An intelligent monitoring method and system for operating state of a mutual inductor

Abstract

The application belongs to the technical field of intelligent monitoring, and particularly relates to a mutual inductor operation state intelligent monitoring method and system, which comprises the following steps: acquiring real-time operation error characteristic data of the mutual inductor and constructing a multi-dimensional feature vector space, calculating the Minkowski distance between each data point and constructing a generalized hybrid distance matrix, determining a near neighbor cutoff distance and calculating the kernel weighted local density of each data point, performing reverse k-nearest neighbor search and calculating a density ratio, combining the reverse neighbor number and the density ratio to generate an isolation coefficient, mapping the isolation coefficient to a statistical distribution model to calculate an outlier probability value, updating an operation abnormality index in combination with the outlier probability mean and fluctuation variance in a sliding time window, and determining that the operation state of the mutual inductor is abnormal when the operation abnormality index exceeds the upper limit of a confidence interval. The application can realize accurate identification and early warning of mutual inductor operation faults and improve the reliability and accuracy of monitoring results.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology. More specifically, this invention relates to an intelligent monitoring method and system for the operating status of a current transformer. Background Technology

[0002] Current transformer condition monitoring primarily employs periodic offline calibration or online comparison based on fixed thresholds. However, current transformers exhibit time-varying and nonlinear error characteristics due to harmonic distortion from nonlinear loads and changes in environmental factors. Existing monitoring methods mainly analyze characteristics such as amplitude and phase deviations based on Euclidean distance or traditional clustering algorithms. However, when constructing the feature space, they cannot determine the changing trend of data points over time, treating data at each moment as independent samples. This makes it impossible to reliably distinguish between normal instantaneous fluctuations and potential slowly changing faults. Furthermore, distance measurement methods are prone to failure in high-dimensional feature spaces and fail to fully utilize the dimensional differences and correlations between different error parameters, making it difficult to accurately detect early, minor fault characteristics of current transformers.

[0003] In the field of density-based anomaly detection, traditional k-nearest neighbor (kNN) algorithms or local anomaly factor (LAB) algorithms can identify some outliers, but they are easily affected by local high-density clusters when dealing with transformer operating data with uneven density distribution. Especially for ambiguous data points at the boundary of normal operation and the critical state of fault, forward nearest neighbor search alone cannot determine the degree of isolation of the data points. Reverse kNN has advantages in identifying outliers in high-dimensional space, but there is no existing technology that combines kernel-weighted local density estimation with the reverse nearest neighbor mechanism. Furthermore, existing anomaly detection mechanisms are mostly based on instantaneous value judgments, which means they cannot smooth out the influence of random noise based on the statistical probability model of the sliding time window, making them highly susceptible to false alarms due to instantaneous disturbances on the power grid side. Summary of the Invention

[0004] To address the technical problems of existing monitoring methods being unable to determine time series trends, making it difficult to distinguish between instantaneous fluctuations and slowly changing faults, and being susceptible to noise interference leading to false alarms, this invention proposes an intelligent monitoring method and system for the operating status of instrument transformers. This method can accurately identify and provide early warnings of instrument transformer operating faults, and improve the reliability and accuracy of monitoring results.

[0005] In a first aspect, the present invention provides an intelligent monitoring method for the operating status of a current transformer, comprising: acquiring real-time operating error feature data of the current transformer and constructing a multi-dimensional feature vector space; calculating the Minkowski distance between each data point in the feature vector space, simultaneously calculating the trend deviation of the data points in the time series and mapping the deviation to a time series penalty factor, using the time series penalty factor to correct the Minkowski distance, and constructing a generalized hybrid distance matrix. The nearest neighbor cutoff distance of each data point is determined based on the generalized mixture distance matrix. The number of neighboring points within the cutoff distance range is counted. The generalized mixture distance of the neighboring points relative to the center point is weighted using the Gaussian kernel function to obtain the kernel-weighted local density of each data point. Perform reverse k-nearest neighbor search on the data points in the multidimensional feature vector space, establish a reverse k-nearest neighbor index list, count the number of reverse neighbors that include the current data point, calculate the ratio of the mean of the kernel-weighted local density of all reverse neighbors to the kernel-weighted local density of the current data point, and generate an isolation coefficient based on density difference by combining the number of reverse neighbors with the ratio. The isolation coefficient is mapped to a preset statistical distribution model to calculate the outlier probability value. The mean outlier probability and fluctuation variance within the sliding time window are combined to update the operating anomaly index of the current transformer. When the value of the operating anomaly index exceeds the upper limit of the set confidence interval, the operating state of the current transformer is determined to be abnormal.

[0006] By adopting the above technical solutions, real-time operational error feature data is obtained to construct a multi-dimensional feature vector space. A generalized mixed distance matrix is ​​constructed by combining the time-series penalty factor to correct the Minkowski distance. The trend change information of the time series is integrated into the spatial distance calculation to enhance the detection capability of data change trends. Gaussian kernel function weighted calculation of kernel-weighted local density is used to more meticulously identify the local distribution characteristics of data points. An isolation coefficient is constructed by combining reverse k-nearest neighbor search and density ratio to solve the detection problem caused by uneven density distribution in high-dimensional space and improve the discrimination of potential outliers. The isolation coefficient is mapped to outlier probability and smoothed using sliding time window statistics to suppress instantaneous noise interference. This anomaly judgment mechanism enables accurate identification and early warning of transformer operation faults and improves the reliability and accuracy of monitoring results.

[0007] Preferably, the step of acquiring the real-time operating error characteristic data of the current transformer and constructing a multi-dimensional feature vector space includes: forming an initial feature vector from the acquired current transformer amplitude deviation sequence, phase deviation sequence and harmonic distortion rate sequence; performing maximum and minimum normalization processing on each dimension of the initial feature vector; mapping the value of each dimension of the feature data to a closed interval between 0 and 1; and combining the normalized feature data of each dimension into a multi-dimensional feature vector space.

[0008] By adopting the above technical solution, the characteristic data of amplitude deviation, phase deviation and harmonic distortion rate are solved and subjected to maximum and minimum normalization to eliminate the order of magnitude difference between different physical dimensions of the transformer, and ensure that each dimension of the feature has equal weight in the distance calculation.

[0009] Preferably, for any two data points in the feature vector space, the absolute value of the difference between the two data points in each feature component is calculated to the power of P. The calculation results in all dimensions are added together and then the P-th power is taken to obtain the Minkowski distance between the two data points.

[0010] By adopting the above technical solution, the P-th power of the absolute value of the numerical difference between any two data points in each feature component of the multidimensional feature vector space is calculated, and the summation and P-th power are taken to obtain the Minkowski distance, which accurately represents the geometric separation degree of different operating states of the mutual inductor in the feature space.

[0011] Preferably, the construction of the generalized mixed distance matrix includes: calculating the first slope of change of the current data point on each feature component relative to the previous time step, and the second slope of change of the data point at the previous time step relative to the corresponding feature component; calculating the absolute value of the difference between the first slope and the second slope on each feature component as the sub-deviation of the dimension; summing the sub-deviations of the current data point on all feature components to obtain the total trend deviation of the current data point; and converting the total trend deviation into a time-series penalty factor; when constructing the generalized mixed distance matrix, for any two data points in the feature vector space, obtaining the time-series penalty factors of the two data points and taking the maximum value of the two, multiplying the maximum value by the Minkowski distance between the two data points to obtain the generalized mixed distance between the two data points.

[0012] By adopting the above technical solution, the first and second slopes of change of the current data point on each feature component are calculated, and the absolute value of the difference is used as the sub-deviation degree to generate the total trend deviation degree. The total trend deviation degree is converted into a time-series penalty factor using an exponential function to correct the Minkowski distance and construct a generalized mixed distance matrix. The abrupt change characteristics of the data are identified by the second-order difference principle to detect the trend of deterioration of the transformer state. The spatial distance of data points with abnormal change trends is increased, making it more difficult for them to find neighbors in density calculation and thus easier to be identified as isolated outliers. This improves the detection sensitivity and early warning speed for the early stage of gradual faults and abrupt states.

[0013] Preferably, determining the nearest neighbor cutoff distance for each data point based on the generalized mixed distance matrix includes: arranging all distance values ​​in the generalized mixed distance matrix in ascending order, and selecting the distance value located at a preset percentage threshold position of the total number as the nearest neighbor cutoff distance.

[0014] By adopting the above technical solution, all distance values ​​in the generalized mixed distance matrix are arranged in ascending order, and the distance value located at the preset percentage threshold of the total number is selected as the nearest neighbor cutoff distance. This achieves the truncation of data with different density distributions, ensuring that each data point has a sufficient number of neighboring points to participate in the local density calculation, avoiding statistical fluctuations due to the neighborhood being too small, and preventing the neighborhood from being too large and smoothing out local detailed features.

[0015] Preferably, the step of using a Gaussian kernel function to perform weighted calculation of the generalized mixture distance between neighboring points and the center point to obtain the kernel-weighted local density of each data point includes: for each center point, traversing all neighboring points within the cutoff distance range of the center point, calculating the quotient of the generalized mixture distance between the neighboring point and the center point divided by the nearest neighbor cutoff distance, calculating the negative of the square of the quotient, calculating the exponential function value of the negative value with the natural constant as the base, and summing the exponential function values ​​corresponding to all neighboring points to obtain the kernel-weighted local density of the center point.

[0016] By adopting the above technical solution, for each center point, the negative of the square of the quotient of the generalized mixed distance divided by the nearest neighbor cutoff distance is calculated by traversing all neighboring points within its cutoff distance range. The exponential function value with the base of the natural constant is then obtained and accumulated to obtain the kernel-weighted local density. The smoothing characteristics of the Gaussian kernel function are used to obtain the density of the local data distribution, avoiding the information loss caused by hard threshold truncation. The local density value is represented as a continuous real number, which more sensitively reflects the degree of clustering of the region where the data point is located, so as to accurately indicate potential abnormal outlier states.

[0017] Preferably, the step of performing a reverse k-nearest neighbor search on data points in the multidimensional feature vector space and establishing a reverse k-nearest neighbor index list includes: for each data point, finding the k nearest neighbors to the data point in the generalized mixed distance matrix; for each found neighbor, recording the data point that initiated the search in the reverse k-nearest neighbor index list of the neighbor, until all data points are traversed to complete the establishment of the reverse k-nearest neighbor index list.

[0018] Preferably, the step of updating the transformer's operational anomaly index by combining the mean outlier probability and the variance of fluctuation within the sliding time window includes: setting the length of the sliding time window to T, calculating the arithmetic mean of all outlier probability values ​​within the range from the current time to the past T-1 time, and calculating the variance of the outlier probability values; multiplying the arithmetic mean by a first weighting coefficient and adding it to the variance by a second weighting coefficient to obtain the transformer's operational anomaly index.

[0019] Secondly, the present invention provides an intelligent monitoring system for the operating status of a current transformer, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned intelligent monitoring method for the operating status of a current transformer is implemented.

[0020] By adopting the above technical solution, a computer program is generated from the above-mentioned intelligent monitoring method for the operating status of a current transformer and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be made based on the memory and the processor for convenient use.

[0021] The beneficial effects of this invention are as follows: This invention corrects the Minkowski distance by using a time-series penalty factor and integrates the trend change information of the time series into the spatial distance calculation to enhance the ability to detect data change trends. At the same time, it uses the Gaussian kernel function to calculate the weighted local density of neighboring points, which more accurately identifies the local distribution characteristics of data points.

[0022] Furthermore, by combining reverse k-nearest neighbor search with density ratio to construct isolation coefficients, not only is the detection problem caused by uneven density distribution in high-dimensional space solved, but the ability to distinguish potential outliers is also improved. By mapping isolation coefficients to outlier probabilities and combining them with sliding window statistics for smoothing and suppressing transient noise interference, a stable anomaly judgment mechanism is achieved, enabling the identification and early warning of transformer malfunctions and improving the reliability and accuracy of monitoring results. Attached Figure Description

[0023] Figure 1 This is a flowchart of an intelligent monitoring method for the operating status of a current transformer according to the present invention; Figure 2 This is a schematic diagram of the accelerating growth curve of trend deviation and time-series penalty factor; Figure 3 This is a schematic diagram of the alarm triggering characteristics of the upper bound of the confidence interval of the current transformer operation anomaly index based on a sliding time window. Detailed Implementation

[0024] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0025] This invention discloses an intelligent monitoring method for the operating status of a current transformer, referring to... Figure 1 This includes steps S1 to S4: S1. Obtain the real-time operating error characteristic data of the mutual inductor and construct a multi-dimensional feature vector space; calculate the Minkowski distance between each data point in the feature vector space, and at the same time calculate the trend deviation of the data points in the time series and map the deviation to the time series penalty factor. Use the time series penalty factor to correct the Minkowski distance and construct a generalized mixed distance matrix.

[0026] In an optional embodiment, a high-precision data acquisition terminal deployed on the secondary side of the instrument transformer synchronously acquires the secondary voltage or current signal of the instrument transformer at a set sampling frequency, and extracts the fundamental amplitude, fundamental phase, and 2nd to nth harmonic components via Fast Fourier Transform. The ratio difference between the instrument transformer's turns ratio and rated parameters is calculated as the amplitude deviation, the phase difference is calculated as the phase deviation, and the total harmonic distortion rate is calculated based on the ratio of each harmonic amplitude to the fundamental amplitude. The amplitude deviation, phase deviation, and harmonic distortion rate acquired at the same time point are combined to form an original feature vector. A minimum-max normalization method is used to map each feature value in the original feature vector to the interval between 0 and 1, eliminating dimensional differences. The normalized feature vector set constitutes a multidimensional feature vector space.

[0027] In the multidimensional feature vector space, the parameter p in the Minkowski distance formula is set to 2. The Euclidean distance between any two data points i and j is calculated as the basic spatial distance. Simultaneously, the time-series values ​​of the data points are extracted, and the predicted value at the current time is calculated using an exponential smoothing prediction algorithm. The absolute value of the difference between the actual monitored value and the predicted value is taken as the trend deviation. An exponential growth function with the trend deviation as the independent variable is constructed as the mapping function, and a time-series penalty factor greater than or equal to 1 is calculated. The larger the trend deviation, the larger the time-series penalty factor. Figure 2 As shown, the basic spatial distance is multiplied by the temporal penalty factor to obtain the corrected generalized mixed distance. This is then used to iterate through all data point pairs to generate a generalized mixed distance matrix composed of the generalized mixed distances.

[0028] In an optional embodiment, the feature data including amplitude deviation, phase deviation, and harmonic distortion rate are solved and normalized to construct a multidimensional feature vector space, including: The acquired transformer amplitude deviation sequence, phase deviation sequence, and harmonic distortion rate sequence are used to form an initial feature vector. Each dimension of the feature data in the initial feature vector is subjected to maximum and minimum normalization processing, and the value of each dimension of the feature data is mapped to a closed interval between 0 and 1. The normalized feature data of each dimension are combined to form a multi-dimensional feature vector space.

[0029] To eliminate the order-of-magnitude differences between the different physical dimensions of the current transformers and ensure that each dimension of features has equal weight in distance calculation, the statistical extreme values ​​of each feature component within the historical sampling period are determined. For any one-dimensional feature data x, the historical maximum value of the data is identified. and minimum value Using the formula Perform a linear transformation. For example, if the amplitude deviation varies from -0.5% to +0.5%, then a measured value of 0% will be mapped to 0.5. For out-of-limit data that may occur during online monitoring, a boundary truncation strategy is adopted, reducing values ​​greater than... The value is mapped to 1, less than The value is mapped to 0, or the statistical extremum update mechanism is triggered. By mapping three-dimensional data containing amplitude deviation, phase deviation, and harmonic distortion rate to a unit hypercube space, a standardized multidimensional feature vector is obtained. .

[0030] In an optional embodiment, calculating the Minkowski distance between data points in the feature vector space includes: For any two data points in the feature vector space, calculate the absolute value of the difference between the two data points in each feature component to the power of P. Add the calculation results in all dimensions and then perform the P-th power operation to obtain the Minkowski distance between the two data points.

[0031] Euclidean distance is preferably used as a specific implementation of Minkowski distance to represent the geometric separation degree of different operating states of the mutual inductor in the feature space. Specifically, suppose there are two data points i and j in the feature vector space, and their corresponding normalized feature vectors are respectively... and , where m is the feature dimension and P is 2. Calculate the squared difference between two points along each dimension. Summing the squared differences across all dimensions yields... The distance value is obtained by taking the square root of the sum. For example, if the coordinates of two points in three-dimensional space are (0.2, 0.2, 0.2) and (0.5, 0.6, 0.2), the calculated result is 0.5. A smaller distance indicates a closer state, while a larger distance indicates differences in their operational characteristics.

[0032] To detect the trend of transformer condition deterioration, the abrupt change characteristics of the data are identified using the second-order difference principle. In an optional embodiment, the trend deviation of data points in the time series is calculated and mapped to a time-series penalty factor. The Minkowski distance is then corrected using the time-series penalty factor to construct a generalized hybrid distance matrix, including: Calculate the first slope of change of the current data point in each feature component relative to the previous data point, and the second slope of change of the data point in the corresponding feature component relative to the data point in the previous time step. The absolute value of the difference between the first and second slopes of change on each feature component is calculated as the sub-deviation of the dimension. The sub-deviations of the current data point on all feature components are added together to obtain the total trend deviation of the current data point. The overall trend deviation is converted into a time-series penalty factor for the data point using an exponential function. The larger the overall trend deviation, the larger the value of the generated time-series penalty factor. When constructing the generalized mixture distance matrix, for any two data points in the eigenvector space, the temporal penalty factors of the two data points are obtained and the maximum value of the two is taken. The maximum value is multiplied by the Minkowski distance between the two data points to obtain the generalized mixture distance between the two data points.

[0033] Let the multidimensional eigenvector at time t be... Where d is the feature dimension, such as amplitude deviation, phase deviation, harmonic distortion rate, etc., and the sampling time interval is... For the m-th eigencomponent, the first slope of change... for The second slope of change for For data points whose initial time is insufficient to calculate the second difference, the total trend deviation is set to 0 by default. The total trend deviation of this data point is the sum of the sub-deviations of each dimension, i.e. .

[0034] Constructing mapping functions, for example As the time-series penalty factor for the corresponding data points, and These are preset adjustment parameters. Specifically, The baseline penalty adjustment coefficient, preferably ranging from 0.1 to 0.5, is used to control the baseline penalty intensity when the data is stable. This is a growth rate adjustment parameter, preferably ranging from 1 to 3, used to control the exponential amplification of the penalty factor when data mutations occur. In actual field deployment, as well as The specific value is determined offline based on the maximum trend deviation in the historical normal operation data of the current transformer. When the data changes smoothly, Dev approaches 0, and the penalty factor approaches 0. Alternatively, a preset baseline value can be used; when a sudden inflection point occurs, Dev increases, and the penalty factor rises exponentially.

[0035] For any two data points i and j in the feature vector space, their corresponding time-series penalty factors are respectively and To ensure the symmetry of the distance matrix, the generalized mixed distance is: This artificially widens the spatial distance between two related points if there are data points with abnormal changing trends. This makes it more difficult for data points with abrupt changes to find neighbors in density calculations, and thus more likely to be identified as isolated outliers, improving the detection sensitivity for the early stages of gradual faults and abrupt changes.

[0036] S2. Determine the nearest neighbor cutoff distance for each data point based on the generalized mixture distance matrix, count the number of neighboring points within the cutoff distance range, and use the Gaussian kernel function to perform weighted calculation on the generalized mixture distance between the neighboring points and the center point to obtain the kernel-weighted local density of each data point.

[0037] In an optional embodiment, all distance values ​​in the generalized mixture distance matrix are sorted in ascending order, and the distance values ​​located in the top K percent of the total number are selected as the globally unified nearest neighbor cutoff distance. For any central data point, all other data points are traversed, and points whose generalized mixture distance to the central point is less than the nearest neighbor cutoff distance are selected as neighborhood points, and the total number of neighborhood points is counted. The generalized mixture distance between the neighborhood point and the central point is substituted into the Gaussian kernel function formula to calculate the weight value corresponding to each neighborhood point. The closer the distance, the higher the weight. The weight values ​​of all neighborhood points of the central point are summed, and the result is the kernel-weighted local density of the data point.

[0038] In an optional embodiment, the nearest neighbor cutoff distance for each data point is determined based on the generalized mixed distance matrix, including arranging all distance values ​​in the generalized mixed distance matrix in ascending order and selecting the distance value located at a preset percentage threshold position of the total number as the nearest neighbor cutoff distance.

[0039] To achieve truncation of data with different density distributions, a parameter selection strategy based on statistical quantiles is adopted. The distance values ​​between all pairs of points calculated from the generalized mixed distance matrix are extracted to construct a one-dimensional distance set. This set is then sorted in ascending order using the quicksort algorithm. In the sorted sequence, the value located at an index position equal to, for example, 20% of the total length is set as the nearest neighbor cutoff distance. For example, if the distance set contains 1000 values, then the 200th value after sorting is taken as... A percentage threshold ensures that each data point has a sufficient number of neighboring points to participate in the local density calculation, avoiding statistical fluctuations caused by an excessively small neighborhood, while also preventing the smoothing away of local details by an excessively large neighborhood. However, those skilled in the art should know that the specific selection of the percentage threshold is related to the noise level of the transformer's operating environment and the data sampling frequency. For operating environments with strong electromagnetic interference and dense transient fluctuations, a larger percentage threshold, such as 15%-20%, is preferred to expand the neighborhood point collection range and enhance the smoothing and anti-interference capabilities against high-frequency noise. For scenarios with stable operating environments and high data signal-to-noise ratios, a smaller percentage threshold, such as 1%-5%, is preferred to avoid over-smoothing and thus accurately preserve local, minor early fault features.

[0040] In an optional embodiment, a Gaussian kernel function is used to weight the generalized mixture distance of neighboring points relative to the center point to obtain the kernel-weighted local density of each data point, including: For each center point, iterate through all neighboring points within the center point's cutoff distance range, calculate the quotient of the generalized mixed distance between the neighboring point and the center point divided by the nearest neighbor cutoff distance, calculate the negative of the square of the quotient, calculate the exponential function value of the negative of the negative with the natural constant as the base, and sum the exponential function values ​​corresponding to all neighboring points to obtain the kernel-weighted local density of the center point.

[0041] The density of local data distribution can be obtained using the smoothing properties of the Gaussian kernel function, avoiding information loss caused by hard thresholding. For a selected center point i, all points satisfying the generalized mixing distance are selected. Let j be a neighboring point. For each neighboring point, calculate the normalized relative distance. Then substitute it into the Gaussian decay formula Calculate the weights. The density weights of neighbors closer to the center point are closer to 1, while the weights of neighbors at the cutoff distance edge decay to 0.368. Kernel-weighted local density of center point i. The sum of the weights of all neighboring points, i.e. Density values ​​are continuous real numbers that can more sensitively reflect the degree of clustering of the area where data points are located. High density values ​​usually correspond to cluster centers in normal operating mode, while low density values ​​indicate possible anomalous outliers.

[0042] S3. Perform reverse k-nearest neighbor search on the data points in the multidimensional feature vector space, establish a reverse k-nearest neighbor index list, count the number of reverse neighbors that include the current data point, calculate the ratio of the mean kernel-weighted local density of all reverse neighbors to the kernel-weighted local density of the current data point, and generate an isolation coefficient based on density difference by combining the number of reverse neighbors and the ratio.

[0043] In an optional embodiment, a k-nearest neighbor search is performed on each data point in the space to determine the k nearest neighbors of each point. The k-nearest neighbor list of all points is traversed to find all points that list the current data point as a neighbor; these points constitute the reverse neighbor set of the current data. A reverse k-nearest neighbor index list containing the ID of each data point and its corresponding reverse neighbor set is established, and the number of elements in the set is counted to obtain the number of reverse neighbors. The kernel-weighted local density of all points in the reverse neighbor set is obtained, and the arithmetic mean of the kernel-weighted local density is calculated. The arithmetic mean of the kernel-weighted local density is divided by the kernel-weighted local density of the data point to obtain the relative density ratio. The isolation coefficient, representing the degree of anomaly of each data point, is calculated using the formula: I = R / (N + 1), where I is the isolation coefficient, R is the relative density ratio, and N is the number of reverse neighbors. When the number of reverse neighbors is 0, the isolation coefficient is directly assigned to the preset maximum isolation coefficient value.

[0044] To identify outliers, in an optional embodiment, a reverse k-nearest neighbor search is performed on the data points in the multidimensional feature vector space to build a reverse k-nearest neighbor index list, including: For each data point, find the k nearest neighbors of the data point in the generalized mixed distance matrix; For each nearest neighbor found, the data point that initiated the search is recorded in the reverse k-nearest neighbor index list of the nearest neighbor, until all data points are traversed and the reverse k-nearest neighbor index list is completed.

[0045] A nearest neighbor parameter k is set. The value of k is determined based on the total number of samples within the current sliding time window in the multidimensional feature vector space. It is fixed as an integer between 10 and 50, according to the average density of clusters formed by the typical normal operation history data of the transformer. For each point in the dataset... Based on the generalized mixed distance sorting, the k nearest neighbor points of a given point are identified. , , , Constructing a reverse index: Initialize an empty list set with a length equal to the total number of samples. For each pair of relations in the forward search process described above... The neighbor is ,Will A unique identifier is appended to the point. The corresponding reverse neighbor list. After traversal, the reverse neighbor list for each data point will contain all other points that consider that point to be its neighbor. Data points in a normal state are usually located within the k-nearest neighbor region of many other points, so they have a large number of reverse neighbors; however, data points in abnormal states of the current transformer rarely become neighbors of other points due to their off-center distribution, resulting in very few or even zero reverse neighbors. If the number of reverse neighbors is zero, the current data point is determined to be in an extremely isolated state, and it is directly assigned the preset maximum isolation coefficient value.

[0046] S4. Map the isolation coefficient to the preset statistical distribution model to calculate the outlier probability value. Combine the mean outlier probability and fluctuation variance within the sliding time window to update the operating anomaly index of the current transformer. When the value of the operating anomaly index exceeds the upper limit of the set confidence interval, the current transformer is determined to be in an abnormal operating state.

[0047] In an optional embodiment, the isolation coefficients calculated from historical normal operation data are fitted to a gamma distribution model. The currently calculated isolation coefficients are input into the cumulative distribution function of this gamma distribution, and the output value is used as the outlier probability value for that data point. A time sliding window of length N is set to store the outlier probability values ​​of the most recent N time points in real time. The arithmetic mean and variance of all outlier probability values ​​within the window are calculated. The weighted summation formula is used to synthesize the mean and variance of the outlier probabilities into a real-time operating anomaly index for the current transformer, where the mean reflects the duration of the anomaly and the variance reflects the severity of the anomaly fluctuation. A confidence interval upper bound of 95% is pre-set based on the baseline data of the current transformer under normal conditions. The operating anomaly index is compared with the upper bound of the confidence interval in real time. Once the operating anomaly index exceeds the upper bound of the confidence interval, such as... Figure 3 It immediately generates an anomaly detection signal and triggers an alarm.

[0048] In an optional embodiment, the operating anomaly index of the current transformer is updated by combining the mean outlier probability and the variance of fluctuation within the sliding time window, including: Set the length of the sliding time window to T, calculate the arithmetic mean of all outlier probabilities from the current time to the past T-1 time, and calculate the variance of the outlier probabilities. The arithmetic mean of the outlier probability values ​​is multiplied by the first weighting coefficient to obtain the first value. Then, the variance of the outlier probability values ​​is multiplied by the second weighting coefficient to obtain the second value. The first value and the second value are added together to obtain the operating anomaly index of the transformer.

[0049] To enhance the robustness of monitoring results and comprehensively assess the severity and volatility of anomalies, a first-in-first-out (FIFO) cache queue of length N is configured. Whenever a new outlier probability value is generated... When the time comes, push the value into the queue and remove the oldest data. Calculate the arithmetic mean of all data in the queue. This reflects the overall level of deviation of the current transformer's operating status within the current time period; it also calculates the variance of the queue data. This reflects the instability of the state. A first weighting coefficient is set. Focusing on abnormal amplitudes, the second weighting coefficient It focuses on fluctuation characteristics and satisfies In steady-state operation monitoring scenarios, the default setting is... The value is 0.6. The value is 0.4. When the monitoring system detects that the power grid is under high-frequency transient disturbance conditions such as frequent switching, in order to reduce the false alarm rate, the first weighting coefficient is dynamically decreased and the second weighting coefficient is increased. For example, it is set to... The value is 0.3. A value of 0.7 is used to enhance the weighting of the judgment of persistent abnormal fluctuation trends. The formula for calculating the anomaly index (Index) is as follows: The operational anomaly index is a comprehensive score. An anomaly alarm is only issued when the probability of a sustained high outlier or drastic fluctuations in probability cause the operational anomaly index to exceed the upper bound of the statistically determined confidence interval, thereby filtering out false alarms caused by single-point noise.

[0050] To verify the effectiveness of this invention, historical operating data of a current transformer in a 110 kV substation over 30 days was selected as the dataset, totaling 43,200 sampling points. This included 42,000 manually calibrated normal operation data points and 1,200 data points of gradual faults caused by early insulation performance degradation. The experimental environment was a high-performance computing workstation with an eight-core CPU and 32GB of RAM.

[0051] A control group and an experimental group were established for comparative testing. The control group only used Euclidean distance to calculate the distance between data points in the feature space, without including the aforementioned time-series penalty factor mechanism, and the subsequent steps remained the same. The experimental group adopted the complete generalized hybrid distance matrix construction method of this invention, which combines Euclidean distance with a time-series penalty factor based on trend deviation, wherein the adjustment parameter... Set to 0.5. Set to 1. The main performance indicators are fault detection rate, false alarm rate, and average response time to gradual faults.

[0052] The control group achieved a fault detection rate of 86.4%, a false alarm rate of 4.2%, and an average response time of 45 minutes. The experimental group achieved a fault detection rate of 97.8%, a false alarm rate of 0.9%, and an average response time of 23 minutes. These experimental data show that the experimental group outperformed the control group in all indicators, particularly in terms of a 22-minute reduction in average response time. The main reason for this is that the control group relied solely on Euclidean distance, which made it difficult to distinguish between the fault point and the statistical fluctuations of normal operation when the amplitude deviation was not yet significant in the early stages of the fault. The experimental group, by inputting a time-series penalty factor, utilized the second-order difference principle to detect the acceleration characteristics of data changes. When the transformer showed an early trend of performance deterioration, even if the amplitude did not exceed the limit, the trend deviation would increase, causing the time-series penalty factor to rise exponentially, thereby widening the generalized mixing distance between the data point and other points. This makes potential fault points more easily exposed as sparse outliers in density estimation and reverse k-nearest neighbor search, thus improving the sensitivity and early warning speed for early gradual faults.

[0053] This invention also discloses an intelligent monitoring system for the operating status of a current transformer, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, an intelligent monitoring method for the operating status of a current transformer according to the present invention is implemented.

[0054] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0055] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for intelligent monitoring of the operating status of a current transformer, characterized in that, include: Obtain real-time operational error characteristic data of the current transformer and construct a multi-dimensional feature vector space; Calculate the Minkowski distance between each data point in the feature vector space, and simultaneously calculate the trend deviation of the data points in the time series and map the deviation into a time series penalty factor. Use the time series penalty factor to correct the Minkowski distance and construct a generalized mixed distance matrix. The nearest neighbor cutoff distance of each data point is determined based on the generalized mixture distance matrix. The number of neighboring points within the cutoff distance range is counted. The generalized mixture distance of the neighboring points relative to the center point is weighted using the Gaussian kernel function to obtain the kernel-weighted local density of each data point. Perform reverse k-nearest neighbor search on the data points in the multidimensional feature vector space, establish a reverse k-nearest neighbor index list, count the number of reverse neighbors that include the current data point, calculate the ratio of the mean of the kernel-weighted local density of all reverse neighbors to the kernel-weighted local density of the current data point, and generate an isolation coefficient based on density difference by combining the number of reverse neighbors with the ratio. The isolation coefficient is mapped to a preset statistical distribution model to calculate the outlier probability value. The mean outlier probability and fluctuation variance within the sliding time window are combined to update the operating anomaly index of the current transformer. When the value of the operating anomaly index exceeds the upper limit of the set confidence interval, the operating state of the current transformer is determined to be abnormal.

2. The intelligent monitoring method for the operating status of a current transformer according to claim 1, characterized in that, The process of acquiring real-time operational error characteristic data of the mutual inductor and constructing a multi-dimensional feature vector space includes: The acquired transformer amplitude deviation sequence, phase deviation sequence, and harmonic distortion rate sequence are used to form an initial feature vector. Each dimension of the feature data in the initial feature vector is subjected to maximum and minimum normalization processing, and the value of each dimension of the feature data is mapped to a closed interval between 0 and 1. The normalized feature data of each dimension are combined to form a multi-dimensional feature vector space.

3. The intelligent monitoring method for the operating status of a current transformer according to claim 2, characterized in that, The calculation of the Minkowski distance between data points in the feature vector space includes: For any two data points in the feature vector space, calculate the absolute value of the difference between the two data points in each feature component to the power of P. Add the calculation results in all dimensions and then perform the P-th power operation to obtain the Minkowski distance between the two data points.

4. The intelligent monitoring method for the operating status of a current transformer according to claim 2, characterized in that, The construction of the generalized hybrid distance matrix includes: Calculate the first slope of change of the current data point in each feature component with respect to the previous time step, and the second slope of change of the data point at the previous time step with respect to the corresponding feature component with respect to the time step before that. The absolute value of the difference between the first and second slopes of change on each feature component is calculated as the sub-deviation of the dimension. The sub-deviations of the current data point on all feature components are summed to obtain the total trend deviation of the current data point. The total trend deviation is then converted into a time-series penalty factor. When constructing the generalized mixture distance matrix, for any two data points in the eigenvector space, the temporal penalty factors of the two data points are obtained and the maximum value of the two is taken. The maximum value is multiplied by the Minkowski distance between the two data points to obtain the generalized mixture distance between the two data points.

5. The intelligent monitoring method for the operating status of a current transformer according to claim 4, characterized in that, The determination of the nearest neighbor cutoff distance for each data point based on the generalized hybrid distance matrix includes: Arrange all distance values ​​in the generalized mixed distance matrix in ascending order, and select the distance value located at the preset percentage threshold of the total number as the nearest neighbor cutoff distance.

6. The intelligent monitoring method for the operating status of a current transformer according to claim 2, characterized in that, The method of using a Gaussian kernel function to weight the generalized mixture distance of neighboring points relative to the center point to obtain the kernel-weighted local density of each data point includes: For each center point, iterate through all neighboring points within the cutoff distance range of the center point, calculate the quotient of the generalized mixed distance between the neighboring point and the center point divided by the nearest neighbor cutoff distance, calculate the negative of the square of the quotient, calculate the exponential function value of the negative value with the natural constant as the base, and sum the exponential function values ​​corresponding to all neighboring points to obtain the kernel-weighted local density of the center point.

7. The intelligent monitoring method for the operating status of a current transformer according to claim 1 or 2, characterized in that, The step of performing a reverse k-nearest neighbor search on data points in the multidimensional feature vector space and establishing a reverse k-nearest neighbor index list includes: For each data point, find the k nearest neighbors of the data point in the generalized mixed distance matrix; For each nearest neighbor found, the data point that initiated the search is recorded in the reverse k-nearest neighbor index list of that nearest neighbor, until all data points are traversed and the reverse k-nearest neighbor index list is established.

8. The intelligent monitoring method for the operating status of a current transformer according to claim 1 or 2, characterized in that, The method of updating the anomaly index of the mutual inductor by combining the mean outlier probability and the variance of fluctuation within the sliding time window includes: Set the length of the sliding time window to T, calculate the arithmetic mean of all outlier probability values ​​from the current time to the past T-1 time, and calculate the variance of the outlier probability values; The arithmetic mean multiplied by the first weighting coefficient and the variance multiplied by the second weighting coefficient are added together to obtain the operating anomaly index of the mutual inductor.

9. An intelligent monitoring system for the operating status of a current transformer, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a method for intelligent monitoring of the operating status of a current transformer according to any one of claims 1-8 is implemented.

Images