Equipment health monitoring method and system based on time-frequency domain characteristics and clustering analysis

Abstract

The invention discloses an equipment health monitoring method and system based on time-frequency domain characteristics and clustering analysis, and relates to the technical field of equipment health monitoring, and the method comprises the following steps: collecting torque data of equipment, and cleaning the torque data to obtain a data set Dataset; carrying out linear detrending processing on the Dataset, recording position indexes of a wave crest and a wave trough, merging the position indexes of the wave crest and the wave trough, arranging the position indexes of the wave crest and the wave trough in sequence to obtain a wave crest position index and a wave trough position index of complete waveform data, and reserving the waveform data of which the distance between the wave crest and the wave trough is greater than a set value to obtain a data set Dataset1; the method comprises the following steps of: extracting time-frequency domain characteristics of Dataset1 to obtain a data set Dataset2, normalizing and clustering the Dataset2, and calculating an outlier and an alarm threshold according to a large class and a small class formed after clustering; according to the equipment health monitoring method and system, the monitoring accuracy and robustness are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment health monitoring, and in particular to an equipment health monitoring method and system based on time-frequency domain features and cluster analysis. Background Art

[0002] In modern handling workshops, handling robots play a vital role. As one of the core equipment of the automated logistics system, the health of the robot is directly related to the efficiency and safety of workshop operations. Due to the frequent operation of the robot, its complex structure and the easy wear of key components such as drive wheels and sensors, the equipment may fail. The abnormal operation and maintenance of the equipment seriously affect the stable operation of the handling workshop.

[0003] The above-mentioned health monitoring process for robots is still in the manual stage. Operators perform regular inspections on robots or repair faults when a robot has already failed. Manual inspections are time-consuming and inefficient. Regular inspections may not be able to detect potential problems in a timely manner. Manual recording of equipment status and fault information may not be accurate or comprehensive enough, which will affect the accuracy of long-term data analysis, trend prediction, and maintenance decisions. Problems can only be discovered when they have developed into failures, resulting in insufficient preventive maintenance measures.

[0004] CN115774652A discloses a health monitoring method, device and medium for group control equipment based on clustering algorithm. The characteristic parameters of the equipment are collected, and the clustering algorithm is used to analyze the sample points in the cluster, and the health curves of normal equipment and abnormal equipment data are generated. Health monitoring is performed according to the health change curve. Since the judgment of normal equipment and abnormal equipment requires original data labels, and the collection of fault data is usually difficult, this method is not conducive to engineering applications. In addition, when taking the mean of the cluster to generate a health curve to monitor the health status of the equipment, some key data features and details may be lost in the process of clustering the data and calculating the mean. The working conditions and performance of each device are unique, and the simplified cluster mean may mask the actual health status of individual devices.

[0005] CN117848704A provides a fault detection method based on signal time-frequency feature hypothesis testing, the extracted feature data is clustered using a clustering algorithm, Gaussian distribution fitting is used for different clusters, hypothesis testing is performed based on whether the distribution of different groups of data is consistent, and signal fault judgment is performed. Since in practical applications, the data may not completely follow the Gaussian distribution, especially in the presence of outliers or biased data distribution, this may lead to inaccurate model fitting. In the scenario of real-time monitoring of signal faults, GMM (Gaussian mixture model) may require more calculation time, which may affect the timeliness of fault detection. Compared with simple clustering algorithms such as K-means, GMM is usually more complex in parameter estimation and model training, and the amount of calculation is also larger. And when performing hypothesis testing, there are usually some prerequisites, such as independence, normality, etc. If these conditions are not met, the test results may be unreliable. Summary of the invention

[0006] Based on the technical problems existing in the background technology, the present invention proposes an equipment health monitoring method and system based on time-frequency domain characteristics and cluster analysis, which improves the accuracy and robustness of monitoring.

[0007] The device health monitoring method based on time-frequency domain features and cluster analysis proposed in the present invention comprises the following steps:

[0008] Step 1: Data collection: Collect the torque data of the equipment and clean it to obtain the data set Dataset;

[0009] Step 2, data labeling: linearly detrend the dataset Dataset, use multiple moving windows to traverse the detrended dataset Dataset0, and record the local maximum of the samples in each moving window; by comparing the sum of the samples under each moving window, take the sample and the minimum moving window length as the optimal moving window, and traverse the detrended dataset Dataset0 with the optimal moving window. If a sample point is confirmed as a local maximum when scanning different moving windows, mark the sample point as a peak and obtain the position index of the peak;

[0010] Step 3: Data labeling and denoising: Remove negative values ​​from all samples in the dataset Dataset0, repeat step 2 to obtain the position index of the trough, merge the position index of the peak and the position index of the trough and arrange them in order to obtain the peak position index and the trough position index of a complete waveform data, retain the waveform data with the peak-to-trough distance greater than the set value to remove the noise data, and obtain the dataset Dataset1;

[0011] Step 4: Data conversion and clustering: Extract the time-frequency domain features of the data set Dataset1, convert the single feature data of the device into multidimensional feature data, obtain the data set Dataset2, normalize the data set Dataset2, set the number of iterations and the number of classes, cluster the normalized data, calculate the outliers and alarm thresholds based on the large and small classes formed after clustering, and generate alarm information when the feature data of the device is greater than the alarm threshold.

[0012] Furthermore, in step 2, the number of the multiple moving windows is s / 2, and the lengths of the multiple moving windows are 5n, n=1, 2, ...s / 2, respectively, where s is the length of the data set Dataset, and n is a parameter.

[0013] Furthermore, in step three, the waveform data with a peak-to-trough distance greater than 5 is retained to remove noise data.

[0014] Furthermore, in step 4, the number of iterations and the number of classes are set, the normalized data is clustered, and the outliers and alarm thresholds are calculated based on the large and small classes formed after clustering, specifically:

[0015] Set the number of classes to 2 and the number of iterations, perform Kmeans clustering on the normalized data, and obtain two types of data;

[0016] According to the number of samples in the class, it is divided into large class Dataset20 and small class Dataset21. The number of samples in the large class n1 is greater than the number of samples in the small class n2;

[0017] According to the large and small categories formed after clustering, outliers and alarm thresholds are calculated.

[0018] Furthermore, in calculating the outliers and alarm thresholds according to the large and small categories formed after clustering, specifically:

[0019] (a1) If the parameter m = n1 / n2, m ≥ 4, then calculate the center distance d from the sample in the small class Dataset21 to the sample in the large class Dataset20 i , get the center distance set D, calculate the mean of the center distance set D and variance δ;

[0020] (a2) If Then the jth sample in the subclass is an outlier, and the outliers of the first subclass are screened out;

[0021] (a3) Calculate the distance between each outlier in the first small class and the class center u of the large class Dataset20 and take the average to get the average value The average As the alarm threshold ζ1.

[0022] Furthermore, in calculating the outliers and alarm thresholds according to the large and small categories formed after clustering, specifically:

[0023] If the parameter m=n1 / n2, m<4, then calculate the upper quartile Q1 and lower quartile Q3 of the distance from each sample of the large category Dataset20 to the class center of Dataset20, and determine the interquartile range IQR=Q3-Q1;

[0024] The points whose distance from each sample of the large category Dataset20 to the class center of Dataset20 is less than (Q1-1.5×IQR) or greater than (Q3+1.5×IQR) are determined as large category outliers;

[0025] Based on steps (a1) to (a2), outliers in subclass Dataset21 are calculated to obtain outliers of the second subclass;

[0026] Aggregate the outliers of the large category and the outliers of the second small category, and calculate the distance of each outlier in the aggregated outliers from the class center u of the large category Dataset20 and average them to get the average value , the average value As the alarm threshold ζ2.

[0028] An equipment health monitoring system based on time-frequency domain features and cluster analysis, including a data acquisition module, a data marking module, a data marking and denoising module, and a data conversion and clustering module;

[0029] The data acquisition module is used to collect the torque data of the equipment and clean it to obtain the data set Dataset;

[0030] The data labeling module is used to perform linear detrending on the dataset Dataset. It uses multiple moving windows to traverse the detrended dataset Dataset0 and records the local maximum of the samples in each moving window. By comparing the sum of the samples under each moving window, the sample and the minimum moving window length are used as the optimal moving window. The detrended dataset Dataset is traversed with the optimal moving window. If a sample point is confirmed as a local maximum when scanned by different moving windows, the sample point is marked as a peak and the position index of the peak is obtained.

[0031] The data labeling and denoising module is used to remove negative values ​​from the samples in the data set Dataset, repeat step 2 to obtain the position index of the trough, merge the position index of the peak and the position index of the trough and arrange them in order to obtain the peak position index and the trough position index of a complete waveform data, retain the waveform data with the peak-to-trough distance greater than the set value to remove the noise data, and obtain the data set Dataset1;

[0032] The data conversion and clustering module is used to extract the frequency domain features of the data set Dataset1, convert the single feature data of the device into multidimensional feature data, obtain the data set Dataset2, normalize the data set Dataset2, set the number of iterations and the number of classes, cluster the normalized data, calculate the outliers and alarm thresholds based on the large and small classes formed after clustering, and generate alarm information when the feature data of the device is greater than the alarm threshold.

[0033] Furthermore, in the data conversion and clustering module, the number of iterations and the number of classes are set, the normalized data is clustered, and the outliers and alarm thresholds are calculated based on the large and small classes formed after clustering, specifically:

[0034] Set the number of classes to 2 and the number of iterations, perform Kmeans clustering on the normalized data, and obtain two types of data;

[0035] According to the number of samples in the class, it is divided into large class Dataset20 and small class Dataset21. The number of samples in the large class n1 is greater than the number of samples in the small class n2;

[0036] According to the large and small categories formed after clustering, outliers and alarm thresholds are calculated.

[0037] Furthermore, in calculating the outliers and alarm thresholds according to the large and small categories formed after clustering, specifically:

[0038] (a1) If the parameter m = n1 / n2, m ≥ 4, then calculate the center distance d from the sample in the small class Dataset21 to the sample in the large class Dataset20 i , get the center distance set D, calculate the mean of the center distance set D and variance δ;

[0039] (a2) If Then the jth sample in the subclass is an outlier, and the outliers of the first subclass are screened out;

[0040] (a3) Calculate the distance between each outlier in the first small class and the class center u of the large class Dataset20 and take the average to get the average value The average As the alarm threshold ζ1;

[0041] (a4) If the parameter m = n1 / n2, m<4, then calculate the upper quartile Q1 and lower quartile Q3 of the distance from each sample of the large class Dataset20 to the class center of Dataset20, and determine the interquartile range IQR = Q3-Q1;

[0042] (a5) The points whose distance from each sample of the large category Dataset20 to the class center of Dataset20 is less than (Q1-1.5×IQR) or greater than (Q3+1.5×IQR) are determined as large category outliers;

[0043] (a6) calculating the outliers in the subclass Dataset21 based on steps (a1) to (a2) to obtain the outliers of the second subclass;

[0044] (a7) Aggregate the outliers of the large category and the outliers of the second small category, and calculate the distance of each outlier in the aggregated outliers from the class center u of the large category Dataset20 and take the average to get the average value The average As the alarm threshold ζ2.

[0045] The advantages of the equipment health monitoring method and system based on time-frequency domain features and cluster analysis provided by the present invention are: the equipment health monitoring method and system based on time-frequency domain features and cluster analysis provided in the structure of the present invention first obtain the torque data of the equipment from the equipment motor, clean the data, use the AMPD method to extract the cleaned valid data, and remove the noise data; then extract the time-frequency domain features of the data, and convert the single torque of the equipment into multidimensional features; finally, use the Kmeans clustering method to divide the data into two categories: large categories and small categories, calculate the abnormal scores of the small category data, determine the outliers according to the abnormal scores, and monitor the health status of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the process of the present invention;

[0047] Figure 2 It is a flow chart of the equipment health monitoring method;

[0048] Figure 3 A schematic diagram of the position index marking of the wave crest;

[0049] Figure 4 A schematic diagram of the position index marking of the trough;

[0050] Figure 5 It is the maximum value diagram of the characteristic data in the time-frequency domain;

[0051] Figure 6 It is the minimum value graph of the feature data in the time-frequency domain;

[0052] Figure 7 It is the mean value diagram of the feature data in the time-frequency domain;

[0053] Figure 8 It is the peak-to-peak diagram of the feature data in the time-frequency domain;

[0054] Fig. 9It is the RMS value diagram of the feature data in the time-frequency domain;

[0055] Fig.10 It is the centroid frequency diagram of the feature data in the time-frequency domain;

[0056] Fig.11 It is the average frequency diagram of the feature data in the time-frequency domain;

[0057] Fig.12 It is the frequency variance diagram of the feature data in the time-frequency domain. DETAILED DESCRIPTION

[0058] Below, the technical solution of the present invention is described in detail through specific embodiments. Many specific details are set forth in the following description to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific implementation disclosed below.

[0059] like Figures 1 to 12 As shown, the device health monitoring method based on time-frequency domain features and cluster analysis proposed in the present invention includes the following steps:

[0060] Step 1: Data collection: Collect the torque data of the equipment and clean it to obtain the data set Dataset;

[0061] The torque data is collected from the motor of the equipment, cleaned, and the missing values ​​are filled to obtain the data set Dataset.

[0062] Step 2: Data labeling: Perform linear detrending on the dataset Dataset, use multiple moving windows to traverse the detrended dataset Dataset0, and record the local maximum of the samples in each moving window; compare the sum of the samples under each moving window, take the sample and the minimum moving window length as the optimal moving window, and use the optimal moving window to traverse the detrended dataset Dataset0. If a sample point is confirmed as a local maximum when scanning different moving windows, mark the sample point as a peak, and get the position index of the peak, such as Figure 3 As shown;

[0063] The data set Dataset0 is first linearly detrended, where the length of Dataset0 is s, and then multiple moving windows (window length is 5n, n=1, 2, ... s / 2) are used to determine the local maximum of each sample (that is, larger than other elements in the current moving window) and record all possible peak positions.

[0064] If a sample point is confirmed as a local maximum when scanning in different moving windows, it means that a sample point corresponds to a local maximum when scanning in some of the previously set moving windows and the best moving window. This means that the sample point is a local maximum in the set composed of different samples, and the sample point is marked as a peak.

[0065] Step 3: Data labeling and denoising: negative values ​​are taken for all samples in the dataset Dataset0, and step 2 is repeated to obtain the position index of the trough. The position index of the peak and the position index of the trough are combined and arranged in order to obtain the peak position index and the trough position index of a complete waveform data. The waveform data with a peak-to-trough distance greater than a set value (for example, 5, where the number of samples per second is a multiple of 5) is retained to remove the noise data and obtain the dataset Dataset1.

[0066] Step 4: Data conversion and clustering: Extract the time-frequency domain features of the dataset Dataset1 and convert the single feature data of the device into multi-dimensional feature data, such as Figures 5 to 12 As shown in the figure, the maximum value, minimum value, mean value, peak-to-peak value, root mean square value, centroid frequency, average frequency, and frequency variance of the time-frequency domain feature data are calculated to obtain the data set Dataset2, normalize the data set Dataset2, set the number of iterations and the number of classes, cluster the normalized data, calculate the outliers and alarm thresholds based on the large and small classes formed after clustering, and generate alarm information when the feature data of the device is greater than the alarm threshold.

[0067] Among them, the number of iterations and the number of classes are set, the normalized data is clustered, and the outliers and alarm thresholds are calculated based on the large and small classes formed after clustering. Specifically: the number of classes is set to 2 and the number of iterations, and Kmeans clustering is performed on the normalized data to obtain two types of data; according to the number of class samples, it is divided into large class Dataset20 and small class Dataset21, and the number of large class samples n1 is greater than the number of small class samples n2; based on the large and small classes formed after clustering, the outliers and alarm thresholds are calculated.

[0068] According to the large and small categories formed after clustering, the outliers and alarm thresholds are calculated as (a1) to (a7):

[0069] (a1) If the parameter m = n1 / n2, m ≥ 4, then calculate the center distance d from the sample in the small class Dataset21 to the sample in the large class Dataset20 i , get the center distance set Calculate the mean of the center distance set D and variance δ;

[0070] in, where xi is the sample in the small class Dataset21, and u is the class center of the large class Dataset20;

[0071] (a2) If Then the jth sample in the subclass is an outlier, and the first subclass outliers {y1,y2,...,y k},1≤k≤n2, where y k is the kth outlier in the first small category of outliers;

[0072] (a3) Calculate the distance between each outlier in the first small class and the class center u of the large class Dataset20 and take the average to get the average value The average As the alarm threshold ζ1;

[0073] (a4) If the parameter m = n1 / n2, m<4, then calculate the upper quartile Q1 and lower quartile Q3 of the distance from each sample of the large class Dataset20 to the class center of Dataset20, and determine the interquartile range IQR = Q3-Q1;

[0074] (a5) The points whose distance from each sample of the large class Dataset20 to the class center of Dataset20 is less than (Q1-1.5×IQR) or greater than (Q3+1.5×IQR) are determined as large class outliers {t1,t2,...,t p},1≤p≤n1,t p is the pth outlier point in the large category of outliers;

[0075] (a6) Based on steps (a1) to (a2), calculate the outliers in the small class Dataset21 and obtain the second small class outliers {y1′, y2′, ..., y q ′},1≤q≤n2,y q ′ is the qth outlier point in the second small category of outliers;

[0076] (a7) Aggregate the large-category outliers and the second-category outliers, and calculate the aggregated outliers {t1, t2, ..., y q ' -1 ,y q The distances of each outlier in ′} from the class center u of the large class Dataset20 are averaged to get the average value The average As the alarm threshold ζ2.

[0077] Through steps one to four, this embodiment uses a device health monitoring method that uses time-frequency domain features and cluster analysis to more comprehensively capture device status information and identify minor changes and trends, which helps to discover potential problems in advance, thereby implementing preventive maintenance and avoiding serious equipment failures. This improves the accuracy of fault detection, reduces false alarms caused by noise or temporary state fluctuations, and improves the stability and reliability of the system. It avoids unnecessary maintenance work, reduces the labor burden, and saves costs and time.

[0078] In addition, this embodiment takes into account the impact of noise data on health monitoring and the impact of extreme abnormal data on overall performance when monitoring the health status of equipment, thereby improving the accuracy and robustness of monitoring.

[0079] Moreover, compared with the patent CN115774652A, the solution of this embodiment is different from that of the patent CN115774652A. From the data point of view, CN115774652A needs to collect the health and fault data of the data, while this embodiment adopts an unsupervised clustering algorithm, which does not need to accumulate fault data and is more conducive to engineering applications. CN115774652A adopts a single feature data, while this embodiment adopts multi-dimensional data in the time-frequency domain, which can reduce the deviation of a single data source and provide more comprehensive equipment health status information. From a methodological point of view, CN115774652A uses cluster mean to generate a health curve to monitor the equipment status, which will conceal the actual health status of individual equipment. The clustering of this embodiment distinguishes between major and minor categories of data, and determines the actual health status after calculating the data anomaly score. Considering the impact of abnormal data on the whole, it is more conducive to examining the actual health status of individual equipment.

[0080] Compared with CN117848704A, this embodiment adopts the Kmeans algorithm, which does not require parameter estimation, is simpler in model training, and requires less calculation. The clustering of this embodiment distinguishes between large and small categories of data, and determines the actual health status after calculating the data abnormality score. Compared with CN117848704A, this embodiment does not require the requirements of data independence and normality, and the monitoring timeliness is higher, the monitoring process is simpler, and the monitoring results are more reliable. In addition, both CN115774652A and CN115774652A do not pre-process the data, and the noise data of the data may have a greater impact on the final result. The data of this embodiment is cleaned and pre-processed (denoising) to avoid the influence of noise data, which is more conducive to health monitoring of the equipment.

[0081] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. The equipment health monitoring method based on time-frequency domain features and cluster analysis is characterized by: The steps include: Step 1: Data collection: Collect the torque data of the equipment and clean it to obtain the data set Dataset; Step 2, data labeling: linearly detrend the dataset Dataset, use multiple moving windows to traverse the detrended dataset Dataset0, and record the local maximum of the samples in each moving window; by comparing the sum of the samples under each moving window, take the sample and the minimum moving window length as the optimal moving window, and traverse the detrended dataset Dataset0 with the optimal moving window. If a sample point is confirmed as a local maximum when scanning different moving windows, mark the sample point as a peak and obtain the position index of the peak; Step 3: Data labeling and denoising: negative values ​​are taken for all samples in the dataset Dataset0, and step 2 is repeated to obtain the position index of the trough. The position index of the peak and the position index of the trough are combined and arranged in sequence to obtain the peak position index and the trough position index of a complete waveform data. The waveform data with the peak-to-trough distance greater than the set value is retained to remove the noise data, and the dataset Dataset1 is obtained. Step 4: Data conversion and clustering: Extract the time-frequency domain features of the data set Dataset1, convert the single feature data of the device into multidimensional feature data, obtain the data set Dataset2, normalize the data set Dataset2, set the number of iterations and the number of classes, cluster the normalized data, calculate the outliers and alarm thresholds based on the large and small classes formed after clustering, and generate an alarm when the time-frequency domain feature data of the device is greater than the alarm threshold.

2. The device health monitoring method based on time-frequency domain features and cluster analysis according to claim 1 is characterized in that: In step 2, the number of the multiple moving windows is s / 2, and the lengths of the multiple moving windows are 5n, n=1, 2, ...s / 2, respectively, where s is the length of the data set Dataset, and n is a parameter.

3. The device health monitoring method based on time-frequency domain features and cluster analysis according to claim 1 is characterized in that: In step three, the waveform data with a peak-to-trough distance greater than 5 is retained to remove noise data.

4. The device health monitoring method based on time-frequency domain features and cluster analysis according to claim 1 is characterized in that: In step 4, set the number of iterations and the number of classes, cluster the normalized data, and calculate the outliers and alarm thresholds based on the large and small classes formed after clustering. Specifically: Set the number of classes to 2 and the number of iterations, perform Kmeans clustering on the normalized data, and obtain two types of data; According to the number of samples in the class, it is divided into large class Dataset20 and small class Dataset21. The number of samples in the large class n1 is greater than the number of samples in the small class n2; According to the large and small categories formed after clustering, outliers and alarm thresholds are calculated.

5. The device health monitoring method based on time-frequency domain features and cluster analysis according to claim 4 is characterized in that: In calculating outliers and alarm thresholds based on the large and small categories formed after clustering, the specific steps are as follows: (a1) If the parameter m = n1 / n2, m ≥ 4, then calculate the center distance d from the sample in the small class Dataset21 to the sample in the large class Dataset20 i , get the center distance set D, calculate the mean of the center distance set D and variance δ; (a2) If Then the jth sample in the subclass is an outlier, and the outliers of the first subclass are screened out; (a3) Calculate the distance between each outlier in the first small class and the class center u of the large class Dataset20 and take the average to get the average value The average As the alarm threshold ζ1.

6. The device health monitoring method based on time-frequency domain features and cluster analysis according to claim 5 is characterized in that: In calculating outliers and alarm thresholds based on the large and small categories formed after clustering, the specific steps are as follows: If the parameter m=n1 / n2, m<4, then calculate the upper quartile Q1 and lower quartile Q3 of the distance from each sample of the large category Dataset20 to the class center of Dataset20, and determine the interquartile range IQR=Q3-Q1; The points whose distance from each sample of the large category Dataset20 to the class center of Dataset20 is less than (Q1-1.5×IQR) or greater than (Q3+1.5×IQR) are determined as large category outliers; Based on steps (a1) to (a2), outliers in subclass Dataset21 are calculated to obtain outliers of the second subclass; Aggregate the outliers of the large category and the outliers of the second small category, and calculate the distance of each outlier in the aggregated outliers from the class center u of the large category Dataset20 and average them to get the average value The average As the alarm threshold ζ2.

7. Equipment health monitoring system based on time-frequency domain features and cluster analysis, characterized in that: It includes data acquisition module, data marking module, data marking and denoising module and data conversion and clustering module; The data acquisition module is used to collect the torque data of the equipment and clean it to obtain the data set Dataset; The data labeling module is used to perform linear detrending on the dataset Dataset. It uses multiple moving windows to traverse the detrended dataset Dataset0 and records the local maximum of the samples in each moving window. By comparing the sum of the samples under each moving window, the sample and the minimum moving window length are used as the optimal moving window. The detrended dataset Dataset0 is traversed with the optimal moving window. If a sample point is confirmed as a local maximum when scanned by different moving windows, the sample point is marked as a peak, and the position index of the peak is obtained. The data marking and denoising module is used to take negative values ​​for all samples in the dataset Dataset0, repeat step 2 to obtain the position index of the trough, merge the position index of the peak and the position index of the trough and arrange them in order to obtain the peak position index and the trough position index of a complete waveform data, retain the waveform data with the peak-to-trough distance greater than the set value to remove the noise data, and obtain the dataset Dataset1; The data conversion and clustering module is used to extract the time-frequency domain features of the data set Dataset1, convert the single feature data of the device into multidimensional feature data, obtain the data set Dataset2, normalize the data set Dataset2, set the number of iterations and the number of classes, cluster the normalized data, and calculate the outliers and alarm thresholds based on the large and small classes formed after clustering. When the time-frequency domain feature data of the device is greater than the alarm threshold, an alarm message is generated.

8. The equipment health monitoring system based on time-frequency domain features and cluster analysis according to claim 7 is characterized in that: In the data conversion and clustering module, set the number of iterations and the number of classes, cluster the normalized data, and calculate the outliers and alarm thresholds based on the large and small classes formed after clustering. Specifically: Set the number of classes to 2 and the number of iterations, perform Kmeans clustering on the normalized data, and obtain two types of data; According to the number of samples in the class, it is divided into large class Dataset20 and small class Dataset21. The number of samples in the large class n1 is greater than the number of samples in the small class n2; According to the large and small categories formed after clustering, outliers and alarm thresholds are calculated.

9. The equipment health monitoring system based on time-frequency domain features and cluster analysis according to claim 8, characterized in that: In calculating outliers and alarm thresholds based on the large and small categories formed after clustering, the specific steps are as follows: (a1) If the parameter m = n1 / n2, m ≥ 4, then calculate the center distance d from the sample in the small class Dataset21 to the sample in the large class Dataset20 i , get the center distance set D, calculate the mean of the center distance set D and variance δ; (a2) If Then the jth sample in the subclass is an outlier, and the outliers of the first subclass are screened out; (a3) Calculate the distance between each outlier in the first small class and the class center u of the large class Dataset20 and take the average to get the average value The average As the alarm threshold ζ1; (a4) If the parameter m = n1 / n2, m<4, then calculate the upper quartile Q1 and lower quartile Q3 of the distance from each sample of the large class Dataset20 to the class center of Dataset20, and determine the interquartile range IQR = Q3-Q1; (a5) The points whose distance from each sample in the large class Dataset20 to the class center of Dataset20 is less than (Q1-1.5×IQR) or greater than (Q3+1.5×IQR) are determined as large class outliers; (a6) calculating the outliers in the subclass Dataset21 based on steps (a1) to (a2) to obtain the outliers of the second subclass; (a7) Aggregate the outliers of the large category and the outliers of the second small category, and calculate the distance of each outlier in the aggregated outliers from the class center u of the large category Dataset20 and take the average to get the average value The average As the alarm threshold ζ2.

Images