Fault diagnosis method based on fine time-shifting multi-scale attention entropy

Abstract

The invention belongs to the technical field of rolling bearing fault diagnosis, and particularly relates to a fault diagnosis method based on fine time-shifting multi-scale attention entropy. The method comprises the steps of performing coarse granularity segmentation on original data by adopting a fine time-shifting sliding window method, calculating attention entropy values of coarse granularity time sequences under various scales, fusing multi-scale feature vectors as extracted fault features, constructing a feature vector set, and inputting the feature vector set into a classifier for fault diagnosis. The method not only effectively solves the problem that the entropy value of a traditional multi-scale entropy method is unstable under a high scale, but also avoids dependence on hyper-parameters in a feature extraction process. Experimental results show that the average classification accuracy of the fault diagnosis method on a bearing data set exceeds 99%, and the rolling bearing fault diagnosis and recognition capability is remarkably improved.

Description

Technical Field

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

[0002] The key mechanical systems of intelligent equipment have been in complex working conditions of high load and strong interference for a long time. The vibration signals in the early stage of the operation of these key mechanical systems often show abnormal fluctuations with weak amplitudes and short durations. However, these potential fault signals are extremely easy to be masked by non-stationary interference information such as background noise and environmental vibration. As a core component of key equipment of high-speed trains, rolling bearings are widely used due to their compact structure and high transmission efficiency. However, small fluctuations in the working state of the bearings will directly affect the operation safety of the equipment. In engineering practice, it is difficult to extract the characteristics of fault signals under complex working conditions, which brings great challenges to fault diagnosis.

[0003] In terms of fault feature extraction, the traditional methods mainly have the following 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 features are easily affected by changes in working conditions and noise interference; Although time-frequency analysis methods (STFT, CWT, WVD, etc.) can characterize the local instantaneous features of signals, their computational complexity is high and expert experience guidance is required; Although adaptive signal decomposition methods (EMD, VMD, etc.) can process non-stationary signals, their performance is severely dependent on parameter settings and problems such as mode mixing are likely to occur.

[0004] In recent years, fault feature extraction methods based on entropy theory have received extensive attention because they can characterize signal complexity from the perspective of nonlinear dynamics. Common indicators include Sample Entropy (SE), Permutation Entropy (PE), Fuzzy Entropy, and Dispersion Entropy (DE), etc. They measure the complexity of data by means of template matching, position rearrangement, fuzzy membership degree, and probability distribution dispersion, etc., providing effective features for early fault diagnosis. However, these methods usually require presetting hyperparameters such as embedding dimension, tolerance threshold, or number of classifications, etc. Their robustness to different working conditions is limited, and their applicability is poor under strong noise backgrounds, easily leading to entropy value distortion. To avoid hyperparameter intervention, Attention Entropy (AE) is proposed to calculate complexity by replacing the original vibration sequence with the adjacent extreme point interval sequence. Its advantage is that it does not require parameter adjustment and is sensitive to faults. However, since it only evaluates signal features at a single scale, its ability to characterize signals at different scales and different resolutions is insufficient, and it is difficult to completely describe the multi-level features of key components.

[0005] To introduce information at different scales, researchers proposed Multiscale Entropy. This method first coarsens (or reconstructs) the signal and then calculates the entropy value at each scale. However, in the coarsening process, the number of samples decreases sharply with the increase of the scale. Especially in the high-scale interval, significant random fluctuations occur, which easily leads to the instability of the entropy value curve. At the same time, the shortening of the sequence also weakens the sensitivity to short-term instantaneous impacts, reducing the reliability of the 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, and needs to be further improved to fully explore multiscale information, 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 (Refined Time-shift Multiscale Attention Entropy, RTSMAE). By improving the deficiencies in the traditional coarsening process, it fully excavates the complexity characteristics of vibration signals at different scales and realizes efficient and stable fault diagnosis of key components.

[0008] A fault diagnosis method based on refined time-shift multiscale attention entropy provided by the present invention includes the following steps:

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

[0010] Step S2, feature extraction: Extract the fault features of the samples based on refined time-shift multiscale 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, refined time-shift coarsening process: Set the scale factor and segment the signal by gradually moving the window;

[0012] Step S22, Attention Entropy Calculation: First, perform key pattern extraction by extracting local maxima and minima in each sample segment and recording their positions. Then, calculate the attention entropy values of the coarse-grained time series at each scale based on the statistical distribution of the extreme value intervals.

[0013] Step S23, Multi-scale Entropy Feature Fusion: Combine the attention entropies at different scales until the scale factor reaches the maximum scale factor to obtain a fine time-shifted multi-scale attention entropy set, forming a feature vector matrix.

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

[0015] Step S21, Fine Time-shifted Coarse-graining Process: For the time series , x i being the signal data at the i-th time point, let the scale factor be , and perform segmented processing on the signal by gradually moving the window, i.e., multi-scale coarse-graining. The process is expressed as:

[0016]

[0017] where is the scale factor, is the coarse-grained sequence at scale , and N is the length of the original time series.

[0018] Step S22, Attention Entropy Calculation: First, define the key patterns by determining the extreme points of the signal, including local maximum points and local minimum points. Use the extreme points to form a key pattern set and record the positions of the key patterns as the basis for attention entropy calculation. Let the signal be , extract all local extremes to form a key pattern set , where each is the position index in the original sequence corresponding to the local extreme.

[0019] For each coarse-grained sequence, calculate the intervals between key patterns and count the frequencies of each interval occurrence to obtain the corresponding probability distribution. Based on the statistical law of the key pattern intervals, calculate the attention entropy value , indicating the -th sequence after fine time-shifted coarse-graining processing at scale factor . Obtain the entropy values of coarse-grained sequences and form a function of attention entropy and scale factor as follows:

[0020]

[0021] where AE （τ） is the scale The attention entropy value below is the scale The occurrence probability of the -th interval below, τ max is the maximum scale factor;

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

[0023]

[0024] In the formula, represents the -th sequence after the refined time-shifted coarse-graining process when the scale factor is , 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 ; Obtain the set of as follows:

[0026]

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

[0028] Step S3, fault diagnosis:

[0029] Select a lightweight classifier and use the extracted multi-scale attention entropy feature vector and the corresponding label as inputs for training to construct a fault diagnosis model; the lightweight classifier includes one or several 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), Naive Bayes (NB);

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

[0031] Advantages of the present invention:

[0032] A fault diagnosis method based on fine time-shifted multi-scale attention entropy provided by the present invention collects vibration signals of key components and divides them into non-overlapping data segments of equal length; then, a fine time-shifting and multi-scale coarse-graining strategy is adopted to construct coarse-grained sequences at different scales, as far as possible retaining the continuity of the signal; next, local extreme points in each segment of the signal are identified, a set of key patterns is constructed, and their positions are recorded, the interval distribution between key patterns is statistically analyzed, the attention entropy at each scale is calculated, and they are fused into a complete multi-scale feature to improve the accuracy and stability of fault recognition; combined with the parameter-free advantage of attention entropy, the sensitivity to faults is improved; finally, the multi-scale feature is input into a classifier for classification and recognition to complete the fault diagnosis task. Compared with traditional methods, the present invention uses a fine coarse-graining strategy of the fine time-shifting sliding method to maintain the sequence length, combines a parameter-free attention entropy metric, stably characterizes the complexity of bearing vibration signals in a multi-scale space, and completes fault diagnosis with a lightweight classifier, which is suitable for the application requirements of complex industrial environments such as high-speed trains; the entire processing flow does not require manual setting of embedding dimensions, similarity tolerances or time-delay parameters, completely avoiding the cumbersome operation of repeatedly adjusting parameters for different devices and working conditions, and greatly improving the universality and usability of the method; an innovative fine time-shifted coarse-graining technology is adopted to effectively solve the problem of instability of traditional multi-scale entropy in the high-scale segment; automatic diagnosis can be realized without pre-inputting theoretical fault characteristic frequencies, effectively avoiding misjudgment problems caused by assembly errors or speed fluctuations. The present invention can be widely applied to the condition monitoring and fault diagnosis of other high-reliability equipment, and has broad application prospects. These advantages make the present invention an efficient, stable and easy-to-popularize fault monitoring solution, providing reliable technical support for the predictive maintenance of industrial equipment. Through experimental verification, the method of the present invention shows superior performance in strong noise scenarios such as high-speed train bogies, and the average classification accuracy on the bearing dataset exceeds 99%, significantly improving the fault diagnosis and recognition ability of rolling bearings, and providing a reliable new solution for the condition monitoring and fault diagnosis of key equipment such as rolling bearings. Description of the Drawings

[0033] Figure 1 It is a 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 It is a schematic diagram of the fine time-shifted coarse-graining process of the present invention;

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

[0036] Figure 4 It is a waveform diagram of signals of different fault types for the CWRU bearing dataset of the present invention;

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

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

[0039] Figure 7 It is the classification result diagram of the original data of the CWRU dataset by Naive Bayes in the specific implementation;

[0040] Figure 8 It is the confusion matrix diagram of different entropy method - Naive Bayes on the CWRU dataset in the specific implementation

[0041] Figure 9 It is the performance evaluation index diagram of RTSMAE-NB on the CWRU dataset in the specific implementation. Specific implementation

[0042] Next, the technical solutions in the present invention will be further described in conjunction with the accompanying drawings. The described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the protection scope of the present invention.

[0043] Example 1

[0044] As Figure 1 shown, a fault diagnosis method based on refined time-shifted multi-scale attention entropy provided in this embodiment includes the following steps:

[0045] Step S1, signal acquisition: Obtain the vibration signals of the rolling bearing in different health states, and divide each signal into 40 groups of independent samples, each sample having a data length of 3000, ensuring no overlap between samples to improve the utilization rate of data;

[0046] Step S2, feature extraction: Use RTSMAE to directly extract the fault features of the samples, calculate the entropy values 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 at different scales;

[0047] The calculation steps of the refined time-shifted multi-scale attention entropy are as follows:

[0048] Step S21, perform the refined time-shifted coarse-graining process:

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

[0050]

[0051] Among them, is the scale factor, is the scale of the coarse-grained sequence, and N is the length of the original time series;

[0052] The present invention uses the fine time-shift method to replace the traditional non-overlapping windowing method, gradually moving the average multi-scale coarse-grained division, and uses this method to obtain the coarse-grained sequences at different scales, maximizing the retention of the continuity and important information of the signal. Figure 2 shows the fine time-shift coarse-graining process described in the present invention. The time series generated by the fine time-shift coarse-graining method described in the present invention at the scale of has a length of . In the traditional multi-scale analysis method, the length of the coarse-grained sequence is (representing the largest positive integer not greater than ), and the data length is much smaller than the RTSMAE method.

[0053] Step S22, attention entropy calculation: As Figure 3 shown, first perform key pattern extraction, and mark the local maxima and minima in each segment of the sample; extract all local extrema to obtain: ; then calculate the attention entropy by the following formula:

[0054]

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

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

[0057]

[0058] In the formula, represents the th sequence after the fine time-shift coarse-graining process when the scale factor is , is the maximum scale factor, is the original time series signal;

[0059] Step S23, multi-scale entropy fusion: Continuously repeat steps S21 and S22 until reaches the maximum scale factor , and obtain the set 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 construct a classification model, and the test set is used to evaluate the generalization performance of the model. By inputting the test set into the trained classifier, the classification result of the bearing health state can be obtained, providing a classification basis for subsequent mechanical fault handling.

[0063] Embodiment 2

[0064] On the basis of Embodiment 1, experimental verification is carried out through real rolling bearing signals:

[0065] In this embodiment, a rolling bearing fault test bench of Case Western Reserve University (CWRU) in the United States is used to obtain vibration signals of rolling bearings in normal and different fault states. A 6205-2RSJEMSKF deep groove ball bearing is selected, with a pitch diameter of 39.04 mm, a rolling element diameter of 7.94 mm, 9 rolling elements, and a contact angle of 0°. The specific acquisition process is as follows: An acceleration sensor is deployed above the bearing housing at the drive end of the test bench, and the bearing vibration acceleration signal is acquired at a sampling frequency of 12 kHz. Different diameter fault points are manufactured on the inner ring, outer ring, and rolling elements of the bearing through electrical discharge machining technology to accurately simulate various fault degree states. Since the outer ring is stationary relative to the sensor, the relative position of the fault point and the load area will directly affect the vibration response. Therefore, the outer ring fault position is divided into three cases: 6 o'clock direction (load area), 3 o'clock (perpendicular to the load area), and 12 o'clock direction (away from the load area). For each type of fault, different diameter fault points are machined through electrical discharge machining, including four fault degrees of 0.007, 0.014, 0.021, and 0.028 inches. Among them, the fault degree with a diameter of 0.007 inches is the lowest and belongs to a weak fault. The larger the fault diameter, the more serious the fault. Different types of vibration signals under the condition of a bearing speed of 1797 r / min are selected: normal signal, weak outer ring fault and weak rolling element fault, and the first three diameters of inner ring faults. Among them, the outer ring fault uses the fault signal in the 6 o'clock direction. A total of 6 types of bearing different state data are selected to verify the effectiveness of RTSMAE combined with different classifier algorithms. The detailed information of the experimental data samples is shown in Table 1:

[0066] Table 1 CWRU Bearing Health State Dataset

[0067]

[0068] Step S1. Preprocess the experimental data: Each time-domain record is intercepted at 10,000 points, and then randomly split into 40 non-overlapping sample segments with a length of 3,000 points, so as to ensure that there are 40×6 = 240 samples for statistics; this setting can reduce the deviation of the entropy statistics caused by the position of random shocks, and also split the long-segment data into small units convenient for quick calculation. Figure 4 (a) shows the time-domain waveform, Figure 4 (b) corresponds to the amplitude spectrum.

[0069] From Figure 4 of (a) and (b), it can be seen that when the rolling bearing vibration signal fails, the corresponding time-domain waveform shows obvious impact characteristics. However, it is difficult to accurately distinguish the health state of the rolling bearing only based on the time-domain waveform and the frequency-domain waveform. 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, and then realize the classification of the fault position or fault degree of the rolling bearing.

[0070] Step S2. Feature extraction: In this embodiment, the maximum scale factor is taken as , and the RTSMAE is used to directly extract the fault characteristics of the samples, calculate the entropy values of each sample at 1-20 scale factors, and form a feature vector matrix. These features can effectively reflect the complexity change of the signal at different scales; it should be noted that in this embodiment 2, when , the length of the coarsened sequence still remains 2,981 points, while the traditional coarsened sequence number only contains 150 points, and the length advantage reaches 19.9 times, significantly improving the stability of the high-scale entropy, proving that the method provided by the present invention considers the connection of the data before and after the breakpoint, thereby reducing the error and avoiding the loss of key information.

[0071] This embodiment considers using 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 dispersion entropy (MDE), and multi-scale fuzzy entropy (MFE), as comparison methods to compare with RTSMAE. After using the t-SNE algorithm to reduce the dimension of the features extracted at 20 scales by multi-scale entropy and the features extracted by single-scale AE, the two-dimensional visualization diagram of the extracted features is as Figure 5 shown. 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 within the same cluster and the larger the inter-class distance between different clusters, the better the feature extraction ability.

[0072] From Figure 5As can be seen from (a) and (g), RTSMAE and MFE have the best feature extraction effect and can well distinguish the features of different types of samples. There are varying degrees of aliasing in the features extracted by other methods. It can also be seen from the figure that compared with the traditional entropy method, the feature extraction capabilities of the RTSMAE, CMAE, and MAE methods extended to the multi-scale space based on the attention entropy are generally better.

[0073] In addition, the solution times of different feature extraction methods were also analyzed, as shown in Table 2.

[0074] Table 2 Solution times of different entropy methods

[0075]

[0076] As can be seen from Table 2, the time required for RTSMAE to extract the features of data of different fault types is 116.63 s, while MFE requires a relatively high time cost, and it takes 21120 s, that is, 5.87 h, to extract features from the dataset. Although CMAE and MAE of the same type as RTSMAE have higher solution efficiency, they do so at the cost of sacrificing the key information of the time series. Although CMAE, MPE, and AE show relatively high computational efficiency, their feature extraction effects are not as good as RTSMAE. Considering the comprehensive visualization feature extraction results and computational efficiency, the RTSMAE proposed in the present invention is a fast and effective feature extraction tool.

[0077] Step S3, Fault diagnosis: Randomly divide the extracted features into a training set and a test set in a 1:1 ratio. Input the features extracted from the training set into the classifier to obtain a training model, and input the test set into the trained classifier to output the recognition results of different health states of the bearing, providing a classification basis for subsequent mechanical fault handling. In order to test the generalization ability of RTSMAE and avoid the influence caused by differences in the classification performance of the classifier, in this embodiment, 8 classifiers, namely 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), are selected to classify the extracted fault features. The parameter settings of the classifier are shown in Table 3 below;

[0078] Table 3 Classification methods and their parameter settings

[0079]

[0080] To objectively compare the model fault diagnosis performance, six commonly used evaluation metrics are adopted, namely Accuracy Rate (ACC), Precision Rate (PRE), Recall Rate (REC), F1 Score (F1), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI). Their value ranges are all from 0 to 1. Generally speaking, the higher the value, the better the classification performance of the model. The equations of the 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. Let be the set of actual labels, be the set of predicted labels. is and the number of sample pairs with the same label in is the number of sample pairs with different labels in the two sets, is all possible combinations of sample pairs. is and the joint probability function of and are respectively and the probability functions of

[0088] The data division is set to randomly shuffle the extracted feature dataset each time, and randomly resample 50 times to reduce the algorithm contingency. 50% is used for the training set, and the remaining 50% is used for the test set. To further verify the performance of RTSMAE in feature extraction, the features extracted by the above eight entropy methods and the original data are input into different classifiers for fault mode recognition. The classification accuracy is shown in Table 4;

[0089] Table 4 Classification Results (%) of Features Extracted from CWRU Dataset by Different Classification Methods

[0090]

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

[0092] As can be seen from Table 4, the average classification accuracy of RTSMAE is the highest, at 99.86%. In addition, when combined with different classifiers, RTSMAE shows the highest average classification accuracy, indicating that the separability of the features extracted by RTSMAE does not depend on the classifier, proving the superior performance of this method in feature extraction. As can be seen from the above table, in the fault diagnosis of different fault types and different fault degrees of bearings, if the original data is directly classified, the average accuracy is 44.78%, and the classification effect is poor. The three methods of RTSMAE, CMAE, and MAE are all based on the multi-scale analysis of attention entropy, and the diagnostic accuracy combined with classifiers is higher than that of the attention entropy in a single-scale space, indicating the necessity and accuracy of extending AE to a multi-scale space. The average accuracies of the remaining traditional multi-scale entropy methods MSE, MPE, MDE, and MFE are 96.67%, 95.41%, 93.67%, and 99.33% respectively, all lower than the proposed RTSMAE. Although MFE also has a high classification accuracy for the features extracted, the low computational efficiency of MFE limits its application. In addition, although the combinations of MAE-KNN, MAE-ELM, MAE-RF, and MSE-RF also show high accuracies on the public dataset, the proposed RTSMAE in this invention has the best classification performance when combined with other classifiers except DT, has good generalization, and a feature extraction method with more stable performance should be selected in practical applications to adapt to different data quality situations.

[0093] The radar chart of the model performance evaluation indicators based on different entropy methods - machine learning classifiers is plotted as Figure 6 shown. As can be seen from Figure 6 it, the evaluation indicators of each method radiate outwards from the center point. The curve of the proposed RTSMAE method is located in the outermost periphery of the radar chart, that is, each performance evaluation indicator is higher than other methods, indicating the superiority of the performance of the proposed RTSMAE. Considering the feature extraction ability and computational ability comprehensively, RTSMAE performs well among the above entropy theory-based methods. Therefore, RTSMAE has both high-efficient feature extraction ability and high-efficient computational ability.

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

[0095] From Figure 7 it can be seen that if no feature extraction is performed on the original data, only the samples with label 0 and label 2 can be correctly classified, and the remaining bearing health states cannot be recognized, indicating that the similarity of the fault data in the original data causes classification confusion. Figure 8 (a) It can be intuitively seen that the RTSMAE-NB method achieves a recognition accuracy of 100%. Figure 8 (b) shows that the CMAE-NB method will misclassify a small number of samples with label 3 as label 1; Figure 8 (c) shows that the MAE-NB method misclassifies label 0 and label 2 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 the NB-MDE has classification confusion between label 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 remaining methods all lead to classification confusion due to the overlap of features.

[0096] The bar charts of six performance evaluation indicators for classifying the features extracted by different entropy methods using the Naive Bayes method are as Figure 9 shown.

[0097] From Figure 9 it can be seen that the performance index of RTSMAE-NB reaches 100% and the error is 0. This result indicates that the method has high stability and reliability. Compared with other entropy methods, the RTSMAE method proposed in the present invention shows significant advantages in fault feature extraction and can capture key feature information in bearing vibration signals more accurately. By combining with the Naive Bayes classifier, this method can not only accurately distinguish the healthy state and fault state of bearings, but also effectively identify different fault types (such as inner race, outer race and rolling element faults) and the severity of faults (such as weak, moderate and severe faults). This excellent classification performance fully proves the practicality and effectiveness of the RTSMAE method in bearing fault diagnosis.

[0098] In summary, the RTSMAE of the present invention exhibits excellent feature stability and diagnostic accuracy. Compared with the existing six typical multi-scale entropy techniques, RTSMAE leads comprehensively in terms of feature clustering clarity, fault classification performance, and operation efficiency, providing a new technical solution for predictive maintenance that is fast, reliable, and can be implemented in real time.

Claims

1. A fault diagnosis method based on fine time-shifted multi-scale attention entropy, characterized in that: The method includes the following steps: Step S1, signal acquisition: Use highly sensitive acceleration sensors installed on rolling bearings to obtain vibration signals of key components in normal and fault states. The fault states include outer ring weak faults, rolling element weak faults, inner ring weak faults, inner ring medium faults, and inner ring severe faults. The outer ring fault positions include the load area, perpendicular to the load area, and away from the load area. Collect bearing vibration signals at a sampling frequency of 10 - 12 kHz. The vibration signal acquisition process strictly follows standardized acquisition frequency and data accuracy requirements, and randomly divides each collected original vibration signal into several independent samples with the same data length and no overlap with each other; Step S2, feature extraction: Extract fault features of the samples based on refined time-shifted multi-scale attention entropy, calculate the entropy value of each sample under a certain scale factor, and form a feature vector matrix. The steps include: Step S21, refined time-shifted coarse-graining process: Set the scale factor and segment the signal by gradually moving the window; Step S22, attention entropy calculation: First, perform key pattern extraction, extract local maximum and minimum values in each segment of the sample and record their positions; then calculate the attention entropy value of the coarse-grained time series at each scale according to the statistical extreme value interval distribution; Step S23, multi-scale entropy feature fusion: Combine the attention entropies at different scales until the scale factor reaches the maximum scale factor to obtain a refined time-shifted multi-scale attention entropy set and form a feature vector matrix; Step S3, fault diagnosis: Select a lightweight classifier, use the extracted multi-scale attention entropy feature vectors and their corresponding labels as inputs for training to construct 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; Input the feature vectors in the test set into the trained classifier to achieve automatic identification 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, characterized in that: The calculation steps of the refined time-shifted multi-scale attention entropy in Step S2 are as follows: Step S21, fine time-shifted coarse-graining process: For the time series , x i is the signal data at the i-th time point. Let the scale factor be . The signal is segmented by gradually moving the window for multi-scale coarse-graining. The process is expressed as: ； wherein, is a scale factor, is the scale coarse-grained sequence, and N is the length of the original time series; Step S22, attention entropy calculation: First, define the key patterns, 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 positions of the key patterns, and use this as the basis for attention entropy calculation. Let the signal be , extract all local extremes, and form the key pattern set as , where each is the position index in the original sequence corresponding to the local extreme; For each coarse-grained sequence, calculate the intervals between key patterns, count the frequencies of each interval occurrence, and obtain the corresponding probability distribution; based on the statistical law of key pattern intervals, calculate the attention entropy value of each coarse-grained sequence , denotes the th sequence after fine time-shifted coarse-graining when the scale factor is , and obtain the entropy values of the coarse-grained sequences, and form the function of attention entropy and scale factor as follows: ； Among them, AE （τ） is the attention entropy value at scale is the occurrence probability of the th interval at scale, τ max is the maximum scale factor; From the above formula, define the refined time-shifted multi-scale attention entropy as RTSMAE(x,τ), and the formula is: ； wherein, represents the -th sequence after fine time-shifted coarse graining when the scaling factor is ; is the maximum scaling 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 ; obtain a set of as follows: ； The obtained multi-scale attention entropy feature vector is .

Images