Adaptive bearing fault diagnosis method based on multi-base wavelet fusion

Abstract

The invention provides a self-adaptive bearing fault diagnosis method based on multi-base wavelet fusion. The objective of the invention is to solve the problems of noise reduction, insufficient feature extraction and low diagnosis precision under noise conditions. A Kaisixi University bearing public data set is used as original data, and Gaussian noise with different SNRs is superposed to simulate various noise intensities. And uniformly carrying out length alignment, down-sampling, equal-length segmentation, division and normalization preprocessing. Then, wavelet bases such as sym4, db4, coif5 and the like are adopted for parallel multi-scale decomposition and reconstruction; and adaptively determining the number of decomposition layers and a threshold strategy according to the noise level, and generating a de-noising branch. And performing weighted fusion on the denoising results of the branches, and performing iterative denoising on the residual error. Signals subjected to noise reduction processing are sent to a double-branch convolution-cycle-attention network, a convolution layer extracts features, an LSTM and a self-attention module capture time sequence changes, and accurate recognition of various bearing faults is achieved. The training adopts a segmented attenuation learning rate and an early stop strategy, and the robustness and generalization ability of different SNR working conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of mechanical fault diagnosis and intelligent sensing technology, and is particularly applicable to the intelligent fault detection and identification of rolling bearings in rotating machinery under noisy conditions. More specifically, it relates to an end-to-end fault diagnosis method based on adaptive signal denoising and multi-scale feature mining using multi-base wavelet fusion. Background Technology

[0002] With the increasing level of industrial intelligence, the reliability and safety of mechanical equipment are receiving growing attention. As a critical component of rotating machinery, rolling bearings are highly susceptible to causing abnormal equipment shutdowns, resulting in economic losses and safety hazards if they malfunction. Therefore, conducting research on intelligent fault diagnosis for bearings has significant theoretical and engineering implications.

[0003] Currently, bearing fault diagnosis is mainly divided into two stages: signal processing and pattern recognition. Traditional methods often employ time-domain, frequency-domain, and time-frequency-domain analysis techniques (such as envelope analysis, short-time Fourier transform, and wavelet transform) combined with statistical features or manual rules to achieve preliminary fault identification. These methods perform well in operating conditions with high signal-to-noise ratios or obvious fault characteristics, but under strong noise and complex operating conditions, the features are easily interfered with, leading to a decrease in diagnostic performance.

[0004] In recent years, end-to-end intelligent diagnostic methods based on deep learning have become a research hotspot, such as convolutional neural networks (CNN), recurrent neural networks (RNN / LSTM), and attention mechanisms. These methods automatically learn multi-layer features from the raw signal, improving diagnostic accuracy and robustness. However, deep learning models still face the following challenges in real-world industrial scenarios:

[0005] Training data is highly sensitive to noise and changes in operating conditions, and is prone to overfitting and decreased generalization ability.

[0006] Strong noise in the original signal can suppress key fault features and affect the model's learning performance;

[0007] Traditional wavelet denoising uses only a single wavelet basis, which is poorly adaptable to diverse signal features and results in significant information loss.

[0008] The noise reduction and diagnostic processes are disconnected, lacking an adaptive and collaborative optimization mechanism, making it difficult to guarantee optimal overall performance.

[0009] Existing research both domestically and internationally has attempted to combine wavelet denoising with deep learning-based diagnostics, but most studies employ fixed parameters, a single wavelet basis, and static network structures, lacking adaptability to varying noise intensities and operating conditions. While some studies have introduced adaptive methods, they have failed to achieve end-to-end optimization integrating denoising and diagnostics, making it difficult to balance signal recovery and fault identification. Achieving high accuracy and strong generalization capability in bearing fault diagnosis under low signal-to-noise ratio and complex operating conditions remains a critical technical challenge that urgently needs to be addressed. Summary of the Invention

[0010] 1. Purpose of the invention

[0011] This invention addresses the problems of low diagnostic accuracy, weak generalization ability, and disconnect between noise reduction and diagnosis processes in existing bearing fault diagnosis methods under varying noise levels and complex operating conditions. It proposes an end-to-end adaptive signal denoising and multi-scale intelligent diagnosis method based on multi-basic wavelet fusion. This method effectively improves signal quality through adaptive fusion of multi-basic wavelet parameters and, combined with a multi-branch convolution-recurrent-attention structure, achieves efficient identification of bearing faults under various types and operating conditions, demonstrating significant innovation and practical value.

[0012] 2. Technical Solution

[0013] The end-to-end fault diagnosis process for the entire system is as follows: Figure 1 As shown. Figure 1 It demonstrates a complete framework from noise signal input, preprocessing, noise reduction, feature extraction, fault diagnosis to adaptive training and inference, providing an overall background for the introduction of subsequent modules.

[0014] The technical solution of this invention includes the following key aspects:

[0015] (1) Adaptive multi-base wavelet fusion noise reduction module

[0016] This module proposes an adaptive noise reduction strategy based on multi-base wavelet fusion to address the noise reduction requirements of bearing vibration signals under different noise levels. Specifically, it includes:

[0017] Input signal preprocessing: The signal is aligned to a uniform length, downsampled, segmented, divided into datasets, and normalized to ensure consistency and comparability in subsequent processing.

[0018] ①Signal and noise model

[0019] y[n]=s[n]+v[n],n=0,…,N-1

[0020] In the formula, y[n] is the nth sample of the observed signal (including noise), s[n] is the nth sample of the real fault signal, v[n] is the nth sample of the additive noise, and N is the total number of samples in the observation sequence (signal length).

[0021] ②Estimation of short-time sample SNR

[0022]

[0023] In the formula For short-time sample signal-to-noise ratio estimation, This represents the signal power estimate obtained after removing the estimated noise power from the average power of the observed signal. `max(,ε)` is used to avoid numerical problems when the numerator is negative or close to zero, and `ε` is a very small positive number. This represents noise power.

[0024] in It can be estimated from the median absolute deviation (MAD) of the wavelet detail coefficients of the first level:

[0025]

[0026] The above estimates are used in this system for configuring switching parameters based on SNR intervals (strong / mid / weak). Where... The noise standard deviation is estimated using the median absolute deviation (MAD) of the level 1 detail coefficients. A constant of 0.6745 is the conversion factor between MAD and standard deviation under a standard normal distribution. 1,k The k-th detail coefficient is obtained from the first layer (finest scale or first layer detail) wavelet transform, and the range of values ​​for k is related to the transform length.

[0027] Parallel Denoising with Multi-Basis Wavelet Basis: Selecting multiple wavelet bases (such as sym4, db4, coif5, etc.) to perform multi-scale decomposition and reconstruction of the signal, resulting in multiple denoising branch results.

[0028] Adaptive parameter selection: Based on the noise level (SNR estimation) of the input signal, the system automatically matches the optimal wavelet decomposition level, threshold rules, noise estimation method and other parameter configurations to achieve adaptive optimization for different noise intensities.

[0029] ③ Wavelet coefficient soft thresholding denoising uses a soft thresholding function:

[0030]

[0031] In the formula d j,k T represents the detail coefficients obtained for the j-th layer; j The threshold is represented by multiplying the baseline threshold by the hierarchical amplification factor α. j :

[0032]

[0033] Commonly used benchmark thresholds include Universal and SURE thresholds:

[0034]

[0035] In the formula The median absolute deviation (MAD) of the detail coefficients at this layer can be used to estimate N. j The number of coefficients in this layer; the unbiased risk estimation formula for the SURE threshold:

[0036] The noise variance is known or estimated as σ 2 The coefficient vector d = (d1, ..., d2) N ), soft threshold estimator η T The SURE expression for (di) = sign(di)max(|di|-T,0) is:

[0037]

[0038] The empirical SURE threshold T can be obtained by finding the minimum value of this function on the candidate T grid. SURE .

[0039] SNR Interval Adaptation Rule: To achieve adaptive noise reduction for different noise intensities, this invention introduces a wavelet basis selection and parameter configuration strategy based on signal-to-noise ratio (SNR) in the multi-basis wavelet fusion module. Specific empirical rules are as follows:

[0040] For extremely low signal-to-noise ratios (SNR < 5dB), low-order orthogonal wavelet bases such as sym4, db4, and coif5 are preferred. These wavelet bases have higher sensitivity to abrupt and impulsive fault characteristics and can maintain the discernibility of sharp anomalies in noisy environments. The parameter configuration adopts the strong mode (Table 1), including a deeper decomposition level, a larger threshold amplification factor, and residual iterative denoising.

[0041] For medium signal-to-noise ratios (5dB≤SNR≤12dB), higher-order orthogonal wavelet bases such as sym8 and coif3 can be added to the lower-order wavelet base. Higher-order wavelets can capture finer-grained feature information and improve the resolution of fault features. The parameter configuration adopts the mid mode, with a moderate number of decomposition levels, a threshold amplification factor of 1.0, and optional residual iteration.

[0042] For high signal-to-noise ratio (SNR>12dB), a biorthogonal wavelet basis (bior3.3) is recommended, as it offers better suppression of structural noise or trend terms. This stage focuses on suppressing spurious fluctuations and removing trend noise to avoid false detections. Parameter configuration uses weak mode, the number of decomposition layers can be appropriately reduced, the threshold amplification factor is kept at 1.0, and Shift processing can be disabled.

[0043] The above rules are automatically completed by the SNR estimation module: the system first performs a fast SNR estimation on the input signal, matches the corresponding configuration (strong / mid / weak) based on the estimated value, and automatically loads the corresponding wavelet basis set and threshold strategy. This mechanism realizes a two-layer adaptation of "noise adaptation - wavelet basis adaptation", which significantly improves the noise reduction robustness and fault feature fidelity under different noise conditions.

[0044] Branch fusion mechanism: Weighted averaging or other fusion strategies are used to improve the signal recovery effect by applying the denoising results of multiple wavelet branches.

[0045] ④ Multi-base fusion mathematical representation and adaptive weights

[0046] Suppose M wavelet bases are used, and the fusion result is a weighted average:

[0047]

[0048] In the formula, M represents the number of wavelet bases used; The reconstructed result after fusion (weighted average) is the final estimate obtained by weighting the reconstructed results of all branches according to their respective weights. The result of reconstructing the m-th branch; w m The weight coefficient (non-negative) of the m-th branch is used for weighted average fusion.

[0049] Adaptive weights use the sample-estimated MSE as a quality measure:

[0050]

[0051] MSE (m) The mean square error of the sample relative to the observation y[n] is estimated for the m-th branch;

[0052] Q m is the quality metric for the m-th branch, i.e., the negative mean square error (the larger the value, the better the branch performance); N is the length of the observation sequence used for sample estimation or calculation of mean square error; in the strong / mid / weak modes, empirical fixed weights or this adaptive weight can be used to achieve online self-adaptation.

[0053] Iterative denoising of residual signal: The residual signal between the initial denoising result and the original signal is further denoised to improve the signal-to-noise ratio.

[0054] ⑤ Iterative formula and termination criterion for residual iterative denoising

[0055] First fusion obtained residual kth iteration:

[0056]

[0057] In the formula The target signal obtained from the first fusion (0th estimate)

[0058] r (0) [n] represents the residual after the first fusion (observation minus the current estimate), which is used for subsequent denoising iterations; For the k-th iteration, the operator For the previous residual r (k-1) [n] The estimated incremental components (residual denoising / detail compensation); This represents a multi-basic wavelet denoising operator or other denoising mapping (which may employ stronger threshold rules, different thresholds, or deeper decomposition levels), whose input is the residual signal and whose output is the correction term to be added to the current estimate; The update formula for the k-th iteration represents the estimated increment for this iteration. Add to the last estimate A new estimate was obtained; r (k) [n] represents the new residual calculated in the k-th iteration, which will be used as the input for the next iteration; the iteration termination condition can be:

[0059] or or k≥K max

[0060] In the formula The relative energy convergence criterion for residuals: when the k-th residual r (k) The normalized 2-norm relative to the original observation y is less than the threshold ε r When the residual is sufficiently small, the iteration can be terminated. The relative convergence criterion for estimating the change is: when the relative change between two consecutive estimates is less than the threshold ε. s The iteration can be stopped when k ≥ K to ensure the results are stable; max

[0061] Termination criterion for maximum iteration count: If the number of iterations reaches a preset upper limit K. max Then it will be forcibly stopped.

[0062] Shift translation processing: Automatically detects platform compatibility and selects whether to use signal translation technology to improve robustness.

[0063] All parameter configurations for this module are shown in Example 1 and Table 1.

[0064] (2) Adaptive multi-scale fault diagnosis network

[0065] This invention employs a dual-branch convolution-recurrent-attention structure to achieve multi-scale feature extraction and classification of denoised signals. The main steps include:

[0066] Input layer: Receives the noise-reduced signal sequence.

[0067] Dual-branch convolutional encoder: extracts features from large and small receptive fields in parallel, adapting to multiple fault modes.

[0068] Feature fusion layer: Multi-scale features are fused using point-by-point multiplication or splicing methods.

[0069] Temporal modeling layer: Introduces LSTM and self-attention mechanism to fully capture the temporal dynamic relationship and key features of the modeling signal.

[0070] Output layer: Fully connected + Softmax to achieve multi-class output.

[0071] The specific network structure and hyperparameter configuration are shown in Example 2 and Tables 3 and 4.

[0072] (3) End-to-end adaptive training and inference process: During the training phase, adaptive learning rate decay, early stopping and validation set monitoring strategies are adopted to improve generalization ability; during the inference phase, the parameters are kept consistent throughout the process to achieve end-to-end integrated diagnosis.

[0073] (4) Learning rate scheduling formula (piecewise decay):

[0074]

[0075] lr0 is the initial learning rate, and every T... drop The learning rate for each epoch is multiplied by a factor γ (0 < γ < 1).

[0076] Simultaneously, a validation-based adaptive descent is employed: if the validation loss does not improve within P validation cycles, then lr = lr·γ.

[0077] 3. Explanation of repeatability and comparability

[0078] To ensure the rigor of diagnostic performance evaluation, this invention employs a strict and reproducible uniform baseline in its controlled experimental design. Because existing literature differs significantly in data sources, fault types, noise models and intensities, preprocessing procedures, sampling rates, sample slicing and partitioning strategies, network structure implementation details, and training hyperparameters, uncontrollable biases can easily be introduced under inconsistent experimental conditions, leading to unreliable conclusions or misleading interpretations. Therefore, the results of existing literature cannot be considered equivalently as directly comparable benchmarks under the same experimental conditions.

[0079] To enhance the credibility of the quantitative assessment of the contribution of this invention, and to analyze the contribution of each module in this invention to diagnostic performance, the following control group was designed while ensuring that the remaining network architecture and training protocol are consistent with this invention:

[0080] Group A (Baseline): The multi-base wavelet adaptive noise reduction module of Example 1 is not used, and the adaptive learning rate mechanism is not used;

[0081] Group B: Use the noise reduction module of Example 1, but without using the adaptive learning rate mechanism.

[0082] Group C: No noise reduction module is used, but an adaptive learning rate mechanism is adopted.

[0083] Group D (Complete Solution of the Invention): Simultaneously adopts the noise reduction module and adaptive learning rate mechanism of Example 1.

[0084] All control groups underwent comparative experiments using the same data sources, standardized preprocessing, network structure, and training protocol. To enhance the reproducibility of the results, key reproducibility information has been clearly documented in the instruction manual.

[0085] 4. Technological Innovation Points

[0086] An adaptive denoising method based on multi-basis wavelet fusion is proposed, which automatically matches parameters and integrates the advantages of multiple wavelet bases to significantly improve signal recovery capability. The denoising and diagnosis are integrated into an end-to-end optimization, and the parameters can be adjusted adaptively and collaboratively to preserve fault features to the maximum extent. A multi-branch convolution-recurrent-attention diagnostic network is used to achieve deep fusion of multi-scale features, which significantly improves the fault identification accuracy under different signal-to-noise ratios and complex working conditions. The entire process is standardized, the model is reproducible, and it has strong applicability.

[0087] 5. Technical Effects

[0088] This invention achieves excellent signal recovery and fault identification results under different signal-to-noise ratios and multiple operating conditions: the average SNR of the denoised signal is improved by more than 5dB, and the RMSE of various faults is reduced by 30-50%; the fault classification accuracy is significantly improved, with an improvement of 6.25% at the lowest SNR; the standard deviation of repeated results in multiple experiments is low, and the method is stable and reliable. Attached Figure Description

[0089] Figure 1 This is the end-to-end fault diagnosis process for the overall system of the present invention;

[0090] Figure 2 This is a detailed schematic diagram of the working structure of the adaptive multi-base wavelet fusion noise reduction module according to an embodiment of the present invention;

[0091] Figure 3 This is the overall structure of the multi-scale fault diagnosis network in an embodiment of the present invention;

[0092] Figure 4 This is the adaptive training and inference process according to an embodiment of the present invention;

[0093] Figure 5 This is a comparison of the diagnostic accuracy results under different SNR and training strategies in the embodiments of the present invention. Detailed Implementation

[0094] Example 1: Adaptive Multi-Base Wavelet Fusion Denoising

[0095] For a detailed explanation of the working principle of the adaptive multi-base wavelet fusion noise reduction module, please refer to the appendix. Figure 2 . Figure 2 The paper demonstrates key steps such as multi-base wavelet parallel denoising, adaptive parameter selection, branch fusion, and residual signal iterative denoising.

[0096] 1. Data Sources and Preprocessing

[0097] The Case Western Reserve University Bearing Data Center (CWRU) dataset was selected. This dataset is a publicly available standard bearing failure dataset that includes various bearing failures (inner ring, outer ring, rolling element of various sizes) and normal operating conditions, with a sampling frequency of 48kHz.

[0098] The experiment selected faults in the 0.014-inch rolling element, inner ring, and outer ring. First, Gaussian white noise with different signal-to-noise ratios (SNRs) was superimposed on the original vibration signal of the bearing, with SNRs of 2dB, 7dB, 12dB, and 17dB, respectively, to simulate various noise environments. Then, all noise signals were preprocessed uniformly. Each type of data was divided into independent samples with a fixed non-overlapping length to avoid overlapping segments of the same data appearing in both the training and test sets. The steps included: unifying the length of each noise intensity signal to 480,000 points, downsampling with a sampling factor of 6, dividing the samples into 250-point segments, obtaining 320 samples from a single signal, dividing the training and test sets in a 7:3 ratio, and inputting the normalized data into the noise reduction module.

[0099] 2. Wavelet basis and parameter configuration

[0100] Six typical wavelet bases were selected, and the key parameters were configured as follows:

[0101] Table 1: Adaptive Multi-Basis Wavelet Denoising Parameter Configuration Table

[0102]

[0103] Parameter Adaptation Rule: In the experiment, the three configurations of strong, mid, and weak were automatically switched according to the SNR interval adaptation rule to verify the noise reduction effect under different noise intensities. For the specific SNR-wavelet basis correspondence, please refer to Section 2-1 'SNR Interval Adaptation Rule' of the technical solution.

[0104] Multi-basis fusion: Each wavelet basis is denoised independently, and the branch results are weighted and averaged point-to-point.

[0105] Residual signal denoising: Under strong noise, the residual signal after initial denoising can be denoised a second time to further improve ΔSNR.

[0106] 3. Noise Reduction Experiment Results

[0107] (1) Reconstruction mean square error (MSE) and root mean square error (RMSE):

[0108]

[0109] (2) Input / output SNR and its improvement:

[0110]

[0111] ΔSNR=SNR out -SNR in

[0112] (3) Peak Signal-to-Noise Ratio (PSNR):

[0113]

[0114] In the formula MAX s Let be the peak amplitude of the nth sample of the real fault signal s[n].

[0115] Table 2: Multi-base wavelet adaptive noise reduction performance under various working conditions

[0116]

[0117] 4. Technical Effect Analysis

[0118] Based on the experimental results in Table 2, the technical advantages of the multi-base wavelet adaptive denoising method of this invention under different fault types and noise levels can be summarized as follows:

[0119] (1) Significantly improves signal quality at extremely low signal-to-noise ratios

[0120] Under extreme conditions with SNR=2dB, the ΔSNR for the three types of faults—rolling element, inner ring, and outer ring—reached +6.11dB, +6.55dB, and +3.24dB, respectively, with an average improvement of about 6dB, significantly suppressing the interference of strong noise on the original signal.

[0121] Meanwhile, the RMSE reduction exceeded 31%, with the RMSE reduction for inner ring and rolling element failures reaching 52.97% and 50.55% respectively, effectively restoring the inherent characteristics of the failure.

[0122] (2) It has strong universality and can adapt to different fault types and noise intensities.

[0123] As the SNR increases from 2dB to 17dB, all indicators, including ΔSNR, ΔRMSE, and ΔPSNR, show significant improvement, and the method remains stable under different fault types. For example, when SNR = 17dB, ΔSNR still reaches 3.75–4.29dB, and RMSE decreases by more than 35%, indicating that this method is not only applicable to extreme noise environments but also effective for medium to high SNR conditions.

[0124] (3) Multibase fusion has obvious advantages and improves the robustness of the system.

[0125] The multi-base wavelet fusion strategy can effectively improve signal quality even when the characteristic signals are weak, such as those from outer ring faults, by combining the feature extraction capabilities of various wavelet bases for different frequency bands and different types of noise.

[0126] (4) Adaptive parameter adjustment and residual signal iteration further improve performance.

[0127] This method automatically selects the optimal parameters (such as the number of decomposition layers, threshold amplification factor, etc.) based on the noise level of the input signal, and effectively deals with noise of different intensities through residual signal iterative processing, ensuring stable and reliable noise reduction effect.

[0128] (5) Provide high-quality input for subsequent intelligent diagnosis

[0129] The denoised signal exhibits significantly improved RMSE, PSNR, and other metrics, providing a high-quality input foundation for feature extraction and classification in subsequent deep learning networks, and effectively enhancing the overall end-to-end diagnostic accuracy and robustness.

[0130] In summary, the multi-base wavelet adaptive noise reduction method of the present invention can significantly improve signal quality under different fault types and noise intensities, and has strong universality and engineering application value, providing a solid foundation for intelligent fault diagnosis under complex working conditions.

[0131] Example 2: Adaptive Multi-Scale Fault Diagnosis

[0132] 1. Data Sources and Sample Division

[0133] The data used was from the CWRU dataset, comprising 10 operating conditions (3 types each for normal operation and faults at different locations / sizes). Gaussian white noise with SNRs of 2dB, 7dB, 12dB, and 17dB was superimposed on each operating condition to enrich the data sample, ensuring a sufficient total data volume. The same data preprocessing method as in Example 1 was used, dividing each data type into independent samples of a fixed, non-overlapping length to avoid overlapping segments of the same data appearing simultaneously in both the training and test sets.

[0134] 2. Network Structure and Parameter Configuration

[0135] Model Structure: The multi-scale diagnostic network includes a multi-scale convolutional encoder branch, an LSTM temporal modeling branch, an attention fusion module, and a classifier (fully connected layer + Softmax). Specific parameters for each layer of the network (kernel size, number of channels, pooling strategy, number of LSTM hidden units, etc.) are detailed in Table 3 of this specification.

[0136] Table 3: Structure of Multi-Scale Convolutional-Recurrent-Attention Diagnostic Network

[0137]

[0138] 3. Training hyperparameter settings

[0139] To ensure fair comparisons among control groups and the reproducibility of experimental results, this embodiment uses the hyperparameter settings shown in Table 4 under a unified training protocol. Specific details are as follows:

[0140] Optimizer and Learning Rate Strategy: The Adam optimizer is used, with an initial learning rate of 0.001. To improve training stability and promote fine convergence, a learning rate decay mechanism is employed—the current learning rate is multiplied by a decay factor of 0.4 every 15 epochs to reduce the step size and suppress the risk of overfitting in the later stages of training.

[0141] Batch size and training epochs: Training uses a mini-batch size of 64, with a total training limit of 60 epochs. This configuration ensures that the model fully learns multi-scale temporal features while balancing training time and computational resource consumption.

[0142] Validation frequency and early stopping strategy: Validation is performed every 30 mini-batches during training to monitor generalization performance; if no performance improvement is seen after 20 consecutive validations, early stopping is triggered as a conservative control measure to prevent overfitting.

[0143] Regularization: A Dropout probability of 0.5 is used to enhance the robustness of the model; other regularization terms (weight decay) are kept consistent across control groups to avoid comparison bias.

[0144] Data partitioning and reproducibility: The data is divided into training and test sets at 70% and 30% respectively. The screening and partitioning process is stratified according to the experimental sequence to avoid data leakage.

[0145] Table 4: Training Hyperparameter Table

[0146]

[0147] 4. Adaptive Training and Inference Process

[0148] The adaptive training and inference process is as follows: Figure 4 As shown. Figure 4 It demonstrates the specific steps of training and inference, including adaptive learning rate, validation monitoring, early stopping strategy, and model weight saving.

[0149] Training phase (adaptive training and early stop)

[0150] Objective: Given a dataset, obtain stable and reproducible model weights through an adaptive learning rate strategy and an early stopping mechanism, thereby avoiding overfitting and improving generalization ability.

[0151] Inputs: Dataset splitting information, initial learning rate, optimizer type (Adam), batch size, maximum number of training epochs, validation frequency, early stopping tolerance, learning rate descent strategy, Dropout ratio, random seed, hardware resources (GPU / CPU), and log requirements.

[0152] Key Process Points: 1. Initialization and Setup: Set a global random seed to ensure reproducible results. Construct the network structure, loss function, and optimizer (Adam, initial learning rate = cfg.initialLearnRate). 2. Execution Device Selection: If the system has a GPU, use it first; otherwise, fall back to CPU. 3. Adaptive Learning Rate Strategy: Use piecewise decay: After every cfg.lrDropPeriod training epochs, multiply the current learning rate by cfg.lrDropFactor. 4. Validation and Early Stopping: Perform validation according to cfg.validationFreq, calculate the validation set loss and metrics. Record the validation results and update the optimal model state. If no improvement in validation loss is observed for cfg.patience consecutive validation epochs, trigger early stopping to end training.

[0153] Model saving and logging: When training reaches optimal validation performance, the corresponding model weights are saved, and the configuration, performance metrics, and training trajectory are recorded. The log includes key metrics such as loss and accuracy for each round / validation, changes in the learning rate, and whether early stopping was triggered.

[0154] Output: Optimal model weights, training curve data, validation metrics, and training results.

[0155] 5. Experimental Design and Result Comparison

[0156] To analyze the contribution of each module in this invention to diagnostic performance, the following control group was designed while ensuring that the remaining network architecture and training protocol are consistent with this invention:

[0157] Group A (Baseline): The multi-base wavelet adaptive noise reduction module of Example 1 is not used, and the adaptive learning rate mechanism is not used;

[0158] Group B: Use the noise reduction module of Example 1, but without using the adaptive learning rate mechanism.

[0159] Group C: No noise reduction module is used, but an adaptive learning rate mechanism is adopted.

[0160] Group D (Complete Solution of the Invention): Simultaneously adopts the noise reduction module and adaptive learning rate mechanism of Example 1.

[0161] Table 5: Comparison of Diagnostic Accuracy under Different SNRs and Training Strategies

[0162]

[0163] 6. Technical Effect Analysis

[0164] For a detailed comparison of diagnostic accuracy under different SNR and training strategies, please refer to the appendix. Figure 5 .

[0165] Based on the experimental data in Table 5, the method of the present invention exhibits superior fault diagnosis performance under different signal-to-noise ratios and different training strategies, as detailed below:

[0166] (1) Robustness is significantly improved under extremely low signal-to-noise ratio. Under extreme noise conditions with SNR=2dB, the diagnostic accuracy of the "adaptive noise reduction + adaptive learning rate training" strategy of this invention (Group D) reaches 95.73%, which is 6.25% higher than the 89.48% of the "noise + fixed learning rate" group (Group A). ​​This indicates that the method has a strong ability to suppress strong noise interference and can effectively ensure the accuracy of fault diagnosis.

[0167] (2) The multi-base wavelet fusion denoising and adaptive training effects are significant. As the SNR increases, the accuracy of group D continues to lead, and the accuracy improvement between group D and group A gradually narrows from 6.25% to 1.49%, reflecting that the advantages of this method are most prominent when the noise impact is most severe. This fully verifies the synergistic gain effect of multi-base wavelet fusion denoising and adaptive learning rate training. The two work together to effectively improve the model's adaptability to complex noise and changes in operating conditions.

[0168] (3) The accuracy of each group is high and the method is highly universal. Under a high signal-to-noise ratio such as SNR=17dB, the accuracy of group D can still reach 98.23%, which is better than the other groups. At the same time, the accuracy of each group exceeds 96%, indicating that the method of the present invention is not only applicable to extreme noise environments, but also applicable to general working conditions, and has strong versatility and stability.

[0169] The entire process of this invention includes core steps such as data generation and preprocessing, denoising and feature fusion, fault diagnosis network, adaptive training stage, and output and evaluation: First, Gaussian white noise with different SNRs is superimposed on the bearing vibration signal to simulate various noise intensities. Length alignment, downsampling, sample segmentation, dataset partitioning, and normalization are then performed to ensure the consistency of the input sample shape. Subsequently, in the denoising and feature fusion stage, an adaptive multi-basis wavelet fusion denoising module automatically selects the optimal parameters based on the current noise level and fuses the denoising results of multiple wavelet bases. If necessary, iterative denoising is performed on the residual signal. Next, in the fault diagnosis network stage, the denoised signal is input into a multi-branch convolutional-recurrent-attention diagnostic network to automatically extract multi-scale temporal features and complete fault classification. In the adaptive training stage, adaptive learning rate decay, early stopping, and validation set monitoring strategies are adopted to improve generalization ability. Finally, the classification results and evaluation indicators are output, and training state information related to adaptive training is provided. For a detailed flowchart, please refer to the abstract appendix. Figure 1 This scheme achieves strong robustness and good generalization through multi-base wavelet adaptive denoising, end-to-end collaborative optimization, and a multi-scale convolutional-recurrent-attention fusion network. It significantly improves signal recovery and diagnostic performance under low signal-to-noise ratio conditions. The entire process is standardized and reproducible, making it suitable for vibration signal denoising and intelligent fault diagnosis of various rotating machinery. It can also be embedded in edge computing devices, industrial gateways, and cloud platforms to achieve online real-time health monitoring and automatic alarms. This method has good engineering scalability and industrial application prospects for different sampling frequencies and sensor data.

Claims

1. A self-adaptive bearing fault diagnosis method based on multi-base wavelet fusion, characterized in that, The method comprises the following steps: a signal preprocessing step: uniform length alignment, down-sampling, sample segmentation and normalization processing are performed on the original vibration signal of the bearing; an adaptive multi-wavelet fusion denoising step: (a) a plurality of wavelet bases (such as sym4, db4, coif5, etc., but not limited to) are used to perform multi-scale decomposition and reconstruction on the preprocessed signal, generating a plurality of denoising branches; (b) according to the noise level of the input signal, the number of wavelet decomposition layers, threshold rules and noise estimation methods are adaptively determined, and the denoising results of each branch are weighted and fused; (c) the residual signal of the preliminary denoising result and the original signal is subjected to iterative denoising to further improve the signal-to-noise ratio; a multi-scale fault diagnosis step: the denoised signal is input into a fault diagnosis network based on a double-branch convolution-cyclic-attention structure, which includes an input layer, a double-branch convolutional encoder (for extracting large and small receptive field features), a feature fusion layer, a time series modeling layer (including LSTM and self-attention modules), and a fully connected output layer, for realizing classification and identification of bearing fault types; an end-to-end adaptive training and inference step: in the training stage, adaptive learning rate decay and early stopping monitoring strategy are used to realize automatic adjustment of network parameters, thereby improving the robustness and generalization ability of the model under low signal-to-noise ratio or complex working conditions; in the inference stage, the key parameter configuration is consistent with that in the training stage to realize end-to-end integrated fault diagnosis.

2. The method of claim 1, wherein, In the wavelet fusion denoising step, the denoising results of each wavelet branch are integrated by point-by-point weighted fusion or other suitable fusion methods to obtain more robust denoising effect.

3. The method of claim 1, wherein, The adaptive selection of the wavelet base is based on a preset parameter configuration group, which includes at least two and more modes (such as weak, medium and strong modes), and the mode used depends on the noise level of the input signal.

4. The method of claim 1, wherein, The double-branch convolutional encoder of the fault diagnosis network is used to extract local features of different scales, which are sent to the time series modeling layer after feature fusion. The layer uses LSTM and self-attention mechanism to capture the time series dynamic changes in the signal, thereby realizing accurate identification of different fault states.

5. The method of claim 1, wherein, In the end-to-end adaptive training and inference step, a piecewise learning rate decay strategy is adopted, that is, the learning rate is automatically reduced after a preset training stage, and an early stopping mechanism is combined to prevent overfitting and improve the generalization performance of the model.

6. The method of claim 1, wherein, The iterative denoising implemented in the denoising step includes re-denoising the residual signal of the preliminary denoising to further improve the signal quality after denoising and provide higher quality input data for subsequent fault diagnosis.

7. The method of claim 1, wherein, This method is suitable for fault diagnosis of bearing vibration signals containing Gaussian white noise interference, and can significantly improve the fault diagnosis performance under low signal-to-noise ratio and complex working conditions by relying on adaptive denoising and multi-scale feature extraction.

Images