An Automatic Detection and Judgment Method and System for Abnormal Phenomena of Electric Energy Meters

Abstract

The present invention discloses a method and system for automatically detecting and determining abnormal phenomena of an electric energy meter. The method adopted by the present invention includes: collecting multi-dimensional power parameter data in real time, mapping the data into a non-linear space, constructing a discriminant matrix, optimizing the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, performing multi-stage dynamic filtering on the eigenvectors of the power parameter data to obtain purified eigenvectors; the purified eigenvectors are processed by an adaptive convolution inverse transform algorithm to reconstruct the time and space characteristics of the signal, generate a frequency-domain signal, and then restore the complete time characteristic signal through inverse transform, obtain the optimal abnormal eigenvector through a multiple iteration convergence algorithm, classify the optimal abnormal eigenvector, and complete automatic detection and determination. The present invention can improve the accuracy of abnormal detection, increase the signal-to-noise ratio, and accurately identify different types of abnormal phenomena.

Description

Technical Field

[0001] The present invention belongs to the technical field of power metering and monitoring, and in particular, to a method and system for automatically detecting and determining abnormal phenomena of an electric energy meter. Background Art

[0002] With the increasing complexity of the power system and the continuous growth of electricity demand, the working environment of the electric energy meter has become more and more diverse and complex. Especially in the context of smart grids and distributed energy systems, the electric energy meter not only needs to accurately measure the electricity consumption of users, but also needs to monitor and transmit multi-dimensional power parameter data in real time, such as voltage, current, active power, reactive power, power factor, harmonic components, etc. These data are of great significance for power companies to optimize grid operation, improve power supply quality and reduce power loss. However, due to factors such as changes in the external environment, aging of power equipment, and external interference, various abnormal phenomena may occur in the actual operation of the electric energy meter, such as measurement errors, data transmission interruptions, abnormal fluctuations in power parameters, etc. Therefore, how to effectively and accurately detect and determine the abnormal phenomena of the electric energy meter has become the focus of attention of power companies and equipment manufacturers.

[0003] The prior art has at least the following technical problems: It cannot dynamically adapt to spectrum changes, resulting in false judgments or missed detections, leading to a low signal-to-noise ratio, being unable to comprehensively reflect the complexity of the electric energy meter data, resulting in inaccurate abnormal determination, lacking an intelligent classification model, being unable to effectively cope with complex and changeable power system abnormalities, having obvious deficiencies in determination ability, and the detection performance gradually degrades during long-term use, making it difficult to maintain a high level of detection and determination ability. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned defects existing in the prior art, and provide a method and system for automatically detecting and determining abnormal phenomena of an electric energy meter, so as to improve the accuracy of abnormal detection, increase the signal-to-noise ratio, and accurately identify different types of abnormal phenomena.

[0005] In the first aspect, the present invention provides a method for automatically detecting and determining abnormal phenomena of an electric energy meter, which includes the steps of:

[0006] S1. Real-time collect multi-dimensional power parameter data, map the data to a non-linear space, construct a discriminant matrix, optimize the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and perform multi-stage dynamic filtering on the eigenvectors of the power parameter data to obtain purified eigenvectors;

[0007] S2. The purified feature vectors are processed by an adaptive convolutional inverse transform algorithm to reconstruct the temporal and spatial features of the signal, generate a frequency-domain signal, and then restore the complete temporal feature signal through an inverse transform. An optimal abnormal feature vector is obtained through a multiple-iteration convergence algorithm, and the optimal abnormal feature vector is classified to complete automatic detection and determination.

[0008] Further, in S1, the collected multi-dimensional power parameter data is preprocessed. To extract the deep non-linear features in the data, non-linear mapping is performed on the preprocessed multi-dimensional data.

[0009] Further, in S1, the mapped data set is used for adaptive weighted discriminant analysis. By constructing a weighted discriminant matrix, normal data and abnormal data are effectively separated in the non-linear space.

[0010] Further, in S1, an objective function is set up to obtain the optimal projection matrix. The within-class scatter matrix is eigen-decomposed to find its corresponding eigenvectors. The eigenvectors correspond to the directions that minimize the within-class scatter; by solving the generalized eigenvalue problem, the directions that maximize the between-class scatter are determined; the eigenvectors with the largest first few eigenvalues are selected as the column vectors of the projection matrix.

[0011] Further, in S1, to purify the feature vectors, a multi-stage dynamic filtering method is adopted. Based on the spectral features of the abnormal data, a frequency-domain filter is designed to perform frequency-domain filtering on each feature vector to remove high-frequency noise and low-frequency fluctuations. After applying the frequency-domain filter, the feature vectors are converted into purified feature vectors.

[0012] Further, in S2, the adaptive convolutional inverse transform algorithm introduces a multiple modulation and adaptive adjustment mechanism in the convolution kernel design and inverse transform process to better adapt to the complex abnormal signals in the electricity meter data; the purified feature vectors are input into the convolution kernel function, and the design of the convolution kernel takes into account the mixed modulation of temporal and frequency features.

[0013] Further, in S2, through convolution operation, the purified feature vectors are processed into a time-domain signal. After obtaining the frequency-domain signal, to restore the complete temporal features of the signal, an inverse transform operation is performed by combining the joint modulation of the frequency domain and the time domain.

[0014] Further, in S2, features are extracted from the reconstructed signal, and the extracted feature vectors are input into the multiple-iteration convergence algorithm in the feature space. Through multiple iterations of optimization, the potential optimal abnormal feature vectors in the signal are extracted.

[0015] Further, in S2, according to the values and feature distributions of the optimal abnormal feature vectors, each feature is analyzed to identify possible abnormal features; by calculating the statistical distribution of each feature in the normal operating state and comparing it with the currently extracted features, if the value of a certain feature deviates from the normal distribution by more than a preset threshold, it is considered that the current feature is abnormal; to further determine the type of abnormal phenomenon, a pre-trained classification model is used to classify the optimal abnormal feature vector.

[0016] In a second aspect, the present invention provides an automatic detection and determination system for abnormal phenomena of an electric energy meter, which includes:

[0017] Feature vector purification unit: Real-time collect multi-dimensional power parameter data, map the data to a non-linear space, construct a discriminant matrix, optimize the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and perform multi-stage dynamic filtering on the feature vectors of the power parameter data to obtain purified feature vectors;

[0018] Automatic detection and determination unit: The purified feature vectors are processed by an adaptive convolutional inverse transform algorithm to reconstruct the time and space features of the signal, generate a frequency domain signal, and then restore the complete time feature signal through inverse transform. The optimal abnormal feature vector is obtained through a multiple iteration convergence algorithm, and the optimal abnormal feature vector is classified to complete automatic detection and determination.

[0019] The beneficial effects of the present invention are:

[0020] 1. By performing deep non-linear mapping on the multi-dimensional features of the original data, the present invention enables abnormal patterns to be fully displayed in the mapped high-dimensional space, and further effectively distinguishes normal and abnormal data through the optimized discriminant matrix, significantly improving the accuracy of abnormal detection.

[0021] 2. Through the multi-stage dynamic filtering method and the adaptive convolutional inverse transform algorithm, the present invention can effectively remove high-frequency noise and low-frequency interference information in the feature vectors, making the remaining signal features more pure and representative, and effectively improving the signal-to-noise ratio.

[0022] 3. During the signal reconstruction process, the adaptive convolutional inverse transform algorithm enables the reconstructed signal to not only retain the original abnormal features but also enhance the time resolution; it can more accurately capture and analyze the occurrence time and development process of abnormal phenomena, contributing to the precise determination of abnormal phenomena.

[0023] 4. The present invention extracts various features from the reconstructed signal, including time domain features, frequency domain features, and non-linear features, comprehensively reflecting the complex characteristics of the electric energy meter data; subsequently, through the multiple iteration convergence algorithm in the feature space, the features are optimized multiple times, and finally the optimal abnormal feature vector is obtained. Description of the Drawings

[0024] Figure 1 It is a flowchart of a method for automatically detecting and determining abnormal phenomena of an electric energy meter according to the present invention;

[0025] Figure 2 It is a structural diagram of a system for automatically detecting and determining abnormal phenomena of an electric energy meter according to the present invention. Detailed Embodiments

[0026] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0027] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs.

[0028] The following specifically describes the specific solutions of a method and a system for automatically detecting and determining abnormal phenomena of an electric energy meter provided by the present invention with reference to the accompanying drawings.

[0029] Referring to the attached Figure 1 , which shows a flowchart of a method for automatically detecting and determining abnormal phenomena of an electric energy meter provided by an embodiment of the present invention. The method includes the following steps:

[0030] S1. Real-time collect multi-dimensional power parameter data, map the data to a non-linear space, construct a discrimination matrix, optimize the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and perform multi-stage dynamic filtering on the eigenvectors of the power parameter data to obtain purified eigenvectors.

[0031] The electric energy meter collects multi-dimensional power parameter data in real time through sensors, including voltage, current, active power, reactive power, power factor, harmonic components, etc. Each sensor is strictly calibrated to ensure the accuracy and stability of the data. The collected raw data is preprocessed, including operations such as signal filtering, noise reduction, and outlier removal. The preprocessing methods used are all existing technologies.

[0032] The preprocessed multi-dimensional data set is denoted as , representing the multi-dimensional parameter vector at the i-th time point, and each It contains all the power parameters at that moment, synchronized by a unified timestamp to ensure the temporal consistency of the data. To extract the deep non-linear features in the data, non-linear mapping is performed on the preprocessed multi-dimensional data, and the mapping function is defined as:

[0033]

[0034] where, represents the dataset after mapping, are the coefficients of the polynomial terms, are the power exponents of the polynomial, is the highest power of the polynomial, are the coefficients of the trigonometric function terms, controlling the weight of each sine term, are the power exponents of the sine function, is the highest power of the sine function. By combining polynomials and logarithmic functions, the original linear data is mapped to a space with complex non-linear features, enabling subsequent detection steps to better identify potential abnormal patterns.

[0035] Adaptive weighted discriminant analysis is performed using the mapped dataset. By constructing a weighted discriminant matrix, normal data and abnormal data are effectively separated in the non-linear space. The discriminant matrix is defined as:

[0036]

[0037] where, is the discriminant matrix, and are the projection matrices, controlling the distribution of data in the non-linear space to maximize the between-class difference and minimize the within-class difference, is the parameter adjusting the weight of the i-th data, is the transpose of the matrix, is the number of data after mapping. By optimizing the following objective function, the optimal projection matrix is obtained:

[0038]

[0039] where, represents the trace of the matrix, and are the within-class and between-class scatter matrices respectively, and their definitions are as follows:

[0040]

[0041]

[0042] where, denotes the mean vector of the k-th class of data, representing the average position of all samples in this class in the feature space; denotes the global mean vector of all data, representing the average position of all samples in the feature space; is the number of data classes, is the number of samples in the k-th class, representing the total number of samples in this class.

[0043] Perform eigenvalue decomposition on the within-class scatter matrix to find its corresponding eigenvectors. The eigenvectors correspond to the directions that minimize the within-class scatter; then, by solving the generalized eigenvalue problem, determine the directions that maximize the between-class scatter; the larger the eigenvalue, the corresponding eigenvector is the more important projection direction; select the eigenvectors with the largest first few eigenvalues as the column vectors of the projection matrix. The column vectors of the projection matrix represent the best projection directions, which can maximize the separation between different classes while maintaining the compactness of the within-class data, and distinguish normal data from abnormal data. By minimizing the objective function, the optimized discriminant matrix can effectively distinguish normal and abnormal data.

[0044] Multiply the optimized discriminant matrix by the mapped data to obtain the feature vectors of each data point in the non-linear space, i.e.:

[0045]

[0046] Feature vector contains the abnormal feature information of each data point in the non-linear space, but the feature vectors still contain noise and irrelevant interference information. Therefore, further purification is needed to extract the most representative abnormal signal features.

[0047] To purify the feature vectors, a multi-stage dynamic filtering method is adopted. Specifically, a frequency-domain filter is designed based on the spectral characteristics of the abnormal data, and each feature vector is filtered in the frequency domain to remove high-frequency noise and low-frequency fluctuations. The frequency response function of the filter is defined as:

[0048]

[0049] where, is the frequency response function of the filter, representing the response amplitude of the filter at different frequencies ; is the control of the filter bandwidth, are the Fourier coefficients, represents the frequency of the signal, is the index variable of the Fourier series, used to represent the order of different frequency components in the filter frequency response, is the factorial. Through the frequency response function, the filter can dynamically adjust its frequency response characteristics to adapt to the spectral characteristics of different eigenvectors, thereby removing unwanted noise components.

[0050] After applying the frequency-domain filter, the eigenvector is converted into a purified eigenvector. The specific purification process can be described by the following convolution formula:

[0051]

[0052] where represents the purified eigenvector, is the filter response function in the time domain, which is related to the frequency response function through Fourier transform; is the time variable; is the offset in the convolution operation. Through the convolution operation, the filter can remove irrelevant high-frequency and low-frequency components in the time domain, retain the representative intermediate-frequency features, and thus obtain the purified eigenvector.

[0053] S2. The purified eigenvector is processed by the adaptive convolution inverse transform algorithm to reconstruct the time and space features of the signal, generate the frequency-domain signal, and then recover the complete time feature signal through the inverse transform. The optimal abnormal eigenvector is obtained through the multiple iteration convergence algorithm, and the optimal abnormal eigenvector is classified to complete automatic detection and determination.

[0054] The purified eigenvector has a higher signal-to-noise ratio and better reflects the abnormal features in the electricity meter data. The purified eigenvector will be further processed. Through the adaptive convolution inverse transform algorithm, the time and space features of the signal are reconstructed; the reconstructed signal not only retains the original abnormal features but also enhances the time resolution, which helps to more accurately determine and analyze abnormal phenomena.

[0055] Specifically, the adaptive convolution inverse transform algorithm introduces a multiple modulation and adaptive adjustment mechanism in the convolution kernel design and inverse transform process to better adapt to the complex abnormal signals in the electricity meter data; the purified eigenvector is input into the convolution kernel function, and the design of the convolution kernel considers the mixed modulation of time and frequency features. The convolution kernel function is defined as:

[0056]

[0057] where represents the convolution kernel function, is the center position of the convolution kernel, controls the expansion coefficient of the convolution kernel, is the periodic parameter of the convolution kernel. The convolution kernel combines the smoothing characteristics of the Gaussian function and the periodic characteristics of the sine and cosine functions, and can be adaptively adjusted to capture different time and frequency features in the signal.

[0058] Through the convolution operation, the purified feature vector is processed into a time-domain signal, and the convolution process formula is as follows:

[0059]

[0060] where, is the frequency-domain signal, representing the representation of the convolved signal in the frequency domain, is the shift parameter in the product operation, representing the translation amount of the signal. The frequency-domain signal contains multi-level time features of the original signal. Thanks to the adaptive modulation of the convolution kernel, it can suppress noise and irrelevant information while retaining the local abnormal features of the signal.

[0061] After obtaining the frequency-domain signal, in order to restore the complete time features of the signal, an inverse transform operation is performed by combining the joint modulation of the frequency domain and the time domain, and the formula is as follows:

[0062]

[0063] where, represents the reconstructed signal, and the modulation term in the inverse transform is used to dynamically adjust the frequency components to ensure that the time features of the signal are retained and enhanced during the reconstruction process. The reconstructed signal contains the complete time features of the abnormal phenomena in the electricity meter data.

[0064] Extract features from the reconstructed signal, such as mean, variance, skewness, kurtosis, etc., and input the extracted feature vector into the multiple iterative convergence algorithm in the feature space. Through multiple iterative optimizations, extract the potential optimal abnormal feature vector in the signal. To achieve this goal, the objective function is defined as follows:

[0065]

[0066] where, is the objective function, is the iterative coefficient matrix, used to control the weight transfer during the iterative process, is the regularization parameter, used to avoid overfitting, is the th feature vector in the iteration,

[0067] Based on the values and feature distributions of the optimal abnormal feature vectors, each feature is analyzed to identify possible abnormal features. Specifically, by calculating the statistical distributions (such as mean and standard deviation) of each feature in the normal operating state and comparing them with the currently extracted features. If the value of a certain feature deviates from the normal distribution by more than a preset threshold, it can be considered that the current feature is abnormal.

[0068] To further determine the type of abnormal phenomenon, a pre-trained classification model is used to classify the optimal abnormal feature vector. The classification model can be a machine learning-based model, such as a support vector machine (SVM), random forest, or neural network, etc. The input of the classification model is the optimal abnormal feature vector, and the output is the possible abnormal type. The final abnormal type is determined by maximizing the output probability of the classifier.

[0069] Refer to the appendix Figure 2 , which shows the structural diagram of an automatic detection and determination system for abnormal phenomena of an electric energy meter provided by another embodiment of the present invention. The system includes:

[0070] Feature vector purification unit: Real-time collect multi-dimensional power parameter data, map the data to a non-linear space, construct a discriminant matrix, optimize the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and perform multi-stage dynamic filtering on the feature vectors of the power parameter data to obtain purified feature vectors;

[0071] Automatic detection and determination unit: The purified feature vectors are processed by an adaptive convolutional inverse transform algorithm to reconstruct the time and space features of the signal, generate a frequency-domain signal, and then restore the complete time feature signal through an inverse transform. The optimal abnormal feature vector is obtained through a multiple iteration convergence algorithm, and the optimal abnormal feature vector is classified to complete automatic detection and determination.

[0072] It should be noted that each unit in the above automatic detection and determination system for abnormal phenomena of an electric energy meter can be implemented in whole or in part by software, hardware, and their combination. The above units can be embedded in the processor of a computer device in hardware form or be independent of it, or can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to the above units. For the specific limitations of an automatic detection and determination system for abnormal phenomena of an electric energy meter, refer to the limitations of an automatic detection and determination method for abnormal phenomena of an electric energy meter in the above text. The two have the same functions and effects and will not be elaborated here.

[0073] The beneficial effects of the present invention are:

[0074] 1. The present invention realizes high-precision detection of potential anomalies in electricity meter data through complex non-linear mapping functions and adaptive weighted discriminant analysis. By performing in-depth non-linear mapping on multi-dimensional features of the original data, abnormal patterns are fully displayed in the mapped high-dimensional space. Further, the optimized discriminant matrix is used to effectively distinguish normal and abnormal data, significantly improving the accuracy of anomaly detection.

[0075] 2. Through the multi-stage dynamic filtering method and the adaptive convolutional inverse transform algorithm, the present invention can effectively remove high-frequency noise and low-frequency interference information in the feature vector. It not only dynamically adjusts the frequency response but also combines the multiple modulation functions of the convolutional kernel, making the remaining signal features purer and more representative, effectively improving the signal-to-noise ratio.

[0076] 3. During the signal reconstruction process of the adaptive convolutional inverse transform algorithm, through the combined modulation of the time domain and the frequency domain, the reconstructed signal not only retains the original abnormal features but also enhances the time resolution; it helps the system capture and analyze the occurrence time and development process of abnormal phenomena more accurately, contributing to the precise determination of abnormal phenomena.

[0077] 4. The present invention extracts various features from the reconstructed signal, including time-domain features, frequency-domain features, and non-linear features, comprehensively reflecting the complex characteristics of electricity meter data; subsequently, through the multiple iterative convergence algorithm in the feature space, the features are optimized multiple times, and finally, the optimal abnormal feature vector is obtained; in order to further refine and determine the types of abnormal phenomena, a classification model based on machine learning is introduced. Through the trained model, the optimal abnormal feature vector can be intelligently classified to accurately identify different types of abnormal phenomena; this classification mechanism helps power companies quickly understand the specific types of anomalies and take corresponding countermeasures.

[0078] The following is an application of the above automatic detection and determination method for electricity meter abnormal phenomena as follows.

[0079] 1. Data collection and preprocessing

[0080] The data collected by the electricity meters monitored in the power system includes parameters such as voltage, current, power factor, and power. Assume that the collected dataset includes 6-dimensional power parameters: voltage , current , active power , reactive power , power factor , harmonic components . The data at each time point is a 6-dimensional vector.

[0081] For example, the power parameters at 5 collected time points are as follows:

[0082]

[0083] Among them, the data at each time point are as follows:

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] These data are collected and stored in real time by sensors. Next, preprocessing is carried out, and simple filtering and denoising algorithms (such as low-pass filtering) are used to remove high-frequency noise and outliers.

[0090] 2. Nonlinear mapping

[0091] To extract deep nonlinear features from the original data, each data point is mapped to a new nonlinear space, and the mapping function used is:

[0092]

[0093] Among them, and , that is, a quadratic polynomial and a first-order sine function are used for mapping. Select , , .

[0094] For the first data point

[0095] , its mapping result is:

[0096]

[0097] For the same mapping is carried out, and the result is:

[0098]

[0099] This gives the mapped data set:

[0100]

[0101] 3. Weighted discriminant analysis

[0102] Weighted discriminant analysis (WDA) is used to optimize the projection matrix W to distinguish normal and abnormal data. First, the within-class scatter matrix needs to be constructed and within-class scatter matrix 。

[0103] 3.1 Calculate the within-class scatter matrix

[0104] Divide the data into 2 classes (normal and abnormal). Assume that:

[0105] Class 1 (normal) contains data points

[0106] Class 2 (abnormal) contains data points

[0107] Calculate the within-class scatter matrix The formula is:

[0108]

[0109] Where, and are the means of Class 1 and Class 2 respectively.

[0110] Mean of Class 1:

[0111]

[0112] Mean of Class 2:

[0113]

[0114] Then calculate the within-class scatter matrix :

[0115] First calculate the contribution of Class 1:

[0116]

[0117] Then calculate the contribution of Class 2:

[0118]

[0119] Therefore, the within-class scatter matrix is:

[0120]

[0121] 3.2 Calculate the between-class scatter matrix

[0122] The between-class scatter matrix The calculation formula is:

[0123]

[0124] Where, is the mean of all data points:

[0125]

[0126] Therefore, the between-class scatter matrix is:

[0127]

[0128] 3.3 Optimize the projection matrix

[0129] Optimize the projection matrix , that is, solve the objective function:

[0130]

[0131] Simplify the calculation of the objective function to:

[0132]

[0133] The objective function then becomes:

[0134]

[0135] Through simplification, we get:

[0136]

[0137] This means that the projection matrix needs to be adjusted in size so that the within-class scatter matrix and the between-class scatter matrix have the smallest possible ratio, thus achieving the goal of maximizing the between-class scatter and minimizing the within-class scatter.

[0138] The optimized projection matrix After iterative adjustment, a suitable projection direction is finally obtained:

[0139]

[0140] 4. Feature purification (detailed calculation process)

[0141] After optimizing the projection matrix , the eigenvectors need to be purified next. The purification process uses a multi-stage dynamic filtering method.

[0142] 4.1 Design a frequency-domain filter

[0143] Select a simple Gaussian filter as the frequency-domain filter, and the frequency response function is:

[0144]

[0145] Select , , and these two parameters are obtained through experimental adjustment. The specific calculation process of the frequency response function is as follows:

[0146] Assume the frequency of the signal (i.e., a sine wave with a period of 1 second), and we get:

[0147]

[0148] 4.2 Filtering Operation

[0149] Filter the eigenvector . Assume there is the following eigenvector before purification:

[0150]

[0151] The eigenvector after filtering can be calculated through a frequency-domain filter:

[0152]

[0153] Assume is a constant signal. The response of the frequency-domain filter is for frequency characteristics, so it can be simply calculated as:

[0154]

[0155] Perform frequency-domain filtering on all eigenvectors in the same way, and we get:

[0156]

[0157] In this way, the purified eigenvector is:

[0158]

[0159] 4.3 Inverse Transformation from Frequency Domain to Time Domain

[0160] The purified eigenvector also needs to be reconstructed in the time domain through an adaptive convolutional inverse transformation algorithm. Assume the selected convolution kernel function is as follows:

[0161]

[0162] Assume the selection of , , and , then the convolution kernel function becomes:

[0163]

[0164] Through the convolution operation, the reconstructed signal is obtained:

[0165]

[0166] 5. Anomaly Detection and Classification

[0167] 5.1 Anomaly Feature Extraction and Multiple Iterative Optimization

[0168] Through the multiple iterative convergence algorithm, the features in the reconstructed signal are further optimized. The objective function is:

[0169]

[0170] Assume the following experimental data and parameters:

[0171] Initial feature vector

[0172] Iterative coefficient

[0173] Regularization parameter

[0174] Number of iterations

[0175] 5.2 Calculate the objective function

[0176] First, for each iteration step, use the first part of the objective function to calculate the difference from the previous iteration result. Considering the first iteration step, substitute into the formula for calculation .

[0177] The first term in the objective function is:

[0178]

[0179] In the first step, substitute and :

[0180]

[0181] Calculate the first term of the objective function:

[0182]

[0183] 5.3 Second part: Regularization term

[0184] The second part of the objective function is the regularization term, which is calculated as follows:

[0185]

[0186] First, calculate each for its 2-norm. Assume for the first eigenvector :

[0187]

[0188] For calculate the function:

[0189]

[0190] Next, for each (assuming is a constant value, similarly calculate its and to perform the regularization term calculation. The total sum of the regularization terms is:

[0191]

[0192] 5.4 Calculate the total value of the objective function

[0193] Combining the first term and the second term, the total value of the objective function is:

[0194]

[0195] 5.5 Iterative optimization

[0196] Adjust the eigenvectors through multiple iterations to gradually approach the optimal abnormal eigenvector. In each iteration, update the value by minimizing the objective function to obtain more accurate abnormal features. In actual operation, use the gradient descent method or other optimization algorithms to solve these iterative steps. Assume that after multiple iterations, the optimal abnormal eigenvector is obtained:

[0197]

[0198] 5.6 Anomaly detection and determination

[0199] After completing the extraction of abnormal features, compare the optimal abnormal eigenvector with the statistical distribution of normal data. Assume the statistical distribution of normal data is:

[0200] The mean of normal data is 0.1969

[0201] The standard deviation of normal data is 0.0035

[0202] Compare the values of each optimal feature vector with the distribution of normal data, and use the following judgment criteria: If the value of a certain feature deviates from the normal mean by more than 3 times the standard deviation, then the feature is determined to be abnormal.

[0203] For , its deviation is , which is less than 3 times the standard deviation (3 * 0.0035 = 0.0105), so this feature is normal.

[0204] For , its deviation is , which is greater than 3 times the standard deviation, so this feature is abnormal.

[0205] For , its deviation is , which is less than 3 times the standard deviation, so this feature is normal.

[0206] For , its deviation is , which is less than 3 times the standard deviation, so this feature is normal.

[0207] For , its deviation is , which is less than 3 times the standard deviation, so this feature is normal.

[0208] According to the above judgment, only is determined to be an abnormal feature. Finally, by combining the analysis of each abnormal feature, it can be concluded that there are abnormal phenomena in the electric energy meter at certain moments. The order of the embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0209] Each embodiment in this specification is described in a progressive manner. For the same or similar parts between each embodiment, reference can be made to each other. The key points of each embodiment are to illustrate the differences from other embodiments.

[0210] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A method for automatically detecting and determining abnormal phenomena of an electric energy meter, characterized in that: Includes steps: S1. Real-time collection of multi-dimensional power parameter data, mapping the data to a nonlinear space, constructing a discriminant matrix, optimizing the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and performing multi-stage dynamic filtering on the feature vector of the power parameter data to obtain a purified feature vector; The discriminant matrix is ​​defined as: in, Represents the mapped data set, is the discriminant matrix, and is the projection matrix, which controls the distribution of data in the nonlinear space so as to maximize the difference between classes and minimize the difference within classes. It is to adjust i The parameters of data weights, is the transpose of the matrix, is the number of data after mapping; The optimal projection matrix is ​​obtained by minimizing the following objective function: in, represents the trace of the matrix, and are the intra-class and inter-class scatter matrices, respectively, which are defined as follows: in, is the mean vector of the k-th class of data, indicating the average position of all samples of this class in the feature space; is the global mean vector of all data, indicating the average position of all samples in the feature space; is the number of categories of data, is the number of samples in the kth class, indicating the total number of samples in this class; S2. The purified feature vector is processed by an adaptive inverse convolution transform algorithm to reconstruct the time and space characteristics of the signal, generate a frequency domain signal, and then restore the complete time feature signal through inverse transform. The optimal abnormal feature vector is obtained through a multiple iterative convergence algorithm, and the optimal abnormal feature vector is classified to complete automatic detection and judgment; The adaptive inverse convolution transform algorithm introduces multiple modulation and adaptive adjustment mechanisms in the convolution kernel design and inverse transform process; the purified feature vector is input into the convolution kernel function, which is defined as: in, represents the convolution kernel function, is the center position of the convolution kernel, Control the expansion factor of the convolution kernel, is the period parameter of the convolution kernel; Through the convolution operation, the purified feature vector is processed into a frequency domain signal. The convolution process formula is as follows: in, is the frequency domain signal, which is the representation of the convolved signal in the frequency domain. It is the shift parameter in the product operation, which indicates the amount of signal translation; After obtaining the frequency domain signal, the inverse transform operation is performed by combining the joint modulation of the frequency domain and the time domain. The formula is as follows: in, Represents the reconstructed signal, the modulation term in the inverse transform Used to dynamically adjust frequency components.

2. The method for automatic detection and determination of abnormal phenomena of electric energy meters according to claim 1 is characterized in that: In S1, the collected multi-dimensional power parameter data is preprocessed, and nonlinear mapping is performed on the preprocessed multi-dimensional data to extract deep nonlinear features in the data.

3. The method for automatic detection and determination of abnormal phenomena of electric energy meters according to claim 1 is characterized in that: In S1, adaptive weighted discriminant analysis is performed using the mapped data set, and a weighted discriminant matrix is ​​constructed to effectively distinguish normal data from abnormal data in a nonlinear space.

4. The method for automatically detecting and determining abnormal phenomena of an electric energy meter according to claim 1, characterized in that: In S1, an objective function is set up to obtain an optimal projection matrix, the intra-class scatter matrix is ​​subjected to eigenvalue decomposition to find its corresponding eigenvector, and the eigenvectors are connected to correspond to the direction in which the intra-class scatter is minimized; by solving the generalized eigenvalue problem, the direction in which the inter-class scatter is maximized is determined; and the eigenvectors with the largest first eigenvalues ​​are selected as column vectors of the projection matrix.

5. The method for automatically detecting and determining abnormal phenomena of an electric energy meter according to claim 1, characterized in that: In S1, in order to purify the feature vector, a multi-stage dynamic filtering method is adopted, a frequency domain filter is designed based on the spectral characteristics of the abnormal data, and each feature vector is frequency-domain filtered to remove high-frequency noise and low-frequency fluctuations. After applying the frequency domain filter, the feature vector is converted into a purified feature vector.

6. The method for automatically detecting and determining abnormal phenomena of an electric energy meter according to claim 1, characterized in that: In S2, features are extracted from the reconstructed signal, and the extracted feature vector is input into a multiple iterative convergence algorithm of the feature space. Through multiple iterative optimizations, the potential optimal abnormal feature vector in the signal is extracted.

7. The method for automatically detecting and determining abnormal phenomena of an electric energy meter according to claim 6, characterized in that: In S2, each feature is analyzed according to the value and feature distribution of the optimal abnormal feature vector to identify possible abnormal features; the statistical distribution of each feature under normal operating conditions is calculated and compared with the currently extracted features. If the value of a certain feature deviates from the normal distribution to a greater extent than a preset threshold, the current feature is considered to be abnormal; in order to further determine the type of abnormal phenomenon, the optimal abnormal feature vector is classified using a pre-trained classification model.

8. An automatic detection and determination system for abnormal phenomena of electric energy meters, used to implement the automatic detection and determination method for abnormal phenomena of electric energy meters as claimed in any one of claims 1 to 7, characterized in that: include: Feature vector purification unit: collects multi-dimensional power parameter data in real time, maps the data into nonlinear space, constructs a discriminant matrix, optimizes the projection matrix by minimizing the objective function to effectively distinguish normal and abnormal data, and performs multi-stage dynamic filtering on the feature vector of the power parameter data to obtain the purified feature vector; Automatic detection and judgment unit: The purified feature vector is processed by an adaptive inverse convolution transform algorithm to reconstruct the time and space characteristics of the signal, generate a frequency domain signal, and then restore the complete time feature signal through inverse transform. The optimal abnormal feature vector is obtained through a multiple iterative convergence algorithm, and the optimal abnormal feature vector is classified to complete automatic detection and judgment.

Images