A substation partial discharge fault early warning method and system

Abstract

The application belongs to the technical field of substation detection and early warning, and provides a substation partial discharge fault early warning method and system, which comprises the following steps: constructing a two-dimensional time sequence evolution graph based on continuously collected pulse phase and amplitude data, detecting structural mutations by calculating phase entropy values cycle by cycle, and intercepting pulse waveform data windows before and after the mutations. Then, rising time and zero-crossing number of the pulses are extracted to construct a waveform morphology feature vector, whether the same physical source before and after the mutations is determined by cosine similarity, so as to exclude the possibility of discharge property change. Finally, independent component analysis is performed on the data window before the mutations, mixed pulses are decomposed into multiple independent source components, and the matching rate of the residual pulses after the mutations and each component is used to distinguish the continuous source and the disappearing source, and a decoupling state report is output. The application can accurately identify the scenario of disappearance of one source in the coexistence of multiple sources, and avoid false reporting of discharge property change or discharge disappearance.

Description

Technical Field

[0001] This invention belongs to the field of substation detection and early warning technology, specifically a method and system for early warning of partial discharge faults in substations. Background Technology

[0002] Online monitoring of partial discharge is an important means of assessing the insulation status of high-voltage power equipment. Existing technologies typically use clustering algorithms (such as K-Means and DBSCAN) to perform unsupervised classification of the pulse characteristics collected by sensors, and then output the discharge type (point discharge, floating discharge, etc.) through pattern recognition, thereby achieving fault early warning.

[0003] However, existing technologies face serious limitations when two or more weak partial discharge sources (such as tip discharges of different phases) occur simultaneously within the same detection area. Because the pulses from different sources overlap in the feature space (amplitude, phase, and time-frequency characteristics), conventional clustering algorithms cannot effectively separate them, forcing all pulses into a single cluster and potentially mislabeling them as a third type (such as floating discharge).

[0004] A more critical issue lies in the dynamic evolution scenario. When one of the partial discharge sources suddenly disappears due to physical reasons (such as particle deposition or tip passivation), the signal collected by the sensor abruptly changes from a mixed state of "source A + source B" to a single-source state of "source A only." At this time, the phase distribution, pulse repetition rate, and other characteristics of the original cluster change drastically. Existing trend tracking modules lack multi-source demixing capabilities and cannot identify whether this change is caused by "source disappearance." Instead, they misjudge it as a change in the properties of the same discharge cluster or its complete disappearance, thus triggering false alarms of "discharge property change" or "discharge disappearance." If maintenance personnel mistakenly believe the alarm and delay maintenance, the remaining real partial discharge sources will continue to develop, potentially eventually leading to an insulation breakdown accident.

[0005] Therefore, the present invention provides a method and system for early warning of partial discharge faults in substations. Summary of the Invention

[0006] The purpose of this invention is to provide a multi-amplifier power supply simulation method based on broadband radio frequency. To solve the above-mentioned problems in the prior art, this invention achieves this through the following technical solution: A substation partial discharge fault early warning method, comprising the following steps: Step 1: Based on the phase amplitude data of each continuously acquired pulse, construct a two-dimensional distribution map and form a time-series evolution map of pulse characteristic distribution through color coding; Step 2: Based on the time-series evolution diagram, perform phase distribution analysis periodically at each power frequency to determine whether a structural abrupt change occurs. If it does, record the pulse waveform data window before and after the structural abrupt change. Step 3: Based on the pulse waveform data window before and after the structural mutation, extract pulse features to construct waveform morphology feature vectors, and determine whether they belong to the same physical source before and after the mutation through similarity analysis; Step 4: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to achieve component source determination, and output a component source decoupling status report.

[0007] As a further aspect of the present invention: In step one, the phase amplitude data acquisition method for each pulse is as follows: The original waveform of partial discharge is continuously acquired by a UHF or ultrasonic sensor installed on high-voltage equipment at a sampling rate of not less than 100 MS / s. When the signal amplitude exceeds the background noise threshold, it is marked as a discharge pulse. For each pulse, the phase angle φ of the occurrence time relative to the power frequency cycle is calculated by synchronously acquiring the zero-crossing point of the power frequency voltage, and the peak value of the pulse waveform is extracted as the amplitude A. Finally, the triplet data of each pulse is output.

[0008] As a further aspect of the present invention: in step one, the process of forming the time-series evolution diagram of the pulse characteristic distribution is as follows: A two-dimensional grid coordinate system is constructed by dividing the phase intervals according to fixed intervals, with the power frequency cycle number as the horizontal axis and the phase angle as the vertical axis. For each output pulse, determine the horizontal coordinate according to the power frequency cycle number it belongs to, and determine the vertical coordinate according to the interval where its phase angle φ is located, and place the pulse into the corresponding grid cell; Calculate the average amplitude of all pulses falling into the grid cells, and fill the grid cells with color according to the color mapping rules. All grid cells are arranged sequentially from left to right on the horizontal axis and from bottom to top on the vertical axis to form a time-series evolution diagram of pulse characteristic distribution.

[0009] As a further aspect of the present invention: in step two, the method for determining structural mutation is as follows: Read the timing evolution graph. The horizontal axis of the timing evolution graph is the power frequency cycle number k, and the vertical axis is the phase interval. For the k-th power frequency cycle, extract the average amplitude of each phase interval to form a vector. Traverse all power frequency cycles to obtain the vector sequence. For each power frequency cycle k, based on the vector sequence, calculate the total amplitude of all intervals within the power frequency cycle, then calculate the proportion of the amplitude of each interval to the total amplitude, and then calculate the phase entropy value of the power frequency cycle based on the obtained proportion; Calculate the phase entropy value for each power frequency cycle, observe the change in phase entropy value over three consecutive power frequency cycles, and set the criteria for determining structural abrupt changes: If the phase entropy value drops sharply from above 2.5 to below 1.5, and the change in phase entropy value is completed within 3 cycles, then a structural mutation is determined to have occurred.

[0010] As a further aspect of the present invention: In step two, the process of recording the pulse waveform data window before and after the structural abrupt change is as follows: Using the power frequency cycle number k where the mutation occurs as the dividing point, take 10 complete power frequency cycles before the mutation as the data window W_pre before the mutation, and take 10 complete power frequency cycles after the mutation as the data window W_post after the mutation. Return to the raw waveform data continuously acquired in step one, extract the waveforms corresponding to the time periods W_pre and W_post, perform pulse detection respectively, and obtain all pulse waveforms and their phase and amplitude information within the two data windows.

[0011] As a further aspect of the present invention: In step three, the process of extracting pulse features and constructing waveform morphology feature vectors based on the pulse waveform data windows before and after the structural mutation is as follows: Read the data window W_pre before the mutation and the data window W_post after the mutation. Each data window contains all pulse waveforms within 10 power frequency cycles. Perform the following operations on each pulse: Locate the start and end points of the pulse waveform, calculate the pulse rise time tr and the number of zero crossings nzc, and obtain a two-dimensional feature vector (tr, nzc) for each pulse. This yields the pulse feature set F_pre for the data window W_pre before the mutation and the pulse feature set F_post for the data window W_post after the mutation. Calculate the average rise time and the average number of zero crossings of pulses within the pulse feature set F_pre, and combine them to obtain the average eigenvector of the pulse feature set F_pre; Calculate the average pulse rise time and the average number of pulse zero crossings within the pulse feature set F_post, and combine them to obtain the average eigenvector of the pulse feature set F_post.

[0012] As a further aspect of the present invention: In step three, the process of determining whether the mutation occurred before and after the mutation belongs to the same physical source through similarity analysis is as follows: Calculate the cosine similarity between the pulse feature set F_pre and the pulse feature set F_post with respect to the average feature vector. If the cosine similarity meets the threshold requirement, it means that they belong to the same physical source before and after the mutation.

[0013] As a further aspect of the present invention: In step four, performing independent component analysis on the pulse waveform data window before the mutation, and decomposing the mixed pulse into multiple independent source components, is as follows: All pulse waveforms are detected from the pre-mutation data window W_pre. Each pulse is truncated to a fixed length of 200 sampling points, and each pulse is represented as a 200-dimensional vector. All pulses are arranged in rows to form an observation matrix X. Using the FastICA algorithm, assuming that the waveforms of each pulse are statistically independent, X is decomposed into K independent source components. Each component corresponds to a hypothetical physical discharge point, and the waveform template of each component is output.

[0014] As a further aspect of the present invention: the process of matching the residual pulse after mutation with each component one by one to determine the component source and outputting a decoupling status report of the component source is as follows: Take the data window W_post after the mutation, detect all pulses from it, retain the complete waveform of each pulse, and mark it as the residual pulse set R; For each decomposed independent source component, extract the waveform template, calculate the waveform similarity between the waveform template and each pulse in the residual pulse set R, count the proportion of pulses in the residual pulse set R whose waveform similarity with the independent source component meets the threshold requirement, and obtain the matching rate. Independent source components with a matching rate greater than 80% are marked as persistent sources, independent source components with a matching rate less than 20% are marked as disappeared sources, and independent source components with a matching rate between 20% and 80% are marked as uncertain sources.

[0015] A substation partial discharge fault early warning system includes the following modules: Temporal evolution analysis module: Based on the phase amplitude data of each continuously acquired pulse, a two-dimensional distribution map is constructed, and a temporal evolution map of pulse characteristic distribution is formed through color coding; The mutation detection and data extraction module: Based on the time-series evolution diagram, it performs phase distribution analysis periodically on a power frequency basis to determine whether a structural mutation has occurred. If it has occurred, it records the pulse waveform data window before and after the structural mutation. Homology discrimination module: Based on the pulse waveform data window before and after the structural mutation, pulse features are extracted to construct waveform morphology feature vectors, and similarity analysis is used to determine whether they belong to the same physical source before and after the mutation; Source separation and decoupling module: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to realize the component source determination, and output the component source decoupling status report.

[0016] The beneficial effects of this invention are as follows: By constructing a time-series evolution diagram and calculating the phase entropy value, the structural abrupt change moment of the transition from multi-source coexistence to single-source dominance can be accurately captured, providing accurate time boundaries for subsequent analysis. By extracting pulse waveform morphological features and calculating cosine similarity, it is possible to effectively determine whether the pulses before and after the abrupt change originate from the same physical source, thereby eliminating the possibility of a change in discharge properties and avoiding false alarms from the source. Under the premise of confirming the same source, the mixed waveform is decomposed into multiple independent source components through independent component analysis. Then, based on the matching rate between the residual pulse after the abrupt change and each component, the persistent source and the disappearing source are distinguished. Finally, a decoupling status report is output, clearly indicating which source has disappeared and which source continues to exist. This conclusion directly replaces the false alarms of discharge property change or discharge disappearance in the prior art, enabling maintenance personnel to accurately grasp the true discharge status, avoiding delays in maintenance due to false alarms leading to the expansion of insulation faults, and significantly improving the early warning reliability of online partial discharge monitoring. Attached Figure Description

[0017] The invention will now be further described with reference to the accompanying drawings.

[0018] Figure 1 This is a flowchart of a substation partial discharge fault early warning method according to an embodiment of the present invention; Figure 2 This is a logic judgment diagram of a substation partial discharge fault early warning method according to an embodiment of the present invention; Figure 3 This is a block diagram of a substation partial discharge fault early warning system according to an embodiment of the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0020] Example 1, please refer to Figures 1-2 As shown in the figure, a substation partial discharge fault early warning method according to an embodiment of the present invention includes the following steps: Step 1: Based on the phase amplitude data of each continuously acquired pulse, construct a two-dimensional distribution map and form a time-series evolution map of pulse characteristic distribution through color coding; In step one, the phase amplitude data of each pulse is acquired in the following way: The original waveform of partial discharge is continuously acquired by using an ultra-high frequency or ultrasonic sensor installed on high-voltage equipment at a sampling rate of not less than 100 MS / s; a background noise threshold is set (e.g., higher than 5 times the standard deviation of the baseline amplitude), and when the signal amplitude exceeds the background noise threshold, it is marked as a discharge pulse; The background noise threshold can be set in the following ways: Take 1000 consecutive sampling points (duration 10 microseconds) to calculate the root mean square value of amplitude RMS_noise, and set the threshold Th to 5 times RMS_noise (on-site measured empirical value, 5 times can filter out more than 99% of white noise interference, less than 3 times is easy to be falsely triggered due to noise fluctuation). For each pulse, the phase angle φ (range 0°-360°) of the occurrence time relative to the power frequency cycle is calculated by synchronously acquiring the zero-crossing point of the power frequency voltage, and the peak value of the pulse waveform is extracted as the amplitude A (unit mV); finally, the triplet data of each pulse (time t, phase φ, amplitude A) is output. The phase angle can be calculated in the following ways: For each detected pulse, calculate the time difference Δt between the pulse start point and the nearest zero-crossing point of the power frequency, and calculate the phase angle according to the formula φ=(Δt / 20ms)×360° (20ms period at 50Hz power frequency). In step one, the method for constructing the two-dimensional distribution map is as follows: A two-dimensional grid coordinate system is constructed with the power frequency cycle number as the horizontal axis (time dimension) and the phase angle 0°-360° as the vertical axis, and the phase interval is divided into 5° intervals. For each output pulse, determine the horizontal coordinate according to its power frequency cycle number and the vertical coordinate according to the interval where its phase angle φ is located, and place the pulse into the corresponding grid cell; each cell summarizes all the pulses that fall into it; In step one, the process of forming a time-series evolution diagram of pulse feature distribution through color encoding is as follows: For each obtained grid cell, calculate the average amplitude of all pulses falling into the grid cell; set the color mapping rules: 0-20mV corresponds to dark blue, 20-40mV corresponds to light blue, 40-60mV corresponds to green, 60-80mV corresponds to orange, and 80-100mV and above corresponds to dark red. Each grid cell is filled with the corresponding color according to the calculated average amplitude; all grid cells are arranged sequentially from left to right on the horizontal axis (time period number) and from bottom to top on the vertical axis (phase interval) to form a complete two-dimensional color spectrum, that is, the time evolution diagram of pulse characteristic distribution; Understandably, the significance of step one lies in transforming the raw, discrete pulse data into a visualized two-dimensional time-series evolution diagram. With time as the horizontal axis, phase as the vertical axis, and color representing amplitude intensity, it visually presents the overall distribution and changing trend of pulse characteristics over time. This diagram provides a structured data foundation for subsequent phase entropy calculations and structural mutation detection, making the originally chaotic pulse sequence quantifiable and analyzable. Step 2: Based on the time-series evolution diagram, perform phase distribution analysis periodically at each power frequency to determine whether a structural abrupt change occurs. If it does, record the pulse waveform data window before and after the structural abrupt change. In step two, based on the time-series evolution diagram, phase distribution analysis is performed periodically at each power frequency to determine whether a structural abrupt change has occurred. Read the time series evolution graph. The horizontal axis of the time series evolution graph represents the power frequency cycle number k (k=1,2,3,…,N), and the vertical axis represents 72 phase intervals (each interval is 5° wide, covering 0° to 360°). For the k-th cycle, extract the average amplitude of each of the 72 intervals to form a vector V_k. Traverse all cycles to obtain the sequence V_1,V_2,…,V_N; where V_N is the vector of the N-th power frequency cycle. For each power frequency cycle k, calculate the total amplitude S_k (i.e., the sum of all components in V_k) of all 72 intervals within the power frequency cycle, and then calculate the proportion of the amplitude of each interval to the total amplitude p_i = (amplitude of the i-th interval) / S_k (if S_k = 0, then all p_i = 0); then calculate the phase entropy value H_k of the power frequency cycle = -∑p_i × log2(p_i), where a larger H_k indicates a more dispersed phase distribution (multi-source coexistence characteristic), and a smaller H_k indicates a more concentrated phase (single-source dominance characteristic). The phase entropy value H_k is calculated cycle by cycle at power frequency. After calculation, the change of the phase entropy value H_k over three consecutive power frequency cycles is observed, and the criteria for determining structural abrupt changes are set: If the phase entropy value H_k drops sharply from above 2.5 to below 1.5, and the change in phase entropy value is completed within 3 cycles, then a structural mutation is determined to have occurred; otherwise, no structural mutation has occurred. If a structural mutation occurs, record the power frequency cycle number k at which the mutation occurs; In step two, the process of recording the pulse waveform data window before and after the structural abrupt change is as follows: Using the power frequency cycle number k where the mutation occurs as the dividing point, take 10 complete power frequency cycles before the mutation (from cycle K-10 to cycle K-1) as the data window W_pre before the mutation, and take 10 complete power frequency cycles after the mutation (from cycle K to cycle K+9) as the data window W_post after the mutation. Return to the raw waveform data continuously acquired in step one, extract the waveforms corresponding to the time periods W_pre and W_post, perform pulse detection respectively, and obtain all pulse waveforms and their phase and amplitude information within the two data windows; Understandably, the significance of step two lies in: based on the time-series evolution diagram generated in step one, quantifying the dispersion of pulse phase distribution by calculating phase entropy values ​​cycle by cycle. When the entropy value drops sharply from high (multi-source coexistence characteristic) to low (single-source dominance characteristic), the occurrence time of structural abrupt change is accurately determined, and pulse waveform data windows before and after the abrupt change are automatically extracted. This step transforms the phenomenon into a quantifiable criterion, providing time boundaries and raw data support for distinguishing between "source disappearance" and "property change." Step 3: Based on the pulse waveform data window before and after the structural mutation, extract pulse features to construct waveform morphology feature vectors, and determine whether they belong to the same physical source before and after the mutation through similarity analysis; In step three, the process of extracting pulse features and constructing waveform morphology feature vectors based on pulse waveform data windows before and after structural mutation is as follows: Read the pre-mutation data window W_pre and post-mutation data window W_post output from step two. Each data window contains all pulse waveforms within 10 power frequency cycles. Perform the following operations on each pulse: Locate the start and end points of the pulse waveform, calculate the pulse rise time tr (the time required for the amplitude to rise from 10% to 90%, in nanoseconds), and calculate the number of zero crossings nzc (the number of times the waveform crosses the baseline, reflecting the oscillation characteristics). Each pulse thus yields a two-dimensional feature vector (tr, nzc), which respectively yields the pulse feature set F_pre for the data window W_pre before the mutation and the pulse feature set F_post for the data window W_post after the mutation; Perform the following operations on the pulse feature set F_pre and the pulse feature set F_post: Calculate the average rise time and the average number of zero crossings of pulses within the pulse feature set F_pre, and combine them to obtain the average eigenvector of the pulse feature set F_pre; Calculate the average pulse rise time and the average number of pulse zero crossings within the pulse feature set F_post, and combine them to obtain the average eigenvector of the pulse feature set F_post; In step three, the process of determining whether the mutation occurred before and after the mutation originates from the same physical source through similarity analysis is as follows: Calculate the cosine similarity between the pulse feature set F_pre and the pulse feature set F_post with respect to the average feature vector. If the cosine similarity meets the threshold (greater than 0.85), it means that the pulses before and after the mutation belong to the same physical source. That is, the remaining pulses after the mutation and a part of the mixed pulses before the mutation come from the same discharge point, thus eliminating the possibility of "change in discharge properties". Otherwise, if it does not meet the threshold, it is determined that they do not belong to the same physical source. At this time, step four is not triggered, and the system maintains the original alarm logic or outputs "feature mutation but not from the same source, manual review is recommended". It should be noted that the 0.85 threshold is set based on statistical analysis of field measurement data: The similarity of waveforms of homogeneous pulses was calculated for pure single-source partial discharge data confirmed from multiple substations. The mean was approximately 0.92, and the standard deviation was approximately 0.04, meaning that over 95% of homogeneous pulse similarities fell between 0.84 and 1.00. Therefore, using 0.85 as the lower limit can cover the vast majority of homogeneous cases. Cross-comparison of heterogeneous pulses showed that the mean similarity was approximately 0.35, and the standard deviation was approximately 0.12, with over 99% below 0.60. 0.85 is significantly higher than this distribution, preventing misjudgment of heterogeneous sources. Simultaneously, the 0.85 value retains an engineering safety margin, tolerating waveform distortion caused by sensor sensitivity fluctuations and slight noise interference. It strikes a balance between a false negative rate below 0.1% and a false negative rate below 0.01%, and conforms to the industry convention of IEEE Std 1434-2014 regarding pulse grouping correlation coefficient thresholds of 0.80-0.90. Understandably, the significance of step three lies in: based on the data windows before and after the mutation extracted in step two, extracting the rise time and number of zero crossings of the pulse to construct a waveform morphology feature vector, and using cosine similarity to determine whether the pulses before and after the mutation originate from the same physical source. This step eliminates the possibility of "change in discharge properties," confirming that the mutation is caused by the disappearance of one of the multiple sources rather than the evolution of a single source itself, thus providing a prerequisite for subsequent source separation and decoupling. Step 4: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to achieve component source determination, and output a component source decoupling status report. In step four, independent component analysis is performed on the pulse waveform data window before the mutation, and the process of decomposing the mixed pulse into multiple independent source components is as follows: Take the data window W_pre before the mutation (i.e., the original waveform data of the 10 power frequency cycles before the mutation occurred) output in step two. All pulse waveforms are detected from W_pre, and each pulse is truncated to a fixed length (200 sampling points). Each pulse is represented as a 200-dimensional vector. All pulses are arranged in rows to form an observation matrix X (the number of rows is the number of pulses, and the number of columns is 200). Using the FastICA algorithm, assuming that the waveforms of each pulse are statistically independent, X is decomposed into K independent source components. Each component corresponds to a hypothetical physical discharge point, and the waveform template (200-dimensional vector) of each component is output. It should be noted that the waveform template is essentially a numerical vector (array) of length 200. It represents the standard pulse waveform of a separated pure partial discharge power source. It corresponds to the inherent waveform characteristics (including rising edge, damped oscillation, overshoot, etc.) of a hypothetical physical discharge point (such as a tip discharge or floating discharge point) at the sensor end. Its output method can be as follows: All pulses are detected from the pre-mutation data window W_pre. Each pulse is truncated to a fixed length (200 sampling points). Assuming there are M pulses in total, an M×200 observation matrix X is constructed, where each row represents the waveform vector of a mixed pulse. Matrix X is input into the FastICA algorithm, which, based on the assumption that the pulse waveforms are statistically independent, iteratively solves by maximizing non-Gaussianity (negative entropy) to find a dismixing matrix W such that: S = W × X; where S is the source matrix obtained after decomposition, with a dimension of K × 200 (K is the estimated number of independent sources). Each row (K rows in total) in the source matrix S is the estimated waveform of an independent source component in that time period. At this time, the algorithm directly outputs each row in S as a 200-dimensional vector. In step four, the process of matching the residual pulse after the mutation with each component one by one to determine the component source and outputting a decoupling status report of the component source is as follows: Take the data window W_post after the mutation output in step two, detect all pulses from it, retain the complete waveform of each pulse (200 sampling points), and call these pulses the residual pulse set R; For each decomposed independent source component, its waveform template is extracted, and the waveform similarity between the waveform template and each pulse in the residual pulse set R is calculated. The proportion of pulses in the residual pulse set R with a waveform similarity greater than 0.85 to the independent source component is counted to obtain the matching rate. Independent source components with a matching rate greater than 80% are marked as persistent sources (still discharging after the mutation), independent source components with a matching rate less than 20% are marked as disappeared sources (stopping discharging after the mutation), and independent source components with a matching rate between 20% and 80% are marked as uncertain sources (requiring manual verification). A decoupling status report is generated based on the matching results, including: each independent source component and its discharge type (identified by calling a pre-classifier through the waveform template, such as tip discharge, floating discharge, etc.). Components with a matching rate greater than 80% are marked as persistent, and components with a matching rate less than 20% are marked as disappeared. The final output conclusion is: the multi-source coexistence state is resolved, a certain source has disappeared, a certain source persists, and the discharge properties have not changed. It should be noted that the 20% and 80% matching rate thresholds are set based on the statistical distribution of field-measured data and engineering tolerance requirements: Retrospective analysis of "one-source disappearance" scenarios confirmed in multiple substations revealed that the residual pulse matching rate of persistent sources in the data window after the abrupt change was generally higher than 85%, with an average of approximately 90%, while the matching rate of disappeared sources was generally lower than 15%, with an average of approximately 8%, showing a clear separation between the two. Setting 80% as the lower limit for persistent source judgment ensures that the identified persistent sources have high confidence (false positive rate less than 0.5%); setting 20% ​​as the upper limit for disappeared source judgment ensures that the identified disappeared sources have indeed stopped discharging (false negative rate less than 1%). The gray area between 20% and 80% corresponds to uncertain situations such as large waveform distortion before and after the abrupt change, insufficient residual pulse count, or similar waveform characteristics of two sources. In these cases, they are automatically marked as uncertain sources and prompt for manual review, avoiding forced erroneous judgments by the system. The above thresholds achieve a balance between engineering safety margin and identification sensitivity, and are consistent with the waveform similarity threshold of 0.85 in step three. Understandably, the significance of step four lies in the following: Assuming the source is confirmed in step three, independent component analysis is performed on the pre-mutation data window to decompose the mixed waveform into multiple independent source components. Then, the matching rate between the residual pulses after the mutation and each component distinguishes between persistent and disappearing sources. Finally, a decoupling report is output, clearly indicating "a certain source has disappeared" and "a certain source continues to exist," directly replacing the false alarms of "change in discharge characteristics" or "discharge disappearance" in existing technologies, thus preventing maintenance personnel from delaying repairs due to false alarms.

[0021] Example 2, please refer to Figure 3 As shown in the embodiment of the present invention, a substation partial discharge fault early warning system includes the following modules: Temporal evolution analysis module: Based on the phase amplitude data of each continuously acquired pulse, a two-dimensional distribution map is constructed, and a temporal evolution map of pulse characteristic distribution is formed through color coding; The mutation detection and data extraction module: Based on the time-series evolution diagram, it performs phase distribution analysis periodically on a power frequency basis to determine whether a structural mutation has occurred. If it has occurred, it records the pulse waveform data window before and after the structural mutation. Homology discrimination module: Based on the pulse waveform data window before and after the structural mutation, pulse features are extracted to construct waveform morphology feature vectors, and similarity analysis is used to determine whether they belong to the same physical source before and after the mutation; Source separation and decoupling module: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to realize the component source determination, and output the component source decoupling status report.

[0022] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for early warning of partial discharge faults in substations, characterized in that: Includes the following steps: Step 1: Based on the phase amplitude data of each continuously acquired pulse, construct a two-dimensional distribution map and form a time-series evolution map of pulse characteristic distribution through color coding; Step 2: Based on the time-series evolution diagram, perform phase distribution analysis periodically at each power frequency to determine whether a structural abrupt change occurs. If it does, record the pulse waveform data window before and after the structural abrupt change. Step 3: Based on the pulse waveform data window before and after the structural mutation, extract pulse features to construct waveform morphology feature vectors, and determine whether they belong to the same physical source before and after the mutation through similarity analysis; Step 4: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to achieve component source determination, and output a component source decoupling status report.

2. The method for early warning of partial discharge faults in substations according to claim 1, characterized in that: In step one, the phase amplitude data of each pulse is acquired in the following way: The original waveform of partial discharge is continuously acquired by a UHF or ultrasonic sensor installed on high-voltage equipment. When the signal amplitude exceeds the background noise threshold, it is marked as a discharge pulse. For each pulse, the phase angle φ of the occurrence time relative to the power frequency cycle is calculated by synchronously acquiring the zero-crossing point of the power frequency voltage, and the peak value of the pulse waveform is extracted as the amplitude A. Finally, the triplet data of each pulse is output.

3. The method for early warning of partial discharge faults in substations according to claim 2, characterized in that: In step one, the process of forming the time-series evolution diagram of the pulse characteristic distribution is as follows: A two-dimensional grid coordinate system is constructed by dividing the phase intervals according to fixed intervals, with the power frequency cycle number as the horizontal axis and the phase angle as the vertical axis. For each output pulse, determine the horizontal coordinate according to the power frequency cycle number it belongs to, and determine the vertical coordinate according to the interval where its phase angle φ is located, and place the pulse into the corresponding grid cell; Calculate the average amplitude of all pulses falling into the grid cells, and fill the grid cells with color according to the color mapping rules. All grid cells are arranged sequentially from left to right on the horizontal axis and from bottom to top on the vertical axis to form a time-series evolution diagram of pulse characteristic distribution.

4. The method for early warning of partial discharge faults in substations according to claim 1, characterized in that: In step two, the method for determining structural mutations is as follows: Read the timing evolution graph. The horizontal axis of the timing evolution graph is the power frequency cycle number k, and the vertical axis is the phase interval. For the k-th power frequency cycle, extract the average amplitude of each phase interval to form a vector. Traverse all power frequency cycles to obtain the vector sequence. For each power frequency cycle k, based on the vector sequence, calculate the total amplitude of all intervals within the power frequency cycle, then calculate the proportion of the amplitude of each interval to the total amplitude, and then calculate the phase entropy value of the power frequency cycle based on the obtained proportion; Calculate the phase entropy value for each power frequency cycle, observe the change in phase entropy value for a consecutive preset number of power frequency cycles, and set the criteria for determining structural abrupt changes: If the phase entropy value drops sharply from above the first phase entropy threshold to below the second phase entropy threshold, and the change in phase entropy value is completed within a preset period, then a structural mutation is determined to have occurred.

5. The method for early warning of partial discharge faults in substations according to claim 4, characterized in that: In step two, the process of recording the pulse waveform data window before and after the structural abrupt change is as follows: Using the power frequency cycle number k where the mutation occurs as the dividing point, take multiple complete power frequency cycles forward as the data window W_pre before the mutation, and take multiple complete power frequency cycles backward as the data window W_post after the mutation. Return to the raw waveform data continuously acquired in step one, extract the waveforms corresponding to the time periods W_pre and W_post, perform pulse detection respectively, and obtain all pulse waveforms and their phase and amplitude information within the two data windows.

6. The method for early warning of partial discharge faults in substations according to claim 1, characterized in that: In step three, the process of extracting pulse features and constructing waveform morphology feature vectors based on the pulse waveform data windows before and after the structural mutation is as follows: Read the data window W_pre before the mutation and the data window W_post after the mutation. Each data window contains all pulse waveforms within the power frequency cycle. Perform the following operations on each pulse: Locate the start and end points of the pulse waveform, calculate the pulse rise time tr and the number of zero crossings nzc, and obtain a two-dimensional feature vector (tr, nzc) for each pulse. This yields the pulse feature set F_pre for the data window W_pre before the mutation and the pulse feature set F_post for the data window W_post after the mutation. Calculate the average rise time and the average number of zero crossings of pulses within the pulse feature set F_pre, and combine them to obtain the average eigenvector of the pulse feature set F_pre; Calculate the average pulse rise time and the average number of pulse zero crossings within the pulse feature set F_post, and combine them to obtain the average eigenvector of the pulse feature set F_post.

7. The method for early warning of partial discharge faults in substations according to claim 6, characterized in that: In step three, the process of determining whether the mutation occurred before and after the mutation originates from the same physical source through similarity analysis is as follows: Calculate the cosine similarity between the pulse feature set F_pre and the pulse feature set F_post with respect to the average feature vector. If the cosine similarity meets the threshold requirement, it means that they belong to the same physical source before and after the mutation.

8. The method for early warning of partial discharge faults in substations according to claim 1, characterized in that: In step four, independent component analysis is performed on the pulse waveform data window before the mutation, and the process of decomposing the mixed pulse into multiple independent source components is as follows: All pulse waveforms are detected from the pre-mutation data window W_pre. Each pulse is truncated to a fixed length of multiple sampling points, and each pulse is represented as a multi-dimensional vector. All pulses are arranged in rows to form an observation matrix X. Using the FastICA algorithm, assuming that the waveforms of each pulse are statistically independent, X is decomposed into K independent source components. Each component corresponds to a hypothetical physical discharge point, and the waveform template of each component is output.

9. A method for early warning of partial discharge faults in substations according to claim 8, characterized in that: The process of matching the residual pulse after mutation with each component one by one to determine the component source and outputting a decoupling status report of the component source is as follows: Take the data window W_post after the mutation, detect all pulses from it, retain the complete waveform of each pulse, and mark it as the residual pulse set R; For each decomposed independent source component, extract the waveform template, calculate the waveform similarity between the waveform template and each pulse in the residual pulse set R, count the proportion of pulses in the residual pulse set R whose waveform similarity with the independent source component meets the threshold requirement, and obtain the matching rate. Independent source components with a matching rate greater than the first preset matching rate are marked as persistent sources, independent source components with a matching rate less than the second preset matching rate are marked as disappeared sources, and independent source components with a matching rate between the second preset matching rate and the first preset matching rate are marked as uncertain sources.

10. A substation partial discharge fault early warning system, characterized in that: Includes the following modules: Temporal evolution analysis module: Based on the phase amplitude data of each continuously acquired pulse, a two-dimensional distribution map is constructed, and a temporal evolution map of pulse characteristic distribution is formed through color coding; The mutation detection and data extraction module: Based on the time-series evolution diagram, it performs phase distribution analysis periodically on a power frequency basis to determine whether a structural mutation has occurred. If it has occurred, it records the pulse waveform data window before and after the structural mutation. Homology discrimination module: Based on the pulse waveform data window before and after the structural mutation, pulse features are extracted to construct waveform morphology feature vectors, and similarity analysis is used to determine whether they belong to the same physical source before and after the mutation; Source separation and decoupling module: If so, perform independent component analysis on the pulse waveform data window before the mutation, decompose the mixed pulse into multiple independent source components, and then match the residual pulse after the mutation with each component one by one to realize the component source determination, and output the component source decoupling status report.

Images