Intelligent meter abnormal reading filtering system based on elastic time window

Abstract

The invention provides an intelligent meter abnormal reading filtering system based on an elastic time window. Self-adaptive recognition and filtering of abnormal reading are achieved through multi-stage modular processing. The system firstly forms an original reading sequence; calculating a fluctuation index of each sampling moment in the original reading sequence based on a preset statistical fluctuation algorithm, and generating a fluctuation sequence; generating a time window parameter set through an elastic adjustment function according to the fluctuation degree sequence; extracting a context subsequence according to the time window parameter set, extracting a time sequence feature and a frequency domain feature in the subsequence, and fusing to form a context fusion feature; calculating an abnormal confidence coefficient sequence based on the abnormal scoring function, performing abnormal filtering on the original reading sequence according to a preset confidence coefficient threshold value, and outputting a filtered reading sequence; according to the system, through fusion of the elastic time window and the multi-dimensional features, the abnormal reading recognition accuracy is effectively improved, and the system is suitable for real-time data quality control of the intelligent meter in a complex power application scene.

Description

Technical Field

[0001] The present invention relates to the technical field of electricity meters, and particularly to an intelligent meter abnormal reading filtering system based on an elastic time window. Background Art

[0002] In the prior art, intelligent meters are widely used in data acquisition and monitoring of power systems, and high-frequency real-time acquisition of various power parameters such as current, voltage, active power, reactive power, and energy consumption is realized through built-in sampling circuits and data upload modules. To improve data utilization efficiency, some systems introduce data processing algorithms such as sliding window analysis, fixed threshold detection, and short-term abnormal smoothing correction to perform preliminary abnormal identification and simple correction on the collected continuous readings, and then provide them for application modules such as power dispatching, load forecasting, and power consumption behavior analysis to call.

[0003] However, the above abnormal identification means in the prior art generally have problems of insufficient sensitivity, poor adaptability, and susceptibility to sudden disturbances. The fixed window and threshold mechanism is difficult to effectively accommodate multi-frequency disturbances, harmonic interference, and metering misdetection phenomena in power grid operation, resulting in coexistence of missed detection and misjudgment of abnormal readings, affecting the accuracy of subsequent data analysis and the stability of system operation. In addition, some methods lack comprehensive modeling of the context environment and temporal dynamic characteristics, and the processing results have insufficient adaptability to actual application scenarios.

[0004] In view of this, there is an urgent need to propose a new intelligent meter abnormal reading filtering scheme that can dynamically adapt to sampling fluctuation characteristics and fully integrate multi-dimensional context features. Summary of the Invention

[0005] This application particularly provides an intelligent meter abnormal reading filtering system based on an elastic time window to improve the accuracy of abnormal reading identification.

[0006] This application provides an intelligent meter abnormal reading filtering system based on an elastic time window, including: An original reading cache module for receiving continuous original readings generated by an intelligent meter under a unified sampling time reference to form an original reading sequence; A fluctuation degree calculation module for receiving the original reading sequence and calculating a fluctuation degree index corresponding to each sampling moment based on a preset statistical volatility algorithm to form a fluctuation degree sequence; A window generation module for receiving the fluctuation degree sequence and generating a set of time window parameters through an elastic adjustment function according to the fluctuation degree sequence, where the set of time window parameters includes a window start time, a window duration, and a window overlap ratio; a feature fusion module, configured to receive the original reading sequence and the time window parameter set, extract a context subsequence from the original reading sequence according to the time window parameter set, extract time series features and frequency domain features based on the context subsequence, and fuse them to form a context fusion feature; A scoring calculation module, configured to receive the context fusion feature and calculate an anomaly confidence sequence based on an anomaly scoring function; The reading filtering module is configured to receive the abnormal confidence sequence, perform abnormality filtering on the original reading sequence according to a preset confidence threshold, and generate a filtered reading sequence.

[0007] The beneficial effects of this application mainly include: (1) By introducing an elastic time window generation mechanism driven by a volatility sequence, the window range can be dynamically adjusted according to the real-time volatility of the sampled data, avoiding the problem of insufficient adaptation to sudden anomalies under the traditional fixed window strategy, thereby improving the accuracy and applicability of anomaly recognition. (2) By using the time series features extracted from the context subsequence and the frequency domain features for fusion analysis, the temporal evolution trend and spectral characteristics of the reading changes can be comprehensively considered, significantly improving the recognition and filtering capabilities of complex disturbances, harmonic interference and intermittent anomalies. (3) By calculating the anomaly confidence sequence based on the anomaly scoring function and combining it with the confidence threshold to control the filtering logic, the sensitivity and stability of anomaly judgment can be effectively balanced, false alarms and missed alarms can be reduced, and the overall quality and credibility of the output data of the smart meter can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 This is a schematic diagram of a smart meter abnormal reading filtering system based on an elastic time window provided in the first embodiment of the present application. DETAILED DESCRIPTION

[0009] The following description sets forth many specific details to facilitate a thorough understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the scope of the present application. Therefore, the present application is not limited to the specific implementations disclosed below.

[0010] The first embodiment of the present application provides a smart meter abnormal reading filtering system based on elastic time window. Figure 1 , which is a schematic diagram of the first embodiment of this application. Figure 1 A first embodiment of the present application provides a smart meter abnormal reading filtering system based on a flexible time window.

[0011] The intelligent meter abnormal reading filtering system based on an elastic time window includes an original reading caching module 101, a fluctuation degree calculation module 102, a window generation module 103, a feature fusion module 104, a scoring calculation module 105, and a reading filtering module 106.

[0012] The original reading caching module 101 is used to receive continuous original readings generated by an intelligent meter under a unified sampling time reference to form an original reading sequence.

[0013] The original reading caching module 101 undertakes the functions of unified acquisition and caching of the original data of the intelligent meter in this system. Its core purpose is to provide stable, continuous, and data input with a unified sampling time reference for subsequent abnormal filtering processing. This module is at the front end of the signal processing link in the system architecture and is directly connected to the sampling output interface of the intelligent meter device. To ensure the accuracy of data acquisition and the system timing consistency, a unified time synchronization mechanism is integrated inside the original reading caching module 101. This time synchronization mechanism can adopt the Network Time Protocol (NTP), the Precision Time Protocol (PTP), or other high-precision time synchronization technologies to ensure that the original readings from multiple meter devices are aligned under a unified time reference, eliminating the time errors introduced by clock drift or sampling offset. Through the time synchronization module, a unified sampling timestamp sequence is generated inside the system, and each sampling data is attached with a unique sampling timestamp to ensure the integrity and continuity of the time series during subsequent analysis.

[0014] The input of the original reading caching module 101 is the real-time power parameter measurement values continuously output by the intelligent meter. The measurement values can include, but are not limited to, multi-dimensional power characteristic values such as active power, reactive power, current, voltage, power factor, total harmonic distortion, etc. In this embodiment, for simplicity of description, the following takes a single measurement quantity, such as the root mean square value of current, as an example for illustration, but the design of this module is fully applicable to the parallel processing of multi-dimensional measurement data. The original reading caching module 101 adopts a dual-buffer architecture, which consists of a real-time acquisition cache unit and a sequential write cache unit. The real-time acquisition cache unit is responsible for receiving the real-time sampling data stream of the intelligent meter at high speed. It is usually organized in the form of a circular buffer, supporting high-speed continuous writing to avoid data reception loss. The sequential write cache unit then sorts and stores the sampling data in the order of the sampling timestamps to form a complete original reading sequence. This sequential write cache has the ability to dynamically expand its capacity. When the buffer is close to full, it can automatically apply for expanded space or release historical data according to a preset data holding time threshold to ensure the stable operation of the system under high-frequency sampling conditions.

[0015] In terms of data integrity guarantee, the original reading cache module 101 is configured with an integrity detection mechanism. Before each piece of sampled data is written into the cache, data legality verification is performed, including numerical range detection, outlier flag detection, and format consistency verification. For example, when the measured current value exceeds the calibration range of the smart meter or there is data format damage, the system will automatically record the abnormal flag bit and retain the abnormal data for further analysis and judgment by the subsequent abnormal filtering logic. Through this mechanism, system anomalies caused by data corruption can be effectively avoided.

[0016] To meet the requirements of high-reliability applications, the original reading cache module 101 supports the ability to continue storing data after power failure. The cache module can be optionally equipped with non-volatile storage media, such as industrial-grade solid-state drives or FRAM storage chips, to quickly restore the historical data that has not been completed for processing after the system power failure or abnormal restart, ensuring data integrity and system continuity. After the system resumes, the cache module first restores the cache state before power failure and fills in the subsequent data collection at the breakpoint in the order of timestamps to ensure the time continuity of the overall data sequence without missing.

[0017] In terms of sampling frequency control, the original reading cache module 101 supports flexible configuration of the sampling frequency. The system can set the sampling interval according to the application scenario requirements, such as 1 second, 500 milliseconds, or a shorter period. The sampling frequency parameter can be remotely configured through the central control module. After the sampling frequency is configured, the cache module drives the sampling process based on the unified clock signal to ensure that each smart meter device outputs readings at a synchronous rhythm, facilitating the subsequent context analysis module to accurately divide and process the data window.

[0018] After the original reading cache module 101 forms the original reading sequence, it provides a complete output of the original reading sequence to the subsequent fluctuation calculation module 102. This original reading sequence is defined as a time series array in terms of data format, and each element contains the sampling timestamp and the corresponding reading value. The data structure is as follows: , where is the timestamp at the th sampling moment, and is the reading value of the th sampling. This data format is a unified format standard within the system, ensuring smooth, efficient, and unambiguous data flow between modules.

[0019] In summary, through unified time reference control, high-speed cache management, data integrity verification, power failure data protection, and format standardization encapsulation, the original reading cache module 101 realizes the stable reception, organization, and sequential output of the original data of smart meters, laying a rigorous and reliable data foundation for the subsequent module's anomaly recognition and filtering.

[0020] The fluctuation degree calculation module 102 is configured to receive the original reading sequence, calculate the fluctuation degree index corresponding to each sampling moment based on a preset statistical fluctuation algorithm, and form a fluctuation degree sequence.

[0021] The function of the fluctuation degree calculation module 102 is to evaluate the fluctuation strength of each sampling moment in sequence according to the time order for the original reading sequence output by the original reading cache module 101, so as to reflect the magnitude of the reading change range in different time periods. The calculation of the fluctuation degree is carried out in a sliding window manner, that is, for each data point in the original reading sequence, the system takes a certain number of adjacent data points forward and backward with this data point as the center to form a local data subsequence. The length of this local subsequence can be set according to the system design requirements. For example, it can include twenty-five data points before and after, so as to form a subsequence consisting of fifty-one points.

[0022] In each subsequence, it is first necessary to calculate the average level of all readings within the subsequence. The system will add up the numerical values of all readings within the subsequence to obtain the total sum, and then divide the total sum by the number of data points within the subsequence to finally obtain the average value of this subsequence. This average value reflects the overall level of the readings during this time period.

[0023] Next, the system starts to calculate the fluctuation degree of this subsequence. The meaning of the fluctuation degree is to describe the deviation degree of each reading relative to this average value. For this purpose, the system will sequentially take out each reading within the subsequence and calculate the difference between it and the average value. Each calculated difference will be squared to eliminate the influence of positive and negative offsets, so that all offset values are shown as positive numbers. The system then accumulates all these squared differences to obtain the total offset.

[0024] In order to convert this total offset into an average offset degree, the system will divide this total offset by the total number of data points in the subsequence to obtain the overall average offset amplitude of the subsequence. Since the offset was squared before, the average offset amplitude obtained at this time is still in square units. Therefore, the system will further take the square root of it to finally obtain the standardized fluctuation degree index. This result is the fluctuation degree value at the current sampling moment, which clearly reflects the overall fluctuation condition of the data points near this sampling moment.

[0025] To prevent individual abnormal readings from interfering with the fluctuation result, the fluctuation calculation module 102 also sets an outlier suppression mechanism. Before entering the fluctuation calculation, the system first calculates the median value of the subsequence, that is, sorts all the readings by size and takes the value at the middle position as the representative value. Then, the system calculates the absolute distance between each reading and this median value, sorts these absolute distances again, and takes the absolute distance at the middle position as the representative value of the overall deviation. If the deviation degree of some readings from the median value far exceeds several times this representative value, these readings will be temporarily excluded from the current fluctuation calculation to avoid overestimating the overall fluctuation due to abnormal jumps of individual readings.

[0026] When processing near the start or end position of the reading sequence, it may not be possible to obtain a complete number of data points before and after the target sampling moment. For this reason, the fluctuation calculation module 102 has the ability of boundary adaptability. When the number of available data points is less than the preset requirement, the system will dynamically reduce the length of the sliding window and only use the actually available data points to participate in the fluctuation calculation, but still ensure that the number of data participating in the calculation is not lower than the minimum safety threshold, so as to balance the calculation stability and the processing integrity of the boundary data.

[0027] To improve the overall calculation efficiency, the fluctuation calculation module 102 adopts a recursive calculation method. When the sliding window moves forward by one data point each time, the system no longer performs a full calculation on all the data in the window, but quickly updates the sum and offset according to the numerical change between the newly entered data point and the data point that has moved out of the window. In this way, when facing a high-frequency data stream, the system can still complete the fluctuation calculation at each sampling moment in an extremely short time, ensuring the real-time processing ability of the system.

[0028] Finally, the system will generate a corresponding fluctuation value for each sampling moment, and organize the fluctuations at all sampling moments into a complete fluctuation sequence according to the time order. This fluctuation sequence not only strictly corresponds to the original reading sequence in terms of data structure, but also has exactly the same time stamp, which is convenient for subsequent modules to directly call and align when generating elastic time windows and feature fusion. Through the above detailed fluctuation calculation logic, the present invention realizes the accurate quantification of the dynamic change characteristics of the original reading sequence, providing a solid and reliable feature input basis for subsequent anomaly filtering.

[0029] For example, for a certain continuous sequence of original readings, the system samples at a sampling frequency of 1 time per second, obtaining the following local sampling data sequence: {10.1, 9.8, 10.3, 10.0, 9.9, 10.2, 10.4, 10.0, 9.7, 10.3, 10.0, 20.5, 10.1, 9.8, 10.2, 10.0, 9.9, 10.3, 10.1, 10.0, 9.8}.

[0030] The system calculates the volatility index at the 12th sampling moment (i.e., the position of the outlier 20.5). It is set that the sliding window length takes 10 data points before and after, and a total of 21 data points form a subsequence.

[0031] First, the system calculates the median of this subsequence as the reference central value before outlier removal. After sorting all the readings, the median is 10.0.

[0032] Then, the system calculates the absolute distance between each reading and the median to form an absolute deviation sequence: {0.1, 0.2, 0.3, 0.0, 0.1, 0.2, 0.4, 0.0, 0.3, 0.3, 0.0, 10.5, 0.1, 0.2, 0.2, 0.0, 0.1, 0.3, 0.1, 0.0, 0.2}.

[0033] After sorting this absolute deviation sequence again, the median absolute deviation (MAD) is 0.2. The outlier removal threshold is set to 5 times MAD (i.e., 1.0). The system identifies the readings with an absolute deviation exceeding 1.0 as outlier readings. In this example, only the reading 20.5 is determined to be an outlier and is not included in the current round of volatility calculation.

[0034] After removing the outlier, the system recalculates the average value based on the remaining 20 valid readings, and its value is 10.015.

[0035] Subsequently, the system squares the difference between each valid reading and this average value, and the sum of the squared deviations is 0.570.

[0036] Dividing this sum by the number of valid samples 20, the average squared deviation is 0.0285.

[0037] Finally, taking the square root of this average squared deviation, the standardized volatility index is 0.169.

[0038] The system outputs the fluctuation index as the fluctuation value at the current 12th sampling moment. The system sequentially performs the above calculation process for each sampling moment of the entire original reading sequence to form a complete fluctuation sequence.

[0039] From the above calculation example, it can be seen that the fluctuation calculation module 102 of the present invention can still stably extract the normal background fluctuation intensity in the presence of short-term abnormal jump readings, effectively avoiding the amplification interference of isolated anomalies on the fluctuation index, and ensuring that the subsequent elastic time window generation module 103 has a reliable input basis when adjusting the window scale.

[0040] Furthermore, the fluctuation calculation module is further configured to: Calculate the median of the current subsequence as the robust central position within each sliding window; Calculate the absolute deviation of each reading relative to the median, and generate an absolute deviation sequence; Extract the median absolute deviation from the absolute deviation sequence, and dynamically calculate the abnormal rejection threshold according to the median absolute deviation, where the abnormal rejection threshold is obtained by coupling and adjusting the adaptive threshold amplification factor with the current fluctuation trend; Identify and temporarily reject the readings that exceed the abnormal deviation range according to the abnormal rejection threshold, and perform a robust standard deviation calculation after the abnormal rejection is completed to obtain a fluctuation index after abnormal interference suppression for subsequent window generation modules to call.

[0041] In the present invention, the fluctuation calculation module does not simply use the traditional standard deviation calculation method, but introduces a robust statistical method to enhance its anti-interference ability in the presence of abnormal jump data. In a specific implementation, the fluctuation calculation module first extracts a local data subsequence including the sampling moment at each target sampling moment through a sliding window mechanism. The length of the sliding window can be flexibly configured according to system settings. For example, 25 data points can be taken before and after, so as to form a subsequence with a length of 51 points, ensuring that each fluctuation calculation is based on sufficient context information.

[0042] After obtaining the subsequence within the current sliding window, the system does not directly calculate the mean value, but first determines the median of the subsequence as the robust central position of the window. The definition of the median is: all sampling readings within the subsequence are arranged in ascending order according to the numerical size. When the total number of data is odd, the median is the reading value at the middle position; when the total number of data is even, the median takes the average of the two middle readings. Since the median is not affected by isolated extreme values compared with the mean value, when there are occasional abnormal mutation readings within the subsequence, the robustness of the central position can still be maintained to prevent overall deviation.

[0043] After determining the robust center, the system sequentially calculates the deviation of each reading in the subsequence from the median. This is the absolute value of the difference between each reading and the median, forming an absolute deviation sequence. This absolute deviation sequence fully reflects the degree of dispersion of each reading around the median, without offsetting each other due to the direction of the deviation (positive or negative).

[0044] After obtaining the absolute deviation sequence, the system further extracts the median absolute deviation as a robust offset statistic. The definition of the median absolute deviation is consistent with the aforementioned median logic: the absolute deviation sequence is sorted in ascending order and the middle value is taken. If the total number is odd, the middle value is taken; if the total number is even, the average of the two middle values is taken. Compared to the standard deviation, the median absolute deviation is more robust and less susceptible to amplification by single extreme outliers, making it a suitable basic indicator for calculating anomaly thresholds.

[0045] After obtaining the median absolute deviation, the system calculates the anomaly rejection threshold based on it. This calculation utilizes dynamic adaptive adjustment logic. The system pre-sets a threshold amplification factor, typically a positive real number greater than 1, such as 3, 4, or 5, to control the sensitivity of anomaly detection. If the amplification factor is set to 4, the initial anomaly detection range is four times the median absolute deviation.

[0046] However, in order to make the anomaly rejection threshold dynamically adaptable within the overall system, the system further introduces the volatility trend coupling adjustment logic. The volatility trend is defined as: during long-term operation, the system calculates the mean and standard deviation of the median absolute deviation in each sliding window, and dynamically records its changing trend. When the overall volatility trend is in an upward state, the system automatically relaxes the value of the threshold amplification factor to prevent excessive misjudgment; when the volatility trend is in a downward state, the threshold amplification factor is appropriately contracted to improve sensitivity. Through this coupling mechanism, the anomaly rejection threshold is determined not only based on the local fluctuation characteristics of the current window, but also takes into account the global background trend, so that it can quickly respond to short-term anomalies and adapt to the overall changes in data in the long term.

[0047] After determining the dynamic anomaly rejection threshold, the system performs anomaly detection on all readings within the current sliding window. Readings with absolute offsets exceeding the threshold are marked as temporarily abnormal and removed from the current volatility calculation. This rejection process does not affect the subsequent feature extraction module or the final output of the filtered reading sequence; it serves only as an intermediate processing step in the volatility calculation.

[0048] After the abnormal readings are removed, the system recalculates the standard deviation based on the remaining valid data points to form a robust standard deviation index as the final volatility output. This robust standard deviation is the volatility index after abnormal interference suppression, and it will be used as an input in the subsequent window generation module for dynamically adjusting the duration and overlapping ratio of the elastic time window. Through the above-mentioned robust volatility calculation logic, the present invention can effectively suppress the amplification effect of isolated mutation interference on volatility estimation in a complex intelligent meter operating environment, and greatly improve the stability and overall abnormal recognition accuracy of the abnormal reading filtering system in a multi-source disturbance scenario.

[0049] For example, within the sliding window corresponding to a certain sampling moment, the system extracts a total of 51 consecutive sampling data points to form the current subsequence. The system first sorts these 51 sampling values in ascending order according to their numerical magnitudes. After sorting, the value of the data point at the 26th position in the sequence is 10.0. Therefore, the system determines that the median of this sliding window is 10.0, serving as the robust central position of this window.

[0050] Next, the system sequentially calculates the absolute value of the difference between each data point in the subsequence and the median 10.0 to form an absolute deviation sequence. For example, some sampling values within the subsequence are 9.8, 10.2, 10.1, 10.0, 11.5, 9.7. Then their corresponding absolute deviation values are 0.2, 0.2, 0.1, 0.0, 1.5, 0.3 respectively. The system calculates the absolute deviations of all 51 data points in this way to form a complete absolute deviation sequence.

[0051] Subsequently, the system sorts the absolute deviation sequence in ascending order again to obtain the median absolute deviation. Since the sequence length is odd, 51, the absolute deviation value at the 26th position is taken as the median absolute deviation. After sorting, the absolute deviation value at the 26th position is 0.3, so the median absolute deviation is 0.3.

[0052] After obtaining the median absolute deviation, the system calculates the abnormal rejection threshold according to the adaptive dynamic adjustment logic. The system presets the threshold amplification factor to 4. At the same time, the current global volatility trend assessment shows that the overall system is in a high volatility stage. After coupling adjustment, the system decides to correct the amplification factor to 4.5. At this time, the system multiplies the median absolute deviation 0.3 by the corrected amplification factor 4.5 to calculate the current abnormal rejection threshold of 1.35.

[0053] Therefore, within this sliding window, any data point with an absolute deviation exceeding 1.35 will be marked as a temporary outlier and excluded from the calculation of the standard deviation of the volatility. For example, in the above subsequence, there is a sampling value of 11.5 with an absolute deviation of 1.5, which exceeds 1.35 and is identified as an abnormal reading and will not be involved in the current round of volatility calculation.

[0054] After completing the outlier exclusion, the system recalculates the standard deviation based on the remaining valid data to obtain the volatility index after abnormal interference suppression. This volatility index is then used as the input basis for generating subsequent time window parameters for the subsequent abnormal reading identification process.

[0055] The window generation module 103 is used to receive the volatility sequence and generate a set of time window parameters through an elastic adjustment function based on the volatility sequence. The set of time window parameters includes the window start time, window duration, and window overlap ratio.

[0056] The core function of the window generation module 103 is to dynamically generate a set of elastic time window parameters according to the volatility sequence output by the volatility calculation module 102. Its generation logic not only directly inherits the real-time nature and volatility characteristics of the volatility sequence but also enables the subsequent feature extraction module to adopt a more contextually matched window for the reading sequence under different volatility states, thereby improving the accuracy and adaptability of abnormal identification.

[0057] After the volatility sequence is input into the window generation module 103, the system first reads the volatility value corresponding to each sampling moment in turn. To determine the specific parameters of each time window, the system has preset the mapping logic of the elastic adjustment function. This logic is based on the following principle: when the volatility is small, it indicates that the readings in the current data sequence change smoothly during this period, and the data is relatively stable and reliable. At this time, the duration of the time window can be appropriately extended to cover more continuous data within the window and form a larger context range; while when the volatility is large, it indicates that there are relatively large fluctuations or potential abnormal risks in the readings in a short period. To avoid including too much unstable data within the window, the duration of the window should be shortened at this time to make the window focus on the data within a shorter local time period, facilitating subsequent modules to focus on fine-grained abnormal features.

[0058] After determining the duration of the time window, the system also needs to determine the start time of each window. The window generation module 103, based on a unified sampling time reference, usually takes the sampling moment where the current volatility index is located as the center position of the window and allocates half of the duration time range forward and backward respectively to form a symmetric time coverage interval. If it encounters the boundary of the data sequence, the system will automatically adjust the window start time to ensure that the actual data points covered by the window are always within the available reading sequence range.

[0059] In addition, the window generation module 103 also needs to determine the overlap ratio between each window. The setting of the overlap ratio can be configured according to the system's requirements for the sensitivity of anomaly detection. Generally, when a higher coverage accuracy is desired for the anomaly detection process, the system will set a higher overlap ratio, causing a large overlapping area between adjacent time windows, so as to ensure that even if an abnormal reading appears only for a short time, there are still multiple windows available to capture and evaluate it. On the contrary, when the system is dealing with relatively stable or high-real-time application scenarios, the overlap ratio can be correspondingly reduced to reduce repeated calculations and improve the overall processing speed and system load balancing ability. The window generation module 103 allows the system to dynamically adjust this overlap ratio parameter, and can support overlap configurations of, for example, 10%, 30%, 50% or even higher ratios.

[0060] In the specific parameter generation process, the window generation module 103 will first input the volatility value into the adjustment function, which contains multiple segments of hierarchical logic inside. For example, when the volatility is in the low value range, the system automatically allocates a relatively large window duration, such as covering more than fifty data points; when the volatility is in the medium range, a medium-length window duration is allocated; and when the volatility exceeds the high value threshold, the window duration is reduced to the minimum coverage range to enhance the response sensitivity to abnormal changes. This multi-segment adjustment logic can not only be set according to the actual characteristics and experience of the power system, but also be adjusted and optimized online according to the historical anomaly recognition effect during the operation of the system.

[0061] The specific parameters of each time window are recorded in a standardized data structure, including information such as the start time, duration, and overlap ratio of the window. These parameters are organized into a complete set of time window parameters after generation, strictly corresponding to each sampling moment in the input volatility sequence one by one, and keeping the timestamps exactly the same. The entire set of time window parameters is then passed to the feature fusion module 104 as a standardized output to guide the subsequent extraction of context subsequences and feature calculations.

[0062] The window generation module 103 plays a key role in dynamic adjustment and real-time adaptation in the overall system architecture. By effectively converting the real-time sampled volatility information at the front end into time window parameters, the system can continuously maintain a reasonable size and resolution ability of the anomaly detection window in the face of various complex data volatility situations such as sudden changes in power load, harmonic disturbances, metering errors, and short-term sampling anomalies. This flexible window generation ability makes the present invention different from the existing fixed window mechanism, and is especially suitable for the accurate identification task of abnormal readings in the complex operating environment of the power system.

[0063] For example, the present invention can provide a function model that dynamically adjusts the time window parameter in combination with the real-time volatility index. The elastic adjustment function takes the volatility index at the current moment as the core input and outputs two core parameters: the window duration and the overlap ratio. The starting moment of the window is still automatically located by the unified sampling benchmark without additional participation of the function. The basic expression of the elastic adjustment function is: ; ; In the above expression, represents the window duration (counted by the number of sampling points) generated at the -th sampling moment, represents the window overlap ratio (expressed as a percentage), represents the volatility index output by the volatility calculation module 102 at the -th sampling moment, represents the upper limit of the maximum window length allowed by the system, represents the adjustment coefficient of the degree of shrinkage of the window length by the volatility, represents the upper limit of the maximum allowed overlap ratio, represents the adjustment coefficient of the expansion speed of the overlap ratio by the volatility.

[0064] The basic logic of the formula is that when the volatility is smaller, approaches , that is, the system allows the use of a longer time window to cover more consecutive readings and form a larger context range; when the volatility increases, the denominator increases accordingly, resulting in decreasing, achieving automatic contraction of the window length. When the volatility changes violently, the system automatically shortens the window, which helps the subsequent module to focus on the mutation characteristics.

[0065] For the overlap ratio part, the preliminary overlap calculation result is obtained by multiplying the volatility by the adjustment coefficient . If the calculated value exceeds the maximum allowed overlap ratio , it is limited by to prevent the overlap ratio from expanding infinitely. This design ensures that in the high-volatility stage, the system appropriately increases the window overlap, which helps to capture short abnormal segments, and avoids excessive computational load in the stable stage.

[0066] The following specifically illustrates the application process of the elastic adjustment function with a set of example parameters. Assume that the system sets = 100 (i.e., the maximum can cover 100 sampling points), = 5, = 60% (i.e., maximum 60% overlap), = 10%. At a certain sampling moment , the fluctuation calculation module 102 obtains = 2.0 (e.g., in standard deviation units). Substituting into the formula for calculation, we get: ; Since the result needs to be rounded up to the complete number of sampling points, and finally the duration of the current time window is determined to be 10 sampling points. Then calculate the overlap ratio: ; Therefore, the time window parameters generated at this moment are that the start moment is located according to the sampling timestamp, the duration is 10 sampling points, and the overlap ratio .

[0067] In another relatively stable scenario, if , substituting we get: ; ; At this time, the window duration automatically expands to 50 sampling points, and the overlap ratio drops significantly to , fully reflecting the current data stability.

[0068] In the present invention, the window start moment is not directly calculated through an elastic adjustment function, but is dynamically arranged in a recursive manner based on the system's unified sampling clock order, through the generated time window duration and the overlap ratio . The specific logic can be expressed as: ; Among them, is the initial sampling moment, and are the duration and overlap ratio obtained through the elastic adjustment function at the current moment respectively. The product term represents the window sliding step size. is the start moment of the next window .

[0069] For example, the system starts from the timestamp seconds. The first sampling moment , through the elastic adjustment function, we get sampling points, . And it is stipulated that the first window starts at .

[0070] The starting time of the second window is: ; If the sampling frequency is 1 Hz, the starting time of the second window is seconds. The third window is recursively derived in sequence, and so on.

[0071] Through the design of the above elastic adjustment function, the present invention realizes the dynamic adjustment of the window coverage range and overlap degree under different fluctuation states, which not only improves the ability to capture short-term anomalies in high-fluctuation scenarios, but also effectively controls computing resources in stable scenarios.

[0072] Furthermore, the window generation module is specifically used for: After receiving the fluctuation degree sequence, calculate the preliminary time window duration and overlap ratio through the elastic adjustment function to form the original time window parameters; For each time window, detect whether the starting time and ending time in the sampling sequence are out of bounds. If there is an out-of-bounds situation, extract the number of actually available data points to form the boundary remaining length index; Input the boundary remaining length index into the boundary compensation function, dynamically calculate the correction coefficient according to the window deviation amount, and perform a decreasing adjustment on the preliminary duration to form the corrected duration; According to the corrected duration, recalculate the overlap ratio to maintain the window sliding balance, form a set of boundary adaptive window parameters, and recursively generate the starting time of the next window according to the corrected parameters to ensure the time continuity and context integrity of the entire time window layout at the sequence boundary.

[0073] In the present invention, the core task of the window generation module is to calculate the duration and overlap ratio of the time window using the elastic adjustment function after obtaining the fluctuation degree sequence, and combine the boundary adaptive mechanism to ensure that the entire time window layout still has a stable context coverage ability at the sequence boundary. The calculation logic of the window generation module highly depends on the fluctuation degree sequence, and calculates the window parameters separately for each sampling moment during the dynamic sliding process. The entire process has complete progression, coupling, and real-time adaptability.

[0074] When the window generation module receives the volatility sequence, it first reads the corresponding volatility index for each sampling moment. The system has pre-set an elastic adjustment function, and the design principle of this function is: when the volatility is small, the system determines that the current data is stable, so the window can cover more data points; when the volatility is large, it indicates that the data changes violently, and the system needs to shorten the window length to improve the ability to capture local changes in time. The implementation method of the elastic adjustment function is: the system first sets a maximum window length, for example, 100 sampling points. When the volatility index is equal to 0, the window duration directly takes the maximum window length. When the volatility gradually increases, the system shortens the window length by shrinking a certain proportion per unit of volatility. For example, assume that 5 sampling points are shrunk per unit of volatility. Then when the volatility at a certain moment is 2.0, the window duration is shortened to the maximum window length minus 5 multiplied by 2, resulting in 90 sampling points. In this way, the system generates a set of preliminary time window duration parameters under different volatilities.

[0075] After calculating the preliminary duration, the system continues to perform the preliminary calculation of the overlap ratio. The design logic of the overlap ratio is: when the volatility is small, the overlap ratio is set low to reduce calculation redundancy; when the volatility is high, to improve the ability to capture short-term anomalies, the overlap ratio is appropriately increased. For example, the system sets the overlap ratio to one-tenth of the volatility. If the current volatility is 0.3, the overlap ratio is 3%. To avoid the infinite amplification of the overlap ratio, the system sets a maximum overlap ratio upper limit of 60%. If the volatility is greater than the critical value, the overlap ratio is fixed at 60% and no longer increases. Through the elastic adjustment function, the system generates the preliminary time window duration and overlap ratio for each sampling moment, forming the original time window parameter set.

[0076] After obtaining the preliminary window parameters, the system further detects the boundary conditions. For each time window, the system calculates its starting sampling moment and the expected ending sampling moment, and determines whether it exceeds the actual available data range of the original reading sequence. If the ending position of a window has exceeded the sequence termination point, or the starting position is earlier than the sequence starting point, it is determined that there is a boundary overstep phenomenon for this window. For the window with overstep, the system counts the number of actually available data points in this window, forming the boundary remaining length index. This index directly reflects the number of effective data points available for this window under the non-overstep condition.

[0077] For each boundary remaining length index, the system introduces a boundary compensation function to correct the window duration. The design logic of the boundary compensation function is as follows: when the window offset out-of-bounds distance is small, the system can maintain the overall window strategy by fine-tuning the window length; when the offset distance is large, the system needs to significantly shorten the window duration to avoid exceeding the data sequence range. Specifically, the system sets a compensation threshold. When the remaining available length is more than 90% of the target duration, the system allows the original duration to be retained without adjustment; when the remaining available length is less than 90% of the target duration, the system directly shortens the duration according to the remaining length. For example, if the initial duration is 50 sampling points and the actual remaining length is 42 sampling points, since 42 is less than 90% of 50, the system directly revises the duration to 42 sampling points.

[0078] After correcting the duration, the system synchronously revises the overlap ratio to maintain the slip balance. The new overlap ratio is deduced by keeping the original number of overlapping sampling points unchanged. For example, if the original duration is 50 and the overlap ratio is 20%, then the number of overlapping sampling points is 10. When the corrected duration is 42, the system keeps the number of overlapping sampling points still 10. Therefore, the new overlap ratio is 10 divided by 42, resulting in approximately 23.8%. Through this dynamic adjustment method, the system can still maintain the smoothness and continuity of context coverage when the window duration changes, avoiding excessive overlap or excessive jumping between windows.

[0079] After both the window duration and the overlap ratio are corrected, the system recursively calculates the start time of the next window based on the final window parameter set. The recursive logic uses the corrected duration multiplied by one minus the corrected overlap ratio as the slip step to deduce the start time of the next window from the start time of the previous window. In this way, the system dynamically generates a set of continuous, complete and boundary self-adaptive optimized time window layouts over the entire time series. The entire window generation process has extremely strong adaptability in a dynamically fluctuating environment, especially in the regions near the start and end of the sequence, and can automatically balance the window length and the integrity of context information, effectively improving the anomaly recognition accuracy and the overall robustness of the system.

[0080] The feature fusion module 104 is used to receive the original reading sequence and the time window parameter set, extract context subsequences from the original reading sequence according to the time window parameter set, respectively extract time series features and frequency domain features based on the context subsequences, and fuse them to form context fusion features.

[0081] The feature fusion module 104 undertakes the core tasks of context information modeling and multi-dimensional feature extraction and fusion in the system of the present invention. The inputs of this module include the original reading sequence output by the original reading cache module 101 and the set of time window parameters output by the window generation module 103. Its final output is the context fusion feature, which is used to provide a complete basis for anomaly discrimination for the subsequent anomaly scoring module.

[0082] After receiving the set of time window parameters, the feature fusion module 104 first extracts the corresponding context subsequence from the original reading sequence according to the start time and duration of each time window. Specifically, the system locates the corresponding timestamp position in the original reading sequence according to the start time recorded in the time window parameters. Subsequently, starting from this start time, the system reads the corresponding number of data points forward or backward until the complete duration range is covered. Each context subsequence strictly maintains time continuity, and the number of data points is exactly the same as the window definition, ensuring the integrity and validity of the data during subsequent feature extraction.

[0083] After completing the extraction of the context subsequence, the feature fusion module 104 executes two feature extraction paths: time series feature extraction and frequency domain feature extraction. Time series feature extraction mainly focuses on the evolution trend and dynamic change pattern of the data in the time dimension. First, the system calculates the overall average level of the data within the subsequence to reflect the basic reading level of the current time period. Subsequently, the system evaluates the fluctuation amplitude of the subsequence and calculates its standard deviation to characterize the data change amplitude within a short period of time. Then, the system calculates the slope change feature of the sequence, that is, analyzes the rising or falling trend of the readings within the subsequence over time. This feature helps to identify slow anomaly states that are gradually rising or falling. In addition, the system also evaluates the skewness and kurtosis of the data, which respectively reflect the asymmetry of the data distribution and the degree of abnormal spikes, further revealing the characteristics of possible extreme outlier readings in the data. Finally, the system evaluates the change amplitude distribution between consecutive sampling points based on the sliding difference method to mine the distribution pattern of mutation signals in time continuity. These time dimension features comprehensively capture the multi-level dynamic information of the data during the time evolution process.

[0084] In the frequency-domain feature extraction path, the feature fusion module 104 performs spectral analysis on the context subsequence. First, the system converts the time-series data into frequency-domain energy distribution information through built-in fast Fourier transform (FFT) processing logic. Through the frequency-domain transformation, the system can observe whether there are periodic oscillation signals or high-frequency interference signals in a specific frequency band in the data. The system extracts multiple frequency-domain feature indicators such as the main energy-concentrated frequency in the spectrum, the total energy distribution, the ratio of low-frequency to high-frequency energy, the intensity of harmonic components, and the high-frequency energy concentration. These frequency-domain features have a strong ability to identify abnormal fluctuation phenomena introduced by harmonic interference, power equipment switching noise, or instantaneous jitter in the signal sampling circuit. The system can also detect the frequency band with abnormally concentrated energy in the spectrum through energy threshold analysis to assist in discriminating systematic oscillation anomalies or interference signals from specific disturbance sources.

[0085] After the respective extraction of the time-series features and frequency-domain features is completed, the feature fusion module 104 enters the fusion processing link. The system standardizes all the feature indicators obtained from the above two feature paths, so that features with different dimensions and value ranges are all converted into a unified numerical scale. The standardization methods can include interval normalization, Z-score normalization, or percentile normalization, depending on the overall feature space design of the system. All the standardized features are integrated into a feature vector and weighted and integrated according to the preset fusion weights. The fusion weights can be set according to empirical rules, model training optimization results, or system adaptation parameters to ensure a reasonable balance between time features and frequency-domain features in the fused feature representation, and each contributes an appropriate proportion of the abnormal discrimination ability. The fused feature vector finally completely represents the comprehensive portrait of the abnormality possibility of the original readings in the current time window in the multi-dimensional feature space.

[0086] After the feature fusion module 104 generates the fused features, it organizes these context-fused features into an ordered feature sequence, strictly aligns them with the time window parameter set, and transfers them as a complete output to the scoring calculation module 105. The entire feature fusion process not only has the ability to dynamically adapt to the local data fluctuation state but also fully absorbs the dual information of the data change trend and frequency anomaly signals, greatly enhancing the comprehensive discrimination accuracy of the subsequent anomaly scoring link.

[0087] Furthermore, the feature fusion module is specifically used for: After receiving the context subsequence, performing a fast Fourier transform on the context subsequence to obtain the corresponding spectrum amplitude sequence; In the spectrum amplitude sequence, according to the preset frequency division rule, extracting the high-frequency interval amplitude subsequence and the low-frequency interval amplitude subsequence, and respectively calculating the total energy of the high-frequency interval and the total energy of the low-frequency interval, and then calculating the high-frequency energy ratio based on the total energy of the high-frequency interval and the overall total energy to form a high-frequency energy ratio indicator; Input the high-frequency energy ratio index into a high-frequency anomaly non-linear weighting function. Among them, the high-frequency anomaly non-linear weighting function calculates a high-frequency weighted anomaly contribution factor based on a set exponential increasing function relationship according to the deviation degree between the high-frequency energy ratio index and the high-frequency energy critical threshold; Use the high-frequency weighted anomaly contribution factor as a high-frequency energy enhancement feature, perform feature splicing and normalization processing with the time series features according to a preset fusion logic to form a multi-dimensional context fusion feature vector, and provide the multi-dimensional context fusion feature vector to a scoring calculation module for performing an anomaly confidence score. Among them, the time series features include sequence mean, standard deviation, slope change feature, skewness feature, and kurtosis feature.

[0088] In the present invention, the core task of the feature fusion module is to fully extract data change features in two dimensions of the time domain and the frequency domain for each context subsequence within each time window, and improve the recognition rate of anomaly changes in the high-frequency energy features through a non-linear enhancement mechanism, and finally form a multi-dimensional context fusion feature vector available for the scoring calculation module. The entire feature fusion process has a complete progressive logic, and the features are highly coupled and complementary to each other, so that the anomaly score has good stability and sensitivity.

[0089] After the feature fusion module receives the context subsequence provided by the window generation module, the system first completes data preprocessing in the time domain to ensure that the length of the context subsequence is consistent with the sampling frequency. After completing the preprocessing, the system performs a fast Fourier transform on the context subsequence. The fast Fourier transform is an algorithm that efficiently converts a time domain sequence into a frequency domain amplitude sequence, and its output result is an amplitude data sequence corresponding to different frequency components, reflecting the energy distribution of each frequency component in the original data.

[0090] After obtaining the spectrum amplitude sequence, the system divides the spectrum into a high-frequency interval and a low-frequency interval according to a preset frequency division rule. The division rule can be set according to the highest sampling frequency of the sampling system and the application scenario. For example, if the system sampling frequency is 100 Hz, the spectrum range is 0 to 50 Hz, and the system can set 0 to 10 Hz as the low-frequency interval and above 10 Hz as the high-frequency interval. The system sequentially extracts the amplitudes corresponding to each frequency point in the high-frequency interval to form a high-frequency interval amplitude subsequence, and at the same time extracts the amplitudes corresponding to each frequency point in the low-frequency interval to form a low-frequency interval amplitude subsequence.

[0091] The system calculates the energy for the amplitude subsequence in the high - frequency range and the amplitude subsequence in the low - frequency range respectively. Specifically, for each frequency point, the square of its amplitude represents the energy contribution at that frequency point. The system accumulates the energies of all frequency points in the high - frequency range to form the total energy of the high - frequency range, and at the same time accumulates the energies of all frequency points in the low - frequency range to form the total energy of the low - frequency range. Subsequently, the system adds the total energy of the high - frequency range and the total energy of the low - frequency range to obtain the overall total energy of the complete context subsequence. The high - frequency energy ratio is defined as the ratio of the total energy of the high - frequency range divided by the overall total energy, forming the high - frequency energy ratio index. The higher the value of this index, the higher the proportion of high - frequency disturbance components in the current subsequence, indicating the possibility of potential harmonic interference, rapid current fluctuations, or short - term abnormal jitters.

[0092] To further improve the recognition sensitivity of high - frequency energy abnormal signals, the system introduces a high - frequency abnormal non - linear weighting function. The basic design idea of the high - frequency abnormal non - linear weighting function is that when the high - frequency energy ratio is close to or exceeds the normal threshold, the abnormal weight should be rapidly increased so that the subsequent scoring calculation module can give priority to the high - frequency disturbance characteristics. The system first sets a high - frequency energy critical threshold, such as 20%. When the high - frequency energy ratio is lower than 20%, the system determines that the overall spectrum energy distribution is within the normal range, and the output of the non - linear weighting function is close to 1. When the high - frequency energy ratio exceeds 20%, the system rapidly amplifies the abnormal contribution according to the preset exponential gain logic. For example, when the high - frequency energy ratio is 25%, the system can design the output of the non - linear function to be 1.5, and when the high - frequency energy ratio rises to 30%, the output of the non - linear function rapidly increases to 2.5. Specifically, this exponential gain logic can be implemented in the system through look - up tables, piece - wise mapping, or fitting models, without complex real - time calculations, ensuring the algorithm operation efficiency.

[0093] After being processed by the non - linear weighting function, the system finally obtains the high - frequency weighted abnormal contribution factor. This factor can be regarded as an amplification and adjustment coefficient for high - frequency energy characteristics, strengthening the representation ability of high - frequency disturbances in the overall abnormal recognition.

[0094] After obtaining the high - frequency weighted abnormal contribution factor, the system fuses it with other feature dimensions. First, the system extracts the time - series features of this context subsequence in the time domain, specifically including sequence mean, standard deviation, slope change feature, skewness feature, and kurtosis feature. The slope change feature is obtained by calculating the slope of the fitting line after linearly fitting the context subsequence; the skewness feature is obtained by calculating the ratio of the third - order central moment to the cube of the standard deviation; the kurtosis feature is obtained by calculating the fourth - order central moment. All these time - domain and frequency - domain features are normalized through standardization to make their dimensions consistent, facilitating subsequent feature splicing and scoring calculations.

[0095] After standardization, the system, according to the preset fusion logic, splices the high-frequency weighted abnormal contribution factor and time series features, including sequence mean, standard deviation, slope change feature, skewness feature, and kurtosis feature, to form a multi-dimensional context fusion feature vector. This fusion feature vector retains the complete information of the original data in terms of time change trend, frequency perturbation component, distribution symmetry, and extreme fluctuation characteristics. The system provides the finally generated multi-dimensional context fusion feature vector to the scoring calculation module as the direct input basis for the scoring calculation module to perform abnormal confidence scoring.

[0096] Through the above feature fusion mechanism, the present invention can accurately capture high-frequency interference, trend mutation, and distribution shift abnormal signals during the operation of intelligent meters, and still has excellent abnormal recognition accuracy and adaptability in a multi-source complex abnormal mixed environment, providing stable and reliable multi-dimensional feature support for the discrimination of subsequent scoring functions.

[0097] The scoring calculation module 105 is used to receive the context fusion feature and calculate the abnormal confidence sequence based on the abnormal scoring function.

[0098] The scoring calculation module 105 undertakes the core task of abnormal discrimination scoring for the fusion feature in the system of the present invention. Its calculation result directly determines the confidence level of whether the data in each time window is judged as normal or abnormal, and is an important decision-making link in the entire abnormal filtering logic chain. The input of the scoring calculation module 105 is the context fusion feature sequence output by the feature fusion module 104, and each fusion feature vector therein completely contains the comprehensive expression of the time series feature and frequency domain feature in this time window.

[0099] When the scoring calculation module 105 receives the fusion feature sequence, the system first performs a standard consistency check on each fusion feature vector to ensure that the input feature format is complete, without missing items, and has a one-to-one timestamp matching relationship with the corresponding time window parameters. The scoring calculation module 105 is pre-integrated with an abnormal scoring function library, which is jointly defined and optimized through statistical learning, historical data training, and expert experience models, and is used to evaluate the abnormal tendency of samples in the fusion feature space.

[0100] The scoring calculation module 105 processes using a comprehensive scoring function. Inside the scoring function, each input feature is evaluated item by item according to the logic of collaborative analysis of multiple feature dimensions. For the time series feature part, the scoring function first examines the deviation amplitude of the average level. When the average level deviates too much from the historical statistical mean range, the scoring function assigns a higher abnormal risk factor to this feature. Secondly, the scoring function analyzes the abnormal inflation degree of the fluctuation amplitude, that is, when the standard dispersion degree is much higher than the empirical interval of the stable operation period of the same period, a higher abnormal tendency score is given to this index. Then, the scoring function evaluates the slope change feature. If the data has a rapid upward or downward trend and exceeds the safe change rate threshold, the contribution degree of its abnormal score is increased. For the skewness and kurtosis indexes, the scoring function combines the distribution laws of the symmetry imbalance degree and the extreme peak degree, identifies local abnormal signals caused by isolated reading jumps or signal burst amplitude transitions, and accumulates the corresponding scoring weights accordingly.

[0101] In the frequency domain feature part, the scoring calculation module 105 analyzes the abnormal concentration phenomenon of high-frequency energy. If there is an obvious aggregation signal in the high-frequency band energy component, and this frequency band should show low-energy characteristics during normal stable operation, it is considered that there may be harmonic interference, electrical flicker or other high-frequency disturbance sources, and the system includes this as a high-priority risk feature in the scoring. The system also analyzes the deviation of the low-frequency dominant frequency distribution. When the position of the main energy frequency band deviates significantly from the normal operating frequency range, it indicates that the data may be affected by equipment switching fluctuations or periodic failures, and the abnormal scoring weight is also increased. For the overall energy change trend of the spectrum, the scoring function evaluates whether the total energy has a significant increase in a short period of time, identifies abnormal states such as potential signal amplification and measurement gain drift, and includes them in the abnormal score contribution.

[0102] In the comprehensive scoring stage, the scoring calculation module 105 sums up the abnormal tendency scoring factors of the above-mentioned each feature dimension through a preset weight coefficient to form a single abnormal confidence value. The weight coefficient inside the scoring function can be set through the system initial model, or can be continuously optimized and updated through the feedback learning and model adaptive mechanism during the actual operation of the system, so as to continuously improve the accuracy and environmental adaptability of the scoring function. The scoring function adopts a non-linear scoring logic in design, avoiding the over-reliance of the scoring result on a single feature dimension caused by simple linear accumulation, and ensuring that the scoring output fully reflects the collaborative abnormal performance among multiple features.

[0103] After the scoring calculation module 105 completes the scoring calculation for each fused feature vector, it organizes the finally obtained anomaly confidence values into an anomaly confidence sequence in chronological order. Each confidence value in this anomaly confidence sequence is within the standardized interval, for example, represented by continuous real numbers between 0 and 1, where a value close to 1 represents a high anomaly tendency, and a value close to 0 represents a very low anomaly risk. The entire anomaly confidence sequence remains strictly synchronized and aligned with the original time window, facilitating the subsequent reading filtering module 106 to directly call and execute the anomaly screening logic.

[0104] Through the above multi-feature, multi-weight, multi-dimensional interactive scoring logic, the scoring calculation module 105 effectively improves the overall accuracy and sensitivity of anomaly recognition. Especially when facing various types of anomaly signals such as high-frequency noise, intermittent jitter, trend drift, and mutation spikes in complex power data, it demonstrates excellent comprehensive adaptation ability and real-time scoring response ability.

[0105] For example, the anomaly scoring function in the scoring calculation module 105 can adopt a comprehensive anomaly scoring function that combines time series features and frequency domain features. Its scoring function adopts a weighted normalization summation form in structure, and the formula is as follows: ; Among them, the final anomaly scoring value is represented by indicated by, is the normalization factor, and its value is equal to the sum of all weight coefficients, that is equal to to the sum of each coefficient.

[0106] In this scoring function, represents the average deviation multiple. Its calculation logic is that the system statistically obtains the average value interval of the original reading sequence under normal operating conditions through long-term historical stable operation data, and then calculates the actual average value of the context subsequence within the current time window. Take the absolute value of the difference between the actual average value and the average value of the normal interval, and divide this difference by the standard deviation of the historical average value to obtain the relative deviation multiple.

[0107] represents the abnormal expansion multiple of the fluctuation amplitude. Its calculation logic is that the system first calculates the standard degree of dispersion within the current window, that is, the sample standard deviation. Then compare this standard deviation with the average value of the standard deviation within the historical normal interval, and the ratio of the two is the abnormal multiple of the fluctuation amplitude.

[0108] Indicates the abnormal multiple of slope change. Its calculation logic is as follows: within the current context subsequence, perform a linear regression fit on the readings along the time axis to calculate the slope of the fitted line. Subsequently, compare this slope with the normal change rate range statistically obtained during long-term stable operation to calculate the relative offset multiple.

[0109] Indicates the abnormal degree of data skewness. The system first calculates the skewness of the data within the context subsequence, and the calculation of skewness uses the standard third-order central moment formula. Subsequently, subtract the mean skewness within the historical stable operation interval from this skewness value, take the absolute value, and then divide by the standard deviation of skewness in the normal fluctuation range to obtain the skewness abnormal multiple. 。

[0110] Indicates the abnormal degree of data kurtosis. The system first calculates the kurtosis within the context subsequence. Its calculation uses the fourth-order central moment formula minus 3 to obtain the standardized kurtosis. Then compare this kurtosis with the average kurtosis in the historical interval to obtain the relative offset multiple.

[0111] Indicates the abnormal multiple of high-frequency energy. In frequency domain analysis, the system first obtains the frequency spectrum through fast Fourier transform. Subsequently, define the high-frequency interval, for example, the frequency band above 5 times the main frequency. Take the ratio of the sum of energies within this frequency band to the sum of total spectrum energies to obtain the high-frequency energy ratio. Compare the high-frequency energy ratio with the average high-frequency energy level during normal operation to obtain the improvement multiple, which is defined as 。

[0112] Indicates the abnormal multiple of low-frequency dominant frequency offset. In frequency domain analysis, the system first locates the dominant frequency component with the maximum energy in the frequency spectrum and records its central frequency. Subsequently, subtract the central frequency of the dominant frequency interval under historical stable operation from this central frequency, take the absolute value of the difference, and normalize it to the standard deviation unit of the normal interval, that is, 。

[0113] Indicates the abnormal multiple of total energy. The system first calculates the total energy of the current window frequency spectrum, and then compares this total energy with the average total energy of the frequency spectrum under historical stable operating conditions to calculate the offset multiple. 。

[0114] Each of the above offset multiple indicators is derived from the context fusion feature extraction output completed by the feature fusion module 104. Therefore, when performing abnormal scoring, the scoring calculation module 105 does not need to repeat feature calculations, but only needs to directly calculate the offset multiple according to the set logic.

[0115] For example, in a specific calculation, the feature offset multiples extracted by the system within the context subsequence are respectively: = 2.5, = 3.0, = 1.8, = 2.2, = 2.0, = 4.5, = 1.6, = 3.8.

[0116] In this scoring model, the weight coefficients configured by the system are respectively: = 10, = 10, = 8, = 5, = 5, = 12, = 8, = 10.

[0117] First, calculate the normalization factor Z, which is the sum of all the above weight coefficients, and get Z = 68. Then substitute all the indicators into the scoring function, calculate the sum of the products of each indicator multiplied by the weight coefficient respectively, and finally divide by Z. The value of the abnormal score S obtained is approximately 0.598.

[0118] This abnormal score value represents the abnormal tendency probability level shown by the readings in the comprehensive feature space within the current time window. When its value is close to 1, it indicates a high abnormal risk, and when its value is close to 0, it indicates basically stable. The scoring calculation module 105 executes the above scoring logic for all the fused feature vectors in sequence to form a complete abnormal confidence sequence for subsequent calls by the reading filtering module 106.

[0119] Furthermore, the scoring calculation module is also used for: After completing each round of abnormal scoring, compare the abnormal confidence sequence output by the scoring calculation module with the filtered reading sequence output by the reading filtering module, calculate the scoring residuals at the corresponding positions of the two at each sampling moment, and form a scoring residual vector; After obtaining the scoring residual vector, count the average offset value, standard deviation, and extreme residual sampling point frequency of the scoring residual vector to form a set of scoring residual characteristic indicators; Input the set of scoring residual characteristic indicators into the weight adjustment module, and dynamically correct the weight coefficients in the scoring function according to the residual contribution of each feature dimension in the abnormal scoring. Among them, for the feature dimension with a higher scoring residual frequency, reduce its corresponding weight coefficient to form a weight suppression correction factor; for the feature dimension with a lower scoring residual frequency and stable offset, increase its corresponding weight coefficient to form a weight enhancement correction factor, and then generate an updated set of scoring weight coefficients; Under the updated set of scoring weight coefficients, the anomaly scoring function is retrained to optimize the discrimination ability of the scoring function in the current context fusion feature vector space based on the training results, and after the training is completed, it is applied to the scoring calculation of subsequent anomaly confidence sequences, realizing the long-term adaptive dynamic optimization of the characteristics of the input data by the scoring calculation module.

[0120] In the present invention, the scoring calculation module not only undertakes the task of generating anomaly confidence scores based on the fusion feature vectors, but also has the ability of self-feedback learning and dynamic weight adjustment, enabling it to continuously optimize the discrimination performance of the scoring function during long-term continuous operation. After each round of anomaly scoring and filtering, the scoring calculation module analyzes the difference between the scoring output and the actual filtering result, and dynamically corrects the weight coefficients of each feature dimension in the scoring function, thereby enhancing the adaptive ability of the overall scoring function in complex scenarios.

[0121] After the scoring calculation module completes the scoring calculation of the current round of context fusion feature vectors and outputs the anomaly confidence sequence, the system compares this anomaly confidence sequence with the filtered reading sequence output by the reading filtering module. Since the reading filtering module performs anomaly elimination based on the anomaly confidence sequence output by the scoring function and a preset confidence threshold, there is a certain difference between the scoring output and the final filtering result. The system sequentially compares the anomaly confidence with the filtering result flag at each sampling moment. If the scoring result at a certain sampling moment deviates from the filtering result (for example, the scoring value is close to the threshold but not eliminated, or the scoring value is low but actually eliminated), it is regarded as a scoring residual at this position. The system forms a scoring residual vector within the entire sequence through this comparison logic, with a corresponding scoring residual value at each sampling moment. The definition of the scoring residual can be realized based on the distance or discrimination difference between the scoring value and the actual anomaly label, and the larger its value, the higher the scoring error.

[0122] After obtaining the scoring residual vector, the system performs statistical analysis on this vector to extract the set of scoring residual feature indicators. First, the system calculates the average offset value of the scoring residual vector, which reflects the offset trend of systematic overestimation or underestimation of the overall scoring system. Subsequently, the system calculates the standard deviation of the scoring residual vector, which is used to describe the dispersion degree of the scoring error and the scoring stability. When the standard deviation of the scoring residual is large, it indicates that the discrimination ability of the scoring function in the current fusion feature space fluctuates. The system also further counts the frequency of extreme residual sampling points, that is, counts the number of sampling moments when the scoring residual value exceeds a preset high error threshold, so as to capture the risk signal of frequent misjudgment of the scoring model in a specific feature dimension.

[0123] After the scoring residual feature index set is extracted, the system inputs this set into the weight adjustment module to perform dynamic correction of the weight coefficients of the scoring function. The weight adjustment logic attributes the scoring error to each feature dimension of the fused feature vector by analyzing the source of the scoring residual contribution. Specifically, the system determines the contribution ratio of each feature dimension to the output of the scoring function through the feature sensitivity matrix recorded during training. If a certain feature dimension frequently dominates the scoring deviation in high-residual samples, the system determines that there is a problem of over-strength contribution of this feature dimension leading to scoring imbalance, and then implements weight suppression correction on its corresponding weight coefficient to reduce its weight value to form a weight suppression correction factor. On the contrary, for the feature dimension with a low frequency of scoring residuals and a stable and reliable scoring result, the system appropriately increases its corresponding weight to form a weight enhancement correction factor, thereby expanding its influence in the scoring function.

[0124] After weight suppression correction and weight enhancement correction, the system forms an updated set of scoring weight coefficients. This set of weight coefficients fully reflects the feedback correction of the current scoring model based on the true filtering results, effectively eliminates the bias trend of the scoring function in specific feature dimensions, and improves the overall discrimination balance. After obtaining the updated weight coefficients, the system immediately retrains the anomaly scoring function based on the historical training data at the current stage and re-optimizes the discrimination boundary position and output response surface of the scoring model. Through this round of training, the discrimination ability of the scoring function is dynamically corrected to a position more suitable for the spatial distribution of the current fused feature vector.

[0125] After completing the training update, the scoring calculation module uses the new version of the scoring function to perform the anomaly scoring task for the subsequent fused feature vectors, and continues to execute a similar dynamic learning and adaptive optimization process in the next round of scoring-filtering-residual analysis closed loop. Through the above-mentioned scoring function residual feedback self-learning mechanism, the present invention realizes the automatic iterative optimization ability of the scoring calculation module in a long-term complex operating environment, enabling the anomaly recognition model to have the continuous self-evolution ability to highly adapt to different anomaly distribution characteristics.

[0126] The reading filtering module 106 is configured to receive the anomaly confidence sequence and perform anomaly filtering on the original reading sequence according to a preset confidence threshold to generate a filtered reading sequence.

[0127] The reading filtering module 106 undertakes the final anomaly reading elimination and filtering function in the system of the present invention, and is the convergence processing unit in the whole anomaly recognition and filtering process. Its output directly determines the purity and data quality stability of the smart meter data before entering the subsequent analysis and application links. The input of the reading filtering module 106 is the anomaly confidence sequence output by the scoring calculation module 105, and at the same time, it also needs to call the original reading sequence provided by the original reading cache module 101 as the basic data source.

[0128] Before entering the specific filtering process, the reading filtering module 106 first strictly verifies the time series consistency of the input data. The system confirms that each scoring result in the abnormal confidence level sequence corresponds one-to-one with the time stamps in the original reading sequence, ensuring that each specific original sampling data point can be accurately located when performing the reading elimination operation, and avoiding any data misalignment or omission.

[0129] After confirming the data sequence consistency, the system performs a filtering determination based on a pre-set abnormal filtering threshold. This abnormal filtering threshold is a predefined standard value, and its value range is consistent with the output interval of the abnormal confidence level scoring. The setting of the threshold can be set according to application scenario experience at the initial stage of system design, such as setting it to a relatively high confidence level standard such as 0.7 or 0.8, or the optimal cut-off point can be determined through statistical analysis based on the abnormal proportion distribution law in the actual operation data, or it can also be dynamically adjusted during the system operation to achieve adaptive optimization. The setting of the threshold directly affects the sensitivity and fault tolerance of the system to abnormal readings. A high threshold will increase the strictness of elimination, which helps to exclude potential abnormal data to the greatest extent, while a low threshold can retain borderline suspicious readings to a certain extent to avoid misdeleting normal fluctuating data.

[0130] In the filtering execution stage of the reading filtering module 106, the system sequentially traverses the abnormal confidence level values at each sampling moment and compares them with the filtering threshold. If the abnormal confidence level value at this sampling moment is lower than or equal to the threshold, it is considered that this data point belongs to normal data under the current analysis logic, and the system retains it intact in the output data sequence; otherwise, if the confidence level value is higher than the threshold, the system determines that there is an abnormal risk in the original reading corresponding to this sampling moment, marks it as an abnormal reading, and eliminates it in the filtering result.

[0131] In order to balance the abnormal recognition accuracy and data continuity, the reading filtering module 106 has a missing compensation logic when performing filtering. For the data points determined to be abnormally eliminated, the system can fill or mark them according to the set strategy. In the simple filling mode, the system can use the normal readings of adjacent sampling points to perform interpolation filling, or use the short-term moving average for dynamic smoothing filling under certain rules to ensure that the continuous data stream in the subsequent analysis link will not be affected by individual elimination operations and cause breakpoints. In the abnormal marking mode, the system can directly retain the elimination position in the output sequence and attach an abnormal label for subsequent modules to perform statistical review and operation and maintenance monitoring analysis on the abnormal occurrence position and frequency when needed.

[0132] After the reading filtering module 106 completes the data filtering process for all sampling moments, a filtered reading sequence will be finally formed. The filtered reading sequence maintains the same timestamp order and data length mapping as the original reading sequence in terms of data structure, and only eliminates the sampling points determined to be high-risk anomalies through scoring, thus significantly improving the purity, smoothness, and credibility of the overall data sequence. As the final output data of the system of the present invention, the filtered reading sequence can be directly called by various application scenarios such as subsequent energy consumption settlement, load forecasting, power grid dispatching, and metering assessment, ensuring the data reliability and abnormal self-correction ability of the entire smart meter system during long-term operation.

[0133] Furthermore, the intelligent meter abnormal reading filtering system based on an elastic time window further includes: A feedback update module, configured to receive the filtered reading sequence and the original reading sequence, calculate the residual vector between the two; update the statistical volatility algorithm parameters in the volatility calculation module based on the residual vector to obtain an updated volatility sequence; and feedback the updated volatility sequence to the window generation module for dynamically updating the time window parameter set.

[0134] In the present invention, the feedback update module undertakes the dynamic adaptive optimization function in the overall system architecture. Its main role is to, after initially completing abnormal filtering, according to the difference information between the filtered result and the original data, real-time correct the parameter settings in the volatility calculation module, thereby improving the adaptability and stability of the volatility sequence in a complex operating environment and further enhancing the real-time adjustment accuracy of the elastic time window.

[0135] After the system completes the aforementioned abnormal reading filtering process, a filtered reading sequence is obtained. The filtered reading sequence is strictly consistent with the original reading sequence in terms of time order and data structure. Only at some sampling moments determined to be abnormal, the reading values have been eliminated or replaced with smooth interpolation or abnormal identifiers according to the filtering logic. When the feedback update module starts the update process, it first synchronously receives the original reading sequence and the filtered reading sequence, and calculates the point-by-point difference between the reading values at the corresponding timestamps of the two to form a residual vector.

[0136] The calculation method of the residual vector is as follows: The system sequentially extracts the reading values at the same timestamp position in the filtered reading sequence and the original reading sequence, calculates the algebraic difference between the two, and forms the residual value at each sampling moment. The entire residual vector is composed of the residual values corresponding to all sampling moments in time order. The residual value reflects the magnitude of the correction of the abnormal filtering logic under the current window parameters and scoring threshold settings. Among them, the moments with larger absolute values of the residuals often correspond to the situation where significant abnormal readings are eliminated or replaced, and the moments with smaller absolute values of the residuals indicate that the data difference before and after filtering is limited and the original data is relatively stable.

[0137] After obtaining the complete residual vector, the feedback update module uses it as the basis for subsequent parameter updates and dynamically adjusts the statistical volatility algorithm parameters in the volatility calculation module. The update logic is jointly modeled based on the total residual, the width of the residual distribution, and the abnormal characteristics in the residual concentration to form a correction amount for the volatility algorithm parameters. Specifically, the system first calculates the overall average residual and the residual standard deviation of the residual vector to respectively evaluate the overall residual offset level and the residual volatility intensity of the system. If the residual average value is higher than the preset normal offset threshold, the system will determine that the current volatility algorithm parameters may have a tendency to underestimate volatility, and correspondingly increase the standard deviation estimation gain coefficient within the window in the volatility calculation. If the residual standard deviation is significantly amplified, the system further analyzes the skewness and kurtosis of the residuals. If there are a large number of isolated high residuals, the system can temporarily relax the outlier rejection threshold to increase the abnormal inclusiveness of the volatility assessment and prevent excessive contraction of the volatility assessment interval.

[0138] When performing parameter updates, the feedback update module does not directly correct each item of the existing volatility sequence. Instead, it reapplies the adjusted algorithm parameters to the original reading sequence and recalculates the volatility index at each sampling moment through a sliding window method to form a new volatility sequence. This updated volatility sequence completely covers all sampling moments, and in terms of numerical performance, it has fully incorporated the residual information reflected by the previous round of anomaly filtering results, realizing the system's real-time learning and rapid adaptation to the dynamic characteristics of current power data.

[0139] After completing the calculation of the new volatility sequence, the feedback update module directly outputs and feeds back the updated volatility sequence to the window generation module. After receiving the new round of volatility sequence, the window generation module regenerates a new set of time window parameters according to the preset elastic adjustment function logic. Since the volatility sequence has been dynamically corrected, the calculation of the duration and overlap ratio of the new round of windows will fully absorb the characteristics of the previous round of filtered residuals, enabling the window layout to have a more accurate window scale self-adaptation ability in subsequent data processing, effectively improving the overall stability, sensitivity, and false judgment suppression ability of anomaly recognition.

[0140] Through the recursive parameter correction mechanism of the feedback update module, the present invention forms a closed-loop dynamic optimization architecture for anomaly detection logic and anomaly feedback logic. The system can continuously self-correct the parameter model according to the real-time data characteristics during long-term operation, and can adapt to the anomaly recognition tasks of smart meters under different seasons, different load characteristics, and different interference scenarios without manual intervention, which is significantly superior to the traditional static configuration method of fixed algorithm parameters.

[0141] Furthermore, the feedback update module is specifically used for: Receiving the latest round of filtered reading sequence and the original reading sequence, and calculating the difference between the two at each sampling point to form a residual vector; Statistically calculate the residual mean, residual standard deviation, and outlier rejection ratio in the residual vector to generate a feedback residual feature set; Dynamically correct the outlier rejection threshold and standard deviation amplification factor in the volatility calculation module according to the feedback residual feature set to achieve real-time expansion or convergence adjustment of the volatility evaluation range; Re-execute the full-sequence volatility calculation under the updated algorithm parameters, and feedback the newly generated volatility sequence to the window generation module to correct the subsequent time window layout, so that the system anomaly recognition process forms a continuous recursive optimization closed loop.

[0142] In the present invention, the feedback update module undertakes an important function of system adaptive recursive optimization. It mainly analyzes the residual information generated after each round of anomaly filtering, and dynamically corrects the parameter settings in the volatility calculation module, so as to realize the real-time learning and gradual evolution of the system in different data characteristic environments during long-term operation. The design logic of the feedback update module is closely nested in the entire anomaly reading filtering process, and its output directly affects the window layout adjustment of the subsequent time window generation module, thus forming a complete dynamic optimization closed loop.

[0143] When the system completes the current round of anomaly filtering, the feedback update module first receives the filtered reading sequence output by the reading filtering module and the original reading sequence output by the original reading cache module. These two groups of data are completely corresponding on the sampling time axis and synchronously contain information of all sampling moments. For each sampling moment, the system calculates the algebraic difference between the filtered reading and the original reading to obtain the residual value at that moment. By sequentially performing this difference calculation within the full sequence range, the system forms a residual vector, where the residual value at each position represents the actual correction amplitude of this filtering and scoring logic at that position.

[0144] After forming a complete residual vector, the system performs statistical processing on it to extract the feedback residual feature set. First, the system calculates the mean of the residual vector, which reflects the cumulative trend of the current overall filtering logic in the offset direction. For example, if the residual mean is continuously positive, it may indicate that the system has a long-term risk of underestimating anomalies. If the residual mean is continuously negative, it may reflect the over-screening phenomenon caused by the system. Secondly, the system calculates the standard deviation of the residual vector. This standard deviation quantifies the dispersion degree of the filtering results at the residual level. The higher the standard deviation value, the greater the adaptive fluctuation of the current scoring and filtering logic for different data segments. The system will further count the outlier rejection ratio, that is, within the full sequence range, the percentage of the number of data points determined to be outliers by the system in the total number of sampling points. The outlier rejection ratio can reflect the filtering intensity level of the system in the overall sample space. By quantifying the residual features in the above three dimensions, the system forms a complete feedback residual feature set, providing a direct data basis for subsequent algorithm parameter adjustment.

[0145] After obtaining the feedback residual feature set, the system enters the algorithm parameter correction phase. First, the system analyzes the current residual mean level. When the residual mean continuously deviates and exceeds the convergence threshold set within the system, the system determines that the outlier rejection threshold in the volatility calculation module may deviate from the optimal range. If the residual mean is positive, the system appropriately increases the outlier rejection threshold and relaxes the outlier recognition sensitivity to avoid mis-screening some boundary samples; if the residual mean is negative, the system appropriately decreases the outlier rejection threshold to enhance the system's ability to capture outlier samples. Second, the system analyzes the current volatility assessment stability based on the residual standard deviation. When the standard deviation is large, it indicates that the volatility estimation is insufficiently adaptable to abnormal volatility signals, and the system appropriately shrinks the standard deviation amplification coefficient to increase the response sensitivity to small changes during volatility calculation; when the standard deviation is small, the system appropriately relaxes the amplification coefficient to alleviate the over-sensitivity to weak normal volatility and smooth the overall assessment curve of the system. Finally, the system performs an overall balance adjustment with reference to the change trend of the outlier rejection ratio. When the outlier rejection ratio is significantly higher than the historical steady-state level of the system, the system globally reduces the convergence intensity of the rejection threshold to ensure data integrity; when the rejection ratio is significantly lower, the system appropriately enhances the rejection sensitivity to prevent missed detections.

[0146] After the above multi-dimensional joint adjustment logic is executed, the system finally forms an updated set of parameters for the outlier rejection threshold and the standard deviation amplification coefficient. This set of parameters comprehensively absorbs the adaptability problems exposed by the previous round of scoring residuals, filtering results, and overall outlier screening behavior, and has a high scene matching ability. The system then uses this updated set of algorithm parameters to re-perform the full-sequence sliding window volatility calculation on the complete original reading sequence. The system performs a robust statistical process within each sliding window according to the newly set rejection threshold and amplification coefficient, estimates the standard deviation again after removing the temporary outlier readings, and thus outputs a new sequence of volatility indicators.

[0147] The system transmits this newly generated volatility sequence in real time to the window generation module as its new input basis. The window generation module recalculates the duration and overlap ratio of each time window based on the updated volatility indicators, and then comprehensively corrects the subsequent time window layout structure. Through this recursive self-learning closed loop, the system of the present invention can continuously track and adapt to the data distribution changes, load switching, disturbance frequency fluctuations, and seasonal operation characteristics that may occur during the long-term operation of the smart meter, so as to maintain the long-term stability and high recognition accuracy of the outlier recognition mechanism in a changing environment.

[0148] Although this application is disclosed above with preferred embodiments, it is not used to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the protection scope of this application should be subject to the scope defined by the claims of this application.

Claims

1. An intelligent meter abnormal reading filtering system based on an elastic time window, characterized in that, Including: An original reading cache module, which is used to receive continuous original readings generated by an intelligent meter under a unified sampling benchmark and form an original reading sequence; A fluctuation degree calculation module, which is used to receive the original reading sequence and calculate the fluctuation degree index corresponding to each sampling moment based on a preset statistical volatility algorithm to form a fluctuation degree sequence; A window generation module, which is used to receive the fluctuation degree sequence and generate a set of time window parameters through an elastic adjustment function according to the fluctuation degree sequence. The set of time window parameters includes a window start time, a window duration, and a window overlap ratio; A feature fusion module, which is used to receive the original reading sequence and the set of time window parameters, extract context subsequences from the original reading sequence according to the set of time window parameters, respectively extract time series features and frequency domain features based on the context subsequences, and fuse them to form context fusion features; A scoring calculation module, which is used to receive the context fusion features and calculate an abnormal confidence level sequence based on an abnormal scoring function; A reading filtering module, which is used to receive the abnormal confidence level sequence and perform abnormal filtering on the original reading sequence according to a preset confidence level threshold to generate a filtered reading sequence.

2. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 1, wherein It also includes: A feedback update module, which is used to receive the filtered reading sequence and the original reading sequence and calculate the residual vector between the two; Based on the residual vector, update the statistical volatility algorithm parameters in the fluctuation degree calculation module to obtain an updated fluctuation degree sequence; Feed back the updated fluctuation degree sequence to the window generation module for dynamically updating the set of time window parameters.

3. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 1, characterized in that, The window generation module is specifically used for: After receiving the fluctuation degree sequence, calculate the preliminary time window duration and overlap ratio through an elastic adjustment function to form original time window parameters; For each time window, detect whether its start time and end time in the sampling sequence are out of bounds. If there is an out-of-bounds situation, extract the number of actually available data points to form a boundary remaining length index; Input the boundary remaining length index into a boundary compensation function, dynamically calculate a correction coefficient according to the window deviation amount, and perform a decreasing adjustment on the preliminary duration to form a corrected duration; According to the corrected duration, recalculate the overlap ratio to maintain the window sliding balance, form a set of boundary adaptive window parameters, and recursively generate the start time of the next window according to the corrected parameters to ensure that the entire time window layout maintains time continuity and context integrity at the sequence boundary.

4. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 1, wherein The fluctuation degree calculation module is also used for: Within each sliding window, calculate the median of the current subsequence as the robust center position; Calculate the absolute deviation of each reading relative to the median and generate an absolute deviation sequence; Extract the median absolute deviation from the absolute deviation sequence and dynamically calculate an abnormal rejection threshold according to the median absolute deviation. Among them, the abnormal rejection threshold is obtained by coupling and adjusting an adaptive threshold amplification factor and the current fluctuation degree trend; Identify and temporarily reject the readings that exceed the abnormal deviation range according to the abnormal rejection threshold, and perform a robust standard deviation calculation after the abnormal rejection is completed to obtain a fluctuation degree index after abnormal interference suppression for subsequent calls by the window generation module.

5. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 1, characterized in that, The specific operations of the feature fusion module are as follows: After receiving the context subsequence, perform a fast Fourier transform on the context subsequence to obtain the corresponding spectral amplitude sequence; In the spectral amplitude sequence, according to the preset frequency division rule, extract the high-frequency interval amplitude subsequence and the low-frequency interval amplitude subsequence, and calculate the total energy of the high-frequency interval and the total energy of the low-frequency interval respectively. Then, calculate the high-frequency energy ratio based on the total energy of the high-frequency interval and the overall total energy to form a high-frequency energy ratio index; Input the high-frequency energy ratio index into the high-frequency anomaly non-linear weighting function. Among them, the high-frequency anomaly non-linear weighting function calculates the high-frequency weighted anomaly contribution factor based on the deviation degree between the high-frequency energy ratio index and the high-frequency energy critical threshold and the set exponential increasing function relationship; Use the high-frequency weighted anomaly contribution factor as the high-frequency energy enhancement feature, and perform feature splicing and normalization processing on the high-frequency energy enhancement feature and the time series feature according to the preset fusion logic to form a multi-dimensional context fusion feature vector, and provide the multi-dimensional context fusion feature vector to the scoring calculation module for performing anomaly confidence scoring. Among them, the time series features include sequence mean, standard deviation, slope change feature, skewness feature, and kurtosis feature.

6. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 1, wherein The scoring calculation module is also used for: After each round of anomaly scoring is completed, compare the anomaly confidence sequence output by the scoring calculation module with the filtered reading sequence output by the reading filtering module, and calculate the scoring residuals at the corresponding positions of the two at each sampling moment to form a scoring residual vector; After obtaining the scoring residual vector, count the average offset value, standard deviation, and extreme residual sampling point frequency of the scoring residual vector to form a scoring residual feature index set; Input the scoring residual feature index set into the weight adjustment module, and dynamically correct the weight coefficients in the scoring function according to the residual contribution of each feature dimension in the anomaly scoring. Among them, for the feature dimension with a higher scoring residual frequency, reduce its corresponding weight coefficient to form a weight suppression correction factor; for the feature dimension with a lower scoring residual frequency and stable offset, increase its corresponding weight coefficient to form a weight enhancement correction factor, and then generate an updated scoring weight coefficient set; Under the updated scoring weight coefficient set, retrain the anomaly scoring function, optimize the discrimination ability of the scoring function in the current context fusion feature vector space according to the training results, and apply it to the scoring calculation of the subsequent anomaly confidence sequence after training to achieve the long-term adaptive dynamic optimization of the input data characteristics by the scoring calculation module.

7. The intelligent meter abnormal reading filtering system based on an elastic time window according to claim 2, characterized in that The specific operations of the feedback update module are as follows: Receive the latest round of filtered reading sequence and the original reading sequence, and calculate the difference between the two at each sampling point to form a residual vector; In the residual vector, count the residual mean, residual standard deviation, and anomaly rejection ratio to generate a feedback residual feature set; Dynamically correct the anomaly value rejection threshold and the standard deviation amplification factor in the fluctuation degree calculation module according to the feedback residual feature set to achieve real-time expansion or convergence adjustment of the fluctuation degree evaluation range; Re - execute the full - sequence volatility calculation under the updated algorithm parameters, and feedback the newly generated volatility sequence to the window generation module to correct the subsequent time - window layout, so that the system anomaly recognition process forms a continuous recursive optimization closed - loop.

Images