Data-driven long-term powered device health state real-time anomaly detection method

Abstract

The application discloses a kind of long-term power-on equipment health state real-time exception detection method based on data driving, which comprises the following steps: step one: the power-on equipment data collected to long-term power-on equipment is processed to obtain observation data, and the observation data is converted or enhanced to obtain effective measurement data;Step two: the effective measurement data in step one is reduced dimension to obtain the data after dimension reduction, and the important variables are selected by random forest model and correlation analysis to the data after dimension reduction;Step three: according to the important variables screened in step two, life characteristics are obtained by time domain feature extraction and frequency domain feature extraction;The life characteristics are smoothed to obtain the life characteristics of power-on equipment;Step four: according to the life characteristics of power-on equipment in step three, the fault factor is obtained by adopting the fault trend prediction algorithm based on multivariate clustering and principal component analysis;Step five: according to the life characteristics of power-on equipment calculated in step three, a fault prediction model is established.The application improves the prediction accuracy and the robustness of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of health state detection of powered equipment, and particularly relates to a long-term powered equipment health state real-time anomaly detection method based on data driving. BACKGROUND

[0002] Long-term powered equipment test data is various data monitored by different sensors in the running process of certain equipment, and some parameter quantities are directly or indirectly related to the running state of the equipment. By analyzing the long-term powered data, a fault prediction and health management system (PHM) is constructed to perform real-time health state evaluation, which can effectively guarantee the normal operation of the system where the equipment is located.

[0003] The core and basis of the long-term powered data health state real-time monitoring based on data driving are fault diagnosis and prediction technology. With the rapid development of test technology, especially the rapid improvement of test information acquisition, transmission and storage capacity, the available state monitoring data (including sensor data) of the target object system, test data and test data increase in series, thereby, the PHM method of the data-driven long-term powered equipment can obtain more support, and more and more attention is paid to the research, various novel algorithms, application expansion and in-depth research in different fields, so that the health monitoring of the data-driven long-term powered equipment has developed rapidly.

[0004] The health monitoring method of the data-driven long-term powered equipment recognizes or learns the health / healthless behavior of the object system from historical data through state monitoring of the object system, converts the original monitoring data into relevant information and behavior models, and predicts the future behavior of the object system. Machine learning and statistical analysis method is the mainstream algorithm of the health monitoring of the data-driven long-term powered equipment, and the existing health detection method has the problems of insufficient prediction accuracy and insufficient robustness of the model. SUMMARY

[0005] The technical problem solved by the application is to overcome the shortcomings of the prior art and provide a long-term powered equipment health state real-time anomaly detection method based on data driving, which improves the prediction accuracy and the robustness of the model.

[0006] The purpose of the application is achieved by the following technical scheme: a long-term powered equipment health state real-time anomaly detection method based on data driving, the method comprising the following steps:

[0007] Step 1: The powered equipment data collected by the long-term powered equipment is subjected to data cleaning, data integration, data transformation and data reduction to obtain observation data, and the observation data is subjected to conversion or enhancement processing to obtain effective measurement data;

[0008] Step two: dimension reduction processing is performed on the effective measurement data in step one to obtain dimension-reduced data, and important variables are selected from the dimension-reduced data by using a random forest model and correlation analysis;

[0009] Step three: life characteristics are obtained from the important variables selected in step two by time domain feature extraction and frequency domain feature extraction; the life characteristics are smoothed to obtain power-on equipment life characteristics;

[0010] Step four: a fault trend prediction algorithm based on multivariate clustering and principal component analysis is used to obtain a fault factor H(i') from the power-on equipment life characteristics in step three;

[0011] Step five: a fault prediction model is established according to the power-on equipment life characteristics calculated in step three.

[0012] In the above data-driven long-term power-on equipment health state real-time anomaly detection method, in step one, the power-on equipment data includes vibration data, environmental data, and electrical data.

[0013] In the above data-driven long-term power-on equipment health state real-time anomaly detection method, in step two, principal component analysis is used to perform dimension reduction processing on the effective measurement data, to screen out a small number of variables that have an impact on faults from multivariate, and to attempt to observe and cluster in a lower dimension. The lower dimension is 2 to 3 dimensions.

[0014] In the above data-driven long-term power-on equipment health state real-time anomaly detection method, in step two, the random forest variable importance mainly includes two evaluation indexes of using a Gini index or out-of-bag data errors.

[0015] In the above data-driven long-term power-on equipment health state real-time anomaly detection method, the Gini index is obtained by the following formula:

[0016]

[0017] wherein GI m is the Gini index, m is the classification, p mk represents the probability of misclassification of k classes into m classes, and K is the total number of classes.

[0018] In the above data-driven long-term power-on equipment health state real-time anomaly detection method, the important variable is obtained by the following formula:

[0019]

[0020] wherein Importance i is the important variable, and errOOB normalerrOOB is the normal classification error rate noise errOOB is the normal classification error rate

[0021] In the above data-driven long-term powered equipment health state real-time anomaly detection method, in step three, the life characteristics include standard deviation, skewness, kurtosis, waveform factor, peak factor, margin factor and L2 norm gain.

[0022] In the above data-driven long-term powered equipment health state real-time anomaly detection method, in step three, the smoothing processing method includes: when extracting high-dimensional features, according to the number of sampling points of the trained sample, the parameter slice of the sliding window is performed, a preset window length and a sliding step are selected, and according to the set sliding window length, the window is slid along the time axis T, and the features of the parameter segments in the window are extracted; the parameter features for clustering anomaly analysis are standard deviation std, kurtosis kurt, skewness skew, waveform factor form, peak factor peak, margin factor margin and L2 norm gain I2; the features extracted in each window are saved as a 7-dimensional vector, and when the sliding window is traversed on the original time sequence parameters, the n parameters F1, F2…Fn form n (T-w, 7) dimensional matrices.

[0023] In the above data-driven long-term powered equipment health state real-time anomaly detection method, in step four, the fault factor H(i') is obtained by the following formula:

[0024]

[0025] Wherein, H(i') represents the fault factor of window i'; j represents the singular value of any sample in the set; # represents the number of elements contained in the set; WindowLength represents the length of the sliding window.

[0026] In the above data-driven long-term powered equipment health state real-time anomaly detection method, in step five, the establishment of the fault prediction model includes the following steps:

[0027] (1) taking the F1, F2…Fn parameters of the normal stage and the fault stage of the existing data in the effective measurement data obtained in step one;

[0028] (2) for the normal stage data, take a sliding window with W1 sampling points, and construct the feature vector I corresponding to each window, set the data label set to 1, and record the normal stage window vector point set as A;

[0029] (3) For the fault stage data, a sliding window is taken at W2 sampling points, and a feature vector I corresponding to each window is constructed, and the data label set is set to 0, and the fault stage window feature vector point set is recorded as B;

[0030] (4) The normal stage window vector point set A and the fault stage window feature vector point set B obtained by each sliding window are normalized to obtain normalized data;

[0031] (5) According to the normalized data, a model KNN_Model is trained by using a KNN method;

[0032] (6) The F1', F2'...Fn' parameters in the effective measurement data obtained in step one are taken as test data;

[0033] (7) According to the sampling frequency of the test data in step (6), a sliding window Wi is constructed in a fixed sampling time, and the feature vector I' in the window is calculated;

[0034] (8) The feature vector group obtained by the sliding window Wi is normalized to obtain a normalized test data feature vector;

[0035] (9) According to the normalized test data feature vector, the trained KNN_Model is called to classify the feature vector, and the following distance formula is used to obtain the five points c1, c2, c3, c4 and c5 closest to the training set feature vector and I', recorded as point set C={c1, c2, c3, c4, c5};

[0036] (10) When the five points in the point set C belong to A, the window Wi is determined as normal data, and the detection value Rwi=1; when the five points in the point set C belong to B, it is determined that the window data has a fault trend, and the detection value Rwi=0; wherein, the majority is greater than or equal to 3;

[0037] (11) Steps (7) to (10) are repeated until all test data sets are traversed, and a detection value list R related to each sliding window is obtained;

[0038] (12) The set N original sliding windows Wi are taken as a judgment window S, and the health factor H in the judgment window S is obtained according to the detection value list R.

[0039] Compared with the prior art, the present application has the following beneficial effects:

[0040] (1) The present application constructs a higher dimensional feature by working condition classification and extraction of time domain feature factors, and improves the fitting ability of the model to complex relationships.

[0041] (2) By defining the fault factors of the sliding window, the present invention effectively extracts fault features from the data, forms reasonable fault criteria, and provides a basis for subsequent fault prediction.

[0042] (3) This invention uses the KMeans model to classify feature vectors, which can efficiently and accurately classify fault feature vectors and normal feature vectors, thereby identifying and classifying potential faults.

[0043] (4) This invention uses the PCA data dimensionality reduction algorithm to map the original n-dimensional features of the dataset to m-dimensional features. While retaining most of the variance in the dimensional features, it selects a few variables that have an impact on the fault from multiple variables, effectively improving the efficiency of fault diagnosis.

[0044] (5) The detection output results of the long-term power supply equipment health status of the present invention are combined with a multivariate clustering model and a supervised learning fault prediction model to ensure the accuracy of the detection results.

[0045] (6) The model structure adopted in this invention is simple and efficient, which reduces the computational resources consumed and alleviates the deployment difficulty and hardware requirements. Attached Figure Description

[0046] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0047] Figure 1 This is a flowchart of a fault trend prediction method based on multivariate clustering and principal component analysis provided in an embodiment of the present invention. Detailed Implementation

[0048] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] Figure 1 This is a flowchart of a fault trend prediction method based on multivariate clustering and principal component analysis provided in an embodiment of the present invention. Figure 1As shown, the embodiment provides a data-driven long-term powered equipment health state real-time anomaly detection method, including the following steps:

[0050] Step one: The vibration data, environmental data, electrical data and other data collected by the long-term powered equipment are cleaned, integrated, transformed and reduced, the dirty data in the data set is removed, the observation data is converted or enhanced, and the effective measurement data is obtained.

[0051] Step two: According to the effective measurement data obtained in step one, first, the principal component (PCA) analysis is used to process the effective measurement data, a few variables that have an impact on failure are selected from the multivariate, and observation and clustering processing are attempted in a lower dimension (2-3 dimensions) visualization, and the cause of the failure is attempted to be explained.

[0052] After dimension reduction, the most important variables are selected by using random forest model and correlation analysis. The main quantitative classification method of random forest variable importance includes two evaluation indexes of using Gini index or out-of-bag data error. The Gini index can be used to evaluate the impurity of variable node splitting, and its calculation formula is:

[0053]

[0054] Where p mk represents the probability of misclassification of k classes into m classes. K is the total number of classes. GI m is the Gini index, and m is the classification.

[0055] Since random forest contains multiple decision trees, the importance of a certain variable VIM i can be measured as:

[0056]

[0057] Where, the importance measurement index VIM i =∑ m∈M GI m -GI l -GI r , M is the set of node occurrences, GI l is the similar class Gini index, GI r is the dissimilar class Gini index, and VIM j is the importance measurement index of variable j. The out-of-bag data error index adds noise to a certain feature label, inputs the noisy data set into the random forest classification result, and measures the importance of the data label. The change can be used as a reference for feature extraction. Its calculation formula is as follows:

[0058]

[0059] where Importance i is the importance of the variable, errOOB normal is the normal classification error rate, errOOB noise is the classification error rate with noise, K is the total number of classes, M is the set of nodes, i is the variable label,

[0060] Step three: the important variables screened out in step two are extracted by time domain feature extraction and frequency domain feature extraction to obtain the corresponding features. The collected features include:

[0061] a) standard deviation (std): measures the dispersion of data within the window. The variance (sample variance) in statistics is the average of the squared values of the difference between each sample value and the average of all sample values.

[0062] b) skewness (kurt): a measure of the direction and degree of skewness of the distribution of data within the window.

[0063] c) skewness (skew): a factor that measures the degree of smoothness of the distribution of data within the window, used to describe the distribution of variables.

[0064] d) waveform factor (form): a dimensionless quantity in alternating current signals, which is the ratio of the root mean square value of the signal to the rectified average value.

[0065] e) peak factor (peak): the ratio of the peak value of the signal to the effective value (RMS), which represents the extreme degree of the peak value in the waveform.

[0066] f) margin factor (margin): the margin factor is the ratio of the peak value of the signal to the root amplitude value.

[0067] g) L2 norm gain (I2): can be simply and intuitively understood as the length of the vector, or the distance from the zero point, or the distance between the corresponding two points.

[0068] In the extraction of high-dimensional features, the parameters of the sliding window should be sliced according to the number of sampling points of the trained samples, and the appropriate range of window length and sliding step should be selected. A longer moving step is used to ensure that the feature parameters extracted in each window segment produce distinguishable changes. According to the set sliding window length, the window is slid along the time axis T, and the parameters of the window are extracted. The parameter features used for clustering anomaly analysis are standard deviation std, kurtosis kurt, skewness skew, waveform factor form, peak factor peak, margin factor margin and L2 norm gain I2. We save the features extracted in each window as a 7-dimensional vector. When the sliding window is completed in the original time sequence parameter, the n parameters F1, F2...Fn form n (T-w, 7) dimensional matrices.

[0069] Step four: according to the power-on equipment life characteristics calculated in step three, the fault trend prediction algorithm based on multivariate clustering and principal component analysis is used to obtain the fault factor H(i'). In order to find the change of parameter characteristics before the occurrence of fault, the fault factor H is constructed to realize the monitoring of parameter change trend before the occurrence of equipment fault. The trained clustering algorithm will obtain the cluster center, and the principal component features constructed by PCA will calculate the distance from the sample point to the cluster center by the following formula, so as to divide the sample points into majority class P and minority class Q,

[0070]

[0071] Wherein, d(x, u) is the distance from the sample point to the cluster center, x is the sample point, u is the cluster center, x i is the sample point with variable mark i, u i is the cluster center with variable mark i, and n is the total number of samples.

[0072] The construction of fault factor is to process the whole sequence by sliding window to obtain a plurality of groups of data, and the mean value of minority class in each group of data processed by sliding window is calculated to obtain the fault factor H.

[0073]

[0074] In the above formula, H(i') represents the fault factor of window i'; j represents the singular value of any sample in the set; # represents the number of elements contained in the set; WindowLength represents the length of the sliding window. Therefore, the value of H represents the ratio of the number of minority classes in the sliding window to the length of the sliding window. The larger the value of H is, the greater the possibility of failure of window i is.

[0075] Step five:

[0076] According to the power-on equipment life characteristics calculated in step three, a fault prediction model is established.

[0077] For the fault prediction algorithm obtained in step four, the establishment of the supervised learning fault prediction model is carried out. According to the normal stage, the degradation stage and the fault stage, the n parameters F1, F2...Fn obtained in step one are divided into n (T-w, 7) dimensional matrices, and the mean and variance of the high-dimensional characteristics of each parameter in each stage are calculated. Finally, the 10 features with the largest difference between the mean and the variance are selected to form a 10-dimensional feature vector I = [I1, I2...I10].

[0078] The steps of the whole model construction and degradation prediction are as follows:

[0079] (1) Take the F1, F2...Fn parameters in the normal stage and the fault stage in the existing data in the effective measurement data obtained in step one; it should be noted that the effective measurement data in step one includes existing data and to-be-measured data.

[0080] (2) For the normal stage data, take a sliding window with W1 sampling points, and construct the feature vector I corresponding to each window. The data label set is set to 1, and the normal stage window vector point set is recorded as A;

[0081] (3) For the fault stage data, take a sliding window with W2 sampling points, and construct the feature vector I corresponding to each window. The data label set is set to 0, and the fault stage window feature vector point set is recorded as B;

[0082] (4) Normalize the normal stage window vector point set recorded as A and the normal stage window vector point set recorded as A obtained by each sliding window to obtain normalized data;

[0083] (5) According to the normalized data, the model KNN_Model is trained by using the KNN method;

[0084] (6) Take the F1', F2'...Fn' parameters in the to-be-tested data of the effective measurement data obtained in step one;

[0085] (7) According to the sampling frequency of the to-be-tested data in step (6), a sliding window Wi is constructed in a fixed sampling time, and the feature vector I' in the window is calculated;

[0086] (8) The feature vector group obtained by the sliding window Wi is normalized to obtain the normalized to-be-tested data feature vector;

[0087] (9) According to the normalized to-be-tested data feature vector, the trained KNN_Model is called to classify the feature vector, and the following distance formula is used to obtain the 5 points closest to the training set feature vector and I', recorded as the point set C = {c1, c2, c3, c4, c5}

[0088]

[0089] wherein I' i is the feature vector of the i-th sliding window of the data to be tested, I i is the feature vector of the i-th sliding window of the existing data,

[0090] (10) When 5 points in the point set C belong to A in majority, the window Wi is determined as normal data, and the detection value Rwi of the window is 1; when 5 points in the point set C belong to B in majority, the window data is determined to have a failure trend, and the detection value Rwi is 0. The formula is as follows:

[0091]

[0092] (11) Repeat steps (7)-(10) until all the data sets to be tested are traversed, and obtain the detection value list R related to each sliding window.

[0093] (12) According to the characteristics of the analyzed data, take the set N original sliding windows Wi as a determination window S to calculate the health factor H in the window.

[0094]

[0095] wherein N is the number of selected original sliding windows, k' is the time sequence number of the currently determined window, and Rwi is the detection value of the currently determined window.

[0096] Seven high-dimensional features are extracted on each parameter quantity, and the calculation method of each feature is as follows:

[0097] Standard deviation (std): measure the dispersion degree of data in the window. The variance (sample variance) in statistics is the average of the square value of the difference between each sample value and the average of all sample values. In many practical problems, it is of great significance to study the variance, that is, the degree of deviation. The standard deviation calculation formula is as follows:

[0098]

[0099] wherein σ is the population variance, X is the population variable, μ is the population mean, and N is the population number.

[0100] Skewness (kurt): a measure describing the skewness direction and degree of data distribution in the window. A numerical feature of the degree of asymmetry of statistical data distribution. Negative numbers are right-skewed, and positive numbers are left-skewed. The skewness calculation formula is as follows

[0101]

[0102] wherein S kSkewness, μ3 and μ2 represent the third and second central moments, respectively, and σ represents the standard deviation.

[0103] Skewness: a measure of how flat or peaked the distribution of data in a window is, used to describe the distribution of a variable. The formula for skewness is as follows:

[0104]

[0105] where μ4 represents the fourth central moment and σ represents the standard deviation.

[0106] Waveform factor: a dimensionless quantity in an alternating current signal, which is the ratio of the root mean square value to the rectified average value. The waveform factor is the ratio of the same power of the direct current signal and the original alternating current signal. The formula for the waveform factor is as follows:

[0107]

[0108] where X rms represents the root mean square value, that is, the sum of the squares of all values, the mean value, and the square root. X arv represents the rectified average value, which refers to the average value of the absolute value of the signal sequence.

[0109] Peak factor: the ratio of the peak value of the signal to the effective value (RMS), which represents the extreme degree of the peak value in the waveform. The formula for the peak factor is as follows:

[0110]

[0111] where X rms represents the root mean square value, that is, the sum of the squares of all values, the mean value, and the square root. X max represents the peak value of the signal.

[0112] Margin factor: the margin factor is the ratio of the peak value of the signal to the root amplitude value. The root amplitude value and the effective value are corresponding, the formula of the root mean square is the arithmetic square root of the average value of the signal square sum, and the root amplitude value is the square of the arithmetic square root of the average value. The formula for the margin factor is as follows:

[0113]

[0114] where X max represents the peak value of the signal. X sra represents the root amplitude value, that is, the square of the average value of the arithmetic square root of each parameter in the sequence.

[0115] L2 norm gain (L2): can be simply and intuitively understood as the length of the vector, or the distance from the zero point, or the distance between the corresponding two points.

[0116]

[0117] The above 7 variables are calculated on the basis of the mean, and are irrelevant to the number of samples calculated, so the window of different sample numbers will not affect the values of these characteristic factors.

[0118] The present application constructs higher dimensional features by working condition classification and extraction of time domain characteristic factors, improves the fitting ability of the model to complex relationships, and effectively peels off the fault features from the data by defining the fault factor of the sliding window, forms a reasonable fault criterion, and provides a basis for subsequent fault prediction. The present application uses the KMeans model to classify the feature vectors, which can efficiently and accurately classify the fault feature vectors and normal feature vectors to identify and classify potential faults. The present application maps the original n-dimensional features of the data set to m-dimensional features through the PCA data dimension reduction algorithm. On the basis of retaining most of the dimensional characteristics of the variance, a small number of variables that have an impact on the fault are selected from the multivariate, effectively improving the efficiency of fault diagnosis. The detection output result of the long-term powered equipment health state of the present application combines the multivariate clustering model and the supervised learning fault prediction model, ensuring the accuracy of the detection result. The model structure of the present application is simple and efficient, reduces the calculation of consumed resources, and reduces the deployment difficulty and hardware requirements.

[0119] Although the present application has been disclosed as above with reference to the preferred embodiments, it is not intended to limit the present application, and any person skilled in the art can make possible changes and modifications to the technical solutions of the present application by using the disclosed methods and technical contents without departing from the spirit and scope of the present application. Therefore, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application, which does not deviate from the technical solutions of the present application, belongs to the protection scope of the technical solutions of the present application.

Claims

1. A data-driven based long-term powered device health state real-time anomaly detection method, characterized in that, The method comprises the following steps: Step one: the data collected by the long-term powered equipment is cleaned, integrated, transformed and reduced to obtain observation data, and the observation data is converted or enhanced to obtain effective measurement data; Step two: the effective measurement data in step one is processed to obtain reduced dimension data, and important variables are selected from the reduced dimension data by using a random forest model and correlation analysis; Step three: the important variables selected in step two are processed by time domain feature extraction and frequency domain feature extraction to obtain life characteristics; and the life characteristics are smoothed to obtain powered equipment life characteristics; Step four: a fault trend prediction algorithm based on multivariate clustering and principal component analysis is used to obtain a fault factor according to the powered equipment life characteristics in step three; Step five: a fault prediction model is established according to the powered equipment life characteristics calculated in step three; In step five, the establishment of the fault prediction model comprises the following steps: (1) taking F1, F2…Fn parameters of the normal stage and the fault stage of the existing data in the effective measurement data obtained in step one; (2) taking a sliding window of W1 sampling points for the normal stage data, and constructing a feature vector I corresponding to each window, setting the data label set as 1, and recording the normal stage window vector point set as A; (3) taking a sliding window of W2 sampling points for the fault stage data, and constructing a feature vector I corresponding to each window, setting the data label set as 0, and recording the fault stage window feature vector point set as B; (4) normalizing the normal stage window vector point set A and the fault stage window feature vector point set B obtained by each sliding window to obtain normalized data; (5) training a model KNN_Model using the KNN method according to the normalized data; (6) taking F1’, F2’…Fn’ parameters in the test data in the effective measurement data obtained in step one; (7) constructing a sliding window Wi within a fixed sampling time according to the sampling frequency of the test data in step (6), and calculating the feature vector I’ in the window; (8) normalizing the feature vector group obtained by the sliding window Wi to obtain the normalized test data feature vector; (9) classifying the feature vector according to the trained KNN_Model according to the normalized test data feature vector, and obtaining the five points c1, c2, c3, c4 and c5 closest to the training set feature vector and I’ according to the distance formula, and recording them as a point set C={c1, c2, c3, c4, c5}; (10) when the five points in the point set C belong to A, the window Wi is determined as normal data, and the detection value Rwi=1; when the five points in the point set C belong to B, it is determined that the window data has a fault trend, and the detection value Rwi=0; wherein, the majority is greater than or equal to 3; (11) repeating steps (7) to (10) until all test data sets are traversed to obtain a detection value list R related to each sliding window; (12) taking the set N original sliding windows Wi as a determination window S, and obtaining the health factor H in the determination window S according to the detection value list R; The health factor H is obtained by the following formula: In the formula, N is the number of original sliding windows selected, k' is the time series number of the current determined window, R wi is the detection value of the current determined window.

2. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step two, the main quantitative classification method of random forest variable importance includes two evaluation indexes of using Gini index or out-of-bag data error; The Gini index is obtained by the following formula: wherein GI m is the Gini index, m is the classification, p mk represents the probability of misclassification of class k into class m, and K is the total number of classes. The important variable is obtained by the following formula: where Importance i is the importance of the variable, errOOB normal is the normal classification error rate, errOOB noise is the classification error rate with noise, K is the total number of classes, M is the set of nodes, and i is the variable label.

3. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step one, the powered equipment data includes vibration data, environmental data and electrical data.

4. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step two, the effective measurement data is processed by dimension reduction using principal component analysis, a few variables that have an impact on failure are screened from the multivariable, and observation and clustering processing are attempted in a lower dimension for visualization; the lower dimension is 2 to 3.

5. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step three, the life characteristics include standard deviation, skewness, kurtosis, waveform factor, peak factor, margin factor and L2 norm gain.

6. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step three, the smoothing processing method includes: when extracting high-dimensional features, the parameter slices of the sliding window are cut according to the number of sampling points of the trained sample, a preset window length and a sliding step length, the window is slid along the time axis T according to the set sliding window length, and the features of the parameter segments in the window are extracted; the parameter features used for clustering anomaly analysis are standard deviation std, kurtosis kurt, skewness skew, waveform factor form, peak factor peak, margin factor margin and L2 norm gain l2; the features extracted in each window are saved as a 7-dimensional vector, and when the sliding window is traversed on the original time sequence parameters, the n parameters F1, F2…Fn form n (T-w, 7) dimensional matrices.

7. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 1, wherein: In step four, the fault factor H(i') is obtained by the following formula: Wherein, H(i') represents the fault factor of window i'; j represents the singular value of any sample in the set; # represents the number of elements contained in the set; WindowLength represents the length of the sliding window.

8. The data-driven based long-term power-on device health state real-time anomaly detection method of claim 5, wherein: Seven high-dimensional features are extracted on each parameter quantity, and the calculation method of each feature is as follows: Standard deviation: measure the dispersion degree of data in the window, the variance in statistics is the average of the square value of the difference between each sample value and the average of all sample values, and the standard deviation calculation formula is as follows: Wherein, σ is the total variance, X is the total variable, μ is the total mean, and N is the total number; Skewness: a measure of the skewness of the data distribution in the window, a numerical feature of the asymmetry of the statistical data distribution, negative numbers are right-skewed, and positive numbers are left-skewed; the skewness calculation formula is as follows: where S k represents skewness, μ3and μ2represent the third and second central moments, respectively, and σ represents the standard deviation; Kurtosis: a factor that measures the smoothness of the data distribution in the window, used to describe the distribution of variables; the kurtosis calculation formula is as follows: Wherein, μ4 represents the fourth order central moment, and σ represents the standard deviation; Waveform factor: a dimensionless quantity in alternating current signal, which is the ratio of the root mean square value of the signal to the rectified average value; the waveform factor is the ratio of the direct current signal of the same power to the original alternating current signal; the waveform factor calculation formula is as follows: where X rms represents the root mean square value, X arv represents the rectified mean value, which refers to the average value of the absolute value of the signal sequence; Peak factor: the ratio of the peak value of the signal to the effective value, representing the extreme degree of the peak value in the waveform; the peak factor calculation formula is as follows: where X rms represents the root mean square value, X max represents the signal peak value; Margin factor: the margin factor is the ratio of the signal peak value to the root amplitude value, and the root amplitude value and the effective value are corresponding, the formula of the root amplitude value is the square root of the average value of the signal square sum, and the root amplitude value is the square of the average value of the arithmetic square root; The margin factor calculation formula is as follows: where X max represents the signal peak value, X sra represents the square root amplitude, i.e. the square of the average of the arithmetic square roots of each parameter in the sequence; L2 norm gain: the length of the vector, or the distance from the vector to the zero point, or the distance between the corresponding two points; The above 7 variables are calculated on the basis of the mean value, and are irrelevant to the number of samples calculated, so the window of different sample numbers will not affect the value of these characteristic factors.

Images