A fault diagnosis method based on fine time-shifted multi-scale attention entropy

Abstract

The present invention belongs to the technical field of rolling bearing fault diagnosis, and specifically relates to a fault diagnosis method based on fine time-shifted multi-scale attention entropy, the steps of which include: using a fine time-shifted sliding window method to perform coarse-grained segmentation on the original data, calculating the attention entropy value of the coarse-grained time series at each scale, fusing multi-scale feature vectors as extracted fault features, constructing a feature vector set and inputting it into a classifier for fault diagnosis. This method not only effectively solves the problem of unstable entropy values of traditional multi-scale entropy methods at high scales, but also avoids dependence on hyperparameters during the feature extraction process. Experimental results show that the average classification accuracy of the fault diagnosis method of the present invention on bearing data sets exceeds 99%, significantly improving the rolling bearing fault diagnosis and identification capabilities.

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing fault diagnosis, and in particular relates to a fault diagnosis method based on fine time-shift multi-scale attention entropy. Background Art

[0002] Critical mechanical systems in intelligent equipment are constantly exposed to complex operating conditions characterized by high loads and strong interference. Early in the operation of these critical mechanical systems, vibration signals often exhibit abnormal fluctuations with weak amplitudes and short durations. However, these potential fault signals are easily obscured by non-stationary interference signals such as background noise and ambient vibration. Rolling bearings, core components of key high-speed train equipment, are widely used due to their compact structure and high transmission efficiency. However, even small fluctuations in the operating state of the bearings can directly impact the safe operation of the equipment. In engineering practice, feature extraction of fault signals under complex operating conditions is difficult, posing a significant challenge to fault diagnosis.

[0003] In terms of fault feature extraction, traditional methods have the following main limitations: although time domain and frequency domain analysis methods (such as peak factor, kurtosis, envelope spectrum, etc.) are simple and easy to use, their statistical characteristics are easily affected by changes in operating conditions and noise interference; although time-frequency analysis methods (STFT, CWT, WVD, etc.) can characterize the local instantaneous characteristics of the signal, their computational complexity is high and requires expert guidance; although adaptive signal decomposition methods (EMD, VMD, etc.) can process non-stationary signals, their performance is heavily dependent on parameter settings and are prone to problems such as modal aliasing.

[0004] In recent years, fault feature extraction methods based on entropy theory have garnered widespread attention due to their ability to characterize signal complexity from a nonlinear dynamics perspective. Common metrics include sample entropy (SE), permutation entropy (PE), fuzzy entropy, and dispersion entropy (DE). These metrics measure data complexity through template matching, permutation, fuzzy membership, and probability distribution dispersion, respectively, providing effective features for early fault diagnosis. However, these methods often require pre-set hyperparameters such as the embedding dimension, tolerance threshold, or number of categories. This leads to limited robustness to varying operating conditions, poor applicability in strong noise environments, and the tendency to distort entropy values. To avoid hyperparameter manipulation, attention entropy (AE) proposes using the interval sequence of adjacent extreme points instead of the original vibration sequence for complexity calculation. Its advantages include requiring no parameter adjustments and being sensitive to faults. However, since it only evaluates signal features at a single scale, it lacks the ability to characterize signals at different scales and resolutions, making it difficult to fully describe the multi-level features of critical components.

[0005] To incorporate information at different scales, researchers have proposed multiscale entropy. This method first coarsens (or reconstructs) the signal and then calculates entropy values at each scale. However, the coarse-graining process results in a sharp decrease in the number of samples as the scale increases. This is particularly true in high-scale intervals, where significant random fluctuations can lead to unstable entropy curves. Furthermore, shortening the sequence reduces sensitivity to short-term transient shocks, reducing the reliability of diagnostic results.

[0006] In summary, the existing entropy method diagnosis technology still faces many problems to be solved in terms of parameter dependence, noise robustness and high-scale stability. Further improvement is needed to fully explore multi-scale information and improve the sensitivity and stability of early fault diagnosis. Summary of the Invention

[0007] To solve the above technical problems, the present invention proposes a fault diagnosis method based on refined time-shift multiscale attention entropy (RTSMAE). By improving the shortcomings of the traditional coarse-grained process, the method fully exploits the complex characteristics of vibration signals at different scales to achieve efficient and stable fault diagnosis of key components.

[0008] The present invention provides a fault diagnosis method based on fine time-shifted multi-scale attention entropy, comprising the following steps:

[0009] Step S1, signal acquisition: Using a high-sensitivity acceleration sensor installed on the rolling bearing, the vibration signals of key components in normal and fault states are acquired. The fault states include outer ring weak fault, rolling element weak fault, inner ring weak fault, inner ring moderate fault, and inner ring severe fault. The outer ring fault locations include the load area, perpendicular to the load area, and far away from the load area. The bearing vibration signal is collected at a sampling frequency of 10-12 kHz. The vibration signal collection process strictly adheres to the standardized collection frequency and data accuracy requirements, and each collected original vibration signal is randomly divided into several independent samples with consistent data length and no overlap, ensuring the independence and representativeness of the samples, which is conducive to the subsequent feature analysis and accurate training and evaluation of the fault diagnosis model.

[0010] Step S2, feature extraction: Extract the fault features of the sample based on refined time-shifted multi-scale attention entropy (RTSMAE), calculate the entropy value of each sample under a certain scale factor, and form a feature vector matrix; the steps include:

[0011] Step S21, fine time shift and coarse graining process: set the scale factor and process the signal in segments by gradually moving the window;

[0012] Step S22, attention entropy calculation: first, perform key pattern extraction, extract local maxima and minima in each sample segment, and record their locations; then calculate the attention entropy value of the coarse-grained time series at each scale based on the statistical extreme value interval distribution;

[0013] Step S23, multi-scale entropy feature fusion: combining the attention entropies at different scales until the scale factor reaches the maximum scale factor, obtaining a fine time-shifted multi-scale attention entropy set, and forming a feature vector matrix;

[0014] Furthermore, the calculation steps of fine time-shifted multi-scale attention entropy are as follows:

[0015] Step S21, fine time shift coarse graining process: for time series , x i is the signal data at the i-th time point, and the scale factor is , the signal is segmented by gradually moving the window, that is, multi-scale coarse-graining, and the process is expressed as:

[0016]

[0017] in, is the scale factor, For scale The coarse-grained sequence, N is the length of the original time series;

[0018] Step S22, attention entropy calculation: First, define the key pattern and determine the extreme points of the signal, including local maximum points and local minimum points; use the extreme points to form a key pattern set, record the key pattern position, and use this as the basis for attention entropy calculation. Suppose the signal is , extract all local extreme values and form a key pattern set as , where each is the position index of the local extreme value in the original sequence.

[0019] For each coarse-grained sequence, calculate the intervals between key patterns and count the frequencies of each interval to obtain the corresponding probability distribution. Based on the statistical laws of key pattern intervals, calculate the attention entropy value of each coarse-grained sequence. , The scale factor is After fine time shift and coarse grain processing sequence, get The entropy value of the coarse-grained sequence is calculated, and the function of attention entropy and scale factor is formed as follows:

[0020]

[0021] Among them, AE （τ） For scale The attention entropy value under For scale Next The probability of occurrence of an interval, τ max is the maximum scale factor;

[0022] Based on the above formula, the fine time-shifted multi-scale attention entropy is defined as RTSMAE(x,τ), and the formula is:

[0023]

[0024] Where, The scale factor is After fine time shift and coarse grain processing A sequence, is the maximum scale factor, is the original time series signal;

[0025] Step S23, multi-scale entropy feature fusion: combine the attention entropies obtained at different scales until Reaching the maximum scale factor ;get Collection As follows:

[0026]

[0027] The obtained multi-scale attention entropy feature vector is .

[0028] Step S3: Fault diagnosis:

[0029] A lightweight classifier is selected to train the extracted multi-scale attention entropy feature vector and the corresponding label as input to build a fault diagnosis model; lightweight classifiers include one or more of K-nearest neighbor (KNN), support vector machine (SVM), extreme learning machine (ELM), logistic regression (SR), BP neural network (BPNN), decision tree (DT), random forest (RF), and naive Bayes (NB);

[0030] The feature vectors in the test set are input into the trained classifier to realize automatic recognition and classification of different health states of key components of rolling bearings.

[0031] Beneficial effects of the present invention:

[0032] The present invention provides a fault diagnosis method based on fine time shift and multi-scale attention entropy. The method collects vibration signals of key components and divides them into non-overlapping data segments of equal length. Fine time shift and multi-scale coarse-graining strategies are then used to construct coarse-grained sequences at different scales, preserving the continuity of the signal as much as possible. Subsequently, local extreme points in each signal segment are identified, a set of key patterns is constructed, and their positions are recorded. The interval distribution between key patterns is statistically analyzed, and the attention entropy at each scale is calculated. The results are then integrated into a complete multi-scale feature to improve the accuracy and stability of fault identification. The parameter-free advantage of attention entropy is combined to enhance the sensitivity to faults. Finally, the multi-scale features are input into a classifier for classification and identification to complete the fault diagnosis task. Compared with traditional methods, the present invention utilizes a fine time-shift sliding method to maintain a fine coarse-graining strategy of sequence length, combined with a hyperparameter-free attention entropy metric, to stably characterize the complexity of bearing vibration signals in multi-scale space, and complete fault diagnosis with a lightweight classifier, which is suitable for application requirements in complex industrial environments such as high-speed trains. The entire processing flow does not require manual setting of embedding dimension, similarity tolerance, or time delay parameters, completely avoiding the tedious operation of repeatedly adjusting parameters for different equipment and working conditions, greatly improving the universality and ease of use of the method. The innovative fine time-shift coarse-graining technology effectively solves the problem of instability of traditional multi-scale entropy in high-scale segments. Automatic diagnosis can be achieved without pre-entering the theoretical fault characteristic frequency, effectively avoiding the problem of misjudgment caused by assembly errors or speed fluctuations. The present invention can be extended to other high-reliability equipment for condition monitoring and fault diagnosis, and has broad application prospects. These advantages make the present invention an efficient, stable, and easy-to-promote fault monitoring solution, providing reliable technical support for predictive maintenance of industrial equipment. Experimental verification shows that the method of the present invention demonstrates superior performance in strong noise scenarios such as high-speed train bogies, with an average classification accuracy of over 99% on bearing datasets, significantly improving the rolling bearing fault diagnosis and identification capabilities, and providing a reliable new solution for condition monitoring and fault diagnosis of key equipment such as rolling bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of the overall process of the fault diagnosis method based on fine time-shifted multi-scale attention entropy of the present invention;

[0034] Figure 2 Schematic diagram of the fine time-shift coarse-graining process of the present invention;

[0035] Figure 3 This is a schematic diagram of the attention entropy calculation method of the present invention;

[0036] Figure 4 This is the signal waveform diagram of different fault types of the CWRU bearing data set of the present invention;

[0037] Figure 5This is a feature visualization diagram of different entropy methods for CWRU dataset samples in a specific implementation manner;

[0038] Figure 6 It is a radar chart of performance indicators of different entropy methods-classifiers on the CWRU dataset in a specific implementation manner;

[0039] Figure 7 This is a diagram showing the classification results of the original data of the CWRU dataset using Naive Bayes in a specific implementation manner;

[0040] Figure 8 The confusion matrix diagram of different entropy methods - Naive Bayes on the CWRU dataset in the specific implementation method is shown in Figure 2.

[0041] Figure 9 This is a performance evaluation index diagram of RTSMAE-NB on the CWRU dataset in a specific implementation method. DETAILED DESCRIPTION

[0042] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings. The embodiments described are only some of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0043] Example 1

[0044] like Figure 1 As shown, this embodiment provides a fault diagnosis method based on fine time-shifted multi-scale attention entropy, including the following steps:

[0045] Step S1, signal acquisition: Acquire vibration signals of rolling bearings in different health states, and divide each signal into 40 independent samples. The length of each sample is 3000, ensuring that there is no overlap between samples to improve data utilization;

[0046] Step S2, feature extraction: RTSMAE is used to directly extract the fault features of the samples. The entropy value of each sample at a scale factor of 1-20 is calculated to form a feature vector matrix. These features can effectively reflect the complexity changes of the signal at different scales.

[0047] The steps for calculating fine time-shifted multi-scale attention entropy are as follows:

[0048] Step S21, performing fine time shift coarse graining process:

[0049] For time series , the signal is segmented by gradually moving the window, that is, multi-scale coarse-graining, and the process is expressed as:

[0050]

[0051] in, is the scale factor, For scale The coarse-grained sequence, N is the length of the original time series;

[0052] The present invention uses a fine time shift method to replace the traditional non-overlapping windowing method, and gradually performs multi-scale coarse-grained division by moving average. This method is used to obtain coarse-grained sequences at different scales, thereby maximally retaining the continuity and important information of the signal. Figure 2 The time series generated by the fine time shift coarse graining method of the present invention is shown. In scale The length of In the traditional multi-scale analysis method, the length of the coarse-grained sequence is (Indicates not greater than The maximum positive integer of ), the data length is much smaller than that of the RTSMAE method.

[0053] Step S22, attention entropy calculation: Figure 3 As shown, first perform key pattern extraction, calibrate local maxima and minima in each sample segment, extract all local extreme values, and obtain: ; Then the attention entropy is calculated by the following formula:

[0054]

[0055] Among them, AE （τ） For scale The attention entropy value under is the probability of occurrence of different intervals.

[0056] Based on the above formula, the fine time-shifted multi-scale attention entropy is defined as RTSMAE (x, τ) formula is:

[0057]

[0058] Where, The scale factor is After fine time shift and coarse grain processing A sequence, is the maximum scale factor, is the original time series signal;

[0059] Step S23, multi-scale entropy fusion: Repeat steps S21 and S22 until Reaching the maximum scale factor ,get Collection As follows:

[0060]

[0061] The obtained multi-scale attention entropy feature vector is .

[0062] Step S3, fault identification: The extracted feature dataset is randomly divided into a training set and a test set in a 1:1 ratio. The training set is used to build the classification model, while the test set is used to evaluate the generalization performance of the model. By inputting the test set into the trained classifier, the health status classification result of the bearing can be obtained, providing a classification basis for subsequent mechanical fault processing.

[0063] Example 2

[0064] Based on Example 1, experimental verification was performed using real rolling bearing signals:

[0065] This example uses a rolling bearing fault test bench at Case Western Reserve University (CWRU) in the United States to acquire vibration signals of rolling bearings under normal and different fault conditions. A 6205-2RSJEMSKF deep groove ball bearing with a pitch diameter of 39.04 mm, a rolling element diameter of 7.94 mm, nine rolling elements, and a contact angle of 0° was selected. The specific acquisition process involves deploying an accelerometer above the bearing seat at the drive end of the test bench to acquire bearing vibration acceleration signals at a sampling frequency of 12 kHz. Fault points of varying diameters were created on the bearing inner and outer rings and rolling elements using electrical discharge machining (EDM) technology to accurately simulate various fault conditions. Because the outer ring is stationary relative to the sensor, the relative position of the fault point to the load zone directly affects the vibration response. Therefore, the outer ring fault location is categorized into three scenarios: 6 o'clock (load zone), 3 o'clock (perpendicular to the load zone), and 12 o'clock (away from the load zone). Each type of fault is electrosparked to produce fault points of different diameters, including four fault levels of 0.007, 0.014, 0.021, and 0.028 inches. The fault level of 0.007 inches is the lowest, which is a weak fault. The larger the fault diameter, the more serious the fault. Different types of vibration signals are selected under the working condition of a bearing speed of 1797 r / min: normal signal, weak fault of the outer ring and weak fault of the rolling element, and faults of the first three diameters of the inner ring. The fault of the outer ring uses the fault signal at the 6 o'clock position. A total of 6 data types with different bearing states are selected to verify the effectiveness of RTSMAE combined with different classifier algorithms. Detailed information on the experimental data samples is shown in Table 1:

[0066] Table 1 CWRU bearing health status dataset

[0067]

[0068] Step S1: Preprocess the experimental data: 10,000 points are captured from each time domain record, and then randomly divided into 40 non-overlapping sample segments of 3,000 points in length. This ensures that there are 40 × 6 = 240 samples for statistics. This setting can reduce the deviation of random impact positions from entropy statistics and also split long segments of data into small units that are easy to calculate quickly. Figure 4 (a) shows the time domain waveform, Figure 4 (b) Corresponding amplitude spectrum.

[0069] Depend on Figure 4 As shown in (a) and (b), when a rolling bearing vibration signal fails, the corresponding time-domain waveform exhibits a relatively obvious impact characteristic. However, it is difficult to accurately distinguish the health status of a rolling bearing based solely on the time-domain and frequency-domain waveforms. Therefore, it is necessary to apply the method of the present invention to extract the fault characteristics of the vibration signal and identify different health states, thereby achieving classification of the rolling bearing fault location or fault severity.

[0070] Step S2, feature extraction: In this embodiment, the maximum scale factor is , RTSMAE is used to directly extract the fault features of the samples, calculate the entropy value of each sample under 1-20 scale factors, and form a feature vector matrix. These features can effectively reflect the complexity changes of the signal under different scales; It is worth noting that in this embodiment 2, when When the length of the coarse-grained sequence is 2981 points, the length of the traditional coarse-grained sequence is only 150 points, with a length advantage of 19.9 times, which significantly improves the stability of high-scale entropy. It proves that the method provided by the present invention takes into account the connection between the data before and after the breakpoint, thereby reducing errors and avoiding the loss of key information.

[0071] This example considers six common multi-scale entropy methods, namely composite multi-scale attention entropy (CMAE), multi-scale attention entropy (MAE), multi-scale attention entropy (MSE), multi-scale permutation entropy (MPE), multi-scale spread entropy (MDE), and multi-scale fuzzy entropy (MFE), as comparison methods to compare with RTSMAE. After using the t-SNE algorithm to reduce the dimensionality of the 20 scale features extracted by multi-scale entropy and the features extracted by single-scale AE, the two-dimensional visualization of the extracted features is shown in the figure below. Figure 5 It should be noted that the clustering ability of the algorithm represents the feature extraction ability of the method: the smaller the intra-class distance between samples in the same cluster, and the larger the inter-class distance between different clusters, the better the feature extraction ability.

[0072] Depend on Figure 5As shown in (a) and (g), RTSMAE and MFE achieve the best feature extraction results, effectively distinguishing the features of different sample types. The features extracted by other methods all suffer from varying degrees of aliasing. The figures also show that compared to traditional entropy methods, RTSMAE, CMAE, and MAE, which extend attention entropy to multi-scale space, generally have better feature extraction capabilities.

[0073] In addition, the solution time of different feature extraction methods is analyzed in Table 2.

[0074] Table 2 Solution time of different entropy methods

[0075]

[0076] As can be seen in Table 2, RTSMAE takes 116.63 seconds to extract features from data of different fault types, while MFE requires a higher time cost, requiring 21,120 seconds, or 5.87 hours, to extract features from the dataset. Although CMAE and MAE, which are similar to RTSMAE, have higher solution efficiency, they do so at the expense of key information in the time series. While CMAE, MPE, and AE exhibit high computational efficiency, their feature extraction performance is inferior to RTSMAE. A comprehensive analysis of the visual feature extraction results and computational efficiency shows that the RTSMAE proposed in this paper is a fast and effective feature extraction tool.

[0077] Step S3, Fault Diagnosis: The extracted features are randomly divided into a training set and a test set in a 1:1 ratio. The features extracted from the training set are input into the classifier to obtain a training model. The test set is input into the trained classifier, and the identification results of different bearing health states are output to provide a classification basis for subsequent mechanical fault processing. In order to test the generalization ability of RTSMAE and avoid the impact of differences in classifier classification performance, this embodiment selects eight classifiers: K-nearest neighbor (KNN), support vector machine (SVM), extreme learning machine (ELM), logistic regression (SR), BP neural network (BPNN), decision tree (DT), random forest (RF), and naive Bayes (NB) to classify the extracted fault features. The classifier parameter settings are shown in Table 3 below.

[0078] Table 3 Classification method and its parameter settings

[0079]

[0080] To objectively compare the fault diagnosis performance of models, six commonly used evaluation metrics are used: Accuracy Rate (ACC), Precision Rate (PRE), Recall Rate (REC), F1 Score (F1), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI). These metrics range from 0 to 1. Generally speaking, higher values indicate better model classification performance. The equations for these six metrics are as follows:

[0081]

[0082]

[0083]

[0084]

[0085]

[0086]

[0087] Among them, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative. is the set of actual labels, is the set of predicted labels. yes and The number of pairs of samples with the same label in , is the number of pairs of samples with different labels in the two sets, are all possible combinations of sample pairs. yes and The joint probability function of and They are and The probability function of .

[0088] The data was partitioned by randomly shuffling the extracted feature dataset each time, resampling it 50 times to reduce algorithmic randomness. 50% of the dataset was used for training, and the remaining 50% for testing. To further verify the performance of RTSMAE in feature extraction, the features extracted by the eight entropy methods and the original data were input into different classifiers for fault pattern recognition. The classification accuracy is shown in Table 4.

[0089] Table 4 Classification results of different classification methods on the features extracted from the CWRU dataset (%)

[0090]

[0091] Note: The average accuracy of each classifier does not include the classification results of the original data

[0092] Table 4 shows that RTSMAE has the highest average classification accuracy, at 99.86%. Furthermore, when combined with different classifiers, RTSMAE exhibits the highest average classification accuracy. This indicates that the separability of the features extracted by RTSMAE is independent of the classifier, demonstrating the superior performance of this method in feature extraction. As shown in the table above, in fault diagnosis of different bearing fault types and fault severity, directly classifying the raw data yields an average accuracy of 44.78%, which is poor. RTSMAE, CMAE, and MAE are all based on multi-scale analysis of attention entropy. Their diagnostic accuracy when combined with classifiers is higher than that of attention entropy in a single scale space, demonstrating the necessity and accuracy of extending AE to multi-scale spaces. The average accuracy of other traditional multi-scale entropy methods, MSE, MPE, MDE, and MFE, is 96.67%, 95.41%, 93.67%, and 99.33%, respectively, all lower than the proposed RTSMAE. Although MFE also achieves high classification accuracy for features extracted, its low computational efficiency limits its application. In addition, although the combination of MAE-KNN, MAE-ELM, MAE-RF and MSE-RF also shows high accuracy on public datasets, the RTSMAE proposed in this invention has the best classification performance when combined with other classifiers except DT, and has good generalization. In practical applications, a feature extraction method with more stable performance should be selected to adapt to different data quality conditions.

[0093] Draw a radar chart of model performance evaluation indicators based on different entropy methods-machine learning classifiers Figure 6 As shown. Figure 6 As can be seen, the evaluation metrics of each method radiate outward from the center, with the curve of the proposed RTSMAE method located at the outermost edge of the radar chart. This means that all performance evaluation metrics are higher than those of the other methods, demonstrating the superior performance of the proposed RTSMAE. Considering both feature extraction and computational power, RTSMAE outperforms the aforementioned entropy-based methods. Therefore, RTSMAE possesses both efficient feature extraction and computational capabilities.

[0094] In addition, it can be seen from Table 4 that among the eight commonly used classifiers selected, the average test accuracy of Naive Bayes is the highest. The Naive Bayes method with the best classification effect is selected to draw the confusion matrix and classification results of the original data classification, as well as the confusion matrix and classification results of the classification of different entropy feature extraction methods. Figure 7 and Figure 8 shown.

[0095] Depend on Figure 7 It can be seen that if feature extraction is not performed on the original data, only samples with labels 0 and 2 can be correctly classified, and the health status of other bearings cannot be identified, indicating that the fault data of the original data are similar, causing classification confusion. Figure 8 (a) It can be seen intuitively that the RTSMAE-NB method achieved 100% recognition accuracy. Figure 8 (b) shows that the CMAE-NB method will cause a small number of samples with label 3 to be misclassified as label 1; Figure 8 (c) shows that the MAE-NB method has labels 0 and 2 misclassified as label 5; Figure 8 (d) shows that the NB-MSE method has classification confusion between label 3 and label 2; Figure 8 (e) shows that the NB-MPE method has classification confusion between label 4 and label 5; Figure 8 (f) shows that NB-MDE exhibits classification confusion between labels 0 and 2; Figure 8 (h) shows that in the classification results of AE, there will be classification confusion between label 1 and label 2, label 3 and label 4, label 5 and label 2. The confusion matrix also shows that RTSMAE can provide sensitive features for the classifier, while the other methods all cause classification confusion due to the overlap between features.

[0096] Draw a histogram of six performance evaluation indicators of the naive Bayes method for classifying features extracted by different entropy methods. Figure 9 shown.

[0097] from Figure 9 As can be seen, the performance index of RTSMAE-NB reached 100% with an error of 0, which shows that the method has high stability and reliability. Compared with other entropy methods, the RTSMAE method proposed in this paper shows significant advantages in fault feature extraction and can more accurately capture key feature information in bearing vibration signals. By combining it with a naive Bayesian classifier, this method can not only accurately distinguish between the healthy and faulty states of bearings, but also effectively identify different fault types (such as inner ring, outer ring, and rolling element faults) and fault severity (such as weak, moderate, and severe faults). This excellent classification performance fully demonstrates the practicality and effectiveness of the RTSMAE method in bearing fault diagnosis.

[0098] In summary, the RTSMAE method demonstrates excellent feature stability and diagnostic accuracy. Compared with six existing typical multiscale entropy techniques, RTSMAE leads the field in feature clustering clarity, fault classification performance, and computational efficiency, providing a fast, reliable, and real-time solution for predictive maintenance.

Claims

1. A fault diagnosis method based on fine time-shifted multi-scale attention entropy, characterized by: The following steps are involved: Step S1, signal acquisition: Using a high-sensitivity acceleration sensor installed on the rolling bearing, obtain vibration signals of key components in normal and fault conditions. Fault conditions include outer ring weak fault, rolling element weak fault, inner ring weak fault, inner ring moderate fault, and inner ring severe fault. Outer ring fault locations include the load area, perpendicular to the load area, and far from the load area. The bearing vibration signal is collected at a sampling frequency of 10-12 kHz. The vibration signal collection process strictly adheres to standardized collection frequency and data accuracy requirements, and each collected original vibration signal is randomly divided into several independent samples with consistent data length and no overlap. Step S2, feature extraction: extracting fault features of samples based on fine time-shift multi-scale attention entropy, calculating the entropy value of each sample under a certain scale factor, and forming a feature vector matrix; the steps include: Step S21, fine time shift and coarse graining process: set the scale factor and process the signal in segments by gradually moving the window; Step S22, attention entropy calculation: first, perform key pattern extraction, extract local maxima and minima in each sample segment, and record their locations; then calculate the attention entropy value of the coarse-grained time series at each scale based on the statistical extreme value interval distribution; Step S23, multi-scale entropy feature fusion: combining the attention entropies at different scales until the scale factor reaches the maximum scale factor, obtaining a fine time-shifted multi-scale attention entropy set, and forming a feature vector matrix; Step S3, fault diagnosis: Select a lightweight classifier, use the extracted multi-scale attention entropy feature vector and the corresponding label as input for training, and build a fault diagnosis model; the lightweight classifier includes one of K-nearest neighbor, support vector machine, extreme learning machine, logistic regression, BP neural network, decision tree, random forest, and naive Bayes; The feature vectors in the test set are input into the trained classifier to realize automatic recognition and classification of different health states of key components of rolling bearings.

2. The fault diagnosis method based on fine time-shifted multi-scale attention entropy according to claim 1 is characterized by: The calculation steps of the fine time-shifted multi-scale attention entropy in step S2 are as follows: Step S21, fine time shift coarse graining process: for time series , x i is the signal data at the i-th time point, and the scale factor is , the signal is segmented by gradually moving the window, that is, multi-scale coarse-graining, and the process is expressed as: ； in, is the scale factor, For scale The coarse-grained sequence, N is the length of the original time series; Step S22, attention entropy calculation: First, define the key pattern and determine the extreme points of the signal, including local maximum points and local minimum points; use the extreme points to form a key pattern set, record the key pattern position, and use this as the basis for attention entropy calculation. Suppose the signal is , extract all local extreme values and form a key pattern set as , where each is the position index of the local extreme value in the original sequence; For each coarse-grained sequence, the intervals between key patterns are calculated, and the frequency of each interval is counted to obtain the corresponding probability distribution; based on the statistical laws of the key pattern intervals, the attention entropy value of each coarse-grained sequence is calculated. , The scale factor is After fine time shift and coarse grain processing sequence, get The entropy value of the coarse-grained sequence is calculated, and the function of attention entropy and scale factor is formed as follows: ； Among them, AE （τ） For scale The attention entropy value under For scale Next The probability of occurrence of an interval, τ max is the maximum scale factor; Based on the above formula, the fine time-shifted multi-scale attention entropy is defined as RTSMAE(x,τ), and the formula is: ； Where, The scale factor is After fine time shift and coarse grain processing A sequence, is the maximum scale factor, is the original time series signal; Step S23, multi-scale entropy feature fusion: combine the attention entropies obtained at different scales until Reaching the maximum scale factor ;get Collection As follows: ； The obtained multi-scale attention entropy feature vector is .

Images