Smart meter reliability evaluation method based on survival analysis

By processing smart meter data using survival analysis, and combining Elastic Net-Cox and a three-parameter Weibull model with a Bayesian optimized random survival forest algorithm, the accuracy and efficiency issues of smart meter reliability assessment are solved, state change transitions are supported, and management level is improved.

CN116522241BActive Publication Date: 2025-11-07HUNAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202310485924.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-04-28
Publication Date
2025-11-07
Estimated Expiration
2043-04-28

AI Technical Summary

Technical Problem

Existing technologies for assessing the reliability of smart meters rely on outdated manuals and complex, time-consuming testing methods, leading to inaccurate assessment results and wasted resources.

Method used

A survival analysis-based approach is employed, using target encoding and box plots to process data, Elastic Net-Cox for dimensionality reduction, and a three-parameter Weibull model and random survival forest algorithm to evaluate the reliability of smart energy meters. This includes Bayesian optimization of the random survival forest algorithm to improve evaluation accuracy.

Benefits of technology

It enables accurate assessment of the reliability of smart energy meters, supports the transition from periodic rotation to state replacement, and improves management level and resource utilization efficiency.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116522241B_ABST
    Figure CN116522241B_ABST
Patent Text Reader

Abstract

The application discloses a kind of wind turbine power characteristics evaluation method based on cloud model and optimal combination weighting based on survival analysis intelligent electric energy meter reliability evaluation method, step S101, electric energy meter original data acquisition and category characteristic numerical processing;Step S102, category characteristics are reduced dimension processing;Step S103, evaluate batch intelligent electric energy meter reliability;Step S104, evaluate the reliability of individual intelligent electric energy meter;The method can realize the reliability analysis of batch intelligent electric energy meter, and the survival rate of each individual intelligent electric energy meter is evaluated, which helps electric power department to find the abnormal condition of intelligent electric energy meter in time, provides theoretical support for the transformation of intelligent electric energy meter from "expiration rotation" to "state replacement", and improves the lean operation and maintenance management level of intelligent electric energy meter.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The application relates to a reliability evaluation method for an intelligent electric energy meter based on survival analysis. BACKGROUND

[0002] As one of important terminal devices of power internet of things, the intelligent electric energy meter plays a key role in comprehensively sensing the state of power grid construction and efficiently processing information. With the high-speed development in China for many years, the currently large number of intelligent electric energy meters in operation have reached the replacement period. The DLT448-2016 Technical Management Regulation for Electric Energy Metering Device clearly specifies that the operation reliability of the electric energy metering device is maintained in the form of field inspection, periodic replacement and sampling inspection, and the JJG596-2012 Electronic AC Electric Energy Meter specifies that the calibration period of the electric energy meter of the first level and the second level is generally 8 years. Due to the technical progress and the improvement of manufacturing technology, the actual service life of the intelligent electric energy meter can generally reach 10-15 years. Based on the JJG596-2012 calibration regulation, a large number of intelligent electric energy meters with good performance are forced to be replaced, which not only wastes resources but also needs power-off processing, thereby bringing a lot of inconvenience to customers. Therefore, it is necessary to carry out the reliability research on the intelligent electric energy meter.

[0003] The reliability evaluation of the intelligent electric energy meter is one of important methods for realizing the conversion of the intelligent electric energy meter from the periodic replacement to the state replacement, and in recent years, many experts have carried out researches on the reliability engineering of the intelligent electric energy meter. At present, some scholars research the reliability of the intelligent electric energy meter based on the stress method of the component, but the method depends on the reliability prediction manual, the updating speed of the manual is generally lagging behind the updating speed of the intelligent electric energy meter, and the accuracy of the final result is affected. Some scholars analyze the service life of the intelligent electric energy meter through the method of the accelerated life test, but the test process is complex and time-consuming and labor-consuming, and the deviation between the test condition and the real operation environment will cause errors in the experimental results.

[0004] Therefore, it is necessary to design a reliability evaluation method for an intelligent electric energy meter based on survival analysis. SUMMARY

[0005] The technical problem to be solved by the application is to provide a reliability evaluation method for an intelligent electric energy meter based on survival analysis, which can realize the reliability analysis of the intelligent electric energy meter and provide theoretical support for the conversion of the intelligent electric energy meter from the periodic replacement to the state replacement.

[0006] The technical solution of the application is as follows:

[0007] A reliability evaluation method for an intelligent electric energy meter based on survival analysis comprises the following steps:

[0008] Step S101, electric energy meter original data acquisition and category feature value processing Collect the basic information, production information and operation information of the smart electric energy meter, use target coding and box plot to perform category feature value processing on the smart electric energy meter original data and eliminate abnormal values;

[0009] The basic information includes voltage, current, accuracy, communication protocol, etc., the production information includes manufacturer, batch, etc., and the operation information includes operation area, state, fault phenomenon, meter age, etc.;

[0010] Step S102, dimension reduction processing is performed on the category features;

[0011] The processed category features are subjected to Elastic Net-Cox dimension reduction processing to screen feature indexes related to the performance of the smart electric energy meter and reduce the influence of redundant features on the calculation efficiency and accuracy of the smart electric energy meter reliability evaluation model;

[0012] Step S103, evaluating the batch smart electric energy meter reliability

[0013] The three-parameter Weibull model is used to evaluate the reliability of the batch smart electric energy meter;

[0014] Step S104, evaluating the reliability of the individual smart electric energy meter

[0015] On the basis of obtaining the reliability of the batch smart electric energy meter, in order to further realize the accurate replacement of the smart electric energy meter, the individual survival rate of some batch smart electric energy meters with low overall reliability is calculated, a Bayesian optimization random survival forest method is proposed to evaluate the survival rate of each smart electric energy meter based on the basic framework of the random survival forest model and the optimization algorithm of Bayesian, and the survival rate curve of the individual smart electric energy meter is obtained, thereby reflecting the reliability of the individual smart electric energy meter.

[0016] Target coding is a coding method based on the category features of the variable value and the category variable of the corresponding dependent variable, which replaces the category features with the combination of the target expected value of the dependent variable for a given specific category value and the target expected value of the dependent variable on all training data, and target coding does not change the number of features;

[0017] X k ' = prior_prob x (1-smoove) + smoove x condition_prob (Formula 1)

[0018] In the formula,

[0019] X k ' represents the coding value of the category k in the category feature X of the smart electric energy meter;

[0020] The prior_prob represents the prior probability of the target variable, which is obtained in Python by prior_prob = train_y.mean(), where train_y represents the labels of the training set in the smart meter data, which is composed of the running time and state of the smart meter.

[0021] The condition_prob represents the conditional probability of the target variable.

[0022]

[0023] The smoove is a reliability coefficient, and the expression is

[0024]

[0025] In the formula, n represents the number of samples with category k in the smart meter category feature X, n + represents the number of samples with category k and positive results in the smart meter category feature X, min_sample_leaf and smoothing are user-defined parameters, min_sample_leaf takes the minimum sample number when calculating the average value of the smart meter category, and the default value is 1, smoothing is a smoothing coefficient used to balance the classification average value and the prior average value, and the larger the value, the stronger the regularization, and the default value is 1.

[0026] The box plot is a statistical chart for displaying data dispersion, which uses five statistical quantities, maximum Q4, upper quartile Q3, median Q2, lower quartile Q1, and minimum Q0, to describe data distribution, and its principle is shown in Figure 2 The running time t of the smart meter in the same batch is taken as the detection object, and the box plot is used to detect and eliminate outliers in batches.

[0027] The expressions of the upper threshold UL and the lower threshold LL are

[0028] UL = Q3 + 1.5 x IQR (Formula 4)

[0029] LL = Q1 - 1.5 x IQR (Formula 5)

[0030] In the formula, IQR represents the difference between the upper quartile Q3 and the lower quartile Q1, i.e. the interquartile range, and the calculation formula is: IQR = Q3 - Q1. The abnormal data greater than the upper threshold UL or less than the lower threshold LL is eliminated.

[0031] In step 102, the features after numerical value are processed by Elastic Net-cox for dimension reduction, and the variable selection of the smart meter fault factors is completed. The Cox model is:

[0032]

[0033] In the formula, h represents a risk function, and t represents the running time of the smart electric energy meter.

[0034] X = (X1, X2, …, X p ) T is a covariate, and is a category feature of the smart electric energy meter obtained through data processing and feature screening in steps 1 and 2, and the category feature includes at least two of the following: a fault phenomenon, a constant, a communication protocol, a current, a sorting person, a meter specification, a wiring mode, a manufacturer code, an accuracy, a type, a metering direction, a rate, a manufacturing standard, a model, a unit for returning, a hardware version, and a user category;

[0035] X i = (X i1 , X i2 , …, X ip ) T is the pth covariate of the ith individual;

[0036] h0(t) is a baseline risk function, and the value of the baseline risk function is a risk function value when all covariates take the value 0 at the t moment;

[0037] ω = (ω1, ω2, …, ω p ) T is a regression coefficient; the regression coefficient is a coefficient to be solved, and is solved through formula 6;

[0038] The parameters of the Cox model are usually calculated by using a partial likelihood function, and the partial likelihood function of the Cox model is as follows:

[0039]

[0040] where δ i is an indicator function, δ i = 0 when an event is censored, δ i = 1 when an event occurs, and H i is a risk set of individuals at the t i moment; the Elastic Net estimation of the Cox model is defined by minimizing the reciprocal of the partial log-likelihood function and adding a suitable penalty term;

[0041]

[0042] In the formula, υ1 and υ2 are non-negative adjustment parameters, the values of υ1 and υ2 are determined to be 0.81 and 0.05 respectively through a 5-fold cross-validation method, ||ω||2 represents an L2 norm, and ||ω||1 represents an L1 norm.

[0043] Dimension reduction principle: the Elastic Net-Cox method compresses the regression coefficient ω of redundant features to 0, so as to remove redundant features for feature selection, and then realizes dimension reduction.

[0044] In step S103, the reliability of the batch smart electric energy meter is evaluated by using the three-parameter Weibull model, and the time node at which the smart electric energy meter enters the loss failure period is calculated. The method of average rank and median rank is introduced to optimize the calculation of reliability, and the gray estimation method is combined to solve the scale parameter, shape parameter and position parameter in the three-parameter Weibull model. According to the above, the reliability function and reliability curve of the batch smart electric energy meter are obtained to reflect the reliability of the batch smart electric energy meter.

[0045] An important evaluation index of the reliability of the smart electric energy meter is the reliability, which represents the probability of completing the specified task under the specified time and conditions. The reliability function of the three-parameter Weibull model is:

[0046]

[0047] In the formula, R is the reliability; t represents the running time of the smart electric energy meter, β is the scale parameter, the larger the scale parameter is, the more dispersed the distribution is, α is the shape parameter, and γ is the position parameter.

[0048] The three parameters of the three-parameter Weibull distribution model are

[0049]

[0050] Calculation process:

[0051] Assuming that T represents the service life of the smart electric energy meter, and L=T-τ represents the remaining life of the smart electric energy meter after time τ, the distribution function and probability density function of the remaining life can be represented as:

[0052]

[0053] The expectation of the remaining life of the smart electric energy meter after time τ is:

[0054]

[0055] For the scale parameter β, shape parameter α and position parameter γ in the Weibull model, the gray estimation method is used for modeling and solving.

[0056] Transforming the formula , the following formula can be obtained:

[0057] After transforming the reliability function, the following formula is obtained:

[0058]

[0059] Let

[0060] x i = ln(-ln(R(t i )))(Equation 9)

[0061]

[0062] Assuming that β = c, 1 / α = -a, γ = b, then equation (10) is transformed into

[0063]

[0064] ln is the natural logarithm with e as the base.

[0065] Assume that (T ti ,x i ) is a time series that meets the gray model, and the scale parameter β, the shape parameter α and the location parameter γ are solved by the gray model theorem. Assuming that the running time of m smart electric energy meters is t1, t2, …, t m , then the original data is T (0) = (t (0) (t1), t (0) (t2), …, t (0) (t m )), and the calculation is performed according to the properties of the gray model.

[0066]

[0067] From this, y and z are calculated, b = z / y, a = y, the parameters a, b are substituted into equation (11) to calculate c, and the three parameters of the three-parameter Weibull distribution model are

[0068]

[0069] Equation (15) is substituted into equation (7) to obtain the reliability function of the smart electric energy meter, and thus the reliability of the batch smart electric energy meter at each time.

[0070] In step 101, to avoid the situation that the numerical values of features differ greatly or the variance of a certain feature is several orders of magnitude larger than that of other features, making some algorithms unable to learn other features, the min-max normalization method is used to normalize the data, and the data is mapped to [0, 1] by transforming the original data, and the min-max normalization formula is:

[0071]

[0072] In the formula, max is the maximum value of a column, min is the minimum value of a column, and the new sequence y1, y2, …, y nis a value between 0 and 1.

[0073] Random survival forest is composed of binary survival trees, and its basic idea is consistent with that of decision tree. When data passes through the split node of the tree, the split criterion is to maximize the survival difference. The specific process of the random survival forest algorithm is as follows:

[0074] (1) Bootstrap is used to randomly extract the original data as sub-sample data, and 63% of the data in the sample is set as experimental data, and 37% of the data is set as out-of-bag data (OOB).

[0075] (2) Each sub-sample grows into a binary survival tree model. At each split node, the split criterion is to maximize the survival difference. The log-rank score is used as the split criterion. When the terminal node is not less than the threshold value set by the algorithm, the tree model is grown. Otherwise, stop growing the tree model.

[0076] (3) The cumulative hazard function (CHF) of each tree is calculated by using the Nelson-Aalen estimation. Then, the average cumulative hazard function of the model can be calculated by using the cumulative hazard function of each tree.

[0077] (4) The prediction accuracy of the model is tested by using the out-of-bag data (OOB). The consistency index (C-index) is used to measure the prediction effect of the model.

[0078] At the same time, in order to improve the accuracy of the model, the Bayesian method is used to adjust the hyperparameters of the random survival forest. The Bayesian adjustment uses Gaussian process, considers the previous parameter information, constantly updates the prior, and has less iteration times and fast speed.

[0079] Beneficial effects:

[0080] The application discloses a kind of reliability evaluation method of intelligent electric energy meter based on survival analysis, it is related to the field of reliability engineering of intelligent electric energy meter, comprising: step S101, from the MDS of power grid company, the basic information such as voltage, current, accuracy, communication protocol of intelligent electric energy meter is collected by intelligent meter quality analysis system, manufacturer, batch and other production information, and operating area, state, fault phenomenon, meter age and other operating information, the original data of intelligent electric energy meter is carried out category characteristic numerical processing and eliminates abnormal value using target coding and box plot;Step S102, the features after processing are processed using Elastic Net-cox, and the feature indexes related to the performance of intelligent electric energy meter are screened, to reduce the influence of redundant features on the calculation efficiency and accuracy of intelligent electric energy meter reliability evaluation model;Step S103, the reliability of batch intelligent electric energy meter is evaluated using three-parameter Weibull model, and the time node that intelligent electric energy meter enters loss failure period is calculated, the method of introducing average rank and median rank is used to optimize the calculation of reliability, and the scale parameter, shape parameter and position parameter in three-parameter Weibull model are solved by combining the method of grey estimation, so that the reliability function and reliability curve of batch intelligent electric energy meter are obtained to reflect the reliability of batch intelligent electric energy meter;Step S104, on the basis of obtaining the reliability of batch intelligent electric energy meter, for further realization of accurate replacement of intelligent electric energy meter, the individual survival rate of some batch intelligent electric energy meter with low overall reliability is calculated, the hyperparameters in random survival forest are optimized using Bayesian as optimization algorithm based on the basic framework of random survival forest model, the method of Bayesian optimization random survival forest is proposed to evaluate the survival rate of each intelligent electric energy meter, and the survival rate curve of intelligent electric energy meter individual is obtained, so as to reflect the reliability of intelligent electric energy meter individual.This method can realize the reliability analysis of batch intelligent electric energy meter, and evaluate the survival rate of each intelligent electric energy meter individual, which helps power department to find the abnormal condition of intelligent electric energy meter in time, provides theoretical support for the change of intelligent electric energy meter from "expiration replacement" to "state replacement", and improves the lean operation and maintenance management level of intelligent electric energy meter.

[0081] The application establishes a reliability analysis model according to the operation state, fault data and attribute information of intelligent electric energy meter, and evaluates the reliability of intelligent electric energy meter based on a data-driven method, which is beneficial to realize accurate replacement and lean management of intelligent electric energy meter.

[0082] The method carries out category feature numerical value processing and removes abnormal values based on a target coding method and a box plot, performs dimension reduction processing on the processed features by using Elastic Net-cox, screens feature indexes related to the performance of the smart electric energy meter, evaluates the reliability of the smart electric energy meter by using a three-parameter Weibull model, calculates the time node of the smart electric energy meter entering the loss failure period, introduces the method of average rank and median rank to optimize the calculation of reliability, solves the scale parameter, shape parameter and position parameter in the three-parameter Weibull model by combining the method of grey estimation, adopts the Bayesian optimization random survival forest algorithm to improve the accuracy of the survival rate evaluation of each smart electric energy meter, and finally proposes a smart electric energy meter reliability evaluation method based on survival analysis, realizes the reliability analysis of the smart electric energy meter, and provides theoretical support for the change of the smart electric energy meter from periodical rotation to state replacement.

[0083] The method carries out category feature numerical value processing and removes abnormal values based on a target coding method and a box plot, performs dimension reduction processing on the processed features by using Elastic Net-cox, screens feature indexes related to the performance of the smart electric energy meter, evaluates the reliability of the smart electric energy meter by using a three-parameter Weibull model, calculates the time node of the smart electric energy meter entering the loss failure period, introduces the method of average rank and median rank to optimize the calculation of reliability, solves the scale parameter, shape parameter and position parameter in the three-parameter Weibull model by combining the method of grey estimation, adopts the Bayesian optimization random survival forest algorithm to improve the accuracy of the survival rate evaluation of each smart electric energy meter, and finally proposes a smart electric energy meter reliability evaluation method based on survival analysis, realizes the reliability analysis of the smart electric energy meter, and provides theoretical support for the change of the smart electric energy meter from periodical rotation to state replacement. BRIEF DESCRIPTION OF DRAWINGS

[0084] Figure 1 It is an implementation flowchart of the present application;

[0085] Figure 2 It is a principle diagram of removing abnormal values by a box plot

[0086] Figure 3 It is a feature after variable screening;

[0087] Figure 4 It is a smart electric energy meter reliability evaluation model;

[0088] Figure 5 It is a reliability evaluation curve of batch smart electric energy meters;

[0089] Figure 6 Random survival forest network construction;

[0090] Figure 7Survival curve of individual smart electric energy meter. DETAILED DESCRIPTION

[0091] The application will be further described in detail below in combination with the drawings and specific examples:

[0092] Example 1

[0093] The technical solutions of the application will be further described below in combination with Figures 1-5 The technical solutions of the application will be further described below in combination with

[0094] The application provides a method for accurately evaluating the reliability of a smart electric energy meter, and proposes a smart electric energy meter reliability evaluation method based on survival analysis. Theoretical analysis and actual measurement results show that, according to the operation state, fault data and attribute information of the smart electric energy meter, the category feature is numerically processed and the abnormal values are removed through a target coding method and a box plot, the processed features are subjected to dimension reduction processing by using Elastic Net-cox, the feature indexes related to the performance of the smart electric energy meter are screened, the reliability of the smart electric energy meter is evaluated by using a three-parameter Weibull model, the time node at which the smart electric energy meter enters the loss failure period is calculated, the method of average rank and median rank is introduced to optimize the calculation of reliability, the scale parameter, shape parameter and position parameter in the three-parameter Weibull model are solved by combining the method of grey estimation, the Bayesian optimization random survival forest algorithm is used to improve the accuracy of the survival rate evaluation of each smart electric energy meter, and finally a smart electric energy meter reliability evaluation method based on survival analysis is proposed, so that the reliability analysis of the smart electric energy meter is realized, and theoretical support is provided for the change of the smart electric energy meter from periodical rotation to state replacement.

[0095] In step S101, the smart meter quality analysis system collects the basic information of the smart electric energy meter such as voltage, current, accuracy and communication protocol, the production information such as manufacturer and batch, and the operation information such as operation area, state, fault phenomenon and meter age from the MDS of the power grid company, and the original data of the smart electric energy meter is subjected to category feature numerical processing and abnormal value removal by using target coding and a box plot.

[0096] The original data and information of the smart electric energy meter need to remove missing and garbled data, and the low-variance features and collinearity features existing in the data are removed for preliminary data dimension reduction. The existing smart electric energy meter data set contains many category feature data such as wiring mode, communication protocol, manufacturer code, disassembly unit, fault phenomenon, manufacturing standard and smart electric energy meter state. The category feature numericalization of the smart electric energy meter is carried out by using a target coding method. The target coding is a coding method based on the category feature of a variable value and the category variable of a corresponding dependent variable. The category feature is replaced by the combination of the target expected value of the dependent variable of a certain specific category value and the target expected value of the dependent variable on all training data. After target coding, the number of features will not change.

[0097] X k = prior_prob x (1 - smoove) + smoove x condition_prob (Formula 1)

[0098] In the formula, X k ' represents the encoding value of the category k in the smart meter category feature X, prior_prob represents the prior probability of the target variable, which is obtained by prior_prob = train_y.mean() in python, where train_y represents the label of the training set in the smart meter data, which is composed of the running time and state of the smart meter, condition_prob represents the conditional probability of the target variable, and smoove is the reliability coefficient, which is expressed as

[0099]

[0100] In the formula, n represents the number of samples of category k in the smart meter category feature X, n + represents the number of samples of category k in the smart meter category feature X with positive results, min_sample_leaf and smoothing are self-defined parameters, min_sample_leaf takes the minimum sample number when calculating the average value of the smart meter category, and the default value is 1, smoothing is a smoothing coefficient used to balance the classification average value and the prior average value, and the larger the value, the stronger the regularization, and the default value is 1.

[0101] To avoid the situation that the numerical value of a feature is greatly different from other features or the variance of a certain feature is several orders of magnitude larger than other features, making some algorithms unable to learn other features, min-max normalization method is used to normalize the data, which maps the data to [0, 1] by transforming the original data, and the min-max normalization formula is:

[0102]

[0103] In the formula, max is the maximum value of a column, min is the minimum value of a column, and the new sequence y1, y2, …, y n are values between 0 and 1.

[0104] As a statistical chart showing data dispersion, the box plot describes data distribution using five statistical quantities: maximum value Q4, upper quartile Q3, median Q2, lower quartile Q1, and minimum value Q0, and considers outliers as extreme values in the data sequence, and the principle is as Figure 2 shown.

[0105] The expressions for the upper threshold (UL) and lower threshold (LL) are as follows:

[0106] UL=Q3+1.5×IQR (Formula 5)

[0107] LL = Q1 - 1.5 × IQR (Formula 6)

[0108] Remove outlier data that is greater than the upper threshold UL or less than the lower threshold LL.

[0109] Step S102: The processed features are dimensionality reduced using Elastic Net-cox to filter feature indicators related to the performance of smart energy meters, thereby reducing the impact of redundant features on the computational efficiency and accuracy of the smart energy meter reliability assessment model.

[0110] The processed features were then subjected to dimensionality reduction using Elastic Net-cox to complete the variable selection for fault factors in smart energy meters. The Cox model is as follows:

[0111]

[0112] In the formula, h represents the risk function, and X = (X1, X2, ..., X...). p ) T As covariates, X i =(X i1 ,X i2 ,…,X ip ) T Let p be the p covariates of the i-th individual, and h0(t) be the baseline hazard function, which takes the value of the hazard function when all covariates are 0 at time t, ω = (ω1, ω2, ..., ω p ) T is the regression coefficient.

[0113] The parameters of the Cox model are usually calculated using the partial likelihood function. The partial likelihood function of the Cox model is:

[0114]

[0115] Where, δ i For the characteristic function, δ is used when the event is censored. i =0, δ when the event occurs i =1,H i For t i The risk set of an individual at any given moment.

[0116] Minimize the negative of the partial log-likelihood function and add an appropriate penalty term to define the ElasticNet estimate of the Cox model.

[0117]

[0118] wherein u1 and u2 are non-negative tuning parameters, and the values of u1 and u2 are determined to be 0.81 and 0.05 respectively by 5-fold cross-validation.

[0119]

[0120] After dimensionality reduction processing by Elastic Net-cox, 17 features were selected from 38 features, and the feature coefficient values after variable screening are shown in Table 2. Figure 3

[0121] In step S103, the reliability of the batch smart electric energy meter is evaluated by using the three-parameter Weibull model, and the time node at which the smart electric energy meter enters the loss failure period is calculated. The method of average rank and median rank is introduced to optimize the calculation of reliability, and the method of gray estimation is combined to solve the scale parameter, shape parameter and location parameter in the three-parameter Weibull model. According to the above, the reliability function and reliability curve of the batch smart electric energy meter are obtained to reflect the reliability of the batch smart electric energy meter.

[0122] An important evaluation index of the reliability of the smart electric energy meter is the reliability, which represents the probability of completing the specified task under the specified time and conditions. The reliability function of the three-parameter Weibull model is:

[0123]

[0124] wherein t represents the running time of the smart electric energy meter, β is the scale parameter, the larger the scale parameter, the more dispersed the distribution, α is the shape parameter, and γ is the location parameter.

[0125] After transformation of the reliability function, the following is obtained:

[0126]

[0127] Let

[0128] x i = In(-In(R(t i )))(formula 13)

[0129]

[0130] Assuming β=c, 1 / α=-a, and γ=b, formula (14) is transformed into

[0131]

[0132] Suppose (T ti , x i ​) is a time series conforming to the grey model, the scale parameter β, the shape parameter α and the location parameter γ are solved by the grey model theorem. Assuming that the running time of m smart meters is t1, t2, …, t m , the original data of which is T (0) = (t (0) (t1), t (0) (t2), …, t (0) (t m )), which is calculated according to the properties of the grey model.

[0133]

[0134]

[0135] Thus, y and z are calculated, b = z / y, a = y, c is calculated by substituting the parameters a and b into formula (15), and the three parameters of the three-parameter Weibull distribution model are

[0136]

[0137] The reliability function of the smart meter can be obtained by substituting formula (19) into formula (11), and thus the reliability of the batch smart meter at each time can be known.

[0138] According to the above analysis, the grey-three-parameter-based Weibull model is as shown in Figure 4 According to the method, the age data of the cleaned smart meter is analyzed, and the scale parameter β, the shape parameter α and the location parameter γ are calculated by the grey estimation method as shown in Table 1.

[0139] Table 1 Parameter values calculated by each method

[0140]

[0141] According to the scale parameter β, the shape parameter α and the location parameter γ obtained in Table 1, the reliability curve of the batch smart meter is obtained according to the established Weibull model for reliability evaluation, as shown in Figure 5 .

[0142] In step S104, on the basis of obtaining the reliability of the batch smart meter, for further precise replacement of the smart meter, the individual survival rate of some batch smart meters with low overall reliability is calculated, a Bayesian optimization random survival forest method is proposed for evaluating the survival rate of each smart meter by taking the random survival forest model as a basic framework and the Bayesian as an optimization algorithm for optimizing the hyperparameters in the random survival forest, and the survival rate curve of the individual smart meter is obtained, thereby reflecting the reliability of the individual smart meter.

[0143] Random survival forest is composed of binary survival trees, and its basic idea is consistent with that of decision tree. When data passes through the split nodes of the tree, the split criterion is to maximize the survival difference. The specific process of random survival forest algorithm is as follows:

[0144] (1) Bootstrap is used to randomly extract the original data as sub-sample data, and 63% of the data in the sample is set as experimental data, and 37% of the data is set as out-of-bag data (OOB).

[0145] (2) Each sub-sample grows into a binary survival tree model. At each split node, the split criterion is to maximize the survival difference. The log-rank score is used as the split criterion. When the terminal node is not less than the threshold value set by the algorithm, the tree model is grown, otherwise the tree model is stopped.

[0146] (3) The cumulative hazard function (CHF) of each tree is calculated using the Nelson-Aalen estimate. Then the average cumulative hazard function of the model can be calculated using the cumulative hazard function of each tree.

[0147] (4) The prediction accuracy of the model is tested by the out-of-bag data (OOB). Random survival forest uses the consistency index (C-index) to measure the prediction effect of the model.

[0148] According to the above analysis, the network construction framework of the proposed random survival forest is shown in Figure 6 .

[0149] At the same time, in order to improve the accuracy of the model, Bayesian method is used to adjust the hyperparameters of random survival forest. Bayesian adjustment uses Gaussian process, considers the previous parameter information, constantly updates the prior, and has less iteration times and faster speed. The search range and optimal value of each parameter are shown in the following table.

[0150] Table 2 Optimization and selection of structure parameters

[0151]

[0152] After determining the parameters of the random survival forest model, two smart electric energy meters are randomly selected to input the model, and the survival curves of the two smart electric energy meters are as follows: Figure 7As shown, the pre-installation operation performance and reliability of different smart electric energy meters are not much different, but as the running time increases, the survival rate curves of different smart electric energy meters will be different, and as the running time increases, the cumulative risk of the No. 0 smart electric energy meter starts to gradually rise when running to 1800 days, and the survival rate is below 0.1 when running to 2000 days, indicating that the No. 0 smart electric energy meter is likely to fail due to a fault event during this period, and should be replaced in time. In the actual operation of the smart electric energy meter, the artificial inspection data shows that the No. 0 smart electric energy meter fails when running to 1930 days, and the No. 1, No. 2 and No. 3 smart electric energy meters have not failed respectively when running to 1939 days, 1898 days and 2020 days, which are right-censored data.

[0153] Although the present application has been disclosed with reference to the preferred embodiments, the embodiments are not intended to limit the present application. Any equivalent changes or modifications made without departing from the spirit and scope of the present application shall also fall within the protection scope of the present application. Therefore, the protection scope of the present application shall be defined by the content of the claims of the present application.

Claims

1. A method for reliability evaluation of smart meters based on survival analysis, characterized in that, The method comprises the following steps: Step S101, raw data collection of the electric energy meter and category feature value processing Collect basic information, production information and operation information of the smart electric energy meter, and perform category feature value processing and abnormal value elimination on the raw data of the smart electric energy meter by using target coding and a box plot; Step S102, dimension reduction processing on the category features; Perform dimension reduction processing on the processed category features by using Elastic Net-Cox, filter feature indexes related to the performance of the smart electric energy meter, and reduce the influence of redundant features on the calculation efficiency and accuracy of the reliability evaluation model of the smart electric energy meter; In step 102, the processed features are subjected to dimension reduction processing by using Elastic Net-Cox, and variable filtering of the fault factors of the smart electric energy meter is completed, and the Cox model is as follows: In the formula, h represents a risk function, and t represents the operation time of the smart electric energy meter; X = (X1, X2,..., X p ) T is a covariant, and is a category feature of the smart electric energy meter obtained through data processing and feature screening in steps 1 and 2, and the category feature includes at least two of the following: failure phenomenon, constant, communication protocol, current, sorting person, electric energy meter specification, wiring mode, manufacturer code, accuracy, type, metering direction, rate, manufacturing standard, model, disassembly return unit, hardware version, and user category; X i = (X i1 , X i2 ,..., X ip ) T pi is the pth covariate for the ith individual; h0(t) is a baseline risk function, and the value of h0(t) is the risk function value when all the covariates are 0 at time t; ω = (ω1, ω2,..., ω p ) T are regression coefficients; The parameters of the Cox model are usually calculated by using a partial likelihood function, and the partial likelihood function of the Cox model is as follows: where δ i is an indicator function, δ i = 0 when the event is censored, δ i = 1 when the event occurs, H i is the risk set of individuals at time t i ; the Elastic Net estimation of Cox model is defined by minimizing the negative log-likelihood function with an appropriate penalty term added; In the formula, v1 and v2 are non-negative adjustment parameters, the values of v1 and v2 are determined to be 0.81 and 0.05 by using a 5-fold cross-validation method, ||ω||2 represents an L2 norm, and ||ω||1 represents an L1 norm; Step S103, evaluation of the reliability of the batch of smart electric energy meters The reliability of the batch of smart electric energy meters is evaluated by using a three-parameter Weibull model; Step S104, evaluation of the reliability of the individual smart electric energy meters On the basis of obtaining the reliability of the batch of smart electric energy meters, in order to further realize accurate replacement of the smart electric energy meters, the individual survival rates of the smart electric energy meters in some batches with low overall reliability are calculated, a random survival forest model is used as a basic framework, a Bayesian optimization algorithm is used to optimize the hyperparameters in the random survival forest, a method based on the Bayesian-optimized random survival forest is proposed to evaluate the survival rate of each smart electric energy meter, and a survival rate curve of the individual smart electric energy meters is obtained, so that the reliability of the individual smart electric energy meters is reflected.

2. The method for intelligent electric energy meter reliability evaluation based on survival analysis according to claim 1, characterized in that, Target coding is an encoding method based on the category features and the category variables of the corresponding dependent variables, the category features are replaced by the combination of the target expected value of the dependent variables for a specific category value and the target expected value of the dependent variables on all training data, and the target coding does not change the number of features; X′ k = prior_prob x (1 - smoove) + smoove x condition_prob (Equation 1) In the formula: X k Xk represents the encoding value of the category k in the smart meter category feature X; prior_prob represents the prior probability of the target variable, in python, the value is obtained by prior_prob=train_y.mean(), wherein train_y represents the label of the training set in the smart electric energy meter data, and is composed of the operation time and the state of the smart electric energy meter; condition_prob represents the conditional probability of the target variable; smoove is a reliability coefficient, and the expression is wherein n represents the number of samples of the class k in the smart meter class feature X, n + represents the number of samples of the class k with positive results in the smart meter class feature X, min_sample_leaf and smoothing are self-defined parameters, min_sample_leaf takes the minimum number of samples when calculating the average value of the smart meter class, and the default value is 1, smoothing is a smoothing coefficient for balancing the classification average value and the prior average value, the greater the value, the stronger the regularization, and the default value is 1.

3. The method for reliability evaluation of smart meter based on survival analysis according to claim 1, characterized in that, As a statistical chart for displaying data dispersion, the box plot describes data distribution by using five statistical quantities, i.e., maximum value Q4, upper quartile Q3, median Q2, lower quartile Q1 and minimum value Q0, and takes the running time t of the smart electric energy meter in the same batch as the detection object to detect and eliminate abnormal values in batches by using the box plot; The expressions of the upper threshold limit UL and the lower threshold limit LL are UL=Q3+1.5×IQR (Formula 4) LL=Q1-1.5×IQR (Formula 5) In the formula, IQR represents the difference between the upper quartile Q3 and the lower quartile Q1, i.e., the quartile range, and the calculation formula is IQR=Q3-Q1. Abnormal data greater than the upper threshold limit UL or less than the lower threshold limit LL are eliminated.

4. The method for reliability evaluation of smart meter based on survival analysis according to claim 1, characterized in that, In step S103, the reliability of the batch smart electric energy meter is evaluated by using the three-parameter Weibull model. An important evaluation index of the reliability of the smart electric energy meter is the reliability, which represents the probability of completing a specified task under specified time and conditions. The reliability function of the three-parameter Weibull model is In the formula, R is the reliability; t represents the running time of the smart electric energy meter, β is the scale parameter, and the larger the scale parameter, the more dispersed the distribution; α is the shape parameter, and γ is the position parameter. The three parameters of the three-parameter Weibull distribution model are

Citation Information

Patent Citations

  • Ischemic stroke adverse outcome risk prediction device integrating epigenetic factors

    CN110797120A

  • Machine learning-based extracellular vesicle classification method and device, equipment and medium

    CN114973245A