Intelligent electric energy meter operation state risk assessment method based on strong association rule identification and multi-model integration

Abstract

The invention discloses an intelligent electric energy meter operation state risk assessment method based on strong association rule identification and multi-model integration. The method comprises the following steps that historical operation data of an intelligent electric energy meter is collected in a multi-source mode, data preprocessing is carried out, and then a comprehensive feature database is constructed; using a strong association rule identification model to mine characteristic factors influencing the intelligent electric energy meter from the comprehensive characteristic database; performing fuzzy reasoning on continuous features in the feature factors by using a fuzzy reasoning system, and performing adaptive optimization in a fuzzy reasoning link to obtain a risk assessment value of the continuous features; calculating local saliency of discrete features in the feature factors in real time in a rolling time window, and performing weighted fusion on the local saliency and the reference saliency through an attenuation factor to obtain a risk assessment value of the discrete features; and carrying out weighted summation on the risk assessment values of the continuous features and the risk assessment values of the discrete features to obtain a final state risk assessment result. According to the invention, the state risk assessment accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power systems, in particular to a smart meter operation state risk assessment method based on strong association rule identification and multi-model integration. BACKGROUND

[0002] In recent years, with the continuous growth of electricity demand, smart meters as the key hub between the power system and the end user to realize energy metering and information interaction, their stable and accurate operation is essential to ensure the reliability of the entire power supply network, maintain the dynamic balance of the power system and promote the optimization of energy allocation. However, smart meters also face various potential problems and challenges during long-term operation, such as inaccurate data collection, communication failure, security vulnerabilities, etc., which can seriously affect the stability and security of the power system. Therefore, accurately and timely positioning the operation status of smart meters is not only related to the stability of the power grid, but also directly affects the energy use experience of users. At present, a considerable number of researchers at home and abroad have conducted relevant research on the operation status assessment of smart meters. These methods, in most cases, only focus on the preliminary collection and simple recording of basic electrical parameters such as voltage, current, and power, and fail to comprehensively investigate the internal and external multi-source potential deep information that affects the operating environment of smart meters. Therefore, these methods can only achieve real-time assessment of the operation health of smart meters in specific environments, and the assessment results of actual smart meter operation status in different real environments are not satisfactory. In this demand background, the association rule identification model capable of effectively identifying high-risk factors from massive data, the fuzzy reasoning system with efficient multi-input operation, and the saliency evaluation system become important tools for accurate state risk assessment of smart meters.

[0003] Currently, the state risk assessment methods using association rule identification, fuzzy reasoning system, and saliency evaluation system operate independently of each other, with a rough structure that cannot comprehensively analyze the collected multi-source input features and achieve joint processing, without considering the influence of input data structure, resulting in low accuracy and high uncertainty of the final state risk assessment. The multi-input operation complexity is high, and the cost is high. The commonly used association rule mining model is the original association rule mining model, which does not consider the distribution characteristics of the collected multi-source input data to reconstruct the important degree evaluation index. The high-risk factors mined by this original model do not have practical reference significance. SUMMARY

[0004] The technical problem to be solved by the present application: In view of the above problems of the prior art, a smart meter operation state risk assessment method based on strong association rule identification and multi-model integration is provided, which can solve the technical problems of low accuracy, high uncertainty, and low real-time efficiency of the existing state risk assessment technology.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is:

[0006] An intelligent electric energy meter operation state risk assessment method based on strong association rule identification and multi-model integration, comprising the following steps:

[0007] Different data sets are obtained by multi-source collection of historical operation data of intelligent electric energy meters, different data sets are respectively preprocessed, and the preprocessed different data sets are constructed into a comprehensive feature database;

[0008] The strong association rule identification model is used to mine feature factors that have an impact on intelligent electric energy meters from the comprehensive feature database through information correlation degree indexes and rule robustness indexes, the information correlation degree indexes include conditional information interaction degree and cross promotion degree, and the rule robustness indexes include time sequence stability degree and enhanced imbalance degree;

[0009] Fuzzy reasoning is performed on the continuous features in the feature factors using a fuzzy reasoning system, and adaptive optimization is performed in the fuzzy reasoning link to obtain the risk assessment value of the continuous features;

[0010] The local saliency of the discrete features in the feature factors is calculated in real time within a rolling time window, and the local saliency and the reference saliency are weighted and fused through a decay factor to obtain the risk assessment value of the discrete features;

[0011] The risk assessment values of the continuous features and the risk assessment values of the discrete features are weighted and summed to obtain the final state risk assessment result.

[0012] Further, when the different data sets are respectively preprocessed, the following steps are included:

[0013] The abnormal values of the current data set are filtered according to the abnormal value satisfaction condition and removed, then the part of the current data set with the removed abnormal values is interpolated, and finally all the input data in the current data set are normalized according to the data magnitude and data range, and are uniformly mapped to the [0, 1] standard interval to realize dimensionless, the formula of the abnormal value satisfaction condition is as follows:

[0014]

[0015] Wherein, x is the value of the input data in the current data set, is the mean value of the input data in the current data set, σ is the standard deviation of the current data set, Q1, Q2 and Q3 respectively represent the value of each quartile point of the current data set;

[0016] Further, the comprehensive feature database includes feature information C1,…,C k corresponding to internal factors in each smart meter operation record k+1 ,…,C n , and feature information C k ,…,C n corresponding to external factors, and further includes an operation state G of each smart meter operation record. When a strong correlation rule recognition model mines feature factors affecting the smart meter from the comprehensive feature database through an information correlation degree index and a rule robustness index, the following steps are included:

[0017] The feature information {C1,…,C k ,…,C n} in the comprehensive feature database is taken as an input variable, and the operation state G in the comprehensive feature database is taken as a target variable.

[0018] Each input variable is taken as a candidate, the support degree of each candidate is calculated, and the candidates with a support degree greater than a minimum support degree are screened to obtain a frequent 1-item set.

[0019] The cross lift degree of each frequent item of the frequent 1-item set and the target variable under a condition variable is calculated, and the condition variable specifically refers to the correlation strength of a specific variable in the analysis variable set to the target state with other variables in the variable set. Then, the frequent items of the frequent 1-item set are sorted according to the cross lift degree to obtain an FR-list.

[0020] An FP-tree is generated according to the FR-list, and the target variable is taken as a leaf node. The condition pattern base of each item is collected by traversing from the leaf node of the FP-tree upwards, and the condition pattern base with a condition information interaction degree greater than zero is retained.

[0021] For the condition pattern base of each frequent item, the cross lift degree of each item in the condition pattern base and the target variable under the condition variable is calculated, and then each item is sorted according to the cross lift degree to obtain a corresponding condition FR-list. Finally, a condition FP-tree is generated according to the condition FR-list, and the frequent pattern of the corresponding frequent item is updated. The condition pattern base of each item is collected by traversing from the leaf node of the condition FP-tree upwards, and the condition pattern base with a condition information interaction degree greater than zero is retained. The above steps are repeated to extract smaller condition pattern bases and build smaller condition FP-trees until it is impossible to build a condition FP-tree.

[0022] The time sequence stability degree of each frequent pattern is calculated, and the frequent pattern with a time sequence stability degree less than a specified value is removed. Then, each frequent item is combined with the corresponding frequent pattern to obtain a frequent pattern set of each frequent item.

[0023] The support, confidence and enhanced imbalance index are calculated for each frequent item set in the frequent pattern set, and the frequent item sets whose support, confidence and enhanced imbalance index are all greater than the corresponding thresholds are selected and merged to obtain strong association rules.

[0024] Furthermore, the formula for cross-lift is as follows:

[0025]

[0026] Where Y is the target variable, X is the input variable, and Sup(X|Z1,Z2,…,Z n ∩Y|Z1,Z2,…,Z n ) represents the factors of all conditional variables {z1,z2,…,z n} occurs, X and Y occur together, Sup(X|z i ) and Sup(Y|z i ) represent the given condition variable factor z i In the case of x and y variable factors in all sample regulations, w i It is z i The weight of , indicating its contribution to the indicator results;

[0027] The formula for conditional information interactivity is as follows:

[0028]

[0029] Where Y is the target variable, X is the input variable, and Z is the conditional variable, Sup(x,y|z i ) represents the factor z under given condition variable i In the case of , the proportion of x variable factors and y variable factors occurring simultaneously in all sample regulations, P(x,y,z) represents the probability of each factor in X, Y, and Z appearing together, Sup(x|z i ) indicates that the x variable is in all the variables containing the conditional variable factor z i The proportion in the sample regulations, Sup(y|z i ) indicates that the y variable is in all the variables containing the conditional variable factor z i The proportion in the sample regulations, CMI(X→Y|Z) represents the information interaction between X and Y under the given conditional variable Z;

[0030] The formula for timing stability is as follows:

[0031]

[0032] Where Y is the target variable, X is the input variable, and T is a time window, Conf t (X,Y) is the confidence between variable X and Y at time t;

[0033] The formula of the enhanced imbalance degree is as follows:

[0034]

[0035] Wherein, Y is a target variable, X is an input variable, and CAII(X→Y) represents the enhanced imbalance level of the association rule X→Y, E represents the type of environmental variables of the smart meter, W c (E) represents the class weight under the specific environment E, N c (E) is the number of class samples reflecting the simultaneous occurrence of variables X and Y under the environment E, Sup(X) represents the support of the X variable, and Sup(Y) represents the support of the Y variable.

[0036] Further, when using a fuzzy inference system to perform fuzzy inference on the continuous features in the feature factors, the method comprises:

[0037] The occurrence density of each continuous feature is calculated, the demarcation points of different health intervals are divided according to the occurrence density distribution for each continuous feature, the membership degrees of data points of each continuous feature in different health intervals are calculated, the fuzzy set corresponding to each continuous feature is obtained, the membership degrees of health intervals at both ends of each continuous feature are calculated using a trapezoidal membership function, and the membership degrees of health intervals at middle positions are calculated using a triangular membership function.

[0038] The fuzzy set of each continuous feature is input into a ladder fuzzy inference system, the ladder fuzzy inference system comprises multiple groups of expert subsystems according to a ladder distribution, each expert subsystem performs fuzzy inference operation on two inputs according to a corresponding fuzzy inference rule table and outputs a corresponding continuous feature health degree.

[0039] Further, when performing adaptive optimization in the fuzzy inference link, the method comprises:

[0040] The occurrence contribution degree of the rule during the operation of each expert subsystem is calculated, and the rules with an occurrence contribution degree lower than a set threshold are removed, and the formula of the occurrence contribution degree of the rule is as follows:

[0041]

[0042] Wherein, ρ i represents the contribution degree of the rule R i , N represents the total number of samples, X j represents the jth input sample, μR i (X j ) represents the rule Ri The fuzzy weight of the input sample X j on the rule R j (X j ) represents the most similar rule R i to the input sample X j The fuzzy weight of the input sample X j on the rule R

[0043] Further, in the adaptive optimization of the fuzzy inference link, including:

[0044] Based on the Bayesian optimization dynamic adjustment of the inference weight between different expert subsystems, ensure that the key features get higher influence in the inference process, and combine the occurrence density calculation of different features to adjust the weight distribution of different features in the inference system, ensure that the weight of different features in the inference system changes dynamically with the data, and the inference weight formula is as follows:

[0045]

[0046] Where ω * represents the optimal inference weight, so that the prediction error L(ω) is minimum, N represents the total number of samples, H pred (ω) represents the predicted health degree calculated based on the current weight ω, H actual represents the actual health degree known in the test;

[0047] The weight formula of different features is as follows:

[0048]

[0049] Where IDF(x i ) represents the occurrence density function value of feature x i , v i is the adaptive weight of feature x i .

[0050] Further, the risk assessment value formula of continuous features is as follows:

[0051]

[0052] Where E is the total number of fuzzy sets existing in the cross of the continuous features in a sample rule in the comprehensive feature database, is the membership function corresponding to the fuzzy set e, x is the feature in the domain of the corresponding fuzzy set e, v i is the adaptive weight of feature x i in the continuous feature.

[0053] Further, the expression of local saliency is as follows:

[0054] Sig ih [t] = Wx ({x ih (τ),y(τ)}|τ∈[t-ΔT,t])

[0055] Among them, x ih (τ) represents the value of the feature ih collected at time τ, y(τ)∈[t-ΔT,t] is the label of whether the corresponding device fails at time τ, 1 represents failure, 0 represents normal, t is the current time, ΔT is the window length; W x () represents the baseline significance weight of the discrete feature, which is expressed as follows:

[0056]

[0057] Among them, s k Represents discrete features, Num(s k ∩g) indicates that the comprehensive feature database also contains discrete features s k The number of sample rules related to fault g, Num(g) represents the number of sample rules of fault occurrence, and n represents the total number of discrete features;

[0058] Furthermore, the risk assessment value formula of discrete features is as follows:

[0059]

[0060] Among them, f2 represents the final risk assessment value composed of all discrete features when a sample regulation of the comprehensive feature database occurs, P k Indicates whether the feature occurs when the sample regulation occurs. If it occurs, P k If it does not happen, P k is 0, S i (t) is the updated saliency, which is expressed as follows:

[0061] S i (t) = γ·W x (s k )+(1-γ)·Sig i [t]

[0062] Among them, Sig i [t] represents the local saliency of feature i, and γ is the attenuation factor.

[0063] Compared with the prior art, the advantages of the present invention are:

[0064] The present invention uses a strong association rule identification model to mine strong association rules from a comprehensive feature database to obtain characteristic factors that affect smart electricity meters. Different models are used to perform risk assessment on different parts of the mined characteristic factors, forming functional complementarity and accurately evaluating the operating status of smart electricity meters.

[0065] The information correlation degree index and the rule robustness index designed in the application are two more effective importance evaluation indexes, which can realize more comprehensive and accurate evaluation of the association rules, and the extracted strong association rules are more thorough, and the quantification of the correlation degree of the association rules becomes more reliable.

[0066] The application performs self-adaptive optimization in the link of continuous feature fuzzy reasoning, can effectively solve the problem of large calculation cost of the traditional fuzzy reasoning system in the multi-input fuzzy derivation process, and further reduces the uncertainty of evaluation.

[0067] The application locally counts the equipment fault data and feature performance in the rolling time window for discrete features, calculates the short-term 'local significance' in real time, and performs weighted fusion with the benchmark significance through the decay factor, so as to realize dynamic and smooth updating of the feature significance, and can more accurately capture the change of the latest fault mode. BRIEF DESCRIPTION OF DRAWINGS

[0068] Figure 1 The flowchart of the method of the embodiment of the application.

[0069] Figure 2 The strong association rule mining flowchart of the embodiment of the application.

[0070] Figure 3 The occurrence density graph of the selected input feature in the embodiment of the application.

[0071] Figure 4 The membership function graph of the input feature in the embodiment of the application.

[0072] Figure 5 The structure diagram of the step fuzzy reasoning system in the embodiment of the application.

[0073] Figure 6 The workflow diagram of the model constructed according to the method of the embodiment of the application.

[0074] Figure 7 The effect comparison diagram of the model constructed according to the method of the embodiment of the application and other models. DETAILED DESCRIPTION

[0075] The application will be further described below in combination with the drawings in the specification and specific preferred embodiments, but the protection scope of the application is not limited thereby.

[0076] In order to solve the technical problems of not joint processing of multi-source input data, not effectively considering heterogeneous characteristics, low accuracy of state risk assessment, high uncertainty, low real-time efficiency and the like in the prior art, the embodiment proposes an intelligent electric energy meter operation state risk assessment method based on strong correlation rule identification and multi-model integration, strong correlation rules are mined from a comprehensive feature database constructed based on multi-source data to obtain feature factors that have an impact on the intelligent electric energy meter, different models are used for risk assessment for different parts of the mined feature factors, and the functions are complementary, so that the operation situation of the intelligent electric energy meter is accurately assessed, and strong real-time performance is achieved.

[0077] As shown in Figure 1 The method of the embodiment includes the following steps:

[0078] S1) Multi-source collection of historical operation data of the intelligent electric energy meter to obtain different data sets, data preprocessing is performed on different data sets respectively, and the preprocessed different data sets are constructed into a comprehensive feature database;

[0079] S2) The strong correlation rule identification model is used to mine feature factors that have an impact on the intelligent electric energy meter from the comprehensive feature database through information correlation degree indexes and rule robustness indexes, the information correlation degree indexes include conditional information interaction degree and cross promotion degree, and the rule robustness indexes include time sequence stability degree and enhanced imbalance degree;

[0080] S3) Fuzzy reasoning is performed on the continuous features in the feature factors using a fuzzy reasoning system, and adaptive optimization is performed in the link of fuzzy reasoning to obtain risk assessment values of the continuous features;

[0081] S4) The local saliency is calculated in real time within a rolling time window for the discrete features in the feature factors, and the local saliency and the reference saliency are weighted and fused through a decay factor to obtain risk assessment values of the discrete features;

[0082] S5) The risk assessment values of the continuous features and the risk assessment values of the discrete features are weighted and summed to obtain a final state risk assessment result.

[0083] The steps will be described in detail below.

[0084] Step S1 of the embodiment specifically includes the following steps:

[0085] S11) Multi-source collection of historical operation data of the intelligent electric energy meter, specifically multi-source collection of historical operation data of the intelligent electric energy meter placed in a real natural environment, thereby obtaining different data sets;

[0086] S12) Data preprocessing is performed on different data sets respectively, specifically data preprocessing is performed through data cleaning, interpolation, normalization and the like, including the following steps:

[0087] S121) Design and construct a data cleaning process.

[0088] The cleaning process is applied to each data set: filter and remove outliers of the current data set according to the outlier satisfaction condition, and the formula of the outlier satisfaction condition is as follows:

[0089]

[0090] Wherein, x is the value of the input data in the current data set, is the mean value of the input data in the current data set, σ is the standard deviation of the current data set, Q1, Q2, Q3 represent the value of each quartile point of the current data set, respectively;

[0091] S122) Design and construct an interpolation process.

[0092] Interpolate the cleaned data part of each data set to achieve data integrity. Specifically, interpolate the part of each data set that has been removed outliers, and when interpolating the part of the current data set that has been removed outliers, the formula is as follows:

[0093]

[0094] Wherein, x new is the new value inserted in the current part removed outliers; x j-1 and x j+1 are the values corresponding to the front and rear sample regulations of the current part removed outliers x j in the current data set according to the order;

[0095] S123) Normalize all input data in each data set according to the data magnitude and data range, and map it to the [0, 1] standard interval to realize dimensionless, and for the current data set, the formula of normalization processing is as follows:

[0096]

[0097] Wherein, x j_new is the jth factor in the current data set after normalization processing, x j is the jth factor in the current data set before normalization processing, x max is the maximum value of each factor in the current data set before normalization processing; x min is the minimum value of each factor in the current data set before normalization processing;

[0098] S13) constructing the preprocessed different data sets into a comprehensive feature database, specifically integrating the normalized data in each data set so that the processed multi-source input features are mapped into a unified sample space to construct a feature database.

[0099] The comprehensive feature database of this embodiment includes feature information C1, ..., C2 corresponding to the internal factors in each operation record of the smart electric energy meter. k And the characteristic information C corresponding to the external factors k+1 ,…,C n , feature information {C1,…,,C n Each feature information in} is composed of a set of feature elements c 1i ,c 2i ,…,c mui Composition, representing the characteristic information C of the i-th factor i The comprehensive feature library also includes each smart energy meter operation record L and the corresponding operation status G, let L = {l1,l2,…,l i ,…,l mu}, indicating that a total of mu operation records of smart energy meters placed in natural environments are collected through sensors. In addition, suppose G is a set containing all target variables, where G={G1,G2,…,G i ,…,G mu},G i Refers to one of the normal or abnormal operation of the smart energy meter. Therefore, the final constructed comprehensive feature database C is as follows:

[0100]

[0101] In the constructed comprehensive feature database C, l i Indicates the number of the corresponding record, that is, starting from the second row of the matrix, each row refers to a record of the operation of the smart energy meter.

[0102] This embodiment constructs a strong association rule identification model and designs two new types of importance evaluation indicators through step S2, so that it can more comprehensively and accurately mine strong association rules related to electricity meter failures from the database, effectively consider potential risk factors to enhance the reliability of identifying the degree of association between various features.

[0103] Strong association rule identification models aim to identify valuable and meaningful connections within massive amounts of input data, enabling the evaluation and prediction of other things based on one or more things. For example, for an antecedent a and a consequent b, a rule pattern like "a→b" can be obtained.

[0104] For the operation record of the smart meter, a set containing all input data can be defined as , wherein . Assuming , , wherein it is assumed that is a target variable, is an input variable, and , a correlation rule can be expressed as: . It can be considered that if the input variable occurs, the target variable will definitely occur.

[0105] The strong correlation rule identification model of the embodiment improves the method of evaluating only two important degree evaluation standards, support and confidence, in the traditional correlation rule identification model, and constructs new evaluation indexes in the face of the real scene of smart meter operation state evaluation to meet more effective strong correlation rule mining. The new evaluation indexes are divided into two categories: information correlation degree index and rule robustness index.

[0106] The traditional support index simply considers the absolute frequency of the item set, which is easy to lead to the bias of those frequent normal state combinations in the mining of correlation rules, and ignores the rules related to rare faults but with great potential value. Although the traditional confidence can show the probability of the occurrence of item set Y under the condition of the occurrence of item set X, it does not fully consider the original occurrence probability of item set Y in the whole data set, which may lead to misjudgment of some seemingly closely related rules but actually with contingency. The two new types of evaluation indexes constructed aim to make up for the shortcomings of traditional support and confidence, so that they are more suitable for smart meter state evaluation, long-term monitoring and anomaly detection, and improve the identification ability of key fault modes and the accuracy of rule mining.

[0107] Among them:

[0108] The information correlation degree index is used to measure the correlation strength and dependency between different variables, which is not only suitable for traditional binary correlation analysis (such as "whether the occurrence of a factor increases the probability of occurrence of another factor"), but also suitable for more complex nonlinear and multivariate interaction. The key information correlation degree index proposed in the embodiment includes conditional information interaction degree and cross lift degree.

[0109] The information interaction degree is suitable for nonlinear correlation calculation of variables, and measures the information sharing degree between two variables, that is, whether the occurrence of one variable can effectively reduce the uncertainty of another variable. However, under the condition of multiple variables, it is easy to be disturbed by the third party variable. Therefore, a new conditional information interaction degree index is constructed to solve the problem that the mutual information degree may be affected by hidden variables and improve the reliability of the rule. The formula of the conditional information interaction degree is as follows:

[0110]

[0111] where Y is the target variable, X is the input variable, and Z is the conditional variable, Sup(x,y|z i ) represents the proportion of x variable factor and y variable factor occurring in all sample records given the condition variable factor z i , P(x,y,z) represents the probability of each factor in X, Y and Z appearing together, Sup(x|z i ) represents the proportion of x variable in all sample records containing the conditional variable factor z i , Sup(y|z i ) represents the proportion of y variable in all sample records containing the conditional variable factor z i , and CMI(X→Y|Z) represents the information interaction degree between X and Y given the conditional variable Z.

[0112] The lift degree can reflect the lift effect of the occurrence of item set X on the occurrence of item set Y. However, the traditional lift degree is only used to measure the binary relationship. Therefore, the new cross-lift degree is constructed, which can be extended to multivariate relationship and used to measure the change of the lift effect between X and Y under different variables Z. The formula of cross-lift degree is as follows:

[0113]

[0114] where Y is the target variable, X is the input variable, and Sup(X|Z1,Z2,…,Z n ∩Y|Z1,Z2,…,Z n ) represents the support degree of X and Y occurring together when the factors {z1,z2,…,z n} of all conditional variables occur, Sup(X|z i ) and Sup(Y|z i ) respectively represent the proportion of x and y variable factors in all sample records given the condition variable factor z i , w i is the weight of z i , representing its contribution to the index result.

[0115] In the construction of conditional mutual information degree index CMI(X→Y|Z) and cross-lift degree CLift(X→Y|z1,z2,…,z n) is not from a new data dimension, but is a part of the input variable set C in the comprehensive feature database. In actual smart meter operation data, each input feature often has dependencies and is not completely independent. While the traditional significance evaluation index calculation method usually assumes that each feature is independent, it calculates a single variable C k impact on the target state G. However, this independence assumption does not hold in real-world applications, and ignoring the linkage of other features can lead to spurious correlations or misjudgments. Therefore, when analyzing the correlation strength of a certain feature variable C k to the target state G, the present scheme introduces other features {C i | i≠k} in the variable set C as conditional variables Z into modeling. By calculating the conditional information interaction degree and cross lift degree, the information contribution of C k and G in the presence of other features Z is measured, and the true causal relationship between the main variable and the target state is accurately identified. For example, in the case where the input variable set includes multiple features such as "device type", "operation period", "alarm marker", "current peak value", etc., when evaluating the correlation between "current peak value" and "fault state", the "device type" can be set as a conditional variable to exclude the structural bias caused by different device architectures on the current feature, making the significance evaluation more reliable and consistent with engineering practice.

[0116] The rule robustness index is mainly used to measure the stability, adaptability and generalization ability of the rule, to ensure that the rule is still applicable when facing different scenarios of smart meter operation. The rule robustness index proposed in this embodiment includes time series stability degree and enhanced imbalance degree.

[0117] The time series stability index is used to measure the stability of a rule at different times to avoid the influence of short-term abnormal data on the rule. The formula of time series stability degree is as follows:

[0118]

[0119] where Y is the target variable, X is the input variable, and T is the time window, Conf t (X, Y) is the confidence between variables X and Y at time t;

[0120] The imbalance degree is reflected in the difference degree of the frequency of each category or item set in the data set. The operation state of the smart meter is greatly affected by external environmental factors (such as temperature, humidity, voltage fluctuation), and the traditional imbalance degree does not consider these factors. Therefore, the embodiment constructs the environmental sensitive imbalance degree (CAII), which takes environmental variables into account. The CAII can run in different environments, and low-frequency but key failure modes will not be ignored in extreme environments. The formula of enhanced imbalance degree is as follows:

[0121]

[0122] wherein Y is the target variable, X is the input variable, and CAII(X→Y) represents the enhanced imbalance level of the association rule X→Y, E represents the type of the environment variable in which the smart meter is located, W c (E) represents the class weight under the specific environment E, N c (E) is the number of class samples reflecting the simultaneous occurrence of variables X and Y under the environment E, Sup(X) represents the support of the X variable, and Sup(Y) represents the support of the Y variable.

[0123] Based on the above, in step S2 of the embodiment, when the strong association rule recognition model is used to mine feature factors that have an impact on the smart meter from the comprehensive feature database through the information correlation degree index and the rule robustness index, an optimized FRFP-Growth algorithm is used, as shown in the following formula: Figure 2 The method includes the following steps:

[0124] First, the support count of each candidate is counted by scanning the comprehensive feature database, and the frequent 1-item set is preliminarily screened according to the set minimum support, and the steps are as follows:

[0125] S51) The feature information {C1,…,C k ,…,C n} in the comprehensive feature database is taken as the input variable, and the running state G in the comprehensive feature database is taken as the target variable;

[0126] S52) Each input variable is taken as a candidate, the support of each candidate is calculated, and the candidate with a support greater than the minimum support is screened to obtain the frequent 1-item set;

[0127] Secondly, the FR-list and the FP-tree are established, and the steps are as follows:

[0128] S53) The cross-promotion degree of each frequent item of the frequent 1-item set and the target variable under the condition variable is calculated, and then each frequent item of the frequent 1-item set is sorted according to the cross-promotion degree to obtain the FR-list; in the embodiment, in order to improve the representativeness of the frequent item set, the cross-promotion degree is introduced as one of the screening indexes on the basis of the support, which measures the cross-promotion degree level of each feature variable and the target variable under the condition variable Z. Then, according to the calculated cross-promotion degree, the item set combination promotion degree is sorted, the candidate with the cross-variable strong association feature is preferentially reserved, and the reservation probability of the rare but high-risk mode is improved;

[0129] S54) generating the FP-tree according to the FR-list and taking the target variable as a leaf node, traversing from the leaf node of the FP-tree upwards, collecting the conditional pattern base of each item, and retaining the conditional pattern base with the conditional information interaction degree greater than zero; in the process of constructing the conditional pattern base, the joint occurrence mode of each candidate path item set and the target feature is combined, the conditional information interaction degree is further introduced, and the real information sharing degree of the candidate path item set and the target feature under the given third variable factor Z is measured. By setting a threshold, only the paths with the information interaction degree significantly greater than zero are retained, the potential false correlation features are removed, and the credibility of the rules is improved;

[0130] S55) constructing a new conditional FP-tree according to the conditional pattern base, until the conditional pattern base is empty or the conditional FP-tree cannot be constructed, specifically, for the conditional pattern base of the current frequent item, the cross-promotion degree of each item in the conditional pattern base and the target variable under the conditional variable is calculated, then each item is sorted according to the cross-promotion degree to obtain the corresponding conditional FR-list, and finally the conditional FP-tree is generated according to the conditional FR-list and the frequent pattern of the current frequent item is updated, the conditional pattern base of each item is collected from the leaf node of the conditional FP-tree upwards, and the conditional pattern base with the conditional information interaction degree greater than zero is retained, the above steps are repeated, smaller conditional pattern bases are extracted, and smaller conditional FP-trees are constructed, until the conditional FP-tree cannot be constructed any more;

[0131] In this embodiment, the prefix path of each frequent item is collected by traversing upwards from the leaf node, re-scanned as a sub-transaction, screened by introducing the conditional information interaction degree to form the conditional pattern base, and then the new FR-list is constructed by secondary screening according to the minimum support degree. Then, the frequent items in the FR-list are sorted in descending order according to the cross-promotion degree, and the key frequent items are selected according to the screening standard to construct the final FP-tree;

[0132] S56) calculating the time sequence stability degree of each frequent pattern respectively, and removing the frequent pattern with the time sequence stability degree less than a specified value, then combining each frequent item with the corresponding frequent pattern to obtain the frequent pattern set of each frequent item, including the pre-post frequent item set containing the feature factor C and the fault factor G at the same time, and the pre-frequent item set formed by the feature factor C; in this embodiment, the time sequence stability degree is introduced as a sorting evaluation index to count the appearance confidence fluctuation of each frequent pattern under different time windows. Only the rule path with stable appearance and small fluctuation in each time window is retained, so as to avoid the overfitting of the model caused by the “occasional strong rule” frequently appearing in a short period, and the long-term applicability of the rule is improved;

[0133] Finally, the frequent item set is mined in the FP-tree and the mining results are merged, and the strong association rules are output, the steps including:

[0134] S57) calculating the support, confidence and enhanced imbalance degree of each frequent item set in the frequent item set set respectively, selecting the frequent item set whose support, confidence and enhanced imbalance degree are all greater than the corresponding threshold, and then merging to obtain strong association rules, realizing the selection of the front frequent item set formed by the characteristic factor C and the item containing the characteristic factor C and the failure factor G in the FP-tree according to the item constraint condition, so as to obtain the set of characteristic variables affecting the operation situation of the smart meter. By mining the front and rear frequent item sets and the front frequent item set, the association rules meeting the item constraint condition are effectively obtained, and the redundancy of the association rules is reduced, thereby improving the efficiency and quality of the mining of the association rules;

[0135] In the output strong association rule stage of the embodiment, the support and confidence of the frequent item set are calculated, and the enhanced imbalance degree index is introduced. The index considers the weight compensation mechanism of the low-frequency but high-risk rule combination under different environmental variable conditions E, so as to avoid that the key risk mode is excluded by the rule screening algorithm due to its rarity. The introduction of the enhanced imbalance degree makes the rule output more comprehensive and sensitive, and is especially suitable for potential anomaly detection of the smart meter under extreme operating environment.

[0136] In summary, the embodiment embeds the information correlation degree index and the rule robustness index into each core stage of the FRFP-Growth, and constitutes a new association rule mining process. The process realizes the accurate modeling of feature-target correlation, the evaluation control of time consistency and the identification and compensation of extreme risk mode, greatly improving the reliability, efficiency and engineering practicability of strong association rule identification.

[0137] The embodiment constructs the occurrence density-adaptive ladder fuzzy reasoning system through step S3, and designs an adaptive optimization mechanism, which can dynamically adjust the reasoning parameters and rules, improve the calculation efficiency, enhance the adaptability of the system to complex environment, and further reduce the uncertainty of the evaluation. It has good flexibility and scalability, and can better adapt to complex and changeable environment, improve the accuracy and reliability of decision-making. In step S3, when the fuzzy reasoning system is used for fuzzy reasoning on the continuous features in the characteristic factor, the following steps are included:

[0138] S31) Preprocessing the multi-input continuous features. Specifically, the occurrence density of each continuous feature is calculated, and the division points of different health intervals are divided according to the occurrence density distribution of each continuous feature.

[0139] Five continuous features are selected for the evaluation of the operational health of the smart meter: average humidity (‰), average temperature (°C), average air pressure (hpa), average electromagnetic interference intensity (μT), and average salt mist concentration (%). Unlike traditional methods that rely solely on probability density functions for feature modeling, this embodiment further constructs an occurrence density function that can directly reflect the importance of each feature in fault occurrence, rather than just staying at the data distribution level. Specifically, first, the probability density function of each feature is calculated based on the input database, and on this basis, a risk weighting mechanism is introduced to construct the corresponding occurrence density function. This function not only considers the probability of occurrence of the feature, but also combines historical data of fault occurrence to ensure that low-frequency but high-risk factors are given sufficient weight in the model. Subsequently, the fuzzy reasoning system is used to fuzz the occurrence density, thereby converting it into a fuzzy health index.

[0140] Traditional fuzzy reasoning systems usually directly use membership functions to fuzz the external environmental features that affect the operational status of smart meters. However, when evaluating the comprehensive impact involving multiple features, it is necessary to build membership functions for different features separately, which not only increases the computational cost, but also cannot accurately depict the actual contribution of each feature to the risk of failure. To solve this problem, this embodiment designs an occurrence density function based on the probability density function to realize the construction of the membership function. To effectively avoid the masking of key failure modes by high-frequency normal features and highlight low-frequency but high-risk feature factors, this embodiment designs a risk weighting mechanism to optimize the probability density function and construct the occurrence density function. Specifically, during density estimation, the fault impact weight R(X), the Kullback-Leibler divergence D KL (P||Q), and the Sigmoid nonlinear weight adjustment term are combined to enhance the weight of low-probability but high-risk factors, so that the final modeling result can more comprehensively reflect the real operational risk of smart meters.

[0141] The occurrence density function IDF(X) is shown in the following formula:

[0142]

[0143] Where P(X) is the probability density function of the feature variable X, reflecting the basic distribution of the feature. R(X) is the fault impact weight function, measuring the impact strength of feature X on fault occurrence. γ is the exponential scaling factor. D KL (P||Q) is the Kullback-Leibler divergence, used to measure the deviation between P(X) and the normal working condition benchmark. is a regularization factor that controls the impact of KL divergence on the overall calculation, preventing overemphasis on abnormal situations. For the nonlinear weight term based on the sigmoid function, it is used to enhance the influence of features in different regions. x c For the quantile center value of the feature variable data, β controls the steepness of the threshold transition, and α is used to adjust the influence of the feature variable.

[0144] In addition, in order to more finely depict the influence range of each feature, the embodiment sets the demarcation point of each continuous feature according to the occurrence density, and divides the continuous feature into four health intervals (rare R, relatively rare RR, ordinary U, and relatively ordinary RU) according to the demarcation point. The importance weight of each interval can be further calculated to enhance the evaluation accuracy of the operation risk of the smart meter. As shown in Figure 3 Figure 3 (a) is the occurrence density diagram of the average humidity, Figure 3 (b) is the occurrence density diagram of the average temperature, Figure 3 (c) is the occurrence density diagram of the average air pressure, Figure 3 (d) is the occurrence density diagram of the average electromagnetic interference intensity, Figure 3 (e) is the occurrence density diagram of the average salt fog concentration. The demarcation points of each continuous feature are shown in Table 1:

[0145] Table 1: Continuous features and their demarcation points

[0146]

[0147] S32) Solving the final fuzzy health degree. Specifically, the membership degree of each continuous feature data point in different health intervals is calculated to obtain the corresponding fuzzy set of each continuous feature.

[0148] Fuzzification is an important concept in fuzzy logic, which is used to map the input precise value to a fuzzy set. In the fuzzification process, the input precise value is converted to the corresponding membership value by using the membership function, i.e. the degree of the value belonging to different fuzzy sets. The membership function is usually a curve or function defined on the range of input values, which describes the relationship between input values and fuzzy sets.

[0149] The embodiment uses a combination of triangular membership function and trapezoidal membership function to realize the fuzzification of the divided health intervals, as shown in Figure 4 The membership degree of the health interval at the two end positions of each continuous feature is calculated using the trapezoidal membership function, and the formula is as follows:

[0150]

[0151] The membership degree of the health interval at the middle position is calculated using the triangular membership function, and the formula is as follows:

[0152] ​

[0153] In the above formula, a, b, c, d are respectively defined as the intersection points of the triangular membership function or the trapezoidal membership function and the x-axis;

[0154] For the data points falling into the common area of adjacent healthy intervals, the calculation method of membership degree is: TC (A,B) = Min(TC A + TC B ), that is, the membership degree takes the minimum value of the membership degrees of the two healthy intervals.

[0155] S33) Constructing a density-adaptive step fuzzy reasoning system for fuzzy reasoning. Specifically, the fuzzy set of each continuous feature is input into the step fuzzy reasoning system, which includes multiple groups of expert subsystems according to step distribution. Each expert subsystem performs two-input fuzzy reasoning operation according to the corresponding fuzzy reasoning rule table and outputs the corresponding continuous feature health degree, while also performing adaptive optimization in the link of fuzzy reasoning.

[0156] Based on the traditional step fuzzy reasoning system with hierarchical structure, this embodiment introduces an adaptive optimization mechanism to improve reasoning efficiency and evaluation accuracy, ensuring efficient calculation of multiple input features and dynamic perception of the operating state of smart meters.

[0157] The fuzzy reasoning system of this embodiment is based on the traditional step fuzzy reasoning system. Fuzzy rules are constructed and hierarchical, multiple input features are combined two by two, and multiple groups of expert subsystems according to step distribution are constructed as shown in Figure 5 The output continuous feature health degree is defined as low (Low), relatively low (RLow), relatively high (RHigh), and high (High) in turn, and the corresponding fuzzy weights are 0.15, 0.35, 0.65, and 0.95, respectively.

[0158] For subsystem 1 constructed by average humidity and average temperature, the corresponding fuzzy reasoning rule table is shown in Table 2:

[0159] Table 2 Fuzzy reasoning rule table of subsystem 1

[0160]

[0161]

[0162] For the fuzzy reasoning between subsequent expert subsystems, the corresponding fuzzy reasoning rule table is shown in Table 3:

[0163] Table 3Table of fuzzy inference rules forthe last subsystems

[0164]

[0165] The fuzzy inference system of the embodiment can also be adaptively optimized in the link of fuzzy inference. In a traditional fuzzy inference system, the fuzzy rules, inference structure and defuzzification method of the system are static, and the parameters in Tables 2 and 3 will always be fixed. This fixed mode is difficult to adapt to changes in actual data distribution and operating environment, which may lead to decreased inference accuracy, increased computational cost and insufficient response capability to abnormal situations. Therefore, the embodiment introduces dynamic adjustment strategies in multiple key links of fuzzy inference to realize an adaptive optimization mechanism, so that the system can be optimized and adjusted according to real-time data distribution and fault modes.

[0166] Specifically, when adaptively optimizing in the link of fuzzy inference, the following is included:

[0167] In terms of fuzzy rule optimization, a pruning method based on occurrence density is proposed. Specifically, the occurrence contribution degree of rules during the operation of each expert subsystem is calculated, and rules with an occurrence contribution degree lower than a set threshold are removed to reduce computational complexity and improve inference efficiency, thereby realizing the screening of efficient rules and the removal of redundant or low-contribution rules. The occurrence contribution degree of rules is calculated by an occurrence density function, and the formula is as follows:

[0168]

[0169] wherein, ρ i represents the contribution degree of rule R i , N represents the total number of samples, X j represents the jth input sample, μR i (X j ) represents the fuzzy weight of rule R i on the input sample X j , and μR j (X j ) represents the fuzzy weight of the rule R i most similar to R j on the input sample X j .

[0170] In terms of inference structure optimization, adaptive inference weight optimization is designed and constructed, specifically, the inference weights between different expert subsystems are dynamically adjusted based on Bayesian optimization to ensure that key features have higher influence in the inference process, and the weight distribution is calculated based on the occurrence density to ensure that the weights of different features in the inference system are dynamically adjusted with data changes, thereby improving the prediction accuracy.

[0171] The inference weight is calculated by minimizing the prediction error, as follows:

[0172]

[0173] where ω * represents the optimal inference weight, such that the prediction error L(ω) is minimized, N represents the total number of samples, H pred represents the predicted health degree calculated based on the current weight ω, H actual represents the actual health degree known in the test.

[0174] The weight of different features is calculated as follows:

[0175]

[0176] where IDF(x i ) represents the occurrence density function value of feature x i , and v i is the adaptive weight of feature x i .

[0177] In addition, in terms of de-fuzzification optimization, the traditional method mostly uses the fixed barycenter method, which is easily disturbed by abnormal data and affects the final evaluation result. The embodiment proposes a weighted de-fuzzification method combined with sliding window smoothing processing, so that the de-fuzzification result is more stable, thereby obtaining the risk evaluation value of continuous features. The formula is as follows:

[0178]

[0179] where E is the total number of fuzzy sets with cross after de-fuzzification of continuous features in a sample regulation in the comprehensive feature database, is the membership function corresponding to the fuzzy set e, x is the feature in the domain of the fuzzy set e, and v i is the adaptive weight of feature x i in the continuous feature.

[0180] The embodiment designs and constructs a time window sliding and decay weighted significance evaluation system through step S4. The significance evaluation system is a commonly used risk evaluation model at present, which mainly quantifies the influence weight of the feature on the fault risk through the proportion of each feature variable in the comprehensive feature database to the smart meter fault event. In order to further improve the time sequence coupling and real-time performance in the smart meter fault risk evaluation, the embodiment designs and constructs a "time window sliding and decay weighted significance evaluation system". The system is used for smart meter risk evaluation of discrete features obtained after strong association rule mining, and can form a function complement with the inference system constructed in step S3. The steps of risk evaluation are as follows:

[0181] The local statistics of the equipment failure data and feature performance are carried out in a rolling time window (such as an hour level or a day level), and a short-term "local significance" is calculated in real time. In this embodiment, the expression of the local significance is as follows:

[0182] Sig ih [t]=W x ({x ih (τ),y(τ)}∣τ∈[t-ΔT,t])

[0183] Wherein, x ih (τ) represents the value of the feature ih collected at time τ, y(τ)∈[t-ΔT,t] is a label indicating whether the equipment fails at time τ, 1 represents a failure, and 0 represents normal, t is the current time, and ΔT is the window length; W x () represents the reference significance weight of the discrete feature, and the expression is as follows:

[0184]

[0185] Wherein, s k represents a discrete feature, Num(s k ∩g) represents the number of sample records in the comprehensive feature database that simultaneously contain the discrete feature s k and the failure g, Num(g) represents the number of sample records in which the failure occurs, and n represents the total number of discrete features.

[0186] Then, the reference significance is weighted and fused through a decay factor, so that the dynamic and smooth updating of the feature significance is realized, the latest failure mode change can be more accurately captured, and finally the risk assessment value of the discrete feature is obtained. The formula is as follows:

[0187]

[0188] S i (t)=γ·W x (s k )+(1-γ)·Sig i [t]

[0189] Wherein, f2 represents the final risk assessment value of all discrete features when a sample record in the comprehensive feature database occurs, P k represents whether the feature occurs when the sample record occurs, if the feature occurs, P k is 1, if the feature does not occur, P k is 0, S i (t) is the updated significance finally obtained, Sig i [t] represents the local significance of the feature i, and γ is a decay factor.

[0190] The state risk prediction based on a single model may cause a series of problems such as low accuracy. In order to improve the evaluation accuracy, in step S5, the real-time state risk evaluation of the smart meter is performed based on the multi-model integrated architecture constructed based on the data structure difference in the foregoing steps, the time window sliding and decay weighted significance evaluation system constructed in step S4, and the occurrence density-adaptive step fuzzy reasoning system constructed in step S3. Through dynamic optimization of the model weight, the optimal evaluation result is achieved.

[0191] In step S5 of the embodiment, when the risk evaluation values of the continuous features and the risk evaluation values of the discrete features are weighted and summed, the evaluation results of the continuous features and the discrete features are first obtained through steps S3 and S4 respectively, and the evaluation results are given corresponding weight coefficients. Then the fitting result is obtained and compared with the actual result for further analysis. Through repeated improvement of the weight, the optimal weight is found, and the evaluation accuracy is improved.

[0192] The weight coefficient formula is as follows:

[0193]

[0194] wherein w is a set composed of the weight coefficients of the models, and n represents the integration of n models.

[0195] The state risk prediction result formula finally composed of the integrated model is as follows:

[0196]

[0197] wherein F represents the final state risk evaluation result of the integrated model. and are the model weights that are continuously dynamically optimized and corrected. f1 and f2 are the state evaluation values of the smart meter of the occurrence density-adaptive step fuzzy reasoning system and the time window sliding and decay weighted significance evaluation system respectively. According to the expert experience, when 0.5 is taken, the integration effect is best.

[0198] The effect of the method of the embodiment is verified through the following test.

[0199] Through the above steps, the model of the integrated strong association rule recognition model and the occurrence density-adaptive step fuzzy reasoning system (MARM-FP-TFIS, MPT) is constructed. The working process of the model is as shown in Figure 6 ​As shown, first, the comprehensive feature database is established according to step S1; then, strong association rules of the comprehensive feature database are screened through the information correlation degree index and the rule robustness index, two important diagnostic criteria, by using the strong association rule identification model according to step S2. Next, the risk of continuous features is evaluated by using the occurrence density-adaptive ladder fuzzy reasoning system according to step S3, and the risk of discrete features is evaluated by using the time window sliding and decay weighted significance evaluation system according to step S4. Finally, the health level of each feature is integrated, and the model weight is dynamically adjusted to realize the optimal evaluation of the running state by integrating the two types of evaluation models, i.e., the time window sliding and decay weighted significance evaluation system and the occurrence density-adaptive ladder fuzzy reasoning system, according to step S5.

[0200] To verify the reliability of the method, two original models are added: one is the traditional ARM model and FP-TFIS model (ARM-PF-TFIS, APT), and the other is the improved association rule model and traditional fuzzy reasoning system (MARM-FIS, MF), which are compared with the model integrating the strong association rule identification model and the occurrence density-adaptive ladder fuzzy reasoning system (MARM-FP-TFIS, MPT).

[0201] The comparison is specifically performed by using the ROC curve (Receiver Operating Characteristic Curve) and the PR curve (Precision-Recall Curve) to compare the evaluation results with the test data. The ROC curve is a commonly used model evaluation tool in machine learning, which shows the trade-off between the true positive rate (TPR) and the false positive rate (FPR) of the model; the PR curve is another index for evaluating the performance of the model, which is particularly suitable for unbalanced data sets and can reflect the ability of the model in detecting positive class samples (abnormal state).

[0202] The accuracy of the evaluation is determined by the area under the curve (AUC). The AUC value ranges from 0 to 1, and the closer the value is to 1, the better the classification performance of the model. The AUC value can well reflect the ability of the model in handling the balance between positive and negative samples. In the present application, the running state evaluation task of the smart electric energy meter faces the possibility of extremely unbalanced proportion of positive and negative samples (normal operation and abnormal operation). Through AUC, the ability of the model to distinguish between positive and negative classes under different thresholds can be comprehensively reflected, rather than being limited to the performance of a single threshold.

[0203] The ROC curves and PR curves of the three models are as shown in Figure 7 , wherein Figure 7 (a) is the ROC curve of the three models, Figure 7(b) are PR curves of three models. AUC comparison of three models is shown in Table 4, wherein, for the ROC curve, the AUC of the MPT model is improved by 10.6% and 14.2% compared with the APT model and the MF model respectively; for the PR curve, the AUC of the MPT model is improved by 15.3% and 23.2% compared with the APT model and the MF model respectively

[0204] Table 4 AUC comparison chart

[0205]

[0206] It can be seen by comparison that the evaluation effect of the model constructed according to the method proposed in the embodiment is obviously better than that of the other two types of models.

[0207] In summary, the application proposes an intelligent electric energy meter operation state risk evaluation method based on strong correlation rule identification and multi-model integration, designs two types of new importance evaluation indexes, constructs a strong correlation rule identification model, improves the evaluation accuracy and the ability to identify potential operation risks; for multi-source input continuous characteristics, a occurrence density-adaptive step fuzzy reasoning system is constructed to solve the intelligent electric energy meter operation condition health degree, realize real-time risk estimation of the intelligent electric energy meter, and reduce the model prediction uncertainty; for multi-source input discrete characteristics, a time window sliding and decay weighted significance evaluation system is designed, and a multi-model integration architecture is constructed according to the data structure difference, and the time window sliding and decay weighted significance evaluation system and the occurrence density-step fuzzy reasoning system are used for real-time state risk evaluation of the intelligent electric energy meter, and the optimal evaluation result is realized by dynamically optimizing the model weight.

[0208] The application has the following beneficial effects:

[0209] 1. Compared with the traditional correlation rule identification model which only uses support and confidence as the important degree evaluation indexes of feature correlation degree, the application designs and builds two more effective important degree evaluation indexes, i.e. information correlation degree index and rule robustness index, so that the model can not only realize more comprehensive and accurate evaluation of the correlation rules, but also make the extracted strong correlation rules more thorough and make the quantification of the correlation degree of the correlation rules more reliable.

[0210] 2. The adaptive optimization mechanism is designed, and the occurrence density and adaptive step fuzzy reasoning system are organically combined, which can effectively solve the problem of high calculation cost of the traditional fuzzy reasoning system in the multi-input fuzzy derivation process, and further reduce the uncertainty of the evaluation. It has good flexibility and scalability, and can better adapt to complex and changeable environment, improve the accuracy and reliability of decision-making.

[0211] 3. The experimental results show that the intelligent electric energy meter evaluation method can accurately evaluate the operation situation of the intelligent electric energy meter, has obvious improvement in accuracy and true positive rate, and has better generalization ability. Higher evaluation value can be realized in a wider application scene, and more powerful support is brought to the management and application of the intelligent electric energy meter.

[0212] The above is only the preferred embodiment of the present application, and the protection scope of the present application is not limited to the above-mentioned embodiments. Any technical solution falling within the concept of the present application shall fall within the protection scope of the present application. It should be noted that, for ordinary skilled persons in the art, some improvements and refinements without departing from the principles of the present application shall also be considered as the protection scope of the present application.

Claims

1. A method for identifying and evaluating the running state risk of a smart meter based on strong correlation rules and multi-model integration, characterized in that, The following steps are involved: The historical operating data of smart electricity meters are collected from multiple sources to obtain different data sets, which are then preprocessed and used to construct a comprehensive feature database. Using a strong association rule identification model to mine characteristic factors that affect smart electricity meters from a comprehensive feature database through information association indicators and rule robustness indicators, the information association indicators include conditional information interaction and cross-lift, and the rule robustness indicators include temporal stability and enhancement imbalance; Performing fuzzy reasoning on the continuous features in the characteristic factors using a fuzzy reasoning system, and performing adaptive optimization in the fuzzy reasoning process to obtain a risk assessment value of the continuous features; Calculating the local significance of the discrete features in the characteristic factors in real time within a rolling time window, and weighting and fusing the local significance with the benchmark significance through an attenuation factor to obtain a risk assessment value of the discrete features; The risk assessment value of the continuous feature is weighted and summed with the risk assessment value of the discrete feature to obtain the final state risk assessment result.

2. The method of claim 1, wherein the method is based on strong association rules to identify risks of the operation state of the smart meter integrated with the multi-model, and wherein the method comprises the steps of: determining a strong association rule set based on the association rule set; and identifying the risks of the operation state of the smart meter integrated with the multi-model based on the strong association rule set. When performing data preprocessing on different data sets, the following steps are included: The outliers in the current dataset are filtered and removed based on the outlier conditions. Then, the part of the current dataset excluding the outliers is interpolated. Finally, all input data in the current dataset are normalized according to the data magnitude and data range, and uniformly mapped to the standard interval of [0, 1] to achieve dimensionless. The formula for the outlier conditions is as follows: where x is the value of the input data in the current data set, is the mean value of the input data in the current data set, σ is the standard deviation of the current data set, and Q1, Q2, Q3 represent the values of each quartile of the current data set, respectively.

3. The method of claim 1, wherein the method is characterized by, The comprehensive feature database includes feature information C1, ..., C corresponding to the internal factors in each operation record of the smart electric energy meter. k And the characteristic information C corresponding to the external factors k+1 ,…,C n , also including the operating status G of each smart energy meter operation record, using the strong association rule identification model to mine the characteristic factors that affect the smart energy meter from the comprehensive feature database through the information correlation index and the rule robustness index, including the following steps: The feature information {C1,..., C k ..., C n} in the comprehensive feature database is taken as an input variable, and the operating state G in the comprehensive feature database is taken as a target variable. Take each input variable as a candidate, calculate the support of each candidate, and select the candidate with support greater than the minimum support to obtain the frequent 1-item set; Calculate the cross lift between each frequent item in the frequent 1-item set and the target variable under the conditional variable. The conditional variable specifically refers to the other variables in the variable set when analyzing the association strength of a specific variable in the variable set with the target state. Then sort the frequent items in the frequent 1-item set according to the cross lift to obtain the FR-list. Generate an FP-tree based on the FR-list and use the target variable as a leaf node. Traverse upward from the leaf node of the FP-tree to collect the conditional pattern bases of each item and retain the conditional pattern bases with a conditional information interaction degree greater than zero. For the conditional pattern base of each frequent item, calculate the cross-lift between each item in the conditional pattern base and the target variable under the conditional variable, then sort each item according to the cross-lift to obtain the corresponding conditional FR-list, and finally generate a conditional FP-tree based on the conditional FR-list and update the frequent pattern of the corresponding frequent item. Traverse upward from the leaf node of the conditional FP-tree, collect the conditional pattern bases of each item, and retain the conditional pattern bases with conditional information interaction greater than zero. Repeat the above steps to extract smaller conditional pattern bases and construct a smaller conditional FP tree until the conditional FP tree can no longer be constructed; Calculate the time sequence stability degree of each frequent pattern respectively, and remove the frequent pattern whose time sequence stability degree is less than a specified value, then combine each frequent item with the corresponding frequent pattern to obtain the frequent pattern set of each frequent item; Calculate the support, confidence and enhanced imbalance degree of each frequent item set in the frequent pattern set respectively, and select the frequent item set whose support, confidence and enhanced imbalance degree are all greater than the corresponding threshold, and then merge to obtain strong association rules.

4. The method of claim 3, wherein the method is characterized by, The formula of cross lift degree is as follows: where Y is the target variable, X is the input variable, and Sup(X|Z1,Z2,…,Z n ) represents the support degree of X and Y jointly occurring when the factor {z1,z2,…,z n} of all conditional variables occurs n Sup(X|z i ) and Sup(Y|z i ) respectively represent the proportion of x and y variable factors in all sample records given the conditional variable factor z i ; w i is the weight of z i , indicating its contribution to the index result; The formula of conditional information interaction degree is as follows: where Y is the target variable, X is the input variable, and Z is the conditional variable, Sup(x, y | z i ) represents the proportion of the x variable factor and the y variable factor occurring simultaneously in all sample records given the conditional variable factor z i , P(x, y, z) represents the probability of each factor in X, Y, and Z appearing together, Sup(x | z i ) represents the proportion of the x variable in all sample records containing the conditional variable factor z i , Sup(y | z i ) represents the proportion of the y variable in all sample records containing the conditional variable factor z i , and CMI(X→Y | Z) represents the information interaction degree between X and Y given the conditional variable Z. The formula of time sequence stability degree is as follows: wherein Y is a target variable, X is an input variable, and T is a time window, Conf t (X, Y) is a confidence between variables X and Y at time t. The formula of enhanced imbalance degree is as follows: where Y is the target variable, X is the input variable, and CAII(X→Y) represents the enhanced imbalance level of the association rule X→Y, E represents the type of environmental variables of the smart meter, W c (E) represents the class weight under the specific environment E, N c (E) is the number of class samples reflecting the simultaneous occurrence of variables X and Y under the environment E, Sup(X) represents the support of the X variable, and Sup(Y) represents the support of the Y variable.

5. The method of claim 1, wherein the method is characterized by, When using the fuzzy reasoning system to perform fuzzy reasoning on the continuous features in the feature factors, the method comprises the following steps: Calculate the occurrence density of each continuous feature respectively, divide the boundary points of different health intervals according to the occurrence density distribution of each continuous feature, and calculate the membership degree of data points of each continuous feature in different health intervals to obtain the fuzzy set corresponding to each continuous feature, the membership degree of the health interval at the two ends of each continuous feature is calculated using the trapezoidal membership function, and the membership degree of the health interval at the middle position is calculated using the triangular membership function; Input the fuzzy set of each continuous feature into the ladder fuzzy reasoning system, the ladder fuzzy reasoning system comprises multiple groups of expert subsystems according to the ladder distribution, each expert subsystem performs fuzzy reasoning operation of two inputs according to the corresponding fuzzy reasoning rule table and outputs the corresponding continuous feature health degree.

6. The method of claim 5, wherein the method is characterized by, When performing adaptive optimization in the fuzzy reasoning link, the method comprises the following steps: Calculate the occurrence contribution degree of the rules in the operation of each expert subsystem, and remove the rules whose occurrence contribution degree is lower than the set threshold, the formula of the occurrence contribution degree of the rules is as follows: wherein p i represents the contribution degree of the rule R i , N represents the total number of samples, X j represents the jth input sample, μR i (X j ) represents the fuzzy weight of the rule R i on the input sample X j , and μR j (X j ) represents the fuzzy weight of the rule R i most similar to R j on the input sample X j .

7. The method of claim 5, wherein the strong association rule-based identification and multi-model integrated smart meter operating state risk assessment method is characterized by, When performing adaptive optimization in the fuzzy reasoning link, the method comprises the following steps: Based on Bayesian optimization, dynamically adjust the reasoning weights between different expert subsystems to ensure that the key features have higher influence in the reasoning process, and combine the occurrence density to calculate the weight distribution of different features to ensure that the weight of different features in the reasoning system is dynamically adjusted with the change of data, the formula of the reasoning weight is as follows: where ω * represents the optimal inference weight, such that the prediction error L(ω) is minimized, N represents the total number of samples, H pred (ω) represents the predicted health based on the current weight ω, H actual represents the actual health known in the test; The weight formula of different features is as follows: wherein, IDF(x i ) represents the occurrence density function value of the feature x i , and v i is the adaptive weight of the feature x i .

8. The method of claim 1, wherein the method is based on strong association rules to identify risks of the operation state of the smart meter integrated with the multi-model, and wherein the method comprises: determining a strong association rule between the operation state of the smart meter and the multi-model; and identifying the risk of the operation state of the smart meter integrated with the multi-model based on the strong association rule. The risk evaluation value formula of the continuous feature is as follows: where E is the total number of fuzzy sets that exist in the cross of the fuzzy sets after the continuous features in the sample regulation of the comprehensive feature database are fuzzified, is the membership function of the corresponding fuzzy set e, x is the feature in the domain of the corresponding fuzzy set e, and v i is the adaptive weight of the feature x in the continuous features. i is the adaptive weight of the feature x in the continuous features.

9. The method of claim 1, wherein the method is based on strong association rules to identify risks of the operation state of the smart meter integrated with the multi-model, and wherein the method comprises: determining a strong association rule between the operation state of the smart meter and the multi-model; and identifying the risk of the operation state of the smart meter integrated with the multi-model based on the strong association rule. The expression of local saliency is as follows: Sig ih [t] = W x ({x ih (τ),y(τ)}∈[t-ΔT,t]) wherein x ih (τ) represents the value of feature ih collected at time τ, y(τ) ∈ [t-ΔT, t] is the label of whether the device fails at time τ, 1 represents failure, 0 represents normal, t is the current time, and ΔT is the window length; W x () represents the reference significance weight of the discrete feature, and the expression is as follows: where s k represents a discrete feature, Num(s k ∩g) represents the number of sample cases containing both the discrete feature s k and the fault g, Num(g) represents the number of sample cases in which the fault occurs, and n represents the total number of discrete features.

10. The method of claim 9, wherein the method is characterized by, The risk evaluation value formula of the discrete feature is as follows: The risk evaluation value formula of the discrete feature is as follows: where f2 represents the final risk assessment value of a sample regulation from the integrated feature database, P k represents whether the feature occurs or not when the sample regulation occurs, P k is 1 if it occurs, and P k is 0 if it does not occur, S i (t) is the last updated saliency, and the expression is as follows: S i (t) = γ · W x (s k )+(1 - γ) · Sig i [t] wherein Sig i [t] represents the local saliency of feature i, and γ is a decay factor.

Images