated loss value is fed back through the backpropagation mechanism, the loss value of each layer parameter is calculated, and the network weights are updated using the SGD optimizer.

7. A rolling bearing fault diagnosis system based on wavelet residual shrinkage network, used to perform the method as described in any one of claims 1-6, characterized in that, Includes the following modules: The data processing and time-frequency conversion module is used to acquire the original vibration signal data file of the rolling bearing, parse the file name and clean the data to construct a standardized vibration signal sample library; read the vibration signal from the sample library to perform STFT and generate a two-dimensional time-frequency spectrum as the model input; The model building module is used to build the WRSN model, which includes a two-dimensional wavelet decomposition convolutional layer WDConv built by the outer product operation and a channel-wise residual shrinking unit containing an adaptive threshold generation subnetwork. The model training module divides the dataset of the two-dimensional time-spectrum graph into a training set, a validation set, and a test set. The WRSN model is trained using the training set. The generalization performance of the model is evaluated after each round of the validation set. The model with the highest validation set accuracy is used as the criterion, and the parameters of the WRSN model with the best performance are saved. The fault diagnosis module is used to load the optimal WRSN model weights that have been trained and saved, receive the two-dimensional time spectrum of the test set and input it into the model for forward propagation inference, the model extracts features and classifies the time spectrum, and outputs the fault type diagnosis result corresponding to the test sample.

Images