Noise robust data-driven fault classification method and system

Abstract

The invention discloses a noise robust data-driven fault classification method and system, and the method comprises the steps: firstly constructing a one-dimensional signal fault classification sample library, then carrying out the data-driven noise robust fault mode classification, and finally constructing a fault signal classification model integrating an enhanced noise reduction function, and achieving the fault classification. The accuracy of the fault diagnosis model under the noise condition is obviously improved, the generalization ability is enhanced, the robustness and the precision are improved, and the method has a great prospect for the practical application of the intelligent fault diagnosis technology.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a noise-robust data-driven fault classification method and system. Background Technology

[0002] The development of machine learning and deep learning has brought significant advantages and advancements to the field of intelligent fault detection. During the operation of systems and components, such as automobiles, aircraft, and factories, aging and failures of various machinery, instruments, and parts are inevitable. Traditional fault detection relies on regular manual maintenance, heavily depending on the experience of maintenance workers, and is often only effective when obvious and observable faults occur. However, by this time, the parts are often beyond repair, and the excessive operating conditions may have already impacted adjacent components, causing even greater losses.

[0003] With the development of data processing, researchers have discovered that various operational data, such as vibration fluctuations, temperature fluctuations, and pressure changes, can reflect the operating patterns of a system, thereby diagnosing system faults, locating fault positions, and monitoring system operation in real time. For example, analyzing vibration signals from mechanical moving parts such as bearings and gears can accurately determine faults such as cracks, corrosion, and impacts. Monitoring multi-point temperature and pressure of pressure vessels can analyze the potential locations of leaks, loosening, and changes in wall thickness. Although such data contains a variety of information, its sheer volume makes quantitative analysis and the establishment of precise mathematical and physical models difficult. Deep learning methods, however, can use multi-layered neural networks to automatically identify the information hidden in complex data, accurately capturing corresponding features in the early stages of fault development, predicting impending fault modes and remaining lifespan, and providing a reference for fault diagnosis and system maintenance. Especially for expensive, precise, or high-risk industries such as aerospace, medical rescue, and nuclear chemistry, where safety and reliability requirements are extremely stringent, timely repair and replacement of components in the early stages of fault development, before affecting system operation and causing significant losses, can greatly extend the overall lifespan of the system.

[0004] Because systems operate in varying environments, data processing struggles with noise contamination. Vibration signals are polluted by adjacent moving equipment and ambient vibrations; pressure signals are affected by the altitude of the installation site; and temperature signals become ineffective with changes in ambient temperature. Furthermore, the amount of data collected during model training is limited, making it difficult to account for diverse environments. Adding a noise processing module to the model to remove noise interference can effectively improve its generalization ability and enhance the robustness of fault diagnosis.

[0005] Typical methods for processing noisy signals include the following: Data cleaning and removal. In the initial data screening process, binning, clustering, outlier detection and removal are common processing methods. These methods usually do not involve complex signal processing theories and model training, but use statistical rules to group the data to identify and process noise and outliers. (1) Binning method can sort and divide adjacent data into several bins, and use the statistics in the bin (mean, median, maximum and minimum values, etc.) to replace the original values, thereby achieving smooth processing of the data. In addition, it can also discretize the data, but the choice of bin size has a great impact on the results: too large will lead to smooth transition and loss of real details; too small will lead to insufficient noise removal. (2) Clustering method can transform the noise removal problem into an outlier detection problem, divide the data into multiple clusters, normal data points belong to one or more dense clusters, and those points that do not belong to any cluster and are far away from the center of any cluster are regarded as noise and are treated as outliers. Clustering method can find clusters of arbitrary shapes and effectively identify noise points in them, but some data that are not in the cluster may be small category data with important significance, and caution is needed when judging. (3) Outlier detection defines outliers by utilizing the statistical characteristics of data distribution. It uses statistical methods to set normal ranges and treats points outside the boundaries as outliers for elimination, correction, or transformation. Its rules are clear and calculations are fast, but it is more likely to cause the loss of important information and features. It is suitable for handling obvious data entry or measurement errors.

[0006] Data filtering and smoothing. In the process of processing time series data or signals, since the data are correlated on the time scale, the statistical characteristics of the "neighboring" data can be used to smooth the noise. Common filtering methods include moving average filtering, median filtering, convolution filtering, channel filtering, etc. The core idea of ​​filtering methods is that effective signals are usually continuously and smoothly changing, while noise signals are more random and fluctuate at high frequencies. By sliding a "filter" window on the data, the statistical value or fitted value of the neighboring data in the window is used to replace the value of the center point, so as to suppress noise and smooth the data. (1) Average filtering uses the arithmetic mean in the window to replace the center value, which will cause the smoothed signal to lag behind the original signal in phase and blur the sharp features. It is suitable for Gaussian noise. The improved weighted average will give different weights to points at different positions. Usually, the weight of the point closer to the center is higher. After reducing the information, the information near the center point can be better preserved. (2) Median filtering uses the median in the window sequence to replace the center value. It is suitable for processing impulse noise. The sorting operation of the data makes the computation cost higher, but it can protect the edge and sharp features of the signal well. (3) Convolution smoothing (Savitzky-Golay filtering) uses a polynomial of a specified order to perform least-squares fitting on all data points within a window, and uses the calculated value to replace the center value. Within a small window, the low-order polynomial can well describe the local trend of the signal and can retain the shape and characteristics of the original signal to a large extent. The selection of parameters such as window size and polynomial order will affect the denoising effect. (4) The theoretical basis of channel filtering is Fourier transform, which decomposes the signal into a superposition of sine waves of different frequencies. Usually, the effective signal is the low-frequency component, and the noise is manifested as the high-frequency component. When processing the signal, the time domain signal is first transformed to the frequency domain through Fourier transform, and the set components (high frequency or low frequency) are attenuated or filtered out by the filter. Then, the time domain is restored through inverse transform to obtain the denoised signal. It is suitable for stationary signals whose frequency components do not change with time.

[0007] Restoring the underlying structure of data. With the development of models and algorithms, more complex algorithms have emerged that can learn or recover the underlying real data structure from noisy data. The emergence of wavelet transform, principal component analysis (PCA), and singular value decomposition (SVD) has greatly improved the efficiency of noise processing and can handle more complex noise types and high-dimensional data. (1) The wavelet transform method solves the defect that Fourier transform cannot handle non-stationary signals. It decomposes the noisy signal into wavelet coefficients with different frequency and time domain characteristics through wavelet basis functions. Then, by setting a threshold, the coefficients representing noise are set to zero or shrunk, while the effective signal coefficients are retained. The denoised signal is then reconstructed by inverse wavelet transform. (2) Both PCA and SVD are denoising methods based on data dimensionality reduction. They believe that data features exist in the direction of large variance and noise exists in the direction of small variance. When processing signals, the data matrix is ​​first centered; then, singular value decomposition (SVD) is performed directly on the matrix, or covariance matrix is ​​calculated followed by eigenvalue decomposition (PCA); the principal components corresponding to the top k largest singular values / eigenvalues ​​are retained, and the remaining components are discarded; finally, the data is reconstructed using the principal components. This method is suitable for processing high-dimensional and multivariate data.

[0008] Machine learning / deep learning methods. Leveraging the ability of deep learning models to extract complex features, these methods eliminate the need for complex data processing. They allow the model to directly learn key features from noisy data and map them to clean data, handling various types of non-stationary noise. The core mechanism for noise reduction is residual learning: compared to having the model directly learn the mapping from complex signals to clean signals, learning noisy residuals is easier, simplifying a complete signal regression task into a relatively simple residual regression task. The residual structure effectively alleviates the vanishing gradient problem and allows for the construction of deeper network structures.

[0009] The development of methods for processing noisy signals has followed an evolutionary process: from linear to nonlinear, from relying on prior assumptions to data-driven adaptive processing, and from handling stationary signals to dealing with complex non-stationary signals. From data cleaning that directly removes outliers to filtering methods that smooth signals, methods based on classical signal processing theory are computationally more efficient, but their ability to handle nonlinear and non-stationary noise is limited. The emergence of complex algorithms, such as wavelet transform, and deep learning methods allows noise to be treated no longer merely as simple statistical perturbations. Instead, it involves mining the intrinsic structure of the data or utilizing powerful nonlinear models to "understand" and "separate" noise, typically meaning a trade-off between higher computational costs and superior denoising results.

[0010] Current research on fault diagnosis models focuses primarily on improving diagnostic accuracy under specific environments. However, when it comes to the deployment and practical application of these models, their versatility and generalization capabilities under varying background noise levels and complex operating conditions still need improvement. Summary of the Invention

[0011] To overcome the shortcomings of existing technologies, this invention provides a noise-robust data-driven fault classification method and system. First, a one-dimensional signal fault classification sample library is constructed. Then, data-driven noise-robust fault mode classification is performed. Finally, a fault signal classification model integrating enhanced noise reduction functions is constructed to achieve fault classification. The fault diagnosis model of this invention significantly improves accuracy under noisy conditions, enhances generalization ability, and improves robustness and precision, showing great promise for the practical application of intelligent fault diagnosis technology.

[0012] The technical solution adopted by this invention to solve its technical problem is as follows: Step 1: Construct a one-dimensional signal fault classification sample library; Step 1-1: Preparation of bearing fault classification dataset; The original dataset is split into training, validation, and test sets in a 1:1:1 ratio. Ten types of faults and normal conditions were selected as the classification results. The faults include two types of internal raceway faults, two types of rolling element faults, and two types of external raceway faults with three different operating conditions. Step 1-2: Add noise to the fault diagnosis signal; Gaussian white noise of the same seed but different intensities was added to the test set to test the model's noise reduction capability; First, calculate the average power of the current signal. If the signal power is detected to be lower than the preset threshold, skip the noise addition process for that sample and directly retain the original signal. The required noise power value is calculated based on the target signal-to-noise ratio and the signal power of the current sample. Based on the calculated required noise power value, a random noise sequence following a Gaussian distribution is generated, with a mean of 0 and a standard deviation equal to the square root of the noise power. The length of the noise sequence is the same as that of the original signal. Finally, the noise sequence is linearly superimposed with the original signal to obtain the final noisy signal; Steps 1-3: Prepare a test sample set for bearing fault classification under different noise environments; Gaussian white noise with SNR values ​​of 20dB, 15dB, 10dB, 5dB, and 0dB was added to the predefined test sets to represent environmental conditions of "near-noise", "light noise", "moderate noise", "heavy noise", and "extreme noise", respectively. Step 2: Data-driven noise-robust fault mode classification; Step 2-1: Data-driven one-dimensional signal fault classification; Design a data-driven bearing vibration signal fault classification framework. In this framework, vibration signal sensors installed at the bearing end collect vibration signals generated during bearing operation, which include fault signals from 10 different parts of the bearing and normal signals when there are no faults. Based on the constructed enhanced noise reduction fault classification model, the different characteristics of vibration signals under different fault conditions are learned to distinguish 11 types of bearing states. Step 2-2: Composite noise reduction for fault diagnosis; Step 2-2-1: Noise reduction at the signal input end; Wavelet thresholding for denoising of noisy signal segments is employed. The model incorporating noise is represented by the following formula: (4) In the formula, The signal contains noise. Useful signal This is a noise signal. The standard deviation of the noise figure; From noisy signals Separating and enhancing useful signals To the greatest extent possible, noise signals Eliminate; Step 2-2-2: Noise reduction at the fault classification model end; Based on wavelet thresholding denoising, a lightweight noise suppression network is set up, consisting of two convolutional layers and two activation layers; In the noise suppression layer, the original signal is first passed through a convolution kernel to... The convolutional layer identifies anomalous noise patterns; then batch normalization and ReLU activation are performed to stabilize the noise feature distribution and preserve significant noise features; finally, a convolutional kernel is used to... The convolutional layer fuses cross-channel noise information; then the Sigmoid activation function is used to generate a probability mask for the noise, which is in the range of [0,1]. When the probability mask approaches 0, it represents a high-probability noise region, and when it approaches 1, it represents a high-probability effective signal region. Finally, the probability mask is multiplied element by element with the original signal to obtain the output after noise suppression. Step 2-2-3: Combined noise reduction at the signal input end and the fault classification model end; First, the noisy signal is input into the noise suppression network. The noise suppression mask generated by two convolutional layers is multiplied with the original signal to obtain the preprocessed signal. Then, wavelet threshold denoising is performed. Wavelet decomposition, noise level estimation, adaptive threshold calculation, threshold processing, and wavelet reconstruction are performed in sequence to obtain the denoised signal. Steps 2-3: Construct a fault signal classification model that integrates enhanced noise reduction functions; A fault diagnosis method based on CNN and LSTM, employing a combined denoising strategy of wavelet thresholding and model denoising, and integrating attention mechanisms with residual networks, is adopted as follows: First, the original one-dimensional signal is directly input into a composite noise reduction module to process the noise in the signal. The composite noise reduction module is a dual noise reduction method consisting of a noise suppression network and wavelet thresholding. Then, the processed signal is input into a CNN, which uses multiple convolutional kernels to extract multi-scale features. Subsequently, residual connections are used to alleviate gradient vanishing, an LSTM network is used to extract temporal features, and an attention mechanism is used to focus on key features. Finally, the features are global average pooled and then input into a fully connected layer for classification. The classification result is output after normalization using the Softmax function.

[0013] Preferably, the original dataset is the CWRU bearing fault detection dataset provided by Case Western Reserve University; The experimental setup used in the CWRU bearing fault detection dataset consists of a 2-horsepower induction motor, a torque sensor, a power sensor, an electronic control system, and the bearing under test. The bearing under test is mounted on the drive (DE) and fan (FE) ends of the motor, and faults are injected into the drive-end bearing. The fault simulation uses electrical discharge machining (EDM) technology to create single-point faults with diameters of 0.007 inches, 0.014 inches, 0.021 inches, and 0.028 inches on the inner raceway, rolling elements, and outer raceway of the bearing, simulating the development stages of the fault from its early stages to its severe stages.

[0014] Preferably, the segmentation of the original dataset refers to segmenting the original data using a sliding window of length 1024.

[0015] Preferably, in steps 1-2: The Gaussian noise expression is as follows: (1) In the formula, the mean Represents the average level of noise; standard deviation Represents the range of fluctuation in noise amplitude; variance The degree to which the noise value deviates from its mean; White noise combined with a Gaussian distribution forms additive white Gaussian noise; the noise signal is linearly superimposed on the source signal, and the signal degradation process is expressed by the following formula: (2) In the formula, For the signal in coordinates The intensity value at that location, For a given value with a mean of 0 and a variance of , The samples are randomly selected from a Gaussian distribution; the intensity of noise is represented by the signal-to-noise ratio (SNR), defined as: (3) In the formula, and These represent signal power and noise power, respectively.

[0016] Preferably, the basis function of the wavelet is db4, and the wavelet decomposition layer is 3.

[0017] A noise-robust, data-driven real-time fault monitoring system, comprising endpoint, edge, and cloud components; The end side includes a vibration signal sensor, a signal sampling module, and a data transmission module; The edge side includes a data preprocessing module, a fault diagnosis module, and an anomaly alarm module; The cloud includes a data storage center, a real-time monitoring and display system, and a model training and optimization center; The terminal side is responsible for data acquisition and is installed at the location of the component that needs to be diagnosed. It is connected to the edge measurement via wired transmission. The edge measurement is responsible for classifying equipment faults and issuing abnormal alarms. It is connected to the cloud via a wireless base station. The cloud device can perform data storage, monitoring and display, and model optimization, and feed back to each module. The vibration signal sensor, as the data source of the system, is a front-end receiving device for acquiring equipment operation signals, converting physical quantities such as mechanical vibration, pressure, and temperature generated during equipment operation into measurable electrical signals. The signal sampling module converts the continuous analog signal from the sensor into a discrete digital signal, and acquires one-dimensional discrete signals of different lengths according to the requirements of subsequent signal processing. The data transmission module completes the wired and wireless transmission functions of the collected device operation signals and various signals generated by the operation status of each module. The data preprocessing module cleans and straightens the raw signal, and performs preliminary outlier processing and data standardization. The fault diagnosis module deploys an anti-noise fault diagnosis model and uses deep learning algorithms to extract relevant fault features from preprocessed vibration signal segments to complete the fault mode classification task. The relevant results are then uploaded to the cloud for data storage center for further processing. The abnormal alarm module can generate alarm information of different levels based on the processing results of the fault diagnosis module and the set alarm threshold; it can activate different alarm mechanisms according to the severity of the fault to ensure that relevant monitoring personnel can obtain equipment fault information in a timely manner and take relevant measures; it can also be associated with the system start-up and shutdown equipment to shut down the equipment before a major fault is judged to occur, so as to avoid more serious losses. The data storage center is responsible for the unified storage and management of various types of data generated by the system, including raw vibration data, diagnostic results, alarm records, and system operation logs. It supports efficient querying and retrospective analysis of historical data. Based on fault classification, it can also predict the remaining service life of the equipment and fault mode based on historical data. The real-time monitoring and display system provides managers with a unified visual interface for equipment status, which intuitively displays the real-time operating status and historical trend changes of each device through charts, curves and topology diagrams. The model training and optimization center continuously iterates and optimizes diagnostic algorithms and updates models based on historical data accumulated in the cloud. Through supervised learning or transfer learning, it uses new data to improve the accuracy and generalization ability of fault diagnosis models, and deploys the optimized models to fault diagnosis modules on the edge side to achieve continuous improvement in system diagnostic performance.

[0018] Preferably, the wired transmission includes signal transmission between end-side devices and edge devices and the wireless base station, as well as signal flow within them.

[0019] Preferably, the wireless transmission includes signal transmission between the wireless base station and the cloud, specifically including one-dimensional vibration signal data, operating status data of each module, model optimization and algorithm iteration data, which are processed by the cloud and interact with each device through the wireless base station.

[0020] The beneficial effects of this invention are as follows: (1) This paper proposes a hybrid model denoising strategy that integrates wavelet threshold denoising and model denoising, and applies it to a basic fault classification model based on CNN and LSTM, incorporating attention mechanisms and residual networks, to construct a general fault classification method in noisy environments. Experimental results show that the denoising method proposed in this paper effectively improves the model's ability to identify and distinguish faults in noisy environments, and significantly improves the diagnostic accuracy of the original model in strong noise environments.

[0021] (2) In the denoising method adopted in this invention, the optimization effect on the model is as follows: hybrid denoising > wavelet threshold denoising > model denoising > base model. Model denoising uses a lightweight noise suppression layer to identify different regions as noise and generate probability masks to achieve denoising. Wavelet denoising separates noise from useful signals and can process noise more finely through thresholding, thus achieving better results than model denoising.

[0022] (3) The proposed noise reduction method was validated using the CWRU bearing dataset. The proposed method can effectively optimize the model performance under complex environments with different noise backgrounds. Moreover, the optimization effect is more obvious under strong interference, which enhances the model's generalization ability, robustness and diagnostic accuracy. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the original signal; Figure 2 A schematic diagram of a noisy signal (SNR=20); Figure 3 This is a schematic diagram of a noisy signal (SNR=15); Figure 4 This is a schematic diagram of a noisy signal (SNR=10); Figure 5 This is a schematic diagram of a noisy signal (SNR=5); Figure 6 This is a schematic diagram of a noisy signal (SNR=0); Figure 7 A schematic diagram of the noise component (SNR=20); Figure 8 A schematic diagram of the noise component (SNR=15); Figure 9 A schematic diagram of the noise component (SNR=10); Figure 10 A schematic diagram of the noise component (SNR=5); Figure 11 This is a schematic diagram of the noise component (SNR=0); Figure 12 A schematic diagram of the vibration signal fault classification framework; Figure 13 This is a schematic diagram of the noise suppression layer structure; Figure 14 This is a schematic diagram of the composite noise reduction signal processing flow. Figure 15 This is a schematic diagram of the signal processing flow; Figure 16(a) is a schematic diagram of model accuracy; Figure 16(b) is a schematic diagram of the confusion matrix in the absence of noise; Figure 16(c) is a schematic diagram of the confusion matrix when SNR=20; Figure 16(d) is a schematic diagram of the confusion matrix when SNR=15; Figure 16(e) is a schematic diagram of the confusion matrix when SNR=10; Figure 16(f) is a schematic diagram of the confusion matrix when SNR=5; Figure 16(g) is a schematic diagram of the confusion matrix when SNR=0; Figure 17 This is a schematic diagram of the ablation experiment results; Figure 18 This is a schematic diagram of the functional framework of a real-time fault monitoring system. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] 1. Construction of a one-dimensional signal fault classification sample library; 1.1 Method for preparing bearing fault classification dataset; To verify the effectiveness of the fault diagnosis model proposed in this invention, the optimization effect of the noise suppression module on fault diagnosis in noisy environments, and the availability of the real-time fault detection system, the CWRU bearing fault detection dataset provided by Case Western Reserve University was used for training and testing.

[0026] The experimental setup used in the CWRU bearing dataset consisted of a 2-horsepower (1.5 kW) induction motor, a torque sensor, a power sensor, an electronic control system, and the bearing under test (SKF6205). The test bearing was mounted on both the drive (DE) and fan (FE) ends of the motor, with fault injection performed on the drive-end bearing. Fault simulation was achieved using electrical discharge machining (EDM) to create single-point faults with diameters of 0.007 inches (0.1778 mm), 0.014 inches (0.3556 mm), 0.021 inches (0.5334 mm), and 0.028 inches (0.7112 mm) on the inner raceway, rolling elements, and outer raceway of the bearing, simulating the development stages of the fault from its early stages to its most severe condition.

[0027] The processing procedure for the original dataset is as follows: the original data is divided using a sliding window of length 1024, and the training set, validation set, and test set are divided in a 1:1:1 ratio, with 2200 samples for each set. Ten fault categories and normal conditions are selected as the classification results. The faults include two types of inner raceway faults, two types of rolling element faults, and six types of outer raceway faults under three different operating conditions. Detailed fault categories are shown in Table 1.

[0028] Table 1. Fault Types

[0029] 1.2 A method for adding noise to fault diagnosis signals was designed and implemented; To avoid the randomness introduced by noise addition, Gaussian white noise of the same seed but different intensities was added to the test set to test the model's noise reduction capability. As one of the most common and important random noise models, Gaussian noise is characterized by its statistical properties following a normal distribution and is widely found in electronic signals, communication channels, and sensor data acquisition. The probability density function of a continuous Gaussian noise random variable is uniquely determined by its mean and variance, and its expression is as follows: (1) In the formula, the mean Representing the average level of noise, Gaussian noise in most engineering contexts is typically represented by zero mean to avoid a constant deviation from the original signal; standard deviation This represents the fluctuation range of the noise amplitude; a larger value indicates a more severe contamination of the signal by the noise. Variance The degree to which a noise value deviates from its mean represents the intensity of the noise power.

[0030] "White noise" means that the power spectral density of the noise is uniformly distributed across the entire frequency domain and is uncorrelated in time. When combined with a Gaussian distribution, it forms Additive White Gaussian Noise (AWGN). The noise signal is linearly superimposed on the source signal, and the signal degradation process can be expressed by the following equation: (2) In the formula, For the signal in coordinates The intensity value at that location, For a given value with a mean of 0 and a variance of , The samples are randomly selected from a Gaussian distribution. The intensity of noise is represented by the signal-to-noise ratio (SNR), measured in decibels (dB), and is defined by the following formula: (3) In the formula, and These represent signal power and noise power, respectively. The lower the SNR value, the stronger the noise power relative to the signal power, and the more serious the noise pollution.

[0031] Based on the fundamental theory of Gaussian white noise, the following process is used to add noise. First, the average power of the current signal is calculated. If the signal power is detected to be lower than a preset threshold (indicating that the signal is invalid or close to silence), the noise addition process for that sample is skipped, and the original signal is directly retained.

[0032] To achieve the preset SNR value, the required noise power value is calculated based on the target signal-to-noise ratio and the signal power of the current sample, ensuring that similar noise addition effects can be obtained for input signals of different intensities.

[0033] Based on the calculated required noise power value, a random noise sequence following a Gaussian distribution is generated, with a mean of 0 and a standard deviation equal to the square root of the noise power. The length of the noise sequence is the same as that of the original signal. Finally, the noise sequence is linearly superimposed on the original signal to obtain the final noisy signal.

[0034] 1.3 Method for preparing bearing fault classification test sample sets under different noise environments; In noisy environments, an SNR value above 20 represents extremely low noise, almost imperceptible, indicating good signal quality. Practical applications involve high-quality sensor data acquisition under ideal conditions, suitable for performance verification under near-noise-free conditions. An SNR between 20 and 10 dB represents low to moderate noise, where the noise is visible and audible, but the signal structure remains clear. This is used to simulate typical industrial environments or signal channels with slight interference; at 10 dB, the signal power is ten times the noise power. An SNR between 10 and 0 dB represents moderate to heavy noise, where noise gradually dominates, the signal is severely distorted, and feature extraction is difficult. This range is used to simulate harsh industrial environments and highly interfering communication environments, and is a key range for differentiating model performance. At 0 dB, the noise power is equal to the signal power, making it difficult to distinguish manually. Extreme noise indicates that the sensor is almost inoperable, used to test the model's extreme performance and failure boundaries.

[0035] To verify the model's anti-interference capability, Gaussian white noise with SNR values ​​of 20dB, 15dB, 10dB, 5dB, and 0dB was used in the predefined test set to represent "near-noise," "light noise," "moderate noise," "heavy noise," and "extreme noise" environmental conditions, respectively. The SNR values ​​after adding noise were calculated, and the results are shown in Table 2.

[0036] Table 2. Results of Noise Addition

[0037] The original signal, the noisy signal, and the noise component in the signal are visualized under each condition, and the results are as follows: Figures 1 to 11 As shown.

[0038] 2. A data-driven method for classifying noise-robust fault modes; 2.1 A data-driven one-dimensional signal fault classification framework was designed; Based on one-dimensional vibration data from vibration signal sensors of moving mechanical parts, this invention designs a data-driven bearing vibration signal fault classification framework, such as... Figure 12As shown.

[0039] In this framework, vibration signal sensors installed at the bearing end collect vibration signals generated during bearing operation, including fault signals from ten different parts of the bearing and normal signals when there are no faults. Based on the constructed enhanced noise reduction fault classification model, the different characteristics of vibration signals under different fault conditions are learned, enabling the differentiation of eleven types of bearing states.

[0040] 2.2 A composite noise reduction method for fault diagnosis was designed and constructed; (1) Noise reduction at the signal input end: A wavelet threshold noise reduction method for noisy signal segments is proposed; The core of wavelet thresholding denoising lies in using wavelet transform to decompose the signal into a series of wavelet basis functions with different time-frequency characteristics, obtaining approximate coefficients representing low-frequency trends and detail coefficients representing high-frequency details and noise. Thresholding is applied to the high-frequency detail coefficients; effective signal features typically correspond to the portion exceeding the threshold, which is preserved or reduced; noise signals typically correspond to the portion below the threshold, which is zeroed or significantly reduced. An inverse transform is then performed on the processed approximate and detail coefficients to obtain the reconstructed and denoised signal. A noisy model can be represented by the following equation: (4) In the formula, The signal contains noise. Useful signal This is a noise signal. This represents the standard deviation of the noise figure. Noise reduction is the process of reducing noise in a noisy signal. Separating and enhancing useful signals To the greatest extent possible, noise signals Eliminate it.

[0041] The key issues in wavelet thresholding denoising lie in the selection of wavelet basis functions, threshold values, and threshold functions, which directly affect the denoising effect. Basis functions determine the quality of signal decomposition and require consideration of factors such as the similarity between wavelet shape and target features, signal smoothness, and symmetry.

[0042] The choice of threshold is related to the balance between noise features and useful signal features. Commonly used thresholds include universal thresholds, Stein unbiased risk estimation, and minimax thresholds. Among them, the universal threshold is theoretically optimal, approaching the maximum and minimum risk for signal coefficients, but excessively long signal lengths can lead to feature loss. It is suitable for preliminary processing in noisy environments. Stein unbiased risk estimation minimizes the mean square error estimate and adapts to signal features. The minimax threshold minimizes the estimation risk in the worst case and has wide applicability.

[0043] The choice of threshold function is a core aspect of wavelet denoising, determining how wavelet coefficients are processed to separate useful signals from noise. Major threshold functions include hard thresholding, soft thresholding, Garrote thresholding, semi-soft thresholding, and SCAD thresholding. Hard thresholding sets the absolute value less than the threshold to zero, while retaining the value greater than the threshold. This results in discontinuities at the threshold points, causing oscillations and blurring in the reconstructed signal. Soft thresholding shrinks the portion greater than the threshold, ensuring overall signal continuity. The reconstructed signal maintains a constant deviation from the original signal.

[0044] (2) Noise reduction at the fault classification model end: A lightweight noise suppression network layer was designed; Building upon wavelet thresholding for noise reduction, the model incorporates a lightweight noise suppression network. This network consists of two convolutional layers and two activation layers, with the following structure: Figure 13 As shown.

[0045] In the noise suppression layer, the original signal is first passed through a convolution kernel to... The convolutional layer identifies anomalous noise patterns; then batch normalization and ReLU activation are performed to stabilize the noise feature distribution and preserve significant noise features. Then, a convolutional kernel is used for... The convolutional layers fuse cross-channel noise information; then, a sigmoid activation function is used to generate a probability mask for the noise, ranging from [0,1]. A probability mask approaching 0 represents a high-probability noise region, and approaching 1 represents a high-probability effective signal region. Finally, the probability mask is multiplied element-wise with the original signal to obtain the noise-suppressed output. The core of this module lies in using a lightweight network to automatically distinguish noise from signals and perform preliminary noise processing, removing isolated noise formed by anomalous signals deviating from the neighborhood's statistical regularity, thus forming a composite noise reduction module with wavelet thresholding. The signal processing flow of the noise reduction part is as follows: Figure 14 As shown.

[0046] (3) A combined noise reduction method for the signal input end and the fault classification model end; For processing noisy signals, the signal is first input into a noise suppression network. The noise suppression mask generated by two convolutional layers is multiplied by the original signal to obtain the preprocessed signal. Then, wavelet thresholding denoising is performed, sequentially involving wavelet decomposition, noise level estimation, adaptive threshold calculation, threshold processing, and wavelet reconstruction to obtain the denoised signal. The wavelet basis function used is "db4", and the wavelet decomposition layer has three levels.

[0047] 2.3 A fault signal classification model integrating enhanced noise reduction functions was designed and constructed; To address the noise problem encountered in practical fault diagnosis applications, this paper proposes a fault diagnosis method based on CNN and LSTM, employing a composite denoising strategy combining wavelet thresholding and model denoising, and integrating attention mechanisms and residual networks. This method aims to diagnose system faults using one-dimensional sensor signals through multiple general-purpose denoising strategies, thereby improving the model's generalization ability and reliability. The signal processing flow of the fault diagnosis model is as follows: Figure 15 As shown.

[0048] First, the original one-dimensional signal is directly input into the composite noise reduction module to process the noise in the signal. Then, the processed signal is input into the CNN for multi-scale feature extraction using multiple convolutional kernels. Next, residual connections are used to alleviate gradient vanishing, LSTM networks are used to extract temporal features, and attention mechanisms are used to focus on key features. Finally, the features are global average pooled and input into a fully connected layer for classification. The classification results are output after normalization using the Softmax function. The parameter changes of the network model are shown in Table 3.

[0049] Table 3. Network Model Parameters

[0050] Figure 16 shows the changes in model accuracy under different noise levels. As the signal-to-noise ratio (SNR) decreases, the model accuracy declines more slowly when the SNR is above 10; however, the accuracy drops significantly when the SNR is below 10. A confusion matrix was plotted using 200 randomly selected samples from the test results, where the horizontal and vertical axes represent the predicted and true labels, respectively. The confusion matrix shows that in the bearing fault classification task, misclassifications are more concentrated in the second group of labels related to outer raceway faults.

[0051] To verify the effectiveness of the noise reduction method proposed in this invention, ablation experiments were conducted on different noise reduction strategies under different signal-to-noise ratio conditions. The results... Figure 17 As shown, the vertical axis represents accuracy.

[0052] Depend on Figure 17As can be seen, the hybrid denoising strategy generally outperforms wavelet denoising, which in turn outperforms model denoising, and model denoising outperforms the performance of the model without a denoising strategy. The differences are even more pronounced in noisy environments. With the decrease in signal-to-noise ratio (SNR), the overall accuracy of each model tends to decline. Without denoising, the accuracy of the base model drops sharply from 96.30% to 29.21% when the SNR decreases from 20dB to 0dB, essentially losing its classification value. Using a hybrid denoising strategy that combines model denoising and wavelet thresholding, the accuracy in a noise-free environment improves to 98.11%, and the accuracy in a 0dB noisy environment improves to 70.07%, verifying that the hybrid denoising strategy significantly improves the model's denoising performance.

[0053] 3. Design and implementation methods of real-time fault monitoring system; 3.1 A functional framework for a one-dimensional signal real-time fault monitoring system was designed; For model deployment applications, this invention designs a functional framework for a real-time fault monitoring system based on one-dimensional signals. The system is deployed on the edge side as follows: (1) a vibration signal sensor, (2) a signal sampling module, and (3) a data transmission module; on the edge side as follows: (4) a data preprocessing module, (5) a fault diagnosis module, and (6) an anomaly alarm module; and on the cloud side as follows: (7) a data storage center, (8) a real-time monitoring and display system, and (9) a model training and optimization center. The functional framework of this system is as follows: Figure 18 ; 3.2 A method for implementing a real-time fault monitoring system is proposed; The real-time fault monitoring system proposed in this invention includes end-side devices, edge devices, and cloud devices. The end-side devices are primarily responsible for data acquisition and are installed at the locations of components requiring fault diagnosis, connected to the edge devices via wired transmission. The edge devices are responsible for classifying device faults and issuing anomaly alerts, connected to the cloud devices via wireless base stations. The cloud devices are capable of data storage, monitoring and display, and model optimization, and then feeding this data back to each device.

[0054] Based on the functional framework, the implementation method of this system and the detailed functions of each module are as follows: (1) Vibration signal sensor; This module, serving as the system's data source, is a front-end receiving device for acquiring equipment operation signals. Its core function is to convert physical quantities such as mechanical vibration, pressure, and temperature generated during equipment operation into measurable electrical signals. This invention utilizes bearing fault vibration signals to verify the effectiveness of the model, typically employing an accelerometer, which can effectively capture various fault characteristics ranging from slow wear to high-frequency impact.

[0055] (2) Signal sampling module; The signal sampling module converts continuous analog signals from sensors into discrete digital signals, acquiring one-dimensional discrete signals of varying lengths according to subsequent signal processing requirements. Utilizing a high-precision analog-to-digital converter improves data acquisition efficiency. Its integrated filter provides signal conditioning and amplification, while also preventing spectral aliasing, ensuring accurate and reliable data for subsequent signal processing.

[0056] (3) Data transmission module; This module primarily handles the wired and wireless transmission of various system-generated signals, including collected equipment operation signals and the operational status of each module. Wired transmission mainly includes signal transmission between end-side and edge devices and the wireless base station, as well as signal flow within these devices. Wireless transmission mainly involves signal transmission between the wireless base station and the cloud, specifically including one-dimensional vibration signal data, operational status data of each module, and model optimization and algorithm iteration data. This data is processed by the cloud and interacts with each device through the wireless base station.

[0057] (4) Data preprocessing module; The data preprocessing module performs necessary cleaning and normalization on the raw signal, including preliminary outlier handling and data standardization. To meet the needs of the subsequent fault diagnosis module, the one-dimensional vibration signal needs to be segmented into 1024 segments. This module can initially eliminate some interference factors generated during the measurement process and convert the data into a standardized format that the algorithm can process.

[0058] (5) Fault diagnosis module; The fault diagnosis module is the core of the entire system. By deploying a noise-resistant fault diagnosis model and utilizing previously proposed and validated deep learning algorithms, it extracts relevant fault features from pre-processed vibration signal segments to complete the fault mode classification task. The results are then uploaded to the cloud for data storage and further processing. This module aims to achieve real-time detection of early-stage faults, assess fault location and severity, and provide reference and technical support for equipment maintenance.

[0059] (6) Abnormal alarm module; This module can generate alarm messages of different levels based on the processing results of the fault diagnosis module and the set alarm thresholds. Depending on the severity of the fault, different alarm mechanisms are activated to ensure that relevant monitoring personnel can obtain equipment fault information in a timely manner and take appropriate measures. It can also be linked to system start-up and shutdown equipment to shut down equipment before a major fault is detected, preventing more significant losses.

[0060] (7) Data storage center; The data storage center is responsible for the unified storage and management of various types of data generated by the system, including raw vibration data, diagnostic results, alarm records, and system operation logs. It supports efficient querying and retrospective analysis of historical data. Based on fault classification, it can also predict the remaining service life of equipment and fault modes based on historical data.

[0061] (8) Real-time monitoring and display system; The real-time monitoring and display system provides managers with a unified visual interface for equipment status. Through charts, graphs, and topology diagrams, it intuitively displays the real-time operating status and historical trend changes of each device. By dynamically monitoring key parameters and providing threshold alerts, managers can comprehensively understand the health status of equipment and promptly detect anomalies.

[0062] (9) Model Training and Optimization Center; The model training and optimization center continuously iterates and optimizes diagnostic algorithms and updates models based on historical data accumulated in the cloud. Through techniques such as supervised learning or transfer learning, it leverages new data to improve the accuracy and generalization ability of fault diagnosis models, and deploys the optimized models to edge-side fault diagnosis modules, thereby continuously improving the system's diagnostic performance.

Claims

1. A noise-robust data-driven fault classification method, characterized in that, Includes the following steps: Step 1: Construct a one-dimensional signal fault classification sample library; Step 1-1: Preparation of bearing fault classification dataset; The original dataset is split into training, validation, and test sets in a 1:1:1 ratio. Ten types of faults and normal conditions were selected as the classification results. The faults include two types of internal raceway faults, two types of rolling element faults, and two types of external raceway faults with three different operating conditions. Step 1-2: Add noise to the fault diagnosis signal; Gaussian white noise of the same seed but different intensities was added to the test set to test the model's noise reduction capability; First, calculate the average power of the current signal. If the signal power is detected to be lower than the preset threshold, skip the noise addition process for that sample and directly retain the original signal. The required noise power value is calculated based on the target signal-to-noise ratio and the signal power of the current sample. Based on the calculated required noise power value, a random noise sequence following a Gaussian distribution is generated, with a mean of 0 and a standard deviation equal to the square root of the noise power. The length of the noise sequence is the same as that of the original signal. Finally, the noise sequence is linearly superimposed with the original signal to obtain the final noisy signal; Steps 1-3: Prepare a test sample set for bearing fault classification under different noise environments; Gaussian white noise with SNR values ​​of 20dB, 15dB, 10dB, 5dB, and 0dB was added to the predefined test sets to be used for environmental conditions of "near-noise", "light noise", "moderate noise", "heavy noise", and "extreme noise", respectively. Step 2: Data-driven noise-robust fault mode classification; Step 2-1: Data-driven one-dimensional signal fault classification; Design a data-driven bearing vibration signal fault classification framework. In this framework, vibration signal sensors installed at the bearing end collect vibration signals generated during bearing operation, which include fault signals from 10 different parts of the bearing and normal signals when there are no faults. Based on the constructed enhanced noise reduction fault classification model, the different characteristics of vibration signals under different fault conditions are learned to distinguish 11 types of bearing states. Step 2-2: Composite noise reduction for fault diagnosis; Step 2-2-1: Noise reduction at the signal input end; Wavelet thresholding for denoising of noisy signal segments is employed. The model incorporating noise is represented by the following formula: (4) In the formula, The signal contains noise. Useful signal This is a noise signal. The standard deviation of the noise figure; From noisy signals Separating and enhancing useful signals To the greatest extent possible, noise signals Eliminate; Step 2-2-2: Noise reduction at the fault classification model end; Based on wavelet thresholding denoising, a lightweight noise suppression network is set up, consisting of two convolutional layers and two activation layers; In the noise suppression layer, the original signal is first passed through a convolution kernel to... Convolutional layers are used to identify abnormal noise patterns; Then, batch normalization and ReLU activation are performed to stabilize the noise feature distribution and retain significant noise features. Then, through a convolution kernel... The convolutional layer fuses cross-channel noise information; then the Sigmoid activation function is used to generate a probability mask for the noise, which is in the range of [0,1]. When the probability mask approaches 0, it represents a high-probability noise region; when it approaches 1, it represents a high-probability effective signal region. Finally, the probability mask is multiplied element by element by the original signal to obtain the output after noise suppression. Step 2-2-3: Combined noise reduction at the signal input end and the fault classification model end; First, the noisy signal is input into the noise suppression network, and the noise suppression mask generated by two convolutional layers is multiplied with the original signal to obtain the preprocessed signal; Then, wavelet threshold denoising is performed, which involves wavelet decomposition, noise level estimation, adaptive threshold calculation, threshold processing, and wavelet reconstruction to obtain the denoised signal. Steps 2-3: Construct a fault signal classification model that integrates enhanced noise reduction functions; A fault diagnosis method based on CNN and LSTM, employing a combined denoising strategy of wavelet thresholding and model denoising, and integrating attention mechanisms with residual networks, is adopted as follows: First, the one-dimensional original signal is directly input into the composite noise reduction module to process the noise in the signal. The composite noise reduction module is a noise suppression network and wavelet threshold noise reduction to perform dual noise reduction on the input noise. Then, the processed signal is input into a CNN, and multi-scale feature extraction is performed using multiple convolutional kernels. Subsequently, residual connections are used to alleviate gradient vanishing, LSTM networks are used to extract temporal features, and attention mechanisms are used to focus on key features. Finally, the features are global average pooled and then input into a fully connected layer for classification. The classification result is output after normalization by the Softmax function.

2. The noise-robust data-driven fault classification method according to claim 1, characterized in that, The original dataset is the CWRU bearing fault detection dataset provided by Case Western Reserve University. The experimental setup used in the CWRU bearing fault detection dataset consists of a 2-horsepower induction motor, a torque sensor, a power sensor, an electronic control system, and the bearing under test. The bearing under test is mounted on the drive (DE) and fan (FE) ends of the motor, and faults are injected into the drive-end bearing. The fault simulation uses electrical discharge machining (EDM) technology to create single-point faults with diameters of 0.007 inches, 0.014 inches, 0.021 inches, and 0.028 inches on the inner raceway, rolling elements, and outer raceway of the bearing, simulating the development stages of the fault from its early stages to its severe stages.

3. The noise-robust data-driven fault classification method according to claim 1, characterized in that, The process of segmenting the original dataset refers to segmenting the original data using a sliding window of length 1024.

4. The noise-robust data-driven fault classification method according to claim 1, characterized in that, In steps 1-2: The Gaussian noise expression is as follows: (1) In the formula, the mean Represents the average level of noise; standard deviation Represents the range of fluctuation in noise amplitude; variance The degree to which the noise value deviates from its mean; White noise combined with a Gaussian distribution forms additive white Gaussian noise; the noise signal is linearly superimposed on the source signal, and the signal degradation process is expressed by the following formula: (2) In the formula, For the signal in coordinates The intensity value at that location, For a given value with a mean of 0 and a variance of , The samples are randomly selected from a Gaussian distribution; the intensity of noise is represented by the signal-to-noise ratio (SNR), defined as: (3) In the formula, and These represent signal power and noise power, respectively.

5. The noise-robust data-driven fault classification method according to claim 1, characterized in that, The wavelet's basis function is db4, and the wavelet decomposition layer has 3 layers.

6. A real-time fault monitoring system employing the data-driven fault classification method as described in claim 1, characterized in that, Including the endpoint, edge, and cloud; The end side includes a vibration signal sensor, a signal sampling module, and a data transmission module; The edge side includes a data preprocessing module, a fault diagnosis module, and an anomaly alarm module; The cloud includes a data storage center, a real-time monitoring and display system, and a model training and optimization center; The terminal side is responsible for data acquisition and is installed at the location of the component that needs to be diagnosed. It is connected to the edge measurement via wired transmission. The edge measurement is responsible for classifying equipment faults and issuing abnormal alarms. It is connected to the cloud via a wireless base station. The cloud device can perform data storage, monitoring and display, and model optimization, and feed back to each module. The vibration signal sensor, as the data source of the system, is a front-end receiving device for acquiring equipment operation signals, converting physical quantities such as mechanical vibration, pressure, and temperature generated during equipment operation into measurable electrical signals. The signal sampling module converts the continuous analog signal from the sensor into a discrete digital signal, and acquires one-dimensional discrete signals of different lengths according to the requirements of subsequent signal processing. The data transmission module completes the wired and wireless transmission functions of the collected device operation signals and various signals generated by the operation status of each module. The data preprocessing module cleans and straightens the raw signal, and performs preliminary outlier processing and data standardization. The fault diagnosis module deploys an anti-noise fault diagnosis model and uses deep learning algorithms to extract relevant fault features from preprocessed vibration signal segments to complete the fault mode classification task. The relevant results are then uploaded to the cloud for data storage center for further processing. The abnormal alarm module can generate alarm information of different levels based on the processing results of the fault diagnosis module and the set alarm threshold; it can activate different alarm mechanisms according to the severity of the fault to ensure that relevant monitoring personnel can obtain equipment fault information in a timely manner and take relevant measures; it can also be associated with the system start-up and shutdown equipment to shut down the equipment before a major fault is judged to occur, so as to avoid more serious losses. The data storage center is responsible for the unified storage and management of various types of data generated by the system, including raw vibration data, diagnostic results, alarm records, and system operation logs. It supports efficient querying and retrospective analysis of historical data. Based on fault classification, it can also predict the remaining service life of the equipment and fault mode based on historical data. The real-time monitoring and display system provides managers with a unified visual interface for equipment status, which intuitively displays the real-time operating status and historical trend changes of each device through charts, curves and topology diagrams. The model training and optimization center continuously iterates and optimizes diagnostic algorithms and updates models based on historical data accumulated in the cloud. Through supervised learning or transfer learning, it uses new data to improve the accuracy and generalization ability of fault diagnosis models, and deploys the optimized models to fault diagnosis modules on the edge side to achieve continuous improvement in system diagnostic performance.

7. The real-time fault monitoring system according to claim 6, characterized in that, The wired transmission includes signal transmission between end-side devices and edge devices and wireless base stations, as well as signal flow within them.

8. The real-time fault monitoring system according to claim 6, characterized in that, The wireless transmission includes signal transmission between the wireless base station and the cloud, specifically including one-dimensional vibration signal data, operational status data of each module, model optimization and algorithm iteration data, which are processed by the cloud and interact with each device through the wireless base station.

Images