A hybrid lifetime prediction method for complex electromechanical devices

Abstract

The application relates to a hybrid life prediction method of a complex electromechanical device, which comprises the following steps: preprocessing full-life data collected by the complex electromechanical device to obtain a labeled training sample set; obtaining a set of training sample subsets in a segmented mode; setting the horizontal and vertical coordinates of the best fitting curve as the 'feature value-weight value' pairs of the training sample subsets; obtaining an improved VPM model by summarizing all the models; obtaining a set of test sample subsets, and predicting each test sample subset according to the improved VPM model to obtain a preliminary prediction result of life evaluation; and obtaining a final life prediction value of the complex electromechanical device. im The application takes the independent variable probability as the weight w of the mean and variance, and reduces the influence of abnormal samples on the classification accuracy; the improved VPM model can be used for calculating the similarity of the probability density curves of a section of data of the measured continuous time sequence, and greatly reduces the influence of abnormal samples.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of life prediction of electromechanical equipment, and particularly relates to a hybrid life prediction method for complex electromechanical equipment. BACKGROUND

[0002] The complex electromechanical equipment roughly experiences a stable operation phase, a performance degradation phase and a final failure phase from starting to work to failure. The conventional due maintenance mode may cause over maintenance of the equipment, thereby increasing the maintenance manpower and the loss of spare parts, and may also cause the system to be unable to safely operate due to untimely maintenance. At present, with the development of intelligent sensing technology and the Internet of Things and the accumulation of information, real-time fault diagnosis and aging assessment of complex electromechanical equipment are possible. Precise life prediction of electromechanical equipment needs to collect long-term historical data of multiple monitoring quantities of the equipment, and the historical data are measured in repeated test experiments under different complex environments and different working conditions. On the basis of the above data, the multiple sensing data are preprocessed, and the life of the whole machine and the operation reliability are evaluated through data mining methods. However, due to the fact that the key components in the real large electromechanical equipment almost never run to complete failure, the full life data are lacking. Therefore, researchers can only make full use of the existing data as much as possible, and thus how to use the training data to obtain more accurate life prediction has been the focus of researchers.

[0003] Discriminant analysis is a statistical method widely used in multivariate analysis. It can be used to classify samples, understand the differences between different categories of samples in certain variables, predict the probability that a new sample belongs to a certain category, and test the correctness of clustering analysis after clustering analysis. Variable prediction model-based class discriminant (VPMCD) is a new multiclass discriminant method that can establish a mathematical model-variable prediction model (VPM) for each class to identify the class without complex iteration. Many studies have shown that the VPMCD classifier has better performance and is a potential multiclass fault diagnosis tool. In the field of life prediction, the VPMCD algorithm also has wide application. However, during the full life data collection process, due to sensor failure or operating state, abnormal data points such as missing sampling, false sampling and false positives may occur, thereby affecting the accuracy of the VPM model and directly affecting the accuracy of the degradation state assessment of the electromechanical system. SUMMARY

[0004] The present application aims to provide a hybrid life prediction method for complex electromechanical equipment, which reduces the influence of abnormal points and improves the efficiency and accuracy of life assessment.

[0005] To achieve the above object, the application adopts the following technical scheme: a hybrid life prediction method of complex electromechanical equipment, which comprises the following sequential steps:

[0006] (1) The full-life data collected by the complex electromechanical equipment is preprocessed, the preprocessed data is divided into a training sample set and a test sample set, the training sample set is pre-segmented and labeled to obtain a labeled training sample set;

[0007] (2) The labeled training sample set is clustered by using a clustering method, and then segmented by using a segmentation mode meeting the accuracy requirement selected according to the accuracy as the standard to obtain a set of training sample subsets under the segmentation mode;

[0008] (3) The probability curve fitting is performed on each feature value of all training sample subsets in the set of training sample subsets, the best fitting curve in the probability curve is selected, and the horizontal and vertical coordinates of the best fitting curve are set as the "feature value-weight value" pair of the training sample subset;

[0009] (4) According to the corresponding relationship between the feature value and the weight value, an equation is established for all feature values in each training sample subset, the parameters of all feature values under four VPM model types are determined, and each model and all models are summarized to obtain an improved VPM model, and the four VPM models are linear model, linear interaction model, quadratic model and quadratic interaction model;

[0010] (5) The test sample set is segmented by using a segmentation mode meeting the accuracy requirement selected according to the accuracy as the standard to obtain a set of test sample subsets, and each test sample subset is predicted according to the improved VPM model to obtain a preliminary prediction result of life evaluation;

[0011] (6) The trend extrapolation method is used to obtain the final life prediction value of the complex electromechanical equipment by comprehensively considering the preliminary prediction result.

[0012] The step (1) specifically comprises the following steps:

[0013] (1a) The interpolation and trend extrapolation method is used to complete the data of the incomplete full-life data, and all full-life data is normalized and preprocessed, and the preprocessed data is divided to obtain a training sample set of M-dimensional life data with a sample number of N and a test sample set wherein, N + represents a positive natural number;

[0014] (1b) pre-segmentation, and the initial length l of the training sample subset is sete =30, the training sample set Segmented into The segment contains l e A set of samples, fix(·) represents the quotient calculation;

[0015] (1c) Taking the sampling time as a reference, add labels 1, 2, 3, ..., C0 to the C0 sets after segmentation one by one. Find the union of the C0 sets with added labels to obtain the labeled training sample set.

[0016] Step (2) specifically includes the following steps:

[0017] (2) Clustering is used to cluster the labeled training sample set, and then the segmentation method that meets the accuracy requirements is selected based on the accuracy rate to obtain the set of training sample subsets under the segmentation method.

[0018] (2a) Set an accuracy threshold Let the number of cluster centers in the FCM algorithm be C = C0;

[0019] (2b) The accuracy of the labeled training sample set was verified using the FCM algorithm, and the results were obtained. e Accuracy at 30 The accuracy metric used is cluster purity, calculated using the following formula:

[0020]

[0021] Where N represents the number of samples in the labeled training sample set, N′ is the number of training samples in the clustering result whose predicted labels are consistent with the true labels, the true labels are the labels added after the training sample set is pre-segmented, and the predicted labels refer to the labels after clustering by the FCM algorithm. The clusters obtained after clustering are defined by the true labels with the highest proportion.

[0022] (2c) If Then reduce the number of cluster centers C to C-1, pre-segment again and add labels, cluster the resulting labeled training sample set again, and verify the accuracy of FCM clustering using formula (1); increase the length l of the training sample subset. e To reduce the number of segments, if the length of the training sample subset is l e >100 is still not enough Then return to step (2a) and lower the accuracy threshold. Accuracy threshold Reduce the number of cluster centers by one percent, reset the number of cluster centers C = C0, and verify the accuracy of FCM clustering again using formula (1). Repeat step (2c) until the result satisfies the condition. The number of segments C of the condition* ;

[0023] like Then the number of segments C that meet the accuracy requirements is determined at this point. * ;

[0024] (2d) obtains C * The set consisting of training sample subsets under each segment in, Let c be the c-th training sample subset in the set of training sample subsets, with label value c.

[0025] Step (3) specifically includes the following steps:

[0026] (3) Perform probability curve fitting on each feature value of all training sample subsets in the set of training sample subsets, select the best fitting curve in the probability curve, and set the horizontal and vertical coordinates of the best fitting curve as the "feature value-weight" pair of the training sample subset.

[0027] (3a) Introduce a probability density function and calculate the similarity of probability density curves for all M feature values ​​of all training sample subsets in the set Ψ of training sample subsets. Use the mean squared error as the criterion for selecting the probability density curve:

[0028]

[0029] In the formula, Y i For the true value, For the predicted value, l e The length of the current training sample subset;

[0030] (3b) Select the probability density curve with the minimum mean square error and store the values ​​within all segments. Yes, that is, the "eigenvalue-weight" pair, where, for The feature value of the m-th dimension of the i-th sample, w im To be based on the selected probability density curve The corresponding ordinate, Let c be the c-th training sample subset in the set of training sample subsets.

[0031] Step (4) specifically includes the following steps:

[0032] (4a) For the set Ψ of the training sample subset, the length is l e All feature values ​​in each training sample subset are used to establish l e An equation, based on the "eigenvalue-weight" relationship between eigenvalues... weight w imFor all eigenvalues Multiply by the weight w im Solve for the undetermined coefficients of all feature values ​​in each training sample subset in the following four VPM models:

[0033] (4a1) The linear model LI is:

[0034]

[0035] When m = n, a0 = a n =0, where, In this context, m and n represent the m-th and n-th dimensional feature values ​​of the samples in the training sample subset, and M is the dimension of the training samples. w im w in They are respectively The weights, a0, a n These are the undetermined coefficients of the linear model;

[0036] (4a2) The linear interaction model L is:

[0037]

[0038] When m = n, b0 = b n =p mn =0, where b0, b n p mn These are the undetermined coefficients of the linear interaction model;

[0039] (4a3) The quadratic model QI is:

[0040]

[0041] When m = n, g0 = g 1n =g 2n =0, where g0, g 1n g 2n These are the undetermined coefficients of the quadratic model;

[0042] (4a4) The secondary interaction model Q is:

[0043]

[0044] When m = n, d0 = d 1n =d 2n =f mn =0, where d0, d 1n d 2n f mn These are the undetermined coefficients for the quadratic interaction model;

[0045] Jointly summing l from each training sample subsete Equations are used to calculate the undetermined coefficients of all feature values in each training sample subset under four VPM models;

[0046] (4b) Feature values after adding weights Select one of the four VPM models as the model of the ith feature value of the cth sample subset Model Select the model type with the minimum sum of squares (SSE) as the model of the ith feature value of the cth sample subset

[0047]

[0048] In the formula, w im is the actual feature value is the corresponding weight is the predicted feature value, and the model type with the minimum sum of squares (SSE) is selected as the model of the ith feature value of the cth sample subset Corresponding model

[0049] (4c) Select the model of all features in all training sample subsets in the training sample set, store the type and parameters of the model, and aggregate all models to obtain an improved VPM model:

[0050]

[0051] In the formula, the improved VPM model is composed of C * M models.

[0052] The step (5) specifically includes the following steps:

[0053] (5) Select a segmentation method that meets the accuracy requirement based on the accuracy as the standard, segment the test sample set to obtain a set of test sample subsets, and predict each test sample subset based on the improved VPM model to obtain a preliminary prediction result of the life assessment;

[0054] (5a) For the test sample set segment the test sample set in a reverse segmentation manner from the end of the data to the front, and the length l e of the test sample subset is consistent with the length of the training sample subset, obtaining a set of test sample subsets, denoted as:

[0055]

[0056] (5b) For the multiple test sample subsets ​​​​According to the improved VPM model, life prediction is respectively carried out, and the test sample subset prediction standard is set as follows:

[0057]

[0058] In the formula, is the true characteristic value of the i-th dimension in the c'th test sample subset, is the predicted characteristic value of the i-th dimension in the c'th test sample subset in the improved VPM model, by solving, the preliminary prediction results of all test sample subsets in are finally obtained, denoted as {c1, c2, …}.

[0059] The step (6) is specifically referring to: using the trend extrapolation method, the preliminary prediction results are comprehensively obtained, and the final life prediction value of the complex mechanical and electrical equipment is obtained, the trend extrapolation method adopts the linear trend extrapolation method, and the formula is:

[0060]

[0061] In the formula, is the final life prediction value, t is a time variable, according to the preliminary prediction results {c1, c2, …} of the test sample subset, a curve is fitted, the coefficients a and b are determined, and the final life prediction value is obtained

[0062] From the above technical solution, the beneficial effects of the present application are: first, when the sample dispersion is large and there are many abnormal samples, four kinds of traditional VPM models cannot describe the relationship between characteristics, and the present application fits the probability density curve of the independent variable on the basis of the traditional VPM model, and the probability of the independent variable is used as the weight w of variance im , so as to reduce the influence of abnormal samples on classification accuracy; second, the traditional VPM classification process is classified respectively for each sample, the improved VPM model is used in the present application, the probability density curve similarity of a segment of data of the measured continuous time sequence can be calculated, so as to greatly reduce the influence of abnormal samples; third, the improved VPM model is used in the present application to jointly determine the multi-sensor data, and misjudgment caused by sensor failure can be effectively excluded. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 The method flowchart of the present application. DETAILED DESCRIPTION

[0064] As Figure 1 shown, a hybrid life prediction method of complex mechanical and electrical equipment, the method comprises the following steps in sequence:

[0065] (1) preprocessing the full-life data collected by complex electromechanical equipment, dividing the preprocessed data into a training sample set and a test sample set, pre-segmenting the training sample set and adding labels to obtain a labeled training sample set;

[0066] (2) clustering the labeled training sample set by using a clustering method, and selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment, to obtain a set of training sample subsets under the segmentation mode;

[0067] (3) fitting a probability curve to each feature value of all training sample subsets in the set of training sample subsets, selecting the best fitting curve in the probability curve, and setting the horizontal and vertical coordinates of the best fitting curve as a “feature value-weight value” pair of the training sample subset;

[0068] (4) establishing an equation for all feature values in each training sample subset according to the corresponding relationship between the feature values and the weight values, determining the parameters of all feature values under four VPM model types, and obtaining each model and summarizing all models to obtain an improved VPM model, wherein the four VPM models are linear model, linear interaction model, quadratic model and quadratic interaction model;

[0069] (5) selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment the test sample set, obtaining a set of test sample subsets, and predicting each test sample subset according to the improved VPM model to obtain a preliminary prediction result of life assessment;

[0070] (6) using trend extrapolation method to obtain the final life prediction value of the complex electromechanical equipment by comprehensively analyzing the preliminary prediction result.

[0071] The step (1) specifically comprises the following steps:

[0072] (1a) using interpolation and trend extrapolation method to complete the missing full-life data, and normalizing and preprocessing all full-life data, dividing the preprocessed data to obtain a training sample set of M-dimensional life data with N samples and a test sample set wherein, N+ represents a positive integer;

[0073] (1b) pre-segmentation, setting the initial length l e = 30 of the training sample subset, segmenting the training sample set into sets containing l e samples, and fix(·) represents quotient calculation;

[0074] (1c) Taking the sampling time as a reference, add labels 1, 2, 3, ..., C0 to the C0 sets after segmentation one by one. Find the union of the C0 sets with added labels to obtain the labeled training sample set.

[0075] Step (2) specifically includes the following steps:

[0076] (2) Clustering is used to cluster the labeled training sample set, and then the segmentation method that meets the accuracy requirements is selected based on the accuracy rate to obtain the set of training sample subsets under the segmentation method.

[0077] (2a) Set an accuracy threshold Let the number of cluster centers in the FCM algorithm be C = C0;

[0078] (2b) The accuracy of the labeled training sample set was verified using the FCM algorithm, and the results were obtained. e Accuracy at 30 The accuracy metric used is cluster purity, calculated using the following formula:

[0079]

[0080] Where N represents the number of samples in the labeled training sample set, N′ is the number of training samples in the clustering result whose predicted labels are consistent with the true labels, the true labels are the labels added after the training sample set is pre-segmented, and the predicted labels refer to the labels after clustering by the FCM algorithm. The clusters obtained after clustering are defined by the true labels with the highest proportion.

[0081] (2c) If Then reduce the number of cluster centers C to C-1, pre-segment again and add labels, cluster the resulting labeled training sample set again, and verify the accuracy of FCM clustering using formula (1); increase the length l of the training sample subset. e To reduce the number of segments, if the length of the training sample subset is l e >100 is still not enough Then return to step (2a) and lower the accuracy threshold. Accuracy threshold Reduce the number of cluster centers by one percent, reset the number of cluster centers C = C0, and verify the accuracy of FCM clustering again using formula (1). Repeat step (2c) until the result satisfies the condition. The number of segments C of the condition * ;

[0082] like Then the number of segments C that meet the accuracy requirements is determined at this point. * ;

[0083] (2d) obtains C * The set consisting of training sample subsets under each segment in, Let c be the c-th training sample subset in the set of training sample subsets, with label value c.

[0084] Step (3) specifically includes the following steps:

[0085] (3) Perform probability curve fitting on each feature value of all training sample subsets in the set of training sample subsets, select the best fitting curve in the probability curve, and set the horizontal and vertical coordinates of the best fitting curve as the "feature value-weight" pair of the training sample subset.

[0086] (3a) Introduce a probability density function and calculate the similarity of probability density curves for all M feature values ​​of all training sample subsets in the set Ψ of training sample subsets. Use the mean squared error as the criterion for selecting the probability density curve:

[0087]

[0088] In the formula, Y i For the true value, For the predicted value, l e The length of the current training sample subset;

[0089] (3b) Select the probability density curve with the minimum mean square error and store the values ​​within all segments. Yes, that is, the "eigenvalue-weight" pair, where, for The feature value of the m-th dimension of the i-th sample, w im To be based on the selected probability density curve The corresponding ordinate, Let c be the c-th training sample subset in the set of training sample subsets.

[0090] Step (4) specifically includes the following steps:

[0091] (4a) For the set Ψ of the training sample subset, the length is l e All feature values ​​in each training sample subset are used to establish l e An equation, based on the "eigenvalue-weight" relationship between eigenvalues... weight w im For all eigenvalues Multiply by the weight w im Solve for the undetermined coefficients of all feature values ​​in each training sample subset in the following four VPM models:

[0092] (4a1) Linear model LI is:

[0093]

[0094] When m = n, a0= a n = 0, where m, n represent the m, n-dimensional feature values of the samples in the training sample subset, M is the dimension of the training sample, w im , w in are the weights of , a0, a n are the undetermined coefficients of the linear model;

[0095] (4a2) Linear interaction model L is:

[0096]

[0097] When m = n, b0= b n = p mn = 0, where b0, b n , p mn are the undetermined coefficients of the linear interaction model;

[0098] (4a3) Quadratic model QI is:

[0099]

[0100] When m = n, g0= g 1n = g 2n = 0, where g0, g 1n , g 2n are the undetermined coefficients of the quadratic model;

[0101] (4a4) Quadratic interaction model Q is:

[0102]

[0103] When m = n, d0= d 1n = d 2n = f mn = 0, where d0, d 1n , d 2n , f mn are the undetermined coefficients of the quadratic interaction model;

[0104] Solve the l e equations in each training sample subset to obtain the undetermined coefficients of all feature values in each training sample subset under the four VPM models;

[0105] (4b) Feature values after adding weights select one of the four VPM models as the model of the ith feature value of the cth sample subset model The model is selected according to the sum of squares and the variance formula is as follows:

[0106]

[0107] In the formula, w im is the actual feature value is the corresponding weight, is the predicted feature value, and the model type that minimizes the sum of squares SSE is selected as the corresponding model

[0108] (4c) Complete the selection of the model for all features of all training sample subsets in the training sample set, store the type and parameters of the model , and aggregate all models to obtain an improved VPM model:

[0109]

[0110] In the formula, the improved VPM model is composed of C * ·M models .

[0111] The step (5) specifically includes the following steps:

[0112] (5) Select a segmentation method that meets the accuracy requirement according to the accuracy rate, segment the test sample set to obtain a set of test sample subsets, and predict each test sample subset according to the improved VPM model to obtain a preliminary prediction result of the life assessment;

[0113] (5a) For the test sample set , a reverse segmentation method is used to segment the test sample set from the end of the data to the front, and the length l e of the test sample subset is consistent with the length of the training sample subset, and a set of test sample subsets is obtained, denoted as:

[0114]

[0115] (5b) According to the improved VPM model, the multiple test sample subsets contained in Ψ' are respectively predicted for life, and the test sample subset prediction standard is set as follows:

[0116]

[0117] In the formula, Let c′ be the true feature value of the i-th dimension in the test sample subset. To obtain the predicted feature value of the i-th dimension in the c-th test sample subset of the improved VPM model, the preliminary prediction results of all test sample subsets in Ψ′ are finally obtained by solving the problem, denoted as {c1,c2,…}.

[0118] Step (6) specifically refers to: using the trend extrapolation method, combining the preliminary prediction results, to obtain the final life prediction value of the complex electromechanical equipment. The trend extrapolation method adopts the linear trend extrapolation method, and its formula is:

[0119]

[0120] In the formula, The final lifespan prediction is given by t, where t is the time variable. Based on the preliminary prediction results {c1, c2, ...} of the test sample subset, a curve is fitted to determine the coefficients a and b, thus yielding the final lifespan prediction.

[0121] Example 1

[0122] For various temperature sensors that are continuously degrading, such as HP outlet total temperature, HPT outlet total temperature, LPC outlet total temperature, LPT outlet total temperature, and fan inlet total temperature, etc.

[0123] Perform preprocessing operations such as segmentation and normalization on all degraded data → perform multiple clustering to obtain a new training set → improve the VPM model for discriminant analysis → obtain lifetime assessment results.

[0124] In summary, this invention, based on the traditional VPM model, fits the probability density curve of the independent variable, using the probability of the independent variable as the weight w of the sum and variance. im This reduces the impact of outlier samples on classification accuracy. Traditional VPM classification processes classify each sample separately. This invention uses an improved VPM model to calculate the similarity of probability density curves for a segment of measured continuous time series data, thereby greatly reducing the impact of outlier samples. This invention uses an improved VPM model to jointly determine multi-sensor data, which can effectively eliminate misjudgments caused by sensor malfunctions.

Claims

1. A hybrid lifetime prediction method for complex electromechanical devices, characterized by: The method comprises the following steps in sequence: (1) preprocessing the whole life data collected by the complex electromechanical equipment, dividing the preprocessed data into a training sample set and a test sample set, pre-segmenting the training sample set and adding labels to obtain a training sample set with labels; (2) clustering the training sample set with labels by using a clustering method, and then selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment, to obtain a set of training sample subsets under the segmentation mode; (3) fitting a probability curve for each feature value of all training sample subsets in the set of training sample subsets, selecting a best fitting curve in the probability curve, and setting the horizontal and vertical coordinates of the best fitting curve as a "feature value-weight value" pair of the training sample subset; (4) According to the corresponding relationship between the eigenvalue and the weight, equations are established for all eigenvalues in each training sample subset, parameters of all eigenvalues under four VPM model types are determined, and each model is obtained according to the standard of variance , and all models are summarized to obtain an improved VPM model, and the four VPM models are linear model, linear interaction model, quadratic model and quadratic interaction model, respectively; (5) selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment the test sample set, to obtain a set of test sample subsets, and predicting each test sample subset according to the improved VPM model to obtain a preliminary prediction result of the life assessment; (6) obtaining a final life prediction value of the complex electromechanical equipment by using a trend extrapolation method and comprehensively analyzing the preliminary prediction result; The step (4) specifically comprises the following steps: (4a) a set of training sample subsets all the feature values in each training sample subset are established equations, according to the "feature value-weight value" pair of feature values the weight value , multiply all the feature values by the weight value , and solve the undetermined coefficients of all the feature values in each training sample subset in the following four VPM models: (4a1) the linear model LI is: ， When Time, where, , In , represent the first , eigenvalues, dimension of the training sample, , , , respectively , weight of, , undetermined coefficient of the linear model; (4a2) the linear interaction model L is: ， When time, wherein, , , are the undetermined coefficients of the linear interaction model; (4a3) the quadratic model QI is: ， When time, wherein, , , are the undetermined coefficients of the quadratic model; (4a4) the quadratic interaction model Q is: ， When Time, where, , , , is the undetermined coefficient of the quadratic interaction model; Solve the equations for the unknown coefficients of all the feature values in each training sample subset under the four VPM models. Solve the equations for the unknown coefficients of all the feature values in each training sample subset under the four VPM models. (4b) Eigenvalues after adding weights select one of the four VPM models as the model for the dimensional eigenvalues of the subset of samples , the model is selected based on the sum of squares and variance, and the sum of squares and variance formula is as follows: ， In the formula, is the actual characteristic value is the corresponding weight value, is the predicted characteristic value, and the model type with the minimum sum of squares of errors (SSE) is selected as is the corresponding model ; (4c) Complete the model for all features of all subsets of training samples in the training sample set. Choice of storage model The type and parameters, and all models The improved VPM model is summarized as follows: ， In the formula, the improved VPM model is composed of one model .

2. The hybrid lifetime prediction method of complex electromechanical devices according to claim 1, characterized in that: The step (1) specifically comprises the following steps: (1a) Interpolation and trend extrapolation method is used to complete the incomplete full life data, and all full life data are normalized pretreated, the pretreated data are divided to obtain training sample set of dimension life data with sample number of and test sample set , wherein, , , represents a positive natural number;​ (1b) pre-segmentation, setting initial length of training sample subset The training sample set is segmented into segments containing a collection of samples, (1c) with reference to the sampling time, adding labels to the segmented set one by one in turn , and taking the union of the labeled set to obtain the labeled training sample set.

3. The hybrid lifetime prediction method of complex electromechanical devices according to claim 1, characterized in that: The step (2) specifically comprises the following steps: (2) clustering the training sample set with labels by using a clustering method, and then selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment, to obtain a set of training sample subsets under the segmentation mode; (2a) setting an accuracy threshold value , the number of clustering centers of the FCM algorithm ; (2b) The accuracy of the labeled training sample set is verified by using the FCM algorithm, and the accuracy under The evaluation index of the accuracy is the clustering purity, and the calculation formula is as follows:​ （1） wherein, represents the number of samples in the labeled training sample set, is the number of training samples in the clustering result whose predicted label is consistent with the true label, the true label is the label added after the training sample set is pre-segmented, and the predicted label refers to the label after the FCM algorithm clustering, and the cluster obtained after clustering is defined as the predicted label of the cluster according to the true label with the largest proportion. (2c) if , then reduce the number of cluster centers , let , pre-segment and add labels again, cluster the labeled training sample set formed again, and verify the accuracy of FCM clustering through formula (1); reduce the number of segments in the manner of increasing the length of the training sample subset , if the length of the training sample subset still does not satisfy , then return to step (2a), reduce the accuracy threshold , reduce the accuracy threshold by one percent, reset the number of cluster centers , and verify the accuracy of FCM clustering again through formula (1), repeat step (2c) until the number of segments that satisfy is obtained ; If , then the segment at this time is the number of segments satisfying the accuracy requirement ; (2d) obtaining a set of training sample subsets under the individual segment wherein, is the i-th training sample subset in the set of training sample subsets, and the label value is , .​ 4. The hybrid lifetime prediction method of complex electromechanical devices according to claim 1, characterized in that: The step (3) specifically comprises the following steps: (3) fitting a probability curve for each feature value of all training sample subsets in the set of training sample subsets, selecting a best fitting curve in the probability curve, and setting the horizontal and vertical coordinates of the best fitting curve as a "feature value-weight value" pair of the training sample subset; (3a) Introducing probability density functions, calculating the similarity of the probability density curves of the feature values of all training sample subsets in the set The similarity of the probability density curves of the feature values of all training sample subsets in the set is calculated, and the mean square error is used as the standard for selecting the probability density curve.​ ， wherein is the true value, is the predicted value, is the length of the current training sample subset; (3b) selecting the probability density curve with the smallest mean square error and storing all "value - weight" pairs, i.e. "eigenvalue - weight" pairs, within the segments, the i-th training sample subset of the set of training sample subsets.​​​​​​​​ 5. The hybrid lifetime prediction method of complex electromechanical devices according to claim 1, characterized in that: The step (5) specifically comprises the following steps: (5) selecting a segmentation mode meeting the accuracy requirement according to the accuracy to segment the test sample set, to obtain a set of test sample subsets, and predicting each test sample subset according to the improved VPM model to obtain a preliminary prediction result of the life assessment; (5a) For the test sample set , segment the test sample set using a reverse segmenting approach that segments from the end of the data forward, the test sample subset length is kept consistent with the training sample subset length, resulting in a collection of test sample subsets, denoted as: ； (5b) to a plurality of test sample subsets contained in the middle , life prediction is carried out according to the improved VPM model respectively, and the test sample subset prediction standard is set as follows: ， wherein is the first segment test sample subset dimensional true eigenvalue, is the first dimensional predicted eigenvalue in the improved VPM model for the segment test sample subset, by solving , the preliminary prediction results for all test sample subsets in are obtained, denoted as 6. The hybrid lifetime prediction method of complex electromechanical devices according to claim 1, characterized in that: The step (6) specifically refers to: obtaining a final life prediction value of the complex electromechanical equipment by using a trend extrapolation method and comprehensively analyzing the preliminary prediction result, wherein the trend extrapolation method is a linear trend extrapolation method, and the formula is: ， wherein is the final life prediction value, is the time variable, according to the preliminary prediction results of the test sample subset , fitting the curve, determining the coefficient , , thus obtaining the final life prediction value .

Images