A high-voltage electrical equipment insulation detection method and system based on a fusion algorithm

Abstract

The application discloses a kind of high-voltage electrical equipment insulation detection method and system based on fusion algorithm, and relates to electrical insulation detection technical field, comprising: obtaining the multi-source detection data generated in the process of high-voltage electrical equipment inspection, and based on the time stamp of each detection channel in multi-source detection data corresponding time alignment is executed, to obtain multi-channel synchronous detection data, multi-source detection data includes electromagnetic detection data, acoustic detection data and thermal imaging detection data;According to multi-channel synchronous detection data, event level segmentation processing is executed, obtains multiple detection events, and for each detection event, extract time range characteristic value, energy distribution characteristic value and morphological structure characteristic value, generate detection event characteristic data;The application realizes fusion diagnosis misjudgment inhibition and reliable discrimination to high-voltage electrical equipment insulation defect under complex working condition by constructing space-time consistent synchronous detection data and quantifying evaluation cross-channel feature homology.

Description

Technical Field

[0001] This invention relates to the field of electrical insulation testing technology, and specifically to a method for insulation testing of high-voltage electrical equipment based on a fusion algorithm. Background Technology

[0002] As the core hub of the power energy transmission and distribution system, the integrity of the insulation structure of high-voltage electrical equipment directly determines the safety and reliability of the power grid operation. Partial discharge, as an important early indicator of insulation performance degradation, has become the mainstream method for comprehensively capturing potential internal defects and assessing the health status of equipment by deploying multiple sensing devices such as ultra-high frequency electromagnetic, ultrasonic, and infrared thermal imaging. This multi-source collaborative monitoring mode utilizes the complementary advantages of different physical field information to reflect the insulation status from multiple dimensions such as electromagnetic radiation intensity, mechanical vibration characteristics, and thermal effect distribution. Compared with single monitoring methods, it significantly improves the sensitivity and coverage of fault detection.

[0003] Currently, in the multi-dimensional state perception process of high-voltage electrical equipment, due to the inherent hardware differences of different modal sensors such as electromagnetic, acoustic, and thermal imaging in terms of physical response speed, signal transmission cable length, and data acquisition card startup delay, the data acquired by each detection channel often have asynchronous time deviations in terms of time reference. Moreover, substation sites are usually filled with complex background noise and random pulse interference. These interference signals are easily confused with weak partial discharge signals, making it difficult to accurately capture the start and end boundaries of effective signals by relying solely on amplitude determination in a single dimension. Consequently, it is impossible to construct synchronous detection data with strict spatiotemporal correspondence, which seriously affects the accurate extraction of subsequent analytical features.

[0004] Secondly, when performing insulation defect diagnosis in complex multi-source interference environments, relying solely on temporal overlap for signal correlation often lacks in-depth consideration of the inherent logical relationships between different physical field characteristics. It is difficult to accurately determine whether multidimensional signals within the same time window truly originate from the same physical defect source. This leads to non-homogeneous random interference signals being easily misclassified as correlated events. Furthermore, without effective feature difference quantification and conflict suppression mechanisms, it is difficult to competitively screen the homology credibility of candidate event clusters. Consequently, when faced with challenging conditions such as multi-parameter concurrent window anomalies and uncertain homology, it is impossible to effectively suppress misjudgments and output reliable diagnostic results. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method for insulation detection of high-voltage electrical equipment based on a fusion algorithm, the method comprising: Acquire multi-source detection data generated during the inspection of high-voltage electrical equipment, and perform time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data, and thermal imaging detection data. Based on the multi-channel synchronous detection data, event-level segmentation processing is performed to obtain multiple detection events. For each detection event, time range feature values, energy distribution feature values, and morphological structure feature values ​​are extracted to generate detection event feature data. Based on the event feature data and multi-channel synchronous detection data, cross-detection channel event association data is constructed, and multiple candidate homogeneous event clusters are calculated based on the event association data. For each candidate homogeneous event cluster, a homogeneity confidence value is calculated. Based on the candidate homologous event clusters, the corresponding homologous confidence values, and event association data, intra-cluster conflict suppression fusion is performed to obtain candidate fusion diagnostic results. Based on the candidate fusion diagnostic results and homologous confidence values, competitive screening among candidate homologous event clusters is performed to generate target fusion diagnostic results.

[0006] Furthermore, the steps for performing time alignment are as follows: Based on the timestamps in the electromagnetic detection data, acoustic detection data, and thermal imaging detection data, the difference between the timestamp of the electromagnetic detection data and the timestamp of the acoustic detection data is calculated to obtain a first synchronization offset value, and the difference between the timestamp of the electromagnetic detection data and the timestamp of the thermal imaging detection data is calculated to obtain a second synchronization offset value. Based on the first synchronization offset value, the timestamps of the acoustic detection data are aligned and shifted to obtain aligned acoustic detection data; Based on the second synchronization offset value, the timestamps of the thermal imaging detection data are aligned and shifted, and multi-channel synchronous detection data is generated based on the electromagnetic detection data, the aligned acoustic detection data, and the aligned thermal imaging detection data.

[0007] Furthermore, the steps for performing event-level segmentation processing are as follows: Based on the signal amplitude change values ​​corresponding to each detection channel in the multi-channel synchronous detection data, amplitude abrupt change segments are identified, and each amplitude abrupt change segment is used as a candidate boundary value for the detection event. Based on the candidate boundary values ​​of the detected events, the signal waveforms before and after the candidate boundary values ​​are subjected to continuity recognition processing to eliminate false detection event boundaries caused by background noise, thereby obtaining effective detection event boundary values. Based on the effective detection event boundary values, the start and end points of each detection event are determined, and time range feature values, energy distribution feature values, and morphological structure feature values ​​are generated based on the signal segments within the detection event to form the detection event feature data.

[0008] Furthermore, the logic for identifying amplitude abrupt change segments is as follows: Based on the signal segments corresponding to the amplitude change values ​​in the multi-channel synchronous detection data, the amplitude gradient value of each signal segment is calculated to form the amplitude gradient distribution value; Based on the amplitude gradient distribution value, identify the local maxima in the amplitude gradient distribution value, and take the position corresponding to each local maxima as a candidate point for amplitude abrupt change. Based on the corresponding positions between amplitude mutation candidate points and amplitude change values, amplitude change values ​​that do not meet the mutation requirements are filtered out to obtain amplitude mutation segments, and the obtained amplitude mutation segments are used as candidate boundary values ​​for detection events.

[0009] Furthermore, the steps for constructing event correlation data across detection channels are as follows: Based on the time range feature values ​​in the detected event feature data and the time information in the multi-channel synchronous detection data, detect event pairs with time overlap are identified to form event time matching pairs; Based on the event time matching pairs, extract the energy distribution feature value and morphological structure feature value of the corresponding detected events, and calculate the difference between the energy distribution feature value and the difference between the morphological structure feature value to generate cross-channel feature difference; Based on cross-channel feature differences, event time matching pairs that do not meet the requirements of cross-detection channels are filtered out, and the filtered event time matching pairs are combined with the corresponding cross-channel feature differences to generate cross-detection channel event association data.

[0010] Furthermore, the steps for generating cross-channel feature differences are as follows: Based on event time matching pairs, the energy distribution feature value corresponding to each detected event is extracted to form a set of energy distribution feature values; Based on the set of energy distribution feature values, calculate the difference in energy distribution feature values ​​between any two detection events with a time matching relationship to generate an energy difference. Based on event-time matching pairs, the morphological and structural feature values ​​corresponding to each detection event are extracted, and the difference between the morphological and structural feature values ​​is calculated. The energy difference and the difference between the morphological and structural feature values ​​are combined to generate a cross-channel feature difference.

[0011] Furthermore, the logic for generating the energy difference is as follows: Based on the set of energy distribution feature values, extract the energy distribution feature values ​​corresponding to the detection events that have time matching relationships to form energy matching feature value pairs; Based on the energy matching feature pairs, calculate the difference between each energy distribution feature value in the energy matching feature pair to form an energy difference set; Based on the set of energy differences, energy differences that do not meet the requirements for generating energy differences are filtered out, and the remaining energy differences are used as energy differences.

[0012] Furthermore, the steps to obtain candidate fusion diagnostic results are as follows: Based on the event association data of each detected event in the candidate homogeneous event cluster, the cross-channel feature difference corresponding to each detected event is extracted to form a cluster feature difference set. Based on the set of feature differences within the cluster, the feature difference distance between detected events is calculated, and detected events with feature difference distances greater than a predetermined threshold are filtered out to obtain a set of consistent events within the cluster. Based on the set of consistent events within the cluster, the fusion feature value of each detected event is calculated, and the fusion feature value is used as the candidate fusion diagnostic result.

[0013] Furthermore, the steps for performing competitive screening among candidate homogeneous event clusters are as follows: Based on the candidate fusion diagnostic results, the fusion feature value corresponding to each candidate homogeneous event cluster is extracted, and the fusion feature value is combined with the corresponding homogeneity confidence value to form a set of competing feature values. Based on the set of competing feature values, calculate the competition score between each candidate homologous event cluster, and remove candidate homologous event clusters whose competition score is less than the predetermined screening criteria to form a competitive retention cluster; Based on the fusion feature values ​​and homology confidence values ​​corresponding to the competing retention clusters, the final fusion feature values ​​of each competing retention cluster are calculated, and the final fusion feature values ​​are used as the target fusion diagnostic results.

[0014] A high-voltage electrical equipment insulation detection system based on a fusion algorithm, used to execute the high-voltage electrical equipment insulation detection method based on the fusion algorithm described above, the system comprising: Data Alignment Module: Acquires multi-source detection data generated during the inspection of high-voltage electrical equipment, and performs time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data, and thermal imaging detection data. Feature extraction module: Based on multi-channel synchronous detection data, it performs event-level segmentation processing to obtain multiple detection events, and extracts time range feature values, energy distribution feature values ​​and morphological structure feature values ​​for each detection event to generate detection event feature data; Event Association Module: Based on the detected event feature data and multi-channel synchronous detection data, it constructs cross-detection channel event association data, calculates multiple candidate homologous event clusters based on the event association data, and calculates a homologous confidence value for each candidate homologous event cluster; Fusion Diagnosis Module: Based on candidate homogeneous event clusters, corresponding homogeneity confidence values, and event association data, it performs intra-cluster conflict suppression fusion to obtain candidate fusion diagnosis results. Based on the candidate fusion diagnosis results and homogeneity confidence values, it performs competitive screening among candidate homogeneous event clusters to generate target fusion diagnosis results.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively eliminates asynchronous errors caused by sensor response differences and hardware startup delays by performing time-stamp-based alignment and translation processing on the acquired electromagnetic detection data, acoustic detection data, and thermal imaging detection data. It constructs multi-channel synchronous detection data that strictly corresponds on the absolute time axis and accurately identifies the boundary values ​​of valid detection events based on the signal amplitude gradient distribution. While eliminating false detection event boundaries caused by background noise, it comprehensively extracts the characteristic data of detection events, including time range, energy distribution, and morphological structure, thus providing a data foundation with spatiotemporal consistency for the deep fusion analysis of multidimensional physical information. Furthermore, this invention generates cross-channel feature differences by calculating the difference between energy distribution feature values ​​and the difference between morphological structure feature values, thereby quantitatively evaluating the physical homology between signals from different detection channels. This leads to the construction of cross-detection channel event correlation data and the generation of candidate homologous event clusters. Based on this, intra-cluster conflict suppression and inter-cluster competition screening are performed. Interference events with excessively high dispersion are eliminated using the intra-cluster feature difference set to obtain a consistent intra-cluster event set. At the same time, the competition retention cluster is locked based on the homology confidence value and competition score value. Finally, the target fusion diagnosis result is output, realizing accurate identification and misjudgment suppression of insulation defects under multi-source interference and multi-parameter anomaly conditions. In summary, this invention solves the problems of multi-parameter window anomalies and uncertain physical homology caused by multi-source interference. By constructing spatiotemporally consistent synchronous detection data and quantitatively evaluating the homology of cross-channel features, it achieves the suppression of misjudgments and reliable identification of insulation defects in high-voltage electrical equipment under complex operating conditions. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a high-voltage electrical equipment insulation detection method based on a fusion algorithm provided in Embodiment 1 of the present invention; Figure 2This is a block diagram of a high-voltage electrical equipment insulation detection method based on a fusion algorithm provided in Embodiment 2 of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0019] Please see Figure 1 As shown, this embodiment discloses a method for insulation detection of high-voltage electrical equipment based on a fusion algorithm. The method includes: S11: Acquire multi-source detection data generated during the inspection of high-voltage electrical equipment, and perform time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data, and thermal imaging detection data. Specifically, the steps for performing time alignment are as follows: S111: Based on the timestamps of the electromagnetic detection data, acoustic detection data, and thermal imaging detection data, calculate the difference between the timestamp of the electromagnetic detection data and the timestamp of the acoustic detection data to obtain a first synchronization offset value, and calculate the difference between the timestamp of the electromagnetic detection data and the timestamp of the thermal imaging detection data to obtain a second synchronization offset value. It should be noted that the high-voltage electrical equipment in this embodiment is specifically a gas-insulated fully enclosed switchgear (GIS) or a large power transformer. For testing purposes, ultra-high frequency electromagnetic sensors, ultrasonic sensors, and infrared thermal imagers are installed at key monitoring points of the equipment (such as basin insulators, transformer tank walls, and cable joints).

[0020] It should be further noted that although the same main control unit sends trigger commands to all the aforementioned sensors when collecting data, physical differences in the response speed of different types of sensors, the length of signal transmission cables, and the startup delay of the acquisition card lead to asynchronous errors in the start time of the data from each channel during actual recording. Therefore, the collected multi-source detection data specifically includes the following: Electromagnetic detection data: records the data sequence of electromagnetic wave signal amplitude changes over time as excited by partial discharge, and its header file contains the start time of electromagnetic recording with nanosecond precision. Acoustic test data: records the data sequence of the amplitude of the mechanical vibration signal caused by partial discharge changing over time, and its data header file contains the start time of the acoustic recording; Thermal imaging detection data: A sequence of thermal images recording the temperature distribution on the surface of the device, with the header file containing the start time of the thermal imaging recording.

[0021] In practical implementation, to eliminate the time error caused by hardware response delay, the electromagnetic detection data with the fastest propagation speed and the most sensitive response is used as the reference time axis. The time deviation between each channel is calculated by numerical subtraction. The specific calculation logic is as follows: First, the recording start time of electromagnetic detection data and the recording start time of acoustic detection data are read, and a subtraction operation is performed. The difference between the recording start time of electromagnetic data and the recording start time of acoustic data is the first synchronization offset value, which reflects the time lag or lead of acoustic signal relative to electromagnetic signal. Secondly, the recording start time of the electromagnetic detection data and the recording start time of the thermal imaging detection data are read, and a subtraction operation is performed. The difference between the electromagnetic recording start time and the thermal imaging recording start time is the second synchronization offset value, which reflects the time lag or lead of the thermal imaging signal relative to the electromagnetic signal.

[0022] S112: Based on the first synchronization offset value, perform alignment and translation processing on the timestamp of the acoustic detection data to obtain aligned acoustic detection data; In practice, an additive correction operation is performed on the timestamp value corresponding to each sampling point in the acoustic detection data. Specifically, the calculated first synchronization offset value is added to the original timestamp value of each data point in the acoustic detection data sequence. Through this translation process, the corrected acoustic detection data is aligned with the electromagnetic detection data in absolute position on the time axis, thereby eliminating the time misalignment caused by hardware startup delay and obtaining aligned acoustic detection data.

[0023] S113: Based on the second synchronization offset value, perform alignment and translation processing on the timestamp of the thermal imaging detection data, and generate multi-channel synchronous detection data based on the electromagnetic detection data, the aligned acoustic detection data, and the aligned thermal imaging detection data. Specifically, the step of generating multi-channel synchronous detection data includes: S113.1: Perform time axis translation correction on thermal imaging data; Using the same processing logic as acoustic data, the calculated second synchronization offset value is accumulated one by one to the time stamp value corresponding to each frame of thermal image in thermal imaging detection data, thereby normalizing the time axis of thermal imaging data relative to the time axis of electromagnetic data and obtaining aligned thermal imaging detection data. S113.2: Data reconstruction and encapsulation based on a unified time window; It should be noted that due to the huge differences in the sampling frequencies of electromagnetic data, acoustic data and thermal imaging data, for example, electromagnetic data is at the gigahertz level while thermal imaging is at the tens of hertz level, direct point-to-point stitching is not possible. Therefore, this embodiment uses a "unified time window" approach to generate synchronous detection data.

[0024] The specific implementation logic is as follows: First, a fixed time window length is set, which is based on the typical duration of a partial discharge event, for example, set as an integer multiple of the power frequency cycle; Secondly, using the time axis of electromagnetic detection data as an index reference, the continuous detection time axis is divided into multiple continuous, non-overlapping time window units. Subsequently, for each time window unit, the following data extraction operations are performed: In the electromagnetic detection data, extract all electromagnetic signal sampling points whose time markers fall within the time window range; In the aligned acoustic detection data, extract all acoustic signal sampling points whose corrected timestamps fall within the time window range; In the aligned thermal imaging detection data, extract all thermal image frames whose corrected timestamps fall within the time window range; Finally, the electromagnetic signal segments, acoustic signal segments, and thermal imaging frames captured within the same time window are combined and encapsulated into an independent multi-channel synchronous detection data packet.

[0025] Through the above steps, the output multi-channel synchronous detection data is represented as a series of data packets arranged in chronological order, and the physical field information contained in each data packet corresponds accurately in absolute time.

[0026] S12: Based on the multi-channel synchronous detection data, perform event-level segmentation processing to obtain multiple detection events, and for each detection event, extract time range feature values, energy distribution feature values, and morphological structure feature values ​​to generate detection event feature data; Specifically, the steps for performing event-level segmentation processing are as follows: S121: Based on the signal amplitude change values ​​corresponding to each detection channel in the multi-channel synchronous detection data, identify amplitude abrupt change segments and use each amplitude abrupt change segment as a candidate boundary value for the detection event; It should be noted that the partial discharge signal of high-voltage electrical equipment usually manifests as a pulse waveform that suddenly appears in a stable background noise. Therefore, identifying the moment when the signal amplitude changes drastically is the basis for signal segmentation; In practice, the following logic is executed independently for each channel (electromagnetic, acoustic, or thermal imaging channel) in the multi-channel synchronous detection data to identify amplitude abrupt change segments; Specifically, the logic for identifying amplitude abrupt change segments is as follows: S121.1: Based on the signal segments corresponding to the amplitude change values ​​in the multi-channel synchronous detection data, calculate the amplitude gradient value of each signal segment to form an amplitude gradient distribution value; In practical implementation, in order to accurately capture the instantaneous change of the signal, it is necessary to calculate the rate of change of the signal amplitude; The specific calculation logic is as follows: For a continuous signal sequence within the detection channel, the amplitude value of the current sampling point and the amplitude value of the previous sampling point are selected sequentially; Perform a subtraction operation to calculate the difference between the amplitude value of the current sampling point and the amplitude value of the previous sampling point; Take the absolute value of the difference and use it as the magnitude gradient value at the current time. Repeat the above process for each data point in the signal sequence to generate an amplitude gradient distribution curve that reflects the degree of drastic change in the signal, i.e., the amplitude gradient distribution value. S121.2: Based on the amplitude gradient distribution value, identify the local maxima in the amplitude gradient distribution value, and take the position corresponding to each local maxima as a candidate point for amplitude abrupt change; In practice, the peak position in the amplitude gradient distribution value usually corresponds to the rising or falling edge of the original signal; The specific identification logic is as follows: a sliding window method is used to scan the amplitude gradient distribution value sequence; Within each sliding window, the gradient value of the center point is compared with the gradient values ​​of its left and right adjacent points. If the gradient value of the center point is greater than the gradient values ​​of both its left and right adjacent points, then the point is determined to be a gradient peak. Record the index of the gradient peak on the time axis and mark it as a candidate point for amplitude abrupt change.

[0027] S121.3: Based on the corresponding positions between amplitude mutation candidate points and amplitude change values, amplitude change values ​​that do not meet the mutation requirements are filtered out to obtain amplitude mutation segments, and the obtained amplitude mutation segments are used as candidate boundary values ​​for detection events.

[0028] It should be noted that random fluctuations in background noise may also produce small gradient peaks, which require secondary screening using amplitude thresholds.

[0029] The specific filtering logic is as follows: Go back to the original signal amplitude sequence and find the original signal amplitude value corresponding to each candidate point of amplitude change; An amplitude reference threshold is set, which can be determined based on the noise floor level of the signal in the current time period, for example, by taking a specific multiple of the average amplitude of the background noise. Compare the original signal amplitude value corresponding to the candidate point with the amplitude reference threshold; If the amplitude value of the original signal is less than the amplitude reference threshold, the candidate point is considered to be a noise fluctuation and is removed. If the original signal amplitude value is greater than or equal to the amplitude reference threshold, the candidate point is retained, and the small time interval in which it is located is marked as an amplitude change segment, which serves as the potential start or end position of the detection event, i.e., the candidate boundary value.

[0030] S122: Based on the candidate boundary values ​​of the detected event, perform continuity recognition processing on the signal waveforms before and after the candidate boundary values ​​to eliminate false detection event boundaries generated by background noise, thereby obtaining effective detection event boundary values; It should be noted that real partial discharge signals have a certain duration and energy decay process, while random pulse interference is often extremely short-lived; Therefore, by checking the continuity of the waveform, spurious signals can be effectively eliminated; The specific implementation logic is as follows: For candidate boundary values ​​that appear in pairs, one is used as the starting candidate and the other as the ending candidate, the signal waveform segment that can be covered by them is extracted. Examine the duration for which the signal amplitude remains above a specific level within the signal waveform segment; The duration is compared with an empirical value for the preset minimum event width; At the same time, check whether the waveform segment has zero-crossing oscillation characteristics (i.e., whether the signal amplitude flips between positive and negative polarities multiple times). If the duration is less than the minimum event width, or if the expected oscillation characteristics (for electromagnetic and acoustic signals) are not detected within the duration, the boundary is determined to be caused by interference and is removed. Conversely, the boundaries that satisfy the requirements of continuity and oscillation characteristics are retained and determined as the effective detection event boundary values.

[0031] S123: Based on the effective detection event boundary values, determine the start and end points of each detection event, and generate time range feature values, energy distribution feature values, and morphological structure feature values ​​based on the signal segments within the detection event to form detection event feature data; In practical implementation, once the effective detection event boundary values ​​are determined, the continuous data stream can be segmented into independent detection event segments, and the time range feature values, energy distribution feature values, and morphological structure feature values ​​can be calculated separately. The calculation logic for the time range feature value is as follows: Read the end time and start time determined from the boundary values ​​of valid detected events; Perform a subtraction operation to calculate the time difference between the end time and the start time, and use it as the event duration; at the same time, record the absolute time position of the start time in the entire inspection cycle, as the event occurrence phase; The duration of an event is combined with the phase in which the event occurs to form a time range characteristic value.

[0032] The calculation logic for the energy distribution characteristic value is as follows: For all signal amplitude sampling points within the detection event, calculate the square of each amplitude value; The energy integral value is obtained by summing the squares of all amplitude values; in another preferred embodiment, the arithmetic mean or root mean square value of all amplitude values ​​within the event segment is calculated. The above energy integral value or root mean square value is used as the energy distribution characteristic value characterizing the intensity of this partial discharge.

[0033] Due to the different types of insulation defects, such as tip discharge, floating discharge, and air gap discharge, the waveform shapes produced are significantly different; The specific logic for extracting morphological and structural feature values ​​is as follows: Rise time ratio: The ratio of the time required for a signal to rise from its starting point to its maximum peak value to the total duration of the event; Waveform steepness: Calculates the maximum rate of change of the amplitude during the rising edge phase of the signal; Oscillation count: The total number of times the waveform crosses the zero-level axis during its duration; Combining the above data on rise time ratio, waveform steepness, and oscillation count forms a morphological structural feature value describing the waveform's geometry.

[0034] S13: Based on the event feature data and multi-channel synchronous detection data, construct cross-detection channel event association data, and calculate multiple candidate homologous event clusters based on the event association data, and calculate the homologous confidence value for each candidate homologous event cluster; Specifically, the steps for constructing event correlation data across detection channels are as follows: S131: Based on the time range feature values ​​in the detected event feature data and the time information in the multi-channel synchronous detection data, identify detected event pairs with time overlap to form event time matching pairs; It should be noted that when partial discharge occurs, electromagnetic waves, sound waves and thermal effects are physically excited by the same source. Although they have different propagation speeds, after the aforementioned time alignment correction, they should show overlap or high proximity on the time axis. In practice, the logic for identifying overlapping execution times is as follows: First, the start and end times of each detected event are extracted from the event feature data; For detection events between different detection channels, such as detection events between electromagnetic channels and acoustic channels, their time spans are compared one by one; Determine whether the time period of a detection event overlaps with the time period of a detection event in another channel, or determine whether the time interval between the center time points of two detection events is less than a preset physical association time window, which is determined based on the device size and the maximum propagation delay. If two detected events overlap in time or the interval between their central time points meets the requirements, then the two detected events are determined to be synchronous in time, and they are combined into an event time matching pair and assigned a unique matching index identifier.

[0035] S132: Based on the event time matching pair, extract the energy distribution feature value and morphological structure feature value of the corresponding detected event, and calculate the difference between the energy distribution feature value and the difference between the morphological structure feature value to generate cross-channel feature difference; Specifically, the steps for generating cross-channel feature differences are as follows: S132.1: Based on event time matching pairs, extract the energy distribution feature value corresponding to each detection event to form a set of energy distribution feature values; In practice, for each event pair identified as time-matched, the energy distribution characteristic value generated in step S12 is read back.

[0036] For example, the energy amplitude of the electromagnetic event and the energy amplitude of the matching acoustic event are extracted, and this set of corresponding energy data is stored in the energy distribution feature value set as the basis data for subsequent difference calculation.

[0037] S132.2: Based on the set of energy distribution feature values, calculate the difference in energy distribution feature values ​​between any two detection events with a time matching relationship to generate an energy difference value; Specifically, the logic for generating the energy difference is as follows: S132.2.1: Based on the set of energy distribution feature values, extract the energy distribution feature values ​​corresponding to the detection events that have time matching relationships to form energy matching feature value pairs; It should be noted that, in order to ensure the rationality of the calculation, it is usually necessary to normalize the energy values ​​of different physical dimensions before forming a pair, for example, by mapping the values ​​to the range of zero to one, in order to eliminate unit differences. The normalized electromagnetic energy value is paired with the acoustic energy value or the thermal imaging energy value to form a data item, and the output is an energy matching feature value pair; S132.2.2: Based on the energy matching eigenvalue pairs, calculate the difference between each energy distribution eigenvalue in the energy matching eigenvalue pairs to form an energy difference set; In practice, numerical subtraction is used to calculate the difference: Subtract the second energy value from the first energy value in the energy matching feature pair and take the absolute value of the result; or, calculate the ratio between the two energy values ​​and use the degree of deviation between this ratio and its ideal ratio, which is usually 1, as the difference. Perform the above calculation on all matching pairs, and summarize all the difference results to form an energy difference set; S132.2.3: Based on the set of energy differences, filter out energy differences that do not meet the requirements for generating energy differences, and use the remaining energy differences as energy differences.

[0038] It should be noted that although signals from the same physical source may behave differently on different sensors, their energy levels are usually positively correlated; that is, strong electromagnetic signals are usually accompanied by strong acoustic signals. If the energy difference is too large, it indicates that the two may only be a coincidence in time rather than having the same physical origin; The specific screening logic is as follows: Set an energy consistency threshold range; Check whether each difference in the set of energy differences falls within the threshold range; If an energy difference exceeds the threshold range, for example, if the electromagnetic signal is extremely strong while the acoustic signal is extremely weak, resulting in a normalized difference close to 1, then the matching relationship is determined to be a false match, and the corresponding energy difference is removed from the set; the values ​​that fall within the threshold range are retained and determined as the final energy difference. S132.3: Based on event time matching pairs, extract the morphological structure feature value corresponding to each detection event, calculate the difference of morphological structure feature value, and combine the energy difference with the morphological structure feature value difference to generate cross-channel feature difference; In practice, after calculating the difference in energy dimension, the difference in waveform morphology dimension is further calculated.

[0039] The specific logic is as follows: extract the morphological and structural feature values ​​of each event in the matched event pair, such as waveform steepness and rise time ratio; Calculate the Euclidean distance or Manhattan distance between the morphological feature vectors of two events, and use the calculated distance value as the difference in morphological feature values. Finally, the energy difference obtained from the previous steps is weighted and summed or vectorized with the difference in morphological and structural feature values ​​calculated here to generate an index that comprehensively reflects the degree of difference in the physical characteristics of the two events, namely the cross-channel feature difference; the weight ratio needs to be set according to the actual situation.

[0040] S133: Based on the cross-channel feature difference, filter out event time matching pairs that do not meet the requirements of cross-detection channel correspondence, and combine the filtered event time matching pairs with the corresponding cross-channel feature difference to generate cross-detection channel event association data; In practice, the final correlation determination is based on the calculated comprehensive difference index: Set a comprehensive correlation threshold, and compare the cross-channel feature difference corresponding to each event with the comprehensive correlation threshold; If the difference in cross-channel characteristics is greater than the comprehensive correlation threshold, it means that although the two events overlap in time, their physical characteristics are very different and they do not belong to the same source discharge. Therefore, their matching relationship is terminated. If the cross-channel feature difference is less than or equal to the comprehensive correlation threshold, then the two events are confirmed to be related events. The confirmed related event ID, its corresponding time information, and the calculated cross-channel feature difference are packaged and stored to generate cross-detection channel event correlation data.

[0041] It should be noted that the logic for generating the comprehensive correlation threshold is as follows: Collect a large amount of historical inspection data of high-voltage electrical equipment, and based on manual experience or disassembly and maintenance results, mark the multi-channel signal pairs in the historical data as "same source positive samples" (confirmed as the same discharge event) and "non-same source negative samples" (confirmed as interference signals from different sources). Secondly, calculate the cross-channel feature difference for all homologous positive samples and non-homologous negative samples respectively, and statistically analyze the distribution probability of the feature difference between these two types of samples in the numerical range. Among them, the difference distribution of homologous positive samples is concentrated in the low numerical range, while the difference distribution of non-homologous negative samples is concentrated in the high numerical range. Finally, select the boundary value between the distribution curves of homologous positive samples and the distribution curves of non-homologous negative samples, or select the upper limit value of the feature difference that can cover the vast majority (e.g., 99%) of homologous positive samples, and determine this value as the comprehensive correlation threshold.

[0042] S14: Based on the candidate homologous event clusters, the corresponding homologous confidence values, and event association data, perform intra-cluster conflict suppression fusion to obtain candidate fusion diagnostic results. Based on the candidate fusion diagnostic results and homologous confidence values, perform competitive screening among candidate homologous event clusters to generate target fusion diagnostic results.

[0043] Specifically, the steps to obtain candidate fusion diagnostic results are as follows: S141: Based on the event association data of each detected event in the candidate homogeneous event cluster, extract the cross-channel feature difference corresponding to each detected event to form a cluster feature difference set; In specific implementation, firstly, for each candidate homogeneous event cluster generated in step S13, all detection events contained in the cluster are traversed (for example, the cluster consists of an electromagnetic pulse event and an acoustic signal event). Subsequently, based on the unique identifier of each detection event, the corresponding event association data is indexed, and the cross-channel feature difference calculated when the detection event is matched with other channel events is read from it; All cross-channel feature differences extracted from the same candidate homogeneous event cluster are summarized to form the intra-cluster feature difference set corresponding to the cluster. This set reflects the degree of dispersion of each physical signal in the cluster in terms of attributes. S142: Based on the feature difference set within the cluster, calculate the feature difference distance between detected events and filter out detected events whose feature difference distance is greater than a predetermined threshold to obtain a consistent event set within the cluster; It should be noted that, due to the possibility of multi-source interference on site (such as the accidental overlap of a real partial discharge signal and a randomly occurring noise signal), pseudo signals from different sources may be mixed into the cluster. In practice, the logic for performing intra-cluster conflict suppression is as follows: Based on the set of feature differences within a cluster, calculate the degree of deviation of each value in the set from the set mean, or calculate the absolute value of the difference between any two values ​​in the set, and define the degree of deviation or the absolute value of the difference as the feature difference distance value.

[0044] Set a predetermined distance tolerance threshold (this threshold is determined by experimental data based on the statistical distribution variance of a large number of historical signals from the same source).

[0045] Compare the feature difference distance value corresponding to each detected event with the distance tolerance threshold; If the feature difference distance value of a certain detected event is greater than the distance tolerance threshold, it means that although the event is classified into the cluster in time, its physical characteristics are significantly conflicted with other events. It is judged as an abnormal interference item and removed from the cluster. Detection events whose feature difference distance value is less than or equal to the distance tolerance threshold are retained to form a cluster of consistent events, thereby ensuring that the signals participating in subsequent fusion have a high degree of physical consistency. S143: Based on the set of consistent events within the cluster, calculate the fusion feature value of each detected event, and use the fusion feature value as the candidate fusion diagnostic result; In practice, feature-level fusion calculations are performed on the cleaned set of intra-cluster consistent events. The specific calculation logic is as follows: extract the original feature data (including energy amplitude, time phase, waveform steepness, etc.) of each detected event in the cluster-consistent event set.

[0046] A weighted average algorithm is used to calculate the fused feature values, where the weights are determined based on the reliability of the sensors in each detection channel (for example, in environments with strong electromagnetic interference, the weights of electromagnetic channel features are appropriately reduced while the weights of acoustic channel features are increased). Alternatively, the maximum likelihood estimation method can be used to calculate the feature values ​​that best represent the true state of the physical event.

[0047] The calculated fusion feature values ​​(such as fusion energy value and fusion phase value) are packaged and used as the candidate fusion diagnosis result corresponding to the candidate homogeneous event cluster; Specifically, the steps for performing competitive screening among candidate homogeneous event clusters are as follows: S144: Based on the candidate fusion diagnosis results, extract the fusion feature value corresponding to each candidate homogeneous event cluster, and combine the fusion feature value with the corresponding homogeneity confidence value to form a set of competing feature values; In practice, when multiple candidate clusters of events from the same source are identified within the same monitoring period (for example, multiple suspected defect sources exist at the same time, or the same defect is misidentified as multiple clusters), data needs to be prepared for competitive screening. The specific operation is as follows: traverse each candidate homogeneous event cluster and extract its fusion feature value calculated in step S143; At the same time, the same source confidence value calculated in step S13 for the cluster is obtained (this value reflects the probability that events within the cluster belong to the same physical source). The fused feature values ​​are paired with the homology confidence values ​​to form a set of competing feature values ​​that describe the overall credibility and feature strength of the cluster.

[0048] S145: Based on the set of competing feature values, calculate the competition score between each candidate homologous event cluster, and filter out candidate homologous event clusters whose competition score is less than the predetermined screening criteria to form a competitive retention cluster; It should be noted that in order to eliminate spurious homogeneous clusters (such as clusters formed by two concurrent random noise coincidences), a competition mechanism needs to be introduced; In practice, the logic for calculating the competition score is as follows: the source confidence value is used as the base coefficient, and the energy amplitude or waveform integrity in the fusion feature value is used as the gain coefficient. Perform multiplication or weighted summation to calculate a numerical value that comprehensively reflects the probability that the cluster is both homologous and a real defect, i.e., the competition score. A predetermined scoring and filtering threshold is set, which is determined based on the system's tolerance for false alarm rate and through experimental data; The competitive score of each candidate homologous event cluster is compared with the scoring screening threshold. If the competitive score is less than the scoring screening threshold, the cluster is deemed to have insufficient credibility and is removed as a pseudo-homologous cluster. Candidate homologous event clusters with competitive scores greater than or equal to the scoring screening threshold are retained and marked as competitive retention clusters.

[0049] S146: Based on the fusion feature value and homology confidence value corresponding to the competing retention clusters, calculate the final fusion feature value of each competing retention cluster, and use the final fusion feature value as the target fusion diagnostic result; In practice, the final diagnostic conclusion is output for the remaining competitive clusters after screening. The specific logic is as follows: If there are multiple competing retention clusters (meaning that there may be multiple insulation defects inside the device at the same time), then the final fusion feature value of each competing retention cluster is calculated separately; The calculation of the final fusion feature value includes: determining the physical type of the defect (such as tip discharge, floating discharge) based on the fusion feature value, determining the severity of the defect based on the energy amplitude, and determining the physical location of the defect based on the time alignment information; The comprehensive data, including defect type, severity, and location information, is output as the target fusion diagnostic result to the user interface or remote monitoring center to reliably identify insulation defects in high-voltage electrical equipment. Example

[0050] Please see Figure 2 As shown, based on a unified inventive concept, this embodiment discloses a high-voltage electrical equipment insulation detection system based on a fusion algorithm, comprising: Data alignment module S21: Acquires multi-source detection data generated during the inspection of high-voltage electrical equipment, and performs time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data and thermal imaging detection data. Feature extraction module S22: Based on multi-channel synchronous detection data, it performs event-level segmentation processing to obtain multiple detection events, and extracts time range feature values, energy distribution feature values ​​and morphological structure feature values ​​for each detection event to generate detection event feature data; Event association module S23: Based on the detected event feature data and multi-channel synchronous detection data, construct cross-detection channel event association data, calculate multiple candidate homologous event clusters based on the event association data, and calculate homologous confidence value for each candidate homologous event cluster; Fusion Diagnosis Module S24: Based on the candidate homogeneous event clusters, the corresponding homogeneous confidence values, and event association data, it performs intra-cluster conflict suppression fusion to obtain candidate fusion diagnosis results. Based on the candidate fusion diagnosis results and homogeneous confidence values, it performs competitive screening among candidate homogeneous event clusters to generate target fusion diagnosis results.

[0051] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for insulation detection of high-voltage electrical equipment based on a fusion algorithm, characterized in that, The method includes: Acquire multi-source detection data generated during the inspection of high-voltage electrical equipment, and perform time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data, and thermal imaging detection data. Based on the multi-channel synchronous detection data, event-level segmentation processing is performed to obtain multiple detection events. For each detection event, time range feature values, energy distribution feature values, and morphological structure feature values ​​are extracted to generate detection event feature data. Based on the event feature data and multi-channel synchronous detection data, cross-detection channel event association data is constructed, and multiple candidate homogeneous event clusters are calculated based on the event association data. For each candidate homogeneous event cluster, a homogeneity confidence value is calculated. Based on the candidate homogeneous event clusters, the corresponding homogeneity confidence values, and event association data, intra-cluster conflict suppression fusion is performed to obtain candidate fusion diagnostic results. Based on the candidate fusion diagnostic results and homogeneity confidence values, competitive screening among candidate homogeneous event clusters is performed to generate target fusion diagnostic results. The steps to obtain candidate fusion diagnostic results are as follows: Based on the event association data of each detected event in the candidate homogeneous event cluster, the cross-channel feature difference corresponding to each detected event is extracted to form a cluster feature difference set. Based on the set of feature differences within the cluster, the feature difference distance between detected events is calculated, and detected events with feature difference distances greater than a predetermined threshold are filtered out to obtain a set of consistent events within the cluster. Based on the set of consistent events within the cluster, the fusion feature value of each detected event is calculated, and the fusion feature value is used as the candidate fusion diagnostic result.

2. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 1, characterized in that, The steps for performing time alignment are as follows: Based on the timestamps in the electromagnetic detection data, acoustic detection data, and thermal imaging detection data, the difference between the timestamp of the electromagnetic detection data and the timestamp of the acoustic detection data is calculated to obtain a first synchronization offset value, and the difference between the timestamp of the electromagnetic detection data and the timestamp of the thermal imaging detection data is calculated to obtain a second synchronization offset value. Based on the first synchronization offset value, the timestamps of the acoustic detection data are aligned and shifted to obtain aligned acoustic detection data; Based on the second synchronization offset value, the timestamps of the thermal imaging detection data are aligned and shifted, and multi-channel synchronous detection data is generated based on the electromagnetic detection data, the aligned acoustic detection data, and the aligned thermal imaging detection data.

3. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 2, characterized in that, The steps for performing event-level segmentation processing are as follows: Based on the signal amplitude change values ​​corresponding to each detection channel in the multi-channel synchronous detection data, amplitude abrupt change segments are identified, and each amplitude abrupt change segment is used as a candidate boundary value for the detection event. Based on the candidate boundary values ​​of the detected events, the signal waveforms before and after the candidate boundary values ​​are subjected to continuity recognition processing to eliminate false detection event boundaries caused by background noise, thereby obtaining effective detection event boundary values. Based on the effective detection event boundary values, the start and end points of each detection event are determined, and time range feature values, energy distribution feature values, and morphological structure feature values ​​are generated based on the signal segments within the detection event to form the detection event feature data.

4. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 3, characterized in that, The logic for identifying amplitude abrupt change segments is as follows: Based on the signal segments corresponding to the amplitude change values ​​in the multi-channel synchronous detection data, the amplitude gradient value of each signal segment is calculated to form the amplitude gradient distribution value; Based on the amplitude gradient distribution value, identify the local maxima in the amplitude gradient distribution value, and take the position corresponding to each local maxima as a candidate point for amplitude abrupt change. Based on the corresponding positions between amplitude mutation candidate points and amplitude change values, amplitude change values ​​that do not meet the mutation requirements are filtered out to obtain amplitude mutation segments, and the obtained amplitude mutation segments are used as candidate boundary values ​​for detection events.

5. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 4, characterized in that, The steps to construct event correlation data across detection channels are as follows: Based on the time range feature values ​​in the detected event feature data and the time information in the multi-channel synchronous detection data, detect event pairs with time overlap are identified to form event time matching pairs; Based on the event time matching pairs, extract the energy distribution feature value and morphological structure feature value of the corresponding detected events, and calculate the difference between the energy distribution feature value and the difference between the morphological structure feature value to generate cross-channel feature difference; Based on cross-channel feature differences, event time matching pairs that do not meet the requirements of cross-detection channels are filtered out, and the filtered event time matching pairs are combined with the corresponding cross-channel feature differences to generate cross-detection channel event association data.

6. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 5, characterized in that, The steps for generating cross-channel feature differences are as follows: Based on event time matching pairs, the energy distribution feature value corresponding to each detected event is extracted to form a set of energy distribution feature values; Based on the set of energy distribution feature values, calculate the difference in energy distribution feature values ​​between any two detection events with a time matching relationship to generate an energy difference. Based on event-time matching pairs, the morphological and structural feature values ​​corresponding to each detection event are extracted, and the difference between the morphological and structural feature values ​​is calculated. The energy difference and the difference between the morphological and structural feature values ​​are combined to generate a cross-channel feature difference.

7. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 6, characterized in that, The logic for generating the energy difference is as follows: Based on the set of energy distribution feature values, extract the energy distribution feature values ​​corresponding to the detection events that have time matching relationships to form energy matching feature value pairs; Based on the energy matching feature pairs, calculate the difference between each energy distribution feature value in the energy matching feature pair to form an energy difference set; Based on the set of energy differences, energy differences that do not meet the requirements for generating energy differences are filtered out, and the remaining energy differences are used as energy differences.

8. The insulation detection method for high-voltage electrical equipment based on a fusion algorithm according to claim 7, characterized in that, The steps for performing competitive screening among candidate clusters of homogeneous events are as follows: Based on the candidate fusion diagnostic results, the fusion feature value corresponding to each candidate homogeneous event cluster is extracted, and the fusion feature value is combined with the corresponding homogeneity confidence value to form a set of competing feature values. Based on the set of competing feature values, calculate the competition score between each candidate homologous event cluster, and remove candidate homologous event clusters whose competition score is less than the predetermined screening criteria to form a competitive retention cluster; Based on the fusion feature values ​​and homology confidence values ​​corresponding to the competing retention clusters, the final fusion feature values ​​of each competing retention cluster are calculated, and the final fusion feature values ​​are used as the target fusion diagnostic results.

9. A high-voltage electrical equipment insulation detection system based on a fusion algorithm, used to execute the high-voltage electrical equipment insulation detection method based on a fusion algorithm as described in any one of claims 1-8, characterized in that, The system includes: Data Alignment Module: Acquires multi-source detection data generated during the inspection of high-voltage electrical equipment, and performs time alignment based on the timestamps corresponding to each detection channel in the multi-source detection data to obtain multi-channel synchronous detection data. The multi-source detection data includes electromagnetic detection data, acoustic detection data, and thermal imaging detection data. Feature extraction module: Based on multi-channel synchronous detection data, it performs event-level segmentation processing to obtain multiple detection events, and extracts time range feature values, energy distribution feature values ​​and morphological structure feature values ​​for each detection event to generate detection event feature data; Event Association Module: Based on the detected event feature data and multi-channel synchronous detection data, it constructs cross-detection channel event association data, calculates multiple candidate homologous event clusters based on the event association data, and calculates a homologous confidence value for each candidate homologous event cluster; Fusion Diagnosis Module: Based on candidate homogeneous event clusters, corresponding homogeneity confidence values, and event association data, it performs intra-cluster conflict suppression fusion to obtain candidate fusion diagnosis results. Based on the candidate fusion diagnosis results and homogeneity confidence values, it performs competitive screening among candidate homogeneous event clusters to generate target fusion diagnosis results.

Images