A multi-dimensional feature extraction method and system for distributed fault recording data
By performing deviation separation and collector screening in distributed fault recording data, combined with random forest feature screening, the problem of false information caused by differences in collector hardware was solved, and more reliable multidimensional feature extraction and fault identification were achieved.
Patent Information
- Application Number
- CN202611544910.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-08
- Publication Date
- 2026-11-06
AI Technical Summary
In existing technologies, multidimensional feature extraction of distributed fault recording data suffers from low reliability, mainly due to false information caused by differences in the hardware of the acquisition device, which affects the accuracy of fault identification.
By separating the deviations based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data, the target deviation of each data acquisition unit is obtained. Based on the target deviation, the data acquisition units are screened, the failure rate differences are statistically analyzed, the maximum number of features of the random forest is adjusted, and feature screening is performed to suppress the interference of hardware differences of the data acquisition units.
This improves the reliability of multi-dimensional feature extraction from distributed fault recording data, avoids false features being mistakenly selected as diagnostic features, and enhances the accuracy of fault identification.
Smart Images
Figure CN123049069A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power distribution network fault identification technology, specifically to a method and system for multidimensional feature extraction of distributed fault recording data. Background Technology
[0002] In distributed power distribution networks, when a ground fault occurs in a distribution line, the time-domain, frequency-domain, and transient components of the fault current all change with the fault. Therefore, multi-dimensional features can be extracted from distributed fault recording data, providing reliable data support for subsequent fault identification based on multi-dimensional features.
[0003] In related technologies, fault waveform data is acquired by multiple data acquisition units deployed on each feeder. The initial multidimensional features extracted from the fault waveform data are then filtered using a fixed number of split features to obtain the final multidimensional features used for diagnosis. However, the hardware differences among the multiple data acquisition units in this method cause the acquired fault waveform data to carry quasi-static biases from the acquisition units, resulting in false information in the filtered multidimensional features due to equipment differences.
[0004] This shows that the reliability of multidimensional feature extraction from distributed fault recording data in related technologies is relatively low. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for multidimensional feature extraction of distributed fault recording data, which solves the technical problem of low reliability of multidimensional feature extraction of distributed fault recording data in related technologies.
[0006] In a first aspect, embodiments of the present invention provide a method for multidimensional feature extraction from distributed fault recording data, the method comprising: Deviation separation is performed based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit. The target deviation is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition unit from the average level of all data acquisition units. Multiple data acquisition units are distributed and deployed on each feeder. Collectors are screened based on the target deviation of multiple collectors to obtain a first set and a second set, wherein the maximum value of the target deviation of multiple collectors in the first set is less than the minimum value of the target deviation of multiple collectors in the second set. Statistical analysis is performed based on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set to obtain the first failure rate and the second failure rate. The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set. The difference between the first failure rate and the second failure rate is analyzed to obtain the difference value, which is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection. The maximum number of features in the random forest is adjusted based on the difference value to obtain the target number of features. Then, the multidimensional features are filtered according to the random forest method and the target number of features to obtain reference features. The target number of features is used to indicate the number of features participating in the optimal split search when a node splits. The target number of features is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.
[0007] In one embodiment, the deviation separation based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit includes: Based on the bus zero-sequence voltage, the fault waveform data collected by each data acquisition unit is separated by deviation to obtain the quasi-static deviation of each data acquisition unit. The quasi-static deviation is used to characterize the inherent measurement offset of the corresponding data acquisition unit. Zero drift correction is performed based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device.
[0008] In one embodiment, the fault recording data includes: zero-sequence current; the deviation separation of the fault recording data collected by each data acquisition unit based on the bus zero-sequence voltage to obtain the quasi-static deviation of each data acquisition unit includes: Analyze the waveform changes of the zero-sequence voltage at the bus to obtain the voltage change rate; Based on the voltage change rate and the zero-sequence current collected by each collector, capacitance estimation is performed to obtain the capacitance coefficient corresponding to each collector. Calculate the product of the capacitance coefficient and the voltage change rate for each data collector to obtain the capacitor current for each data collector. For each data collector, the corresponding capacitor current is removed from the zero-sequence current it collects to obtain the residual current for each data collector. The quasi-static deviation of each data acquisition device is obtained by averaging the residual current corresponding to each data acquisition device and the residual current accumulated by each data acquisition device in historical fault events.
[0009] In one embodiment, the zero-drift correction based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device includes: The drift value is obtained by weighting the quasi-static deviation of multiple data acquisition devices based on the historical data volume of multiple data acquisition devices. The drift value is used to characterize the average level of the inherent measurement offset of all data acquisition devices. The difference between the quasi-static deviation and the drift value of each data acquisition unit is calculated to obtain the target deviation of each data acquisition unit.
[0010] In one embodiment, the collector screening based on the target deviation of multiple collectors to obtain a first set and a second set includes: Arrange the target deviations of multiple data acquisition devices in ascending order to obtain the target deviation sequence; Multiple differences are obtained by calculating adjacent differences based on the target deviation sequence; The average of the two target deviations corresponding to the largest difference among multiple differences is determined as the segmentation threshold; Analyze the relative magnitudes of the target deviation and segmentation threshold of multiple data collectors, identify the data collectors with target deviations less than the segmentation threshold as the first set, and identify the data collectors with target deviations greater than the segmentation threshold as the second set.
[0011] In one embodiment, the statistical analysis performed on the historical data volume and corresponding feeder failure counts of multiple data collectors in the first set, and the historical data volume and corresponding feeder failure counts of multiple data collectors in the second set, to obtain a first failure rate and a second failure rate, includes: Calculate the ratio of the number of feeder faults corresponding to each data collector to its historical data volume to obtain the feeder fault frequency corresponding to each data collector. Based on the historical data volume of multiple collectors in the first set, the feeder fault frequency corresponding to multiple collectors in the first set is weighted and averaged to obtain the first fault rate; Based on the historical data volume of multiple collectors in the second set, the feeder fault frequencies corresponding to multiple collectors in the second set are weighted and averaged to obtain the second fault rate.
[0012] In one embodiment, analyzing the difference between the first failure rate and the second failure rate to obtain the difference value includes: Calculate the absolute difference between the first failure rate and the second failure rate to obtain the inter-group failure rate difference; The difference in failure rate between groups is normalized based on the feeder failure frequency corresponding to multiple data acquisition units to obtain the difference value.
[0013] In one embodiment, adjusting the maximum number of features in the random forest based on the difference value to obtain the target number of features includes: The maximum number of features in the random forest is scaled based on the difference value and the preset minimum number of features to obtain the target number of features. The target number of features is negatively correlated with the difference value and positively correlated with the preset minimum number of features.
[0014] In one embodiment, before performing feature filtering on the multidimensional features according to the random forest method and the target feature number to obtain reference features, the method further includes: Feature extraction is performed on fault recording data collected by multiple acquisition devices to obtain multidimensional features. The multidimensional features include at least two of the following: time domain features, frequency domain features, time-frequency domain features, sequence component features, and transient features.
[0015] Secondly, embodiments of the present invention provide a multi-dimensional feature extraction system for distributed fault recording data, the system comprising: The deviation separation module is used to separate deviations based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit. The target deviation is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition unit from the average level of all data acquisition units. Multiple data acquisition units are distributed and deployed on each feeder. The grouping and screening module is used to screen multiple collectors based on their target deviations to obtain a first set and a second set, wherein the maximum value of the target deviations of multiple collectors in the first set is less than the minimum value of the target deviations of multiple collectors in the second set. The first analysis module is used to perform statistical analysis based on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set, to obtain a first failure rate and a second failure rate. The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set. The second analysis module is used to analyze the difference between the first failure rate and the second failure rate to obtain the difference value, wherein the difference value is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection. The feature selection module is used to adjust the maximum number of features in the random forest based on the difference value to obtain the target number of features. It then selects multidimensional features based on the random forest method and the target number of features to obtain reference features. The target number of features is used to indicate the number of features participating in the optimal split search when a node splits. The target number of features is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.
[0016] The present invention has the following beneficial effects: This invention separates deviations based on bus zero-sequence voltage, historical data volume of each data acquisition unit, and fault waveform data to obtain the target deviation for each data acquisition unit. Based on the target deviations of multiple data acquisition units, the units are sieved to obtain a first set and a second set. Subsequently, statistical analysis is performed on the historical data volume and corresponding feeder fault counts of multiple data acquisition units in the first set, and the historical data volume and corresponding feeder fault counts of multiple data acquisition units in the second set, to obtain a first fault rate and a second fault rate. Then, the difference between the first fault rate and the second fault rate is analyzed to obtain the difference value. Based on the difference value, the maximum number of features in the random forest is adjusted to obtain the target number of features. Finally, the random forest is used to... This invention uses the method and target feature number to filter multidimensional features and obtain reference features. It considers the interference of the inherent measurement offset of the acquisition device on the intrinsic features of the fault during the feature filtering process. By using the bus zero-sequence voltage as a common reference, the fault waveform data of each acquisition device is separated by deviation, and the difference value is obtained. Based on this, the maximum number of features when splitting random forest nodes is adaptively adjusted, so that the probability of features carrying offset information (i.e., inherent measurement offset) being selected in each split is reduced. This effectively suppresses the interference of hardware difference of the acquisition device, avoids the problem of false features being mistakenly selected as the final diagnostic features, and improves the reliability of multidimensional feature extraction of distributed fault waveform data. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating a multi-dimensional feature extraction method for distributed fault recording data provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a multi-dimensional feature extraction system for distributed fault recording data provided in an embodiment of the present invention. Detailed Implementation
[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.
[0019] In related technologies, fault waveform data is acquired by multiple data acquisition units deployed on various feeders. The initial multidimensional features extracted from the fault waveform data are then filtered using a fixed number of split features to obtain the final multidimensional features used for diagnosis. However, different data acquisition units have hardware differences, meaning they possess inherent measurement biases. This inherent measurement bias refers to a fixed deviation value that is systematically higher or lower than the measured values due to factors such as manufacturing calibration errors, installation location differences, and temperature drift. This inherent measurement bias causes the acquired fault waveform data to be superimposed with the quasi-static bias unique to each data acquisition unit at the source. This quasi-static bias propagates step-by-step along the feature extraction link to various dimensions such as time domain, frequency domain, and transient features. Consequently, the extracted multidimensional features, in addition to reflecting fault information, also couple with hardware difference components of the data acquisition units unrelated to the fault.
[0020] When a fixed number of splitting features is used for screening, random forests randomly select a fixed number of features from all features for optimal splitting search at each node split. This means that features carrying inherent measurement offset information always participate in each node split with equal probability, leading to false information in the selected multidimensional features due to equipment differences. Ultimately, the selected feature subset cannot accurately reflect the fault state. Therefore, the reliability of multidimensional feature extraction from distributed fault waveform data using existing technologies is low.
[0021] This invention provides a method and system for multidimensional feature extraction of distributed fault waveform data to improve the reliability of multidimensional feature extraction. The method separates the deviations based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation for each data acquisition unit. Based on the target deviations of multiple data acquisition units, data acquisition units are screened to obtain a first set and a second set. Subsequently, statistical analysis is performed on the historical data volume and corresponding feeder fault counts of multiple data acquisition units in the first set, and the historical data volume and corresponding feeder fault counts of multiple data acquisition units in the second set, to obtain a first fault rate and a second fault rate. The difference between the first and second fault rates is then analyzed to obtain the difference value. Based on the difference value, the maximum number of features in the random forest is adjusted to obtain the target number of features. Finally, multidimensional features are screened according to the random forest method and the target number of features to obtain reference features. This effectively suppresses interference from hardware differences in the data acquisition units, avoids the problem of false features being mistakenly selected as the final diagnostic features, and improves the reliability of multidimensional feature extraction of distributed fault waveform data.
[0022] This invention is applicable to scenarios in distributed power distribution networks where multidimensional features need to be extracted from fault recording data and selected for optimal performance. For example, it is suitable for ground fault location and diagnosis scenarios where each data acquisition device exhibits inherent measurement offset differences. The specific scheme of a multidimensional feature extraction method and system for distributed fault recording data provided by this invention is described below with reference to the accompanying drawings.
[0023] This invention proposes a multidimensional feature extraction method for distributed fault recording data. The method provided by this invention can be executed by a multidimensional feature extraction system for distributed fault recording data, specifically by the analysis device within that system. Please refer to [link to relevant documentation]. Figure 1 The diagram illustrates a flowchart of a multi-dimensional feature extraction method for distributed fault recording data according to an embodiment of the present invention. The method includes the following steps: Step S1: Based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data, deviation separation is performed to obtain the target deviation of each data acquisition unit.
[0024] The target bias is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition unit from the average level of all data acquisition units, with multiple data acquisition units distributed across each feeder.
[0025] Bus zero-sequence voltage refers to the vector sum of the three-phase voltages at the bus and the corresponding zero-sequence component. This bus zero-sequence voltage can be obtained by a voltage transformer installed at the bus. After the voltage transformer collects the bus zero-sequence voltage, it can transmit it to the analysis device via a data bus or communication network. For example, the sampling frequency of the voltage transformer can be 12800Hz.
[0026] The fault waveform data includes zero-sequence current. The busbar connects to multiple feeders, each equipped with a data acquisition unit (e.g., a non-contact Hall effect sensor). The data acquisition unit collects the zero-sequence current of its corresponding feeder in real time. When a ground fault is detected, each data acquisition unit records the zero-sequence current it has collected, resulting in fault waveform data for each unit. This fault waveform data covers the zero-sequence current for three power frequency cycles before the fault and seven power frequency cycles after the fault, providing a complete record of the transient and steady-state zero-sequence current data before and after the fault. After recording, each data acquisition unit can upload the fault waveform data to the analysis device via a data bus or communication network. The sampling frequency of multiple data acquisition units can be 12800Hz. The power frequency cycle refers to the time required for alternating current to complete one complete waveform change in the power system; for example, in a 50Hz power grid, the power frequency cycle is 0.02 seconds.
[0027] Specifically, the deviation separation based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit includes: Based on the bus zero-sequence voltage, the fault waveform data collected by each data acquisition unit is separated by deviation to obtain the quasi-static deviation of each data acquisition unit. The quasi-static deviation is used to characterize the inherent measurement offset of the corresponding data acquisition unit. Zero drift correction is performed based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device.
[0028] Quasi-static deviation refers to the inherent bias of a data acquisition unit, caused by factors such as manufacturing calibration errors, installation location differences, and temperature drift, resulting in a systematic overestimation or underestimation of the true value. This fixed deviation is an inherent, quasi-static, and long-term hardware characteristic of the data acquisition unit, and it does not change with the occurrence of faults. For example, if data acquisition unit 1 is overly sensitive, and its measured zero-sequence current is always 0.5A higher than the true value, while data acquisition unit 2 is overly sluggish, and its measured zero-sequence current is always 0.3A lower than the true value, then the quasi-static deviation of data acquisition unit 1 is "+0.5A", and the quasi-static deviation of data acquisition unit 2 is "-0.3A".
[0029] The process of separating the fault recording data acquired by each data acquisition unit based on the bus zero-sequence voltage to obtain the quasi-static deviation of each data acquisition unit includes: Analyze the waveform changes of the zero-sequence voltage at the bus to obtain the voltage change rate; Based on the voltage change rate and the zero-sequence current collected by each collector, capacitance estimation is performed to obtain the capacitance coefficient corresponding to each collector. Calculate the product of the capacitance coefficient and the voltage change rate for each data collector to obtain the capacitor current for each data collector. For each data collector, the corresponding capacitor current is removed from the zero-sequence current it collects to obtain the residual current for each data collector. The quasi-static deviation of each data acquisition device is obtained by averaging the residual current corresponding to each data acquisition device and the residual current accumulated by each data acquisition device in historical fault events.
[0030] The zero-sequence current actually measured by the data acquisition unit includes the feeder-to-ground capacitance current (hereinafter referred to as capacitance current), quasi-static deviation, fault transient current, and noise current. The instantaneous value of the capacitance current is proportional to the rate of change of the bus zero-sequence voltage. Therefore, capacitance estimation, unaffected by quasi-static deviation, can be performed by projecting the zero-sequence current onto the direction of the voltage change rate. Specifically, for each data acquisition unit, the zero-sequence current and corresponding voltage change rate of all sampling points within the fault recording window (the first 3 power frequency cycles and the last 7 power frequency cycles, a total of 10 power frequency cycles) are multiplied by an inner product to obtain the first inner product result for that data acquisition unit. The inner product of the voltage change rate of all sampling points within the fault recording window and the data acquisition unit itself is also calculated to obtain the second inner product result. The analysis device can calculate the ratio of the first inner product result to the second inner product result for that data acquisition unit to obtain the capacitance coefficient corresponding to that data acquisition unit. The second inner product result is used to normalize the first inner product result for each data acquisition unit. The unit of capacitance coefficient is farad (F).
[0031] For each sampling time, the capacitance coefficient corresponding to each data acquisition unit is multiplied by the voltage change rate to obtain the capacitive current of each data acquisition unit at each sampling time. Subsequently, for each data acquisition unit, the analysis device can calculate its zero-sequence current at each sampling time and the difference between it and its capacitive current at each sampling time. The capacitive current can then be discarded to obtain its residual current at each sampling time.
[0032] Fault transient current is a high-frequency, impulsive component that appears at the instant a ground fault occurs, forming an alternating positive and negative oscillating waveform near the fault moment. Within the fault recording window, the fault transient current decays rapidly over time, and its positive and negative half-waves are essentially symmetrical. In other words, the positive and negative half-cycles of the fault transient current waveform are basically symmetrical within the fault recording window, and its long-time window integral value approaches zero, equivalent to the average value of the fault transient current within the fault recording window approaching zero. Noise current is generated by random interference factors such as thermal noise and electromagnetic interference from the data acquisition unit. Its positive and negative deviations have roughly equal probabilities, meaning that the mean value of the noise current also approaches zero within the fault recording window. Based on this, the analysis device can perform long-term averaging of the residual current accumulated from multiple historical fault events for each data acquisition unit to remove fault transient current and noise current from the residual current, thereby obtaining the quasi-static deviation of each data acquisition unit. By using the residual current accumulated by each data acquisition unit in multiple historical fault events for long-term averaging, the number of samples involved in the averaging can be increased, allowing the positive and negative components of fault transient current and noise current to cancel each other out more fully with a larger sample size. The more historical fault events involved in the long-term averaging, the more fully the fault transient and noise are canceled out, and the more accurate the estimation of quasi-static bias.
[0033] For example, the quasi-static bias of the u-th data collector It can be represented as: In the formula, This is an index for fault events. This represents the average number of historical fault events processed by the u-th data collector. This represents the total number of all sampling points within the fault recording window. For the u-th collector in the th... The residual current at the t-th sampling time in the fault event.
[0034] When a ground fault is detected for the first time, the analysis device can perform mean processing on the residual current at each sampling time within the fault recording window to obtain the quasi-static deviation of the data acquisition unit.
[0035] The process of performing zero-drift correction based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device includes: The drift value is obtained by weighting the quasi-static deviation of multiple data acquisition devices based on the historical data volume of multiple data acquisition devices. The drift value is used to characterize the average level of the inherent measurement offset of all data acquisition devices. The difference between the quasi-static deviation and the drift value of each data acquisition unit is calculated to obtain the target deviation of each data acquisition unit.
[0036] The analysis device can count the number of times each data acquisition unit uploads fault waveform data. This number represents the historical data volume of that data acquisition unit. For example, if a data acquisition unit uploads fault waveform data 50 times, its historical data volume is 50 records. Among multiple feeders, some feeders were put into operation earlier, resulting in more fault events recorded by their corresponding data acquisition units; others were put into operation later, resulting in fewer fault events recorded by their corresponding data acquisition units. Therefore, the number of times different data acquisition units upload fault waveform data varies, meaning the historical data volume of each data acquisition unit is not identical. Data acquisition units with larger historical data volumes have more effective samples for deviation separation, making them less affected by finite sample random fluctuations when calculating quasi-static deviation, thus resulting in higher accuracy in quasi-static deviation estimation. Therefore, by also using the historical data volume of each data acquisition unit as an input parameter for deviation separation, data acquisition units with larger historical data volumes can be given higher weight in drift value calculation due to their more accurate quasi-static deviation. This allows the drift value to more accurately reflect the true average level of the inherent measurement offset of all data acquisition units, thereby improving the accuracy of the target deviation calculation for each data acquisition unit.
[0037] For example, the target deviation of the u-th collector It can be represented as: In the formula, Let be the quasi-static deviation of the u-th data collector. This is the drift value. This represents the total number of data collectors. Let be the historical data volume of the j-th collector. Let be the quasi-static deviation of the j-th data collector.
[0038] In this embodiment of the invention, if the historical data volume of a certain collector is 0, the historical data volume of that collector is set to the initial base of 1, so as to avoid the problem that the calculation cannot be performed normally due to the denominator being zero when calculating the drift value by weighted averaging the quasi-static deviations of multiple collectors based on the historical data volumes of multiple collectors.
[0039] Step S2: Based on the target deviation of multiple collectors, the collectors are screened to obtain the first set and the second set.
[0040] Among them, the maximum value of the target deviation among multiple collectors in the first set is less than the minimum value of the target deviation among multiple collectors in the second set.
[0041] Dividing the data collectors into two groups based on the magnitude of the target deviation allows us to distinguish between data collectors with relatively small target deviations (i.e., the first set) and data collectors with relatively large deviations (i.e., the second set). This provides a basis for subsequently calculating the first failure rate of the feeders corresponding to the data collectors in the first set and the second failure rate of the feeders corresponding to the data collectors in the second set.
[0042] Specifically, the collector screening based on the target deviation of multiple collectors to obtain a first set and a second set includes: Arrange the target deviations of multiple data acquisition devices in ascending order to obtain the target deviation sequence; Multiple differences are obtained by calculating adjacent differences based on the target deviation sequence; The average of the two target deviations corresponding to the largest difference among multiple differences is determined as the segmentation threshold; Analyze the relative magnitudes of the target deviation and segmentation threshold of multiple data collectors, identify the data collectors with target deviations less than the segmentation threshold as the first set, and identify the data collectors with target deviations greater than the segmentation threshold as the second set.
[0043] Optionally, the analysis device may also determine any value between the two target deviations corresponding to the largest difference among multiple differences as the segmentation threshold.
[0044] For example, the target deviation sequence obtained from the target deviations of 6 collectors is 0.10, 0.12, 0.15, 0.18, 0.80, and 0.85. After sorting, the calculated differences are 0.02, 0.03, 0.03, 0.62, and 0.05, respectively. Among them, 0.62 is the largest difference, corresponding to the two target deviations of 0.18 and 0.80. Then, the average value of 0.18 and 0.80, 0.49, can be used as the segmentation threshold. Collectors with target deviations less than 0.49 are assigned to the first set, and collectors with target deviations greater than 0.49 are assigned to the second set.
[0045] By screening based on the average of the two target deviations corresponding to the largest difference among multiple differences, the difference in target deviation between the first set and the second set after screening can be maximized. This enables clear grouping of the collectors according to the size of the target deviation, making the first set and the second set highly distinguishable in the dimension of target deviation. This provides a reliable statistical basis for quantifying the degree of misleading effect of inherent measurement bias on feature selection.
[0046] Step S3: Statistical analysis is performed on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set to obtain the first failure rate and the second failure rate.
[0047] The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set.
[0048] Specifically, the statistical analysis based on the historical data volume and corresponding feeder fault count of multiple collectors in the first set, and the historical data volume and corresponding feeder fault count of multiple collectors in the second set, to obtain the first fault rate and the second fault rate, includes: Calculate the ratio of the number of feeder faults corresponding to each data collector to its historical data volume to obtain the feeder fault frequency corresponding to each data collector. Based on the historical data volume of multiple collectors in the first set, the feeder fault frequency corresponding to multiple collectors in the first set is weighted and averaged to obtain the first fault rate; Based on the historical data volume of multiple collectors in the second set, the feeder fault frequencies corresponding to multiple collectors in the second set are weighted and averaged to obtain the second fault rate.
[0049] When a ground fault occurs (i.e., a fault event occurs), all data acquisition units will collect and upload fault waveform data. However, not all feeders corresponding to data acquisition units are faulty lines in a fault event. The number of feeder faults corresponding to a data acquisition unit refers to the number of times the feeder corresponding to that data acquisition unit itself has a fault, which is used to reflect the fault frequency of that feeder.
[0050] The analysis device stores the identifiers of the faulty feeders after each fault event is identified. The analysis device can match the identifier of the feeder corresponding to the data acquisition unit with the identifier of the faulty feeder to obtain the number of faults for the feeder corresponding to the data acquisition unit. The identifier is used to uniquely locate the feeder among multiple feeders.
[0051] Some feeders were put into operation relatively early, resulting in a larger volume of historical data for their corresponding data collectors, leading to higher statistical reliability in fault rate estimations. Conversely, some feeders were put into operation relatively late, resulting in a smaller volume of historical data for their corresponding data collectors, making fault rate estimations more susceptible to fluctuations due to random factors. Directly averaging or comparing the fault counts of each data collector in the set would give equal weight to collectors with later operation and smaller historical data volumes compared to those with earlier operation and larger historical data volumes, causing the statistical results to deviate from the true level. Therefore, a weighted average using historical data volume as the weight ensures that data collectors with larger historical data volumes and higher statistical reliability have a higher weight in the fault rate calculation within their respective sets, thus making the calculation results more accurately reflect the average fault frequency of the feeders corresponding to the data collectors in the set.
[0052] For example, the first failure rate It can be represented as: In the formula, For the first set The Middle The amount of historical data collected by each collector. For the first set The Middle The feeder fault frequency corresponding to each data collector. For the first set The Middle The ratio of the number of feeder failures corresponding to each data collector to its historical data volume. As mentioned above, in this embodiment of the invention, if the historical data volume of a certain data collector is found to be 0, then the historical data volume of that data collector is set to the initial base of 1.
[0053] The method by which the analysis device obtains the second failure rate is similar to the method by which it obtains the first failure rate, and will not be described in detail here.
[0054] Step S4: Analyze the difference between the first failure rate and the second failure rate to obtain the difference value.
[0055] The difference value is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection.
[0056] By comparing the difference between the first failure rate and the second failure rate, it can be determined whether there is a statistical coupling between the inherent measurement offset of the data collector and the fault state. If the difference between the first failure rate and the second failure rate is significant (i.e., the larger the difference), it indicates that there is a coupling between the inherent measurement offset (i.e., deviation information) and the fault state (i.e., fault information). In other words, the target deviation of the data collector alone can distinguish the failure rate of the feeder corresponding to the data collector. This indicates that the offset information has the ability to distinguish between faulty and non-faulty samples in random forest node splitting, thus posing a risk of misrepresenting fault information in feature selection. This coupling can cause the offset information to misrepresent fault information in feature selection. If the difference between the first failure rate and the second failure rate is not significant (i.e., the smaller the difference), it indicates that the inherent measurement offset has a small impact on feature selection.
[0057] Specifically, the analysis of the difference between the first failure rate and the second failure rate to obtain the difference value includes: Calculate the absolute difference between the first failure rate and the second failure rate to obtain the inter-group failure rate difference; The difference in failure rate between groups is normalized based on the feeder failure frequency corresponding to multiple data acquisition units to obtain the difference value.
[0058] For example, the difference value It can be represented as: In the formula, The highest failure rate, This is the second highest failure rate. This represents the difference in failure rates between groups. This represents the maximum value among the feeder fault frequencies corresponding to multiple data acquisition units. This represents the minimum feeder fault frequency among multiple data acquisition units. By normalizing the difference in fault rates between groups, the resulting difference values can be mapped to the [0,1] interval.
[0059] In this embodiment of the invention, when the target deviation of multiple data collectors is all 0, it is impossible to separate the first set and the second set, and the analysis device can directly set the difference value to 0. Also, when the feeder failure rate corresponding to all data collectors is the same, then... and If they are equal, the analysis device can also directly set the difference value to 0.
[0060] Step S5: Adjust the maximum number of features in the random forest based on the difference value to obtain the target number of features, and perform feature filtering on the multidimensional features according to the random forest method and the target number of features to obtain reference features.
[0061] Among them, the target feature number is used to indicate the number of features participating in the optimal split search when a node splits. The target feature number is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.
[0062] In random forests, at each node split, the feature with the largest number of features is first randomly selected from all features to form a candidate subset. Then, the optimal splitting feature is searched within this subset. The maximum number of features controls the number of features randomly selected each time. If the maximum number of features is a fixed value, then regardless of the degree of coupling between bias information and fault information, features carrying bias information always participate in each split with the same probability. When the difference value is large, it means that the coupling degree is high, and features carrying bias information are more likely to be selected as splitting features due to their statistical association with the fault label.
[0063] Based on this, the target feature number is obtained by adjusting the maximum feature number of the random forest based on the difference value. Then, the multidimensional features are filtered according to the random forest method and the target feature number to obtain reference features. When the difference value is large, the maximum feature number can be narrowed to reduce the probability that features carrying bias information are selected and dominate the split, so that the true fault features have more chances to win in nodes that are not disturbed by bias features. When the difference value is small, the maximum feature number is maintained to maintain the normal randomness and generalization ability of the model, thereby achieving adaptive optimization of feature selection, so that the selected reference features can better reflect the intrinsic information of the fault.
[0064] Specifically, the process of adjusting the maximum number of features of the random forest based on the difference value to obtain the target number of features is as follows: the maximum number of features of the random forest is scaled based on the difference value and the preset minimum number of features to obtain the target number of features, wherein the target number of features is negatively correlated with the difference value and positively correlated with the preset minimum number of features.
[0065] For example, the number of target features It can be represented as: In the formula, This is the maximum number of features in the random forest. This maximum number of features is the default value. For example, this maximum number of features can be 80% of the total number of features. The preset minimum number of features is used, for example, the value range of the preset minimum number of features can be [total number of features × 10%, total number of features × 20%]. The preset minimum number of features is used to ensure that there are still enough features to participate in the competition in each node split when the difference value is large, so as to prevent the decision tree from overfitting or severely degrading its discriminative power due to too few features in the candidate subset.
[0066] The difference value. The preset gain sharpness index controls the rate at which the number of target features decreases as the difference value increases. For example, the preset gain sharpness index is set to 2. It is understood that the values of the maximum number of features, the preset minimum number of features, and the preset gain sharpness index in the random forest of this embodiment are exemplary and can be set according to the actual application scenario. This invention does not impose specific limitations on these values. This allows the target feature number to be smaller when the difference value is large (i.e., the statistical coupling between the inherent measurement bias of the collector and the fault state is strong), reducing the probability that features carrying bias information will be extracted and dominate the split; and the target feature number to be larger when the difference value is small (i.e., the coupling is weak), maintaining the normal randomness and generalization ability of the model.
[0067] Optionally, if the current failure event is not the first failure event, i.e., historical difference value records exist, the analysis device can perform weighted calculations based on the historical difference values obtained from each historical failure event to obtain a weighted average difference value. Then, based on the weighted average difference value, the maximum number of features in the random forest is adjusted to obtain the target number of features. The implementation method of adjusting the maximum number of features in the random forest based on the weighted average difference value to obtain the target number of features is similar to the implementation method of adjusting the maximum number of features in the random forest based on the difference value to obtain the target number of features, and will not be elaborated here. For example, weights can be assigned decreasing weights from most recent to oldest, i.e., historical difference values with later failure times have greater weights, and historical difference values with earlier failure times have smaller weights, and the sum of the weights of multiple historical difference values equals 1. This improves the sensitivity of the weighted average difference value to recent changes in system state, allowing the smoothed difference value to reflect the current level of confusion while also considering the stability of historical trends.
[0068] Before performing feature filtering on the multidimensional features based on the random forest method and the target feature number to obtain reference features, the method further includes: Feature extraction is performed on fault recording data collected by multiple acquisition devices to obtain multidimensional features. The multidimensional features include at least two of the following: time domain features, frequency domain features, time-frequency domain features, sequence component features, and transient features.
[0069] The analysis device pre-stores a feature extraction algorithm library, which includes various feature extraction algorithms. For example, time-domain feature extraction algorithms (such as time-domain analysis methods based on statistics) are used to extract time-domain features such as amplitude, peak value, distortion rate, mean, and variance of zero-sequence current from fault recording data. Frequency-domain feature extraction algorithms (such as Fast Fourier Transform) are used to extract frequency-domain features such as harmonic content and spectral energy distribution from fault recording data. Time-frequency domain feature extraction algorithms (such as wavelet transform algorithms) are used to extract time-frequency domain features such as wavelet packet coefficients and instantaneous frequency from fault recording data. Sequence component feature extraction algorithms (such as the instantaneous symmetric component method) are used to extract the amplitude and phase features of the zero-sequence component from fault recording data. Transient feature extraction algorithms (such as transient feature analysis methods based on wavelet transform and information entropy) are used to extract transient features such as current change rate and transient duration from fault recording data. The analysis device can call the algorithms in the feature extraction algorithm library to calculate various features of the fault recording data uploaded by each acquisition device in turn, so as to obtain a multi-dimensional feature set composed of multiple features.
[0070] In summary, this embodiment of the invention separates deviations based on bus zero-sequence voltage, historical data volume of each data acquisition device, and fault waveform data to obtain the target deviation of each data acquisition device. Based on the target deviations of multiple data acquisition devices, data acquisition devices are screened to obtain a first set and a second set. Subsequently, statistical analysis is performed on the historical data volume and corresponding feeder fault counts of multiple data acquisition devices in the first set, and the historical data volume and corresponding feeder fault counts of multiple data acquisition devices in the second set, to obtain a first fault rate and a second fault rate. Then, the difference between the first fault rate and the second fault rate is analyzed to obtain the difference value. Based on the difference value, the maximum number of features in the random forest is adjusted to obtain the target number of features. Finally, based on the random forest... The forest method and target feature number are used to filter multidimensional features to obtain reference features. This embodiment of the invention considers the interference of the inherent measurement offset of the acquisition device on the intrinsic features of the fault during the feature filtering process. By using the bus zero-sequence voltage as a common reference, the fault waveform data of each acquisition device is separated by deviation, and the difference value is obtained. Based on this, the maximum number of features when the random forest node splits is adaptively adjusted, so that the probability of the feature carrying the offset information (i.e., the inherent measurement offset) being selected in each split is reduced. This effectively suppresses the interference of hardware difference of the acquisition device, avoids the problem of false features being mistakenly selected as the final diagnostic features, and improves the reliability of multidimensional feature extraction of distributed fault waveform data.
[0071] This invention also proposes a multi-dimensional feature extraction system for distributed fault recording data. Please refer to [link / reference]. Figure 2 , Figure 2 This diagram illustrates the structure of a multi-dimensional feature extraction system 200 for distributed fault recording data according to an embodiment of the present invention. The system includes: The deviation separation module 201 is used to perform deviation separation based on the bus zero-sequence voltage, the historical data volume of each data acquisition device and the fault waveform data to obtain the target deviation of each data acquisition device. The target deviation is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition device from the average level of all data acquisition devices. Multiple data acquisition devices are distributed and deployed on each feeder. The grouping and screening module 202 is used to screen the collectors based on the target deviation of multiple collectors to obtain a first set and a second set, wherein the maximum value of the target deviation of multiple collectors in the first set is less than the minimum value of the target deviation of multiple collectors in the second set. The first analysis module 203 is used to perform statistical analysis based on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set, to obtain a first failure rate and a second failure rate. The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set. The second analysis module 204 is used to analyze the difference between the first failure rate and the second failure rate to obtain the difference value, wherein the difference value is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection. The feature selection module 205 is used to adjust the maximum number of features of the random forest based on the difference value to obtain the target number of features, and to select the multidimensional features according to the random forest method and the target number of features to obtain reference features. The target number of features is used to indicate the number of features participating in the optimal split search when the node splits. The target number of features is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.
[0072] It should be noted that the system provided in the above embodiments is only an example of the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device can be divided into different functional modules to complete all or part of the functions described above. In addition, the multi-dimensional feature extraction system for distributed fault waveform data and the multi-dimensional feature extraction method for distributed fault waveform data provided in the above embodiments belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.
[0073] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
Claims
1. A method for multi-dimensional feature extraction of distributed fault recording data, characterized in that, The method includes: Deviation separation is performed based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit. The target deviation is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition unit from the average level of all data acquisition units. Multiple data acquisition units are distributed and deployed on each feeder. Collectors are screened based on the target deviation of multiple collectors to obtain a first set and a second set, wherein the maximum value of the target deviation of multiple collectors in the first set is less than the minimum value of the target deviation of multiple collectors in the second set. Statistical analysis is performed based on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set to obtain the first failure rate and the second failure rate. The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set. The difference between the first failure rate and the second failure rate is analyzed to obtain the difference value, which is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection. The maximum number of features in the random forest is adjusted based on the difference value to obtain the target number of features. Then, the multidimensional features are filtered according to the random forest method and the target number of features to obtain reference features. The target number of features is used to indicate the number of features participating in the optimal split search when a node splits. The target number of features is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.
2. The multi-dimensional feature extraction method of distributed fault recording data according to claim 1, characterized in that, The deviation separation is performed based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit, including: Based on the bus zero-sequence voltage, the fault waveform data collected by each data acquisition unit is separated by deviation to obtain the quasi-static deviation of each data acquisition unit. The quasi-static deviation is used to characterize the inherent measurement offset of the corresponding data acquisition unit. Zero drift correction is performed based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device.
3. The multi-dimensional feature extraction method of distributed fault recording data according to claim 2, characterized in that, The fault recording data includes: zero-sequence current. The fault recording data collected by each data acquisition unit is separated based on the bus zero-sequence voltage to obtain the quasi-static deviation of each data acquisition unit, including: Analyze the waveform changes of the zero-sequence voltage at the bus to obtain the voltage change rate; Based on the voltage change rate and the zero-sequence current collected by each collector, capacitance estimation is performed to obtain the capacitance coefficient corresponding to each collector. Calculate the product of the capacitance coefficient and the voltage change rate for each data collector to obtain the capacitor current for each data collector. For each data collector, the corresponding capacitor current is removed from the zero-sequence current it collects to obtain the residual current for each data collector. The quasi-static deviation of each data acquisition device is obtained by averaging the residual current corresponding to each data acquisition device and the residual current accumulated by each data acquisition device in historical fault events.
4. The multi-dimensional feature extraction method of distributed fault recording data according to claim 2, characterized in that, The zero-drift correction is performed based on the quasi-static bias of multiple data acquisition devices and the historical data volume of multiple data acquisition devices to obtain the target bias of each data acquisition device, including: The drift value is obtained by weighting the quasi-static deviation of multiple data acquisition devices based on the historical data volume of multiple data acquisition devices. The drift value is used to characterize the average level of the inherent measurement offset of all data acquisition devices. The difference between the quasi-static deviation and the drift value of each data acquisition unit is calculated to obtain the target deviation of each data acquisition unit.
5. The multi-dimensional feature extraction method of distributed fault recording data according to claim 1, characterized in that, The collector screening based on the target deviation of multiple collectors yields a first set and a second set, including: Arrange the target deviations of multiple data acquisition devices in ascending order to obtain the target deviation sequence; Multiple differences are obtained by calculating adjacent differences based on the target deviation sequence; The average of the two target deviations corresponding to the largest difference among multiple differences is determined as the segmentation threshold; Analyze the relative magnitudes of the target deviation and segmentation threshold of multiple data collectors, identify the data collectors with target deviations less than the segmentation threshold as the first set, and identify the data collectors with target deviations greater than the segmentation threshold as the second set.
6. The method for multidimensional feature extraction of distributed fault recording data according to claim 1, characterized in that, The statistical analysis is performed on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set, to obtain a first failure rate and a second failure rate, including: Calculate the ratio of the number of feeder faults corresponding to each data collector to its historical data volume to obtain the feeder fault frequency corresponding to each data collector. Based on the historical data volume of multiple collectors in the first set, the feeder fault frequency corresponding to multiple collectors in the first set is weighted and averaged to obtain the first fault rate; Based on the historical data volume of multiple collectors in the second set, the feeder fault frequencies corresponding to multiple collectors in the second set are weighted and averaged to obtain the second fault rate.
7. The method for multidimensional feature extraction of distributed fault recording data according to claim 1, characterized in that, The analysis of the difference between the first failure rate and the second failure rate yields the difference value, including: Calculate the absolute difference between the first failure rate and the second failure rate to obtain the inter-group failure rate difference; The difference in failure rate between groups is normalized based on the feeder failure frequency corresponding to multiple data acquisition units to obtain the difference value.
8. The method for multidimensional feature extraction of distributed fault recording data according to claim 1, characterized in that, The process of adjusting the maximum number of features in the random forest based on the difference value to obtain the target number of features includes: The maximum number of features in the random forest is scaled based on the difference value and the preset minimum number of features to obtain the target number of features. The target number of features is negatively correlated with the difference value and positively correlated with the preset minimum number of features.
9. The method for multidimensional feature extraction of distributed fault recording data according to claim 1, characterized in that, Before performing feature filtering on the multidimensional features based on the random forest method and the target feature number to obtain reference features, the method further includes: Feature extraction is performed on fault recording data collected by multiple acquisition devices to obtain multidimensional features. The multidimensional features include at least two of the following: time domain features, frequency domain features, time-frequency domain features, sequence component features, and transient features.
10. A multi-dimensional feature extraction system for distributed fault recording data, characterized in that, The system includes: The deviation separation module is used to separate deviations based on the bus zero-sequence voltage, the historical data volume of each data acquisition unit, and the fault waveform data to obtain the target deviation of each data acquisition unit. The target deviation is used to characterize the degree of deviation of the inherent measurement offset of the corresponding data acquisition unit from the average level of all data acquisition units. Multiple data acquisition units are distributed and deployed on each feeder. The grouping and screening module is used to screen multiple collectors based on their target deviations to obtain a first set and a second set, wherein the maximum value of the target deviations of multiple collectors in the first set is less than the minimum value of the target deviations of multiple collectors in the second set. The first analysis module is used to perform statistical analysis based on the historical data volume and corresponding feeder failure count of multiple collectors in the first set, and the historical data volume and corresponding feeder failure count of multiple collectors in the second set, to obtain a first failure rate and a second failure rate. The first failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the first set, and the second failure rate is used to characterize the average failure frequency of multiple feeders corresponding to multiple collectors in the second set. The second analysis module is used to analyze the difference between the first failure rate and the second failure rate to obtain the difference value, wherein the difference value is used to characterize the degree to which the inherent measurement offset of multiple acquisition devices misleads feature selection. The feature selection module is used to adjust the maximum number of features in the random forest based on the difference value to obtain the target number of features. It then selects multidimensional features based on the random forest method and the target number of features to obtain reference features. The target number of features is used to indicate the number of features participating in the optimal split search when a node splits. The target number of features is negatively correlated with the difference value. The multidimensional features are obtained by feature extraction based on fault recording data.