Unsupervised health index construction and performance degradation evaluation method for rolling bearing

Abstract

The invention belongs to the technical field of rolling bearing health index construction and performance degradation evaluation, and discloses a rolling bearing health index construction and performance degradation evaluation method based on unsupervised learning, and the method integrates rolling bearing operation state data samples, and carries out the data preprocessing, feature extraction and feature screening. And in combination with a multi-scale time sequence convolutional network-multi-head self-attention mechanism method, a model capable of evaluating the performance degradation of the rolling bearing in an unsupervised manner is constructed. Through the method, the health state and the degradation trend of the rolling bearing can be monitored and maintained more accurately and more effectively, so that accidental shutdown events are reduced in actual production, and the safety and the reliability of equipment in the operation process are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing health assessment, and particularly relates to a method for constructing unsupervised health indicators and evaluating performance degradation of rolling bearings. Background Art

[0002] As a key component of industrial rotating machinery, rolling bearings are of great significance for the normal operation of mechanical equipment. However, under the influence of long-term operation and harsh working environments, their health conditions will gradually deteriorate, eventually resulting in failures, causing huge economic losses and production safety risks to enterprises. During operation, bearings often bear high loads and are in harsh environments, which makes the formation of defects usually a gradual process. In addition, the operating state of bearings is affected by various factors, resulting in weak fault signals that are easily masked by high-noise signals, thus making it extremely difficult to distinguish early failures of bearings. Therefore, effectively and real-time monitoring and accurately evaluating the degradation state of rolling bearings have become the key to enhancing the operating reliability of equipment, avoiding unexpected shutdowns, and reducing maintenance costs.

[0003] In the patent "A method and device for evaluating the health state of the spindle bearing of a manufacturing equipment" (CN202110192571.X) applied by Nanjing University of Aeronautics and Astronautics, a method for evaluating the health state of bearings is proposed. This method first collects the original vibration signals of the bearings in real time, preprocesses the original vibration signals, then performs demodulation processing on the processed data to obtain the squared envelope signal of the processed vibration signals, constructs a bearing health assessment model based on this signal, and finally uses this health assessment model to evaluate the health state of the bearings. The deficiencies of this method are that the data collection requirements are relatively high, the dependence on expert experience and knowledge is relatively strong, the adaptive ability of the model is relatively weak, and the degree of intelligence is insufficient. Summary of the Invention

[0004] Aiming at the problems existing in the prior art, the present invention provides a method for constructing unsupervised health indicators and evaluating performance degradation of rolling bearings.

[0005] The present invention is implemented as follows. A method for constructing unsupervised health indicators and evaluating performance degradation of rolling bearings constructs a high-precision rolling bearing health assessment model through the combination of a deep learning model, the multi-scale temporal convolutional network (MSTCN) and the multi-head self-attention mechanism (MHSA); uses unsupervised learning to extract the time domain, frequency domain and power spectral features of the rolling bearing vibration signal, and screens out the optimal features through monotonicity and trendiness to construct a health index (HI); when training the model, uses the method of combining grid search with cross-validation to take the average to determine the best parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolution kernel size and the number of dilation convolution factors. Therefore, the convolution kernel range is set from 2 to 5, and the range of the dilation factor number is from 0 , 2 1 , 2 2 , 2 3 , 2 4 to 0 , 2 1 , 2 2 , 2 3 , 2 4 , 2 5 , 2 6 selects the optimal parameters as its parameters, and performs iterative optimization through the reconstruction error as the loss function to achieve accurate degradation evaluation; finally, through the smoothed health state degradation curve, the health state of the rolling bearing can be accurately monitored, providing effective support for equipment predictive maintenance.

[0006] Furthermore, the method includes the following steps:

[0007] S1: Collection of rolling bearing data;

[0008] S2: Data preprocessing;

[0009] S3: Feature extraction;

[0010] S4: Feature screening;

[0011] S5: Construction of a rolling bearing health assessment model;

[0012] S6: Smoothing processing of the rolling bearing health state degradation curve.

[0013] Furthermore, the collection of the rolling bearing data monitors the data of the equipment operation state through a sensor monitoring system. The system consists of a digital display, a PCB 352C33 type acceleration sensor, and a test rolling bearing for collecting and gathering the changes of temperature, horizontal vibration signal and vertical vibration signal during the operation process, including health state data and degradation state data.

[0014] Further, S2 specifically includes: data denoising. The specific operation is to perform downsampling on the original vibration signal sampled at high frequency, intercept the original signal with a sliding window, calculate the average value of the data in each window as a new observed data sequence, reduce the total amount of data while denoising, thereby improving the data quality and the subsequent model calculation efficiency.

[0015] Further, S3 specifically includes: extracting features from the processed data, including time-domain features, frequency-domain features, and power spectrum features.

[0016] 1) The time-domain features include:

[0017] Mean value, root mean square value, standard deviation, root amplitude, absolute average amplitude, skewness, kurtosis, variance, peak-to-peak value, maximum value, minimum value, peak value, waveform index, peak index, impulse index, margin index, skewness index, kurtosis index.

[0018] 2) The frequency-domain features include:

[0019] Average frequency, centroid frequency, mean square frequency, frequency variance.

[0020] 3) Power spectrum features:

[0021] Sum of power spectra, peak-to-peak value of power spectra, mean value of power spectra, standard deviation of power spectra, peak value of power spectra, kurtosis of power spectra.

[0022] Further, S4 specifically includes: constructing a mixed index based on monotonicity and trendiness according to the rolling bearing degradation mechanism to screen the extracted features.

[0023] 1) Monotonicity characterizes the rising or falling trend of a feature. As the bearing degrades, this process is irreversible, and the signal features should theoretically show a monotonically increasing or decreasing trend. Monotonicity is measured by the absolute difference between the positive and negative derivatives of the feature. The specific formula is as follows:

[0024]

[0025] Among them, T is the length of the feature sequence, and dH represents the difference between each value of the feature sequence and the previous value. S mon = 1 indicates that the feature shows complete monotonicity in the time series;

[0026] 2) Trendiness measures the linear correlation degree of a feature with time. As the working time of the bearing increases, the wear degree of the bearing becomes more and more serious, and the signal features should theoretically show a certain correlation with time. The specific formula of trendiness is as follows:

[0027]

[0028] where T is the length of the feature sequence, and x i is the i-th value of the feature sequence, t i is the mean value of the feature sequence, is the cumulative working time corresponding to the i-th value, and t is the mean value of the time series.

[0029] 3) The hybrid index uses the mean of the monotonicity value and the trendiness value as the criterion for the optimal feature. The specific formula of the hybrid index is as follows:

[0030]

[0031] After obtaining the final scores of each feature, the top q features are selected as the input features of the model, where the value of q can be determined according to actual needs.

[0032] Furthermore, the S5 specifically includes:

[0033] Construct a model based on the Multi-Scale Temporal Convolutional Network with Multi-Head Self-Attention Mechanism (MSTCN-MHSA). The input sequence first passes through the Temporal Convolutional Network (TCN) layers composed of multiple convolutions with different sizes and dilation factors respectively to extract the degradation features of the bearing, and then the data of all TCN layers are concatenated; the obtained multi-scale data is input into the multi-head self-attention mechanism to remove the redundant information in the multi-scale data and retain the useful information. After that, the data obtained by the multi-head self-attention mechanism MHSA is connected with the multi-original scale data in a residual connection; the data after the residual connection is input into the fully connected layer network for fusion to construct the Health Index (HI), and finally the result is output, that is, the health index corresponding to the input sequence sample;

[0034] When training the MSTCN-MHSA model, an unsupervised learning method is adopted, and the features extracted and screened from the rolling bearing samples in the healthy state are used as the input. First, initialize the model parameters, and use the method of combining grid search and cross-validation to take the average value to determine the optimal parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolution kernel size and the number of dilation convolution factors. Therefore, set the convolution kernel range from 2 to 5 and the number of dilation factors range from [2 0 , 2 1 , 2 2 , 2 3 , 2 4 to [2 0 , 2 1 , 2 2 , 23 , 2 4 , 2 5 , 2 6 Select the optimal parameters as its parameters. Subsequently, calculate the reconstruction error between the model input and the model output, use this error as the loss function, and gradually optimize the model parameters through backpropagation. Repeat this process until the reconstruction error converges to a small range, and then save the final model parameters. Subsequently, input the feature data of the test bearing samples into the trained MSTCN-MHSA model to generate the health indicators of the rolling bearing.

[0035] Further, the S6 specifically includes:

[0036] The health index (HI) curve obtained in step S5 may have high-frequency noise. In order to obtain a more reliable health index, that is, to reduce the burrs and spikes of the curve and make the HI curve smoother, the moving average method is used to smooth the degradation curve of the rolling bearing to filter out the high-frequency noise in the data and retain the main low-frequency trend.

[0037] Another object of the present invention is to provide an unsupervised learning-based rolling bearing health index construction and performance degradation evaluation system for implementing the above-mentioned unsupervised learning-based rolling bearing health index construction and performance degradation evaluation method, including:

[0038] Data collection module 1, used for collecting rolling bearing data;

[0039] Data preprocessing module 2, connected to the data collection module 1, used for preprocessing rolling bearing data;

[0040] Feature extraction module 3, connected to the data preprocessing module 2, used for extracting time-domain features, frequency-domain features, and power spectrum features from the bearing vibration signal;

[0041] Feature screening module 4, connected to the feature extraction module 3, used for screening the extracted features and selecting appropriate features;

[0042] Rolling bearing health assessment model construction module 5, connected to the data preprocessing module 2, the feature extraction module 3, and the feature screening module 4, used for constructing a bearing health assessment model;

[0043] Smoothing processing module 6, connected to the rolling bearing health assessment model construction module 5, used for smoothing the bearing health index.

[0044] Another object of the present invention is to provide a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the above-mentioned random access method of the adaptive distributed queue.

[0045] The technical solution of the present invention solves the problems of difficult accurate prediction of the health state of rolling bearings, low data processing efficiency, and poor model adaptability in the industrial application of rolling bearing health monitoring and performance degradation assessment, achieving significant technological progress, which is specifically manifested in the following aspects:

[0046] 1. Unsupervised health assessment enables more accurate degradation prediction

[0047] Problems of the prior art: Most of the traditional rolling bearing health monitoring is based on supervised learning and relies on a large amount of labeled fault data. However, in actual industrial applications, bearing fault data is usually scarce, making it difficult to meet the requirements of supervised learning and resulting in inaccurate health assessment results.

[0048] The present invention adopts an unsupervised learning method that does not require fault-labeled data. By constructing an MSTCN-MHSA model, it automatically extracts and screens key features, and can generate accurate health indicators (Health Index, HI) using health state data. This unsupervised approach significantly improves the adaptability of the model, enabling high-precision degradation prediction even in the absence of fault data.

[0049] 2. Multi-dimensional feature extraction and screening improve the accuracy and reliability of the model

[0050] Problems of the prior art: In the existing rolling bearing monitoring methods, single features or simple feature combinations are difficult to fully reflect the health state of bearings, and feature redundancy and data noise often lead to inaccurate prediction results.

[0051] The present invention extracts multi-dimensional features from the time domain, frequency domain, and power spectrum, constructs a hybrid index using monotonicity and trendiness, and screens out key features highly correlated with bearing degradation. This feature selection process is based on the degradation mechanism, ensuring the representativeness and effectiveness of the data input into the model, effectively eliminating the interference of redundant features, and improving the accuracy and reliability of the health assessment model.

[0052] 3. Multi-scale temporal convolutional network and self-attention mechanism optimize data processing efficiency

[0053] Problems of the prior art: In traditional bearing health monitoring systems, the ability of deep learning models to extract temporal features and suppress noise is weak. When processing large-scale temporal data, it often leads to slow model training speed and difficulty in meeting the high real-time requirements of industrial sites.

[0054] The present invention adopts a combined structure of a multi-scale temporal convolutional network (MSTCN) and a multi-head self-attention mechanism (MHSA). The MSTCN layer can effectively extract multi-scale temporal features and capture different frequencies and temporal dependencies of vibration data; while the MHSA mechanism improves the anti-interference ability of the model by removing redundant information and retaining effective features. This combined structure significantly improves the data processing efficiency, shortens the model training time, and meets the requirements of real-time health assessment in industrial applications.

[0055] 4. The smoothing process of the health state degradation curve improves the visualization of the degradation trend

[0056] Problems of the prior art: The results of traditional bearing health assessment fluctuate greatly, and the health state curve is unstable, resulting in the degradation trend being difficult to observe, which affects the prediction of the remaining life of the bearing.

[0057] The present invention smooths the health state degradation curve by the moving average method, effectively reducing the high-frequency fluctuations of the curve, making the health state curve more stable and facilitating the intuitive observation of the bearing degradation trend. The smoothed degradation curve can provide a clear health change trend, providing an intuitive basis for accurately predicting the remaining service life of the bearing, which is helpful for predictive maintenance and fault prevention.

[0058] 5. The noise reduction data preprocessing improves the computational efficiency and accuracy of the model

[0059] Problems of the prior art: There is a lot of noise in the traditional rolling bearing data processing process, resulting in a decrease in the training accuracy of the model and low computational efficiency, and it is easily affected by environmental noise in the industrial field, and the prediction effect is not ideal.

[0060] The present invention adopts a data noise reduction method in the data preprocessing stage. By calculating the mean value of every M pieces of data, it not only removes the data noise but also significantly reduces the total amount of data. The noise reduction preprocessing reduces the interference of data fluctuations on the model, improves the computational efficiency, ensures that the model is more stable and accurate during operation, and adapts to the data environment in industrial applications.

[0061] The present invention solves the problems of the traditional technology in aspects such as data annotation, feature extraction, model training, and degradation trend visualization in the field of rolling bearing health monitoring and performance degradation assessment, significantly improving the accuracy, real-time performance, and stability of health assessment, providing a technical guarantee for the predictive maintenance of equipment. These technological advancements not only improve the robustness and adaptability of the system but also extend the service life of the equipment, having broad industrial application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a flow chart of the method for constructing an unsupervised rolling bearing health index and evaluating performance degradation of the present invention;

[0063] Figure 2 This is the network framework diagram of the multi-scale temporal convolutional network - multi-head self-attention mechanism model of the present invention;

[0064] Figure 3 This is the degradation curve of the health state of the rolling bearing based on MSTCN-MHSA in the embodiment of the present invention;

[0065] Figure 4 The unsupervised rolling bearing health index construction and performance degradation evaluation system in the embodiment of the present invention;

[0066] In the figure: 1. Data collection module; 2. Data preprocessing module; 3. Feature extraction module; 4. Feature screening module; 5. Rolling bearing health assessment model construction model; 6. Smoothing processing module. Specific implementation manners

[0067] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0068] The following are two specific application embodiments:

[0069] Embodiment 1: Health monitoring of rolling bearings in wind turbines

[0070] Application scenario: In a wind power generation unit, the rolling bearing, as a core transmission component, bears large mechanical stress and environmental stress. Once a failure occurs, it will cause the generator set to shut down or even be damaged. Since wind power generation units are usually located in remote areas, real-time monitoring and predictive maintenance are particularly important.

[0071] 1. Data collection and preprocessing: The sensor system installed at the bearing of the wind turbine collects temperature and vibration data in real time, and smooths the signal through a data denoising method to reduce noise interference.

[0072] 2. Feature extraction and screening: Extract time-domain, frequency-domain and power spectrum features from the collected vibration signals, including features such as peak-to-peak value, root mean square value, and average frequency. According to monotonicity and trendiness, screen out the features highly related to the bearing health state and construct an input feature set.

[0073] 3. Health index modeling and monitoring: Input the screened feature data into the MSTCN-MHSA model to generate health indexes. Dynamically adjust the health indexes through the reconstruction error of the model to generate a real-time health degradation curve.

[0074] 4. Smoothing the health degradation curve: Use the moving average method to smooth the generated degradation curve to obtain a stable trend curve of the bearing health change. Combine the changes in the degradation curve to judge the health status of the rolling bearing and predict potential failure risks.

[0075] This method realizes the real-time health monitoring of the rolling bearings of wind turbines, predicts potential failures in advance, avoids sudden shutdowns, provides guarantee for the stable operation of wind turbine generators, and extends the service life of the equipment.

[0076] Example 2: Degradation assessment of the rolling bearings of high-speed train axles

[0077] The rolling bearings of high-speed trains carry the task of high-speed operation of the trains, and their failures will directly affect the safety of the trains. Therefore, the rolling bearings of high-speed train axles need continuous monitoring and accurate degradation assessment to ensure the safety and operation efficiency of the trains.

[0078] 1. Data collection and preprocessing: Install vibration sensors and temperature sensors at the axle positions of the trains to collect temperature and vibration data during operation in real time. Remove environmental interference through data denoising methods and extract key data.

[0079] 2. Multi-dimensional feature extraction and screening: Perform time-domain and frequency-domain analysis on the vibration signals to extract features such as mean value, standard deviation, centroid frequency, and power spectrum peak value. Use monotonicity and trend evaluation indicators for screening, retain the features most relevant to the degradation process, and improve the prediction accuracy of the model.

[0080] 3. Degradation assessment model construction and training: Input the screened features into the MSTCN-MHSA model, use unsupervised learning methods to train the model, optimize the model parameters through reconstruction error, and generate health indicators. The generated health indicators can dynamically reflect the wear state of the bearings.

[0081] 4. Degradation curve smoothing and fault prediction: Use the moving average method to smooth the curve in the health state degradation curve, minimize the high-frequency fluctuations to observe the overall trend. According to the changes in the degradation curve, judge the health status of the bearings in a timely manner and predict the possible degradation speed and fault time.

[0082] This method provides a high-precision prediction means for the health management of the rolling bearings of high-speed trains. The system can continuously monitor and evaluate the bearing status without affecting the train operation, help the train maintenance team detect the degradation trend of the bearings in time, and thus plan replacement or maintenance to ensure the train's driving safety and efficient operation.

[0083] Such as Figure 1As shown in the figure, an unsupervised health index construction and performance degradation evaluation method for rolling bearings provided by an embodiment of the present invention includes the following steps:

[0084] S1, Rolling bearing data collection:

[0085] Monitor the data of the equipment operation state through a sensor monitoring system, which consists of a digital display, a PCB352C33 type acceleration sensor, a test rolling bearing, etc. for collecting and gathering the changes of various physical quantities (such as temperature and vibration) during the operation process, including health state data and degradation state data.

[0086] S2, Data preprocessing:

[0087] Data noise reduction and outlier processing. The specific processing method is to intercept the original signal with a sliding window, calculate the average value of every M pieces of data, reduce the total amount of data while reducing noise, and improve the model calculation efficiency.

[0088] S3, Feature extraction:

[0089] Extract the characteristic parameters related to the degradation process of the rolling bearing from the processed data samples, including time domain characteristics, frequency domain characteristics, and power spectrum characteristics, as shown in Table 1 specifically.

[0090] Table 1 Calculation formulas for each feature

[0091]

[0092]

[0093] S4, Feature screening:

[0094] According to the degradation mechanism of the rolling bearing, construct a hybrid index based on monotonicity and trend to screen the extracted features.

[0095] 1) Monotonicity characterizes the rising or falling trend of a feature. As the bearing degrades, this process is irreversible, and the signal characteristics should theoretically show a monotonically increasing or decreasing trend. Monotonicity is measured by the absolute difference between the positive derivative and the negative derivative of the feature, and the specific formula is as follows:

[0096]

[0097] Among them, T is the length of the feature sequence, and dH represents the difference between each value of the feature sequence and the previous value. S mon =1 indicates that the feature is completely monotonic in the time series.

[0098] 2) Trend measures the linear correlation degree of a feature with time. As the working time of the bearing increases, the wear degree of the bearing becomes more and more serious, and the signal feature should theoretically show a certain correlation with time. The specific formula for trend is as follows:

[0099]

[0100] where T is the length of the feature sequence, x i is the i-th value of the feature sequence, t i is the mean of the feature sequence, is the cumulative working time corresponding to the i-th value, and t is the mean of the time sequence.

[0101] 3) The mixed index uses the mean of the monotonicity value and the trend value as the screening criterion for the optimal feature. The specific formula for the mixed index is as follows:

[0102]

[0103] After obtaining the final scores of each feature, the top q features are selected as the input features of the model, where the value of q can be determined according to actual needs.

[0104] S5. Construct a rolling bearing evaluation model:

[0105] Construct a model based on the multi-scale temporal convolutional network - multi-head self-attention mechanism, as Figure 2 shown. The input sequence first passes through three temporal convolutional network (TCN) layers composed of different convolutional sums and dilation factors respectively to extract the degradation features of the bearing, and then the data of the three TCN layers are concatenated; the obtained multi-scale data is input into the multi-head self-attention mechanism to remove the redundant information in the multi-scale data and retain the useful information. After that, the data obtained by the multi-head self-attention mechanism (MHSA) is connected with the original multi-scale data in a residual connection; the data after the residual connection is input into the fully connected layer network for fusion to construct HI, and finally the result is output, that is, the health index corresponding to the input sequence sample;

[0106] When training the MSTCN-MHSA model, an unsupervised learning method is adopted, and the features extracted and screened from the rolling bearing samples in the healthy state are used as the input. First, initialize the model parameters, and use the method of combining grid search with cross-validation to take the average to determine the optimal parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolutional kernel size and the number of dilation convolutional factors. Therefore, set the convolutional kernel range from 2 to 5 and the dilation factor range from [2 0 , 2 1 , 2 2 , 2 3 , 2 4 to [2 0 , 2 1 , 22 , 2 3 , 2 4 , 2 5 , 2 6 Select the optimal parameters as its parameters. Then, calculate the reconstruction error between the input and the model output, use this error as the loss function, and gradually optimize the model parameters through backpropagation. Repeat this process until the reconstruction error converges to a small range, and then save the final model parameters. Subsequently, input the feature data of the test bearing samples into the trained MSTCN-MHSA model to generate the health indicators of the rolling bearing.

[0107] S6, Smoothing processing of the rolling bearing health indicator:

[0108] The health indicator (HI) curve obtained in step S5 may have high-frequency noise. To obtain a more reliable rolling bearing health state degradation curve, that is, to reduce the burrs and spikes of the curve and make the HI curve smoother, the moving average method is used to smooth the rolling bearing health indicator. The moving average method is a low-pass filter that can filter out the high-frequency noise of the data and retain the main low-frequency trend. Its expression is as follows:

[0109]

[0110] In the formula, y N is the input data, and N is the window size.

[0111] In this rolling bearing health indicator construction and performance degradation evaluation system based on unsupervised learning, each module plays a key role. The following is a detailed explanation of each module and related charts.

[0112] The working principle of the present invention realizes the evaluation of the health state and the prediction of performance degradation of the rolling bearing through steps such as data preprocessing, feature extraction and screening, and construction and optimization of the unsupervised learning model. The specific steps are as follows:

[0113] 1. Data collection and preprocessing

[0114] During the operation of the rolling bearing, the sensor monitoring system (including acceleration sensors and temperature sensors) collects data in real time under the operating state. These data contain health state and degradation state information. To reduce noise interference and improve the effectiveness of the data, the preprocessing step uses a noise reduction method to calculate the mean of every M pieces of data, which not only reduces the amount of data but also improves the calculation efficiency of the model.

[0115] 2. Multidimensional feature extraction

[0116] Extract the key features of time domain, frequency domain and power spectrum from the preprocessed data. The time domain features include mean, root mean square value, standard deviation, peak-to-peak value, etc. The frequency domain features include average frequency, center frequency, etc. The power spectrum features include power spectrum sum, power spectrum peak, etc. These multi-dimensional features can reflect the physical property changes of the bearing under different health states, providing rich information for subsequent modeling.

[0117] 3. Feature Screening

[0118] To improve the recognition accuracy of the model for the bearing degradation state, the feature screening process selects key features through monotonicity and trend indicators. Monotonicity is used to measure the irreversible change trend of feature values over time, indicating the monotonic increasing or decreasing trend of features; trend is used to measure the linear correlation degree between features and time to reflect the progress of bearing wear. Finally, features with the optimal degradation information are screened out through the mean of monotonicity and trend scores, providing reliable degradation features for model input.

[0119] 4. Construction of Health Index Model

[0120] Construct an unsupervised health assessment model based on the multi-scale temporal convolutional network (MSTCN) and the multi-head self-attention mechanism (MHSA). MSTCN captures multi-scale features through convolutional kernels and dilation factors of different scales, extracting different frequency components and time dependence relationships of the data. In the model, the method of combining grid search with cross-validation and taking the average is used to determine the optimal parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolutional kernel size and the number of dilation convolutional factors. Set the convolutional kernel range from 2 to 5 and the number of dilation factors range from 0 , 2 1 , 2 2 , 2 3 , 2 4 to 0 , 2 1 , 2 2 , 2 3 , 2 4 , 2 5 , 2 6 and select the optimal parameters as its parameters. Then, input the extracted multi-scale features into the MHSA module, and use the self-attention mechanism to remove redundant information and retain key features, thereby reducing data noise interference. Finally, the multi-scale features and attention features are input into the fully connected layer after residual connection to generate a health index (HI) to characterize the health state of the rolling bearing.

[0121] 5. Unsupervised Model Training

[0122] During the training process of the model, it does not rely on labels and directly uses the characteristics of rolling bearings in the normal state as input, and optimizes the model parameters through the reconstruction error. The model is repeatedly trained to minimize the reconstruction error, taking this error as the loss function of unsupervised learning, and using backpropagation and gradient descent to optimize the model until the reconstruction error converges to a small range. The trained model has the ability to identify the degraded state and can generate health indicators for new rolling bearing data.

[0123] 6. Smoothing the health state degradation curve

[0124] During the evaluation process, the generated health indicators are plotted over time as a health state degradation curve. To eliminate the influence of noise and data fluctuations, the moving average method is used to smooth the curve to form a stable degradation curve. This curve can clearly reflect the health change trend of the rolling bearing and achieve accurate monitoring of the bearing degradation process.

[0125] Based on the above working principle, the present invention constructs an unsupervised health assessment model, which can accurately predict the degradation state of rolling bearings and provides important technical support for the health monitoring and predictive maintenance of equipment.

[0126] Figure 3 Shows the results obtained by the rolling bearing health indicator construction model based on MSTCN-MHSA (Multi-Scale Temporal Convolution - Multi-Head Self-Attention Mechanism) in this embodiment. Through the chart, the differences between the health indicators constructed by different models can be seen to evaluate the evaluation performance of each model.

[0127] Figure 4 Is the framework diagram of the entire system, including the following modules:

[0128] Data collection module 1: This module is responsible for collecting the operating state data of rolling bearings.

[0129] Data preprocessing module 2: Preprocesses the collected rolling bearing data, denoises the original data and reduces the data volume.

[0130] Feature extraction module 3: Extracts the features related to the rolling bearing degradation process from the processed data samples according to the rolling bearing degradation mechanism.

[0131] Feature screening module 4: Screens the extracted features with a mixed index composed of monotonicity and trendiness to further obtain the features that can fully reflect the rolling bearing degradation process.

[0132] Rolling bearing health assessment model construction module 5: Uses deep learning algorithms, such as temporal convolutional network and multi-head self-attention mechanism, to train a rolling bearing health assessment model based on the rolling bearing health state feature samples and conduct health assessment on the rolling bearing data samples.

[0133] Smoothing processing module 6: Smooth the rolling bearing health indicators output by the model to reduce the influence of noise.

[0134] These modules and charts together constitute a complete unsupervised rolling bearing health indicator construction and performance degradation assessment system. As an optimized solution of the embodiment of the present invention, the following are the specific implementation solutions for each step:

[0135] S1: Rolling bearing data collection

[0136] Use a sensor monitoring system to monitor and collect the operating state data of the rolling bearing in real time. These data include temperature, horizontal vibration signal, vertical vibration signal, etc.

[0137] S2: Data preprocessing

[0138] The sliding average method is used to slide the window of the collected signal data. The window size is M. By calculating the average value Mean of the data within the window to replace the original data points, noise reduction and smoothing processing are achieved.

[0139] S3: Feature extraction

[0140] Extract the key feature parameters affecting the degradation process of the rolling bearing from the processed data samples, mainly including time-domain features, frequency-domain features, and power spectrum features.

[0141] S4: Feature screening

[0142] Screen the extracted features according to the hybrid index constructed by monotonicity and trend. After obtaining the final score of each feature, select the top q features as the input features of the model, where the value of q can be determined according to actual needs.

[0143] S5: Construct a rolling bearing health assessment model

[0144] Use deep learning algorithms (such as temporal convolutional network, multi-head self-attention mechanism, etc.) to train a rolling bearing health assessment model based on the historical health state data of the rolling bearing. The input of the model is the selected top q features, and the output is the health indicator HI of the bearing.

[0145] S6: Smooth the rolling bearing health indicators

[0146] Smooth the bearing health indicators output by the model. The moving average method can be used to reduce the influence of noise.

[0147] These steps are integrated together to form a complete rolling bearing health index construction and performance degradation evaluation system. Through this system, the degradation state analysis of rolling bearings can be realized, and corresponding maintenance decisions can be made, greatly improving the working efficiency and safety of mechanical equipment.

[0148] As Figure 4 shown, an embodiment of the present invention provides a rolling bearing health index construction and performance degradation evaluation system based on deep learning for implementing the rolling bearing health index construction and performance degradation evaluation method based on deep learning. The system includes:

[0149] Data collection module 1: This module is responsible for collecting the operating state data of rolling bearings.

[0150] Data preprocessing module 2: Preprocesses the collected rolling bearing data, denoises the original data, and reduces the data volume.

[0151] Feature extraction module 3: Extracts features related to the rolling bearing degradation process from the processed data samples according to the rolling bearing degradation mechanism.

[0152] Feature screening module 4: Screens the extracted features with a mixed index composed of monotonicity and trendiness to further obtain features that can fully reflect the rolling bearing degradation process.

[0153] Rolling bearing health assessment model construction model 5: Uses deep learning algorithms, such as temporal convolutional network and multi-head self-attention mechanism, trains a rolling bearing health assessment model according to the rolling bearing health state feature samples, and conducts a health assessment on the rolling bearing data samples.

[0154] Smoothing processing module 6: Smooths the rolling bearing health index output by the model to reduce the influence of noise.

[0155] An embodiment of the present invention provides a computer device, which includes a memory and a processor. The memory stores a computer program. When the computer program is executed by the processor, the processor executes the steps of the rolling bearing health index construction and degradation state evaluation method based on unsupervised learning.

[0156] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated designed hardware. Those of ordinary skill in the art can understand that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or included in processor control code, such as provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and their modules of the present invention can be implemented by hardware circuits of programmable hardware devices such as very large scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or can be implemented by software executed by various types of processors, or can be implemented by a combination of the above hardware circuits and software, such as firmware.

[0157] The present invention provides a specific application of a method for constructing unsupervised health indicators and evaluating performance degradation of rolling bearings, as follows:

[0158] Step 1, rolling bearing data collection;

[0159] The data of the embodiment comes from the data collected by the rolling bearing accelerated life test bench of Xi'an Jiaotong University. The accelerated life test bench consists of a digital display, an acceleration sensor, a test rolling bearing, etc., and can carry out accelerated life tests of rolling bearings under different working conditions to obtain full life cycle data. In the test, two PCB 352C33 type accelerometers were placed at the 90° position on the housing of the bearing under test, that is, one was installed in the horizontal direction and the other was installed in the vertical direction. The sampling frequency was 25.6 kHz, the sampling time for each time was 1.28 s, the sampling interval was 1 min, and the number of sampling points for each time was 32768. A total of 15 test bearings were used in the test. Every 5 bearings were divided into a group, and a total of three groups were divided. Each group of bearings operated normally under the same working conditions until failure.

[0160] Step 2, data preprocessing;

[0161] The original horizontal vibration signal of the rolling bearing is processed using the moving average method, and the average value is calculated every 300 pieces of data, reducing the data volume to 1 / 300 of the original data.

[0162] Step 3, feature extraction;

[0163] Feature extraction is performed on the processed data, mainly including:

[0164] 1) Time domain features include:

[0165] Mean value, root mean square value, standard deviation, root amplitude, absolute average amplitude, skewness, kurtosis, variance, peak-to-peak value, maximum value, minimum value, peak value, waveform index, peak index, impulse index, margin index, skewness index, kurtosis index.

[0166] 2) Frequency domain features include:

[0167] Average frequency, center frequency, mean square frequency, frequency variance.

[0168] 3) Power spectrum features include:

[0169] Sum of power spectrum, peak-to-peak value of power spectrum, mean value of power spectrum, standard deviation of power spectrum, peak value of power spectrum, kurtosis of power spectrum.

[0170] Step 4, feature screening;

[0171] Use the mixed index Cri as the comprehensive index of the optimal features for screening, and calculate the corresponding Cri value for each feature. As shown in Table 2.

[0172] Table 2 Ranking of comprehensive scores of each feature

[0173]

[0174] Step 5, construct a bearing health assessment model based on MSTCN-MHSA;

[0175] Use the first 15 feature parameters as the input of the health assessment model, and the corresponding HI index as the model output; the first 30% of the data is the health state data of the rolling bearing and is used as the training set, and the remaining data is used as the validation set. Use the training set to train the model, and then use the validation set to evaluate the model performance;

[0176] Use the method of grid search combined with cross-validation to take the average value to determine the best parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolution kernel size and the number of dilated convolution factors. Therefore, set the convolution kernel range from 2 to 5 and the number of dilation factors range from [2 0 , 2 1 , 2 2 , 2 3 , 2 4 to [2 0 , 2 1 , 2 2 , 2 3 , 2 4 , 2 5 , 2 6 . Finally, the optimal parameters obtained through experiments are that the convolution kernel sizes are 2, 3, and 5 respectively, and the dilated convolution factor combination is [2 0 , 2 1 , 2 2,2 3 ,2 4 , [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ,2 5 and [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ,2 5 ,2 6 , and the result verifies that the features obtained from different receptive fields can indeed significantly help in constructing the Health Index (HI). The loss function selects the reconstruction error data z t The corresponding reconstruction error e t The expression of is as shown in the formula, and the training objective of the model is to minimize the overall reconstruction error E = ∑e t .

[0177] e t = ||z t - z' t ||

[0178] The learning rate is set to 0.0001, the optimizer selects Adam, and the number of iterations is set to 30.

[0179] To verify the advancement of the method proposed in the present invention, comparative experiments are conducted with the classical LSTM-ED model and the AE model. To avoid significant differences in model performance caused by differences in the number of model parameters, the network parameters of the three models are adjusted so that the number of parameters of the three networks is similar. The parameters of each model are shown in Table 3.

[0180] Table 3 Main parameter settings of each model

[0181]

[0182] To make the results more reliable, the monotonicity, trend, and robustness of the three models on each sample were calculated and added together as the final result, as shown in Table 4. From the results in the table, it can be seen that the method proposed in the present invention obtained relatively high monotonicity scores on most datasets, especially on bearing datasets such as Data 1_1 and Data 1_2 where the trend changes are relatively regular, and the performance was particularly prominent. In addition, although all three methods achieved good scores overall, indicating that these methods have a certain degree of reliability in bearing degradation state extraction, in the dataset of Data 3_2 with low trend and unclear monotonicity, the method proposed in the present invention still performed excellently, further demonstrating its advantages in constructing health indicators. Secondly, the health indicators constructed by the method proposed in the present invention showed high robustness in all datasets, indicating its significant advantages in anti-interference ability. These results show that the method proposed in the present invention can stably construct high-quality health indicators under different datasets and complex working conditions, and has broad practical application potential.

[0183] Table 4 Score table of different HI construction methods in XJTU - SY dataset

[0184]

[0185] Step 6, smoothing the degradation curve of the rolling bearing health state;

[0186] In order to obtain a more reliable degradation curve of the rolling bearing health state, that is, to reduce the burrs and spikes of the curve and make the HI curve smoother, the moving average method is used to smooth the degradation curve of the rolling bearing health state. Figure 3 Degradation curves based on the MSTCN - MHSA model on 6 bearing datasets.

[0187] Example 1: Health assessment and degradation prediction of rolling bearings in industrial mechanical equipment

[0188] In industrial equipment such as compressors, fans, and motors, rolling bearings are crucial components that directly affect the normal operation of the equipment. The present invention proposes a technical solution based on a deep learning model, which can collect a large amount of data from operation data, vibration signals, temperature sensor data, and maintenance records, and input it into the deep learning model for training and analysis. Through this model, the health state of the rolling bearing can be evaluated in real time, helping maintenance personnel formulate maintenance plans, reducing equipment downtime, and improving production efficiency.

[0189] Example 2: Health assessment and degradation prediction of rolling bearings in wind turbines

[0190] The rolling bearings of wind turbines need to withstand complex load conditions for a long time, so their health status is crucial for power generation efficiency and system safety. The present invention can collect big data from the vibration signals, rotational speeds, temperature data of the generator, as well as environmental conditions, etc., and input them into a deep learning model for training. The output of the model can evaluate the health status of the bearings in real time, helping the operator to perform maintenance in a timely manner, reducing the failure risk and downtime of the generator, and improving the overall operational efficiency of the wind farm.

[0191] Embodiment 3: Health Assessment and Degradation Prediction of Rolling Bearings in Aero-engines

[0192] The rolling bearings of aero-engines operate under high-temperature, high-speed, and high-load conditions, and health monitoring is crucial. The present invention establishes an accurate evaluation model by acquiring vibration data, temperature data, and oil analysis data during the operation of the engine, in combination with deep learning technology. Through this model, the health status of the rolling bearings of aero-engines can be evaluated, helping aviation maintenance personnel to take preventive measures in a timely manner, avoiding aviation accidents caused by bearing failures, and improving flight safety and the service life of the engine.

[0193] Embodiment 4: Health Assessment and Degradation Prediction of Rolling Bearings in Rail Transit

[0194] In the rail transit system, the health status of the rolling bearings of trains directly affects the safety and operation efficiency of trains. This technical solution collects data such as vibration, temperature, and rotational speed during the operation of the train, combines the information of the surrounding environment and loads along the way, and inputs them into a deep learning model for training. By evaluating the health status of the rolling bearings in real time, maintenance personnel can arrange maintenance or replacement of the bearings in advance, avoid sudden failures during train operation, and ensure the safety and continuous operation of the transportation system.

[0195] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention shall be covered by the protection scope of the present invention.

Claims

1. A method for constructing unsupervised health indicators and evaluating performance degradation of rolling bearings, characterized in that: A high-precision rolling bearing health assessment model was constructed by combining the deep learning model multi-scale time series convolutional network (MSTCN) and the multi-head self-attention mechanism (MHSA). This method uses unsupervised learning to extract the time domain, frequency domain and power spectrum features of the rolling bearing vibration signal, and selects the optimal features through monotonicity and trend to construct a health index (HI). During model training, the grid search combined with the cross-validation average method is used to determine the optimal parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolution kernel size and the number of dilation convolution factors, so the convolution kernel range is set to 2 to 5 and the dilation factor range is [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ] to [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ,2 5 ,2 6 ] The optimal parameters are selected as its parameters, and iterative optimization is performed by reconstructing the error as the loss function to achieve accurate degradation assessment; finally, the health status of the rolling bearing can be accurately monitored through the smoothed health status degradation curve.

2. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: Specifically include: S1: rolling bearing data collection; S2: data preprocessing; S3: Feature extraction; S4: Feature screening; S5: Construct a rolling bearing health assessment model; S6: Smoothing of rolling bearing health degradation curve.

3. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The rolling bearing data collection monitors the data under the operating status of the equipment through a sensor monitoring system. The system consists of a digital display, a PCB 352C33 acceleration sensor, and a test rolling bearing, and is used to collect and collect changes in temperature, horizontal vibration signals, and vertical vibration signals during operation, including health status data and degradation status data.

4. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The S2 specifically includes: Data denoising, data denoising, the specific operation is to downsample the original vibration signal with high frequency sampling, perform sliding window interception on the original signal, calculate the average value of the data in each window as the new observation data sequence, reduce the total amount of data while reducing the noise, and improve the calculation efficiency of the model.

5. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The S3 specifically includes: extracting features from the processed data, including time domain features, frequency domain features and power spectrum features, wherein the time domain features include: mean, root mean square value, standard deviation, root amplitude, absolute average amplitude, skewness, kurtosis, variance, peak-to-peak value, maximum value, minimum value, peak value, waveform index, peak index, pulse index, margin index, skewness index, kurtosis index; the frequency domain features include: average frequency, center of gravity frequency, mean square frequency, frequency variance; the power spectrum features include: power spectrum sum, power spectrum peak-to-peak value, power spectrum mean, power spectrum standard deviation, power spectrum peak value, power spectrum kurtosis.

6. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The S4 specifically includes: constructing a hybrid index based on monotonicity and trend according to the rolling bearing degradation mechanism to screen the extracted features; 1) Monotonicity characterizes the rising or falling trend of a feature. As the bearing degrades, this process is irreversible. The signal feature should theoretically show a monotonic rising or monotonic falling trend. Monotonicity is measured by the absolute difference between the positive and negative derivatives of the feature. The specific formula is as follows: Where T is the length of the feature sequence, dH represents the difference between each value of the feature sequence and the previous value; S mon =1 indicates that the feature is completely monotonic in the time series; 2) Trendiness measures the linear correlation between a feature and time. As the working time of the bearing increases, the degree of bearing wear becomes more and more serious. In theory, the signal feature should show a certain correlation with time. The specific formula for trendiness is as follows: Among them, T is the length of the feature sequence, x i is the i-th value of the feature sequence, t i is the mean of the feature sequence, is the cumulative working time corresponding to the i-th value, and t is the mean of the time series; 3) The hybrid index takes the mean of the monotonicity value and the trend value as the standard of the optimal feature. The specific formula of the hybrid index is as follows: After obtaining the final score of each feature, the first q features are selected as the input features of the model, where the q value can be determined according to actual needs.

7. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The S5 specifically includes: A multi-scale temporal convolutional network-multi-head self-attention mechanism model is constructed. The input sequence first passes through a temporal convolutional network layer composed of multiple convolutions of different sizes and expansion factors to extract the degradation characteristics of the bearing, and then the data of all TCN layers are spliced; the obtained multi-scale data is input into the multi-head self-attention mechanism to remove redundant information in the multi-scale data and retain useful information. Then, the data obtained by the multi-head self-attention mechanism MHSA is residually connected with the multi-original scale data; the residually connected data is input into the fully connected layer network for fusion to construct the health index (HealthIndex, HI), and finally the result is output, that is, the health index corresponding to the input sequence sample; Unsupervised learning is used to train the MSTCN-MHSA model. The features extracted and screened from the rolling bearing samples in a healthy state are used as input. First, the model parameters are initialized. The grid search combined with the cross-validation average method is used to determine the optimal parameters of the multi-scale time series convolutional network (MSTCN). The main parameters include the convolution kernel size and the number of dilation convolution factors. Therefore, the convolution kernel range is set to 2 to 5 and the dilation factor range is [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ] to [2 0 ,2 1 ,2 2 ,2 3 ,2 4 ,2 5 ,2 6 ]Select the optimal parameters as its parameters; then, calculate the reconstruction error between the input and the model output, use the error as the loss function, and gradually optimize the model parameters through back propagation; repeat this process until the reconstruction error converges to a smaller range, and save the final model parameters; then, input the test bearing sample feature data into the trained MSTCN-MHSA model to generate the health indicators of the rolling bearing.

8. The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning as claimed in claim 1, characterized in that: The S6 specifically includes: using a moving average method to smooth the rolling bearing degradation curve to filter out high-frequency noise in the data.

9. A rolling bearing health indicator construction and performance degradation assessment system based on unsupervised learning, which implements the rolling bearing health indicator construction and performance degradation assessment method based on unsupervised learning as described in any one of claims 1 to 8, characterized in that: The rolling bearing health index construction and performance degradation assessment method based on unsupervised learning includes: Data collection module 1, used for rolling bearing data collection; The data preprocessing module 2 is connected to the data collecting module 1 and is used for preprocessing the rolling bearing data; The feature extraction module 3 is connected to the data preprocessing module 2 and is used to extract the time domain features, frequency domain features and power spectrum features from the bearing vibration signal; The feature screening module 4 is connected to the feature extraction module 3 and is used to screen the extracted features and select appropriate features; A rolling bearing health assessment model building module 5 is connected to the data preprocessing module 2, the feature extraction module 3 and the feature screening module 4, and is used to build a bearing health assessment model; The smoothing processing module 6 is connected to the rolling bearing health assessment model building module 5 and is used for smoothing the bearing health indicators.

10. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the rolling bearing health indicator construction and performance degradation assessment method based on unsupervised learning as described in any one of claims 1-8.

Images