A method and system for identifying and grading the pollution flashover discharge phase of a polluted insulator

Abstract

The present application belongs to the technical field of high-voltage external insulation state monitoring and image intelligent recognition, and particularly relates to a method and system for recognizing and grading early warning of a pollution insulator pollution flashover discharge phase, wherein the method is based on a double-band ultraviolet image, can take advantage of the high sensitivity response of the early weak discharge of the solar blind ultraviolet channel and the good characterization ability of the UVA channel to the arc expansion process in the middle and later stages, and can perform stable phase division and risk grading on the pollution flashover development process through time sequence constraint weighting modeling, phase confidence joint discrimination, adaptive time sequence smoothing and risk index early warning mechanism.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage external insulation condition monitoring and intelligent image recognition technology, specifically relating to a method and system for identifying and classifying the flashover stage of polluted insulators. Background Technology

[0002] Insulators operate in complex environments such as humidity, salt spray, industrial dust, and chemical pollution for extended periods, and their surfaces are prone to accumulating contaminant layers. When the contaminant layer absorbs moisture, the surface conductivity of the insulator increases, leakage current increases, and the local electric field distribution becomes distorted, which can easily induce surface partial discharge and further develop into surface arcing, leading to flashover accidents in severe cases.

[0003] The flashover process of contaminated insulators is usually not instantaneous, but rather evolves from weak discharge and partial discharge to arc expansion and eventual flashover breakdown. Therefore, objectively identifying the stage of flashover discharge and assessing its risk level can provide a basis for differentiated operation and maintenance measures such as condition monitoring, key inspections, live-line cleaning, and planned outage maintenance.

[0004] In existing technologies, the main methods for monitoring the condition of polluted insulators include leakage current method, visible light observation method, and solar-blind ultraviolet imaging method. Although the leakage current method is widely used, it is greatly affected by environmental humidity, grounding conditions, and electromagnetic interference, and it is difficult to intuitively characterize the spatial distribution of discharge. The visible light observation method is not sensitive enough for weak discharges. Solar-blind ultraviolet imaging is more sensitive to weak signals in the early stage of discharge, but has limited characterization of the later arc expansion process.

[0005] Another type of UVA imaging can clearly reflect the development process of the electric arc. If combined with solar-blind ultraviolet imaging, it can achieve a more comprehensive capture of discharge information.

[0006] For example, the invention patent with publication number CN120652233A, entitled "Method and System for Detecting Insulation Discharge Defects Based on Solar-Blind Ultraviolet and UVA Dual Bands," provides a collaborative detection mechanism for solar-blind ultraviolet and UVA dual bands. Combined with an improved contrast-limited adaptive histogram equalization algorithm and wavelet fusion technology, it significantly enhances anti-interference capabilities in complex environments and expands the dynamic range of weak-light images by more than 40%, enabling clear identification of weak discharge morphology and effectively preserving discharge details. However, its judgment of flashover discharge remains at the level of whether discharge exists, lacking a unified criterion based on image quantification features. Therefore, it is difficult to achieve stable identification and graded early warning of the flashover evolution process. If only the detection of discharge is used as the basis for power outage maintenance, it can easily lead to over-maintenance or unnecessary power outages, failing to meet the economic and targeted requirements of engineering sites. Summary of the Invention

[0007] Purpose of the invention: This invention addresses the problems in existing technologies for monitoring flashover discharge of polluted insulators, such as incomplete single-day-blind ultraviolet imaging information, blurred stage boundaries, reliance on experience for identification, and difficulty in supporting differentiated operation and maintenance decisions. It proposes a method for identifying and classifying flashover discharge stages in polluted insulators, and also provides a system for identifying and classifying flashover discharge stages in polluted insulators.

[0008] Technical solution: On the one hand, the present invention provides a method for identifying and classifying the flashover stage of polluted insulators, the method comprising: A dual-band ultraviolet image or video frame sequence of polluted insulator discharge is simultaneously acquired using a solar-blind ultraviolet imaging channel and a UVA imaging channel to form a dual-band image sequence to be identified. The dual-band image sequence is then preprocessed to obtain a standardized dual-band six-dimensional feature vector, including: the total pixel area of ​​all discharge spot regions in the solar-blind ultraviolet image, the pixel area of ​​the single largest discharge spot region in the solar-blind ultraviolet image, the number of connected regions of independent discharge spots in the solar-blind ultraviolet image, the total pixel area of ​​all discharge spot regions in the UVA image, the pixel area of ​​the single largest discharge spot region in the UVA image, and the number of connected regions of independent discharge spots in the UVA image. Based on the standardized dual-band six-dimensional feature vector, the dual-band area response ratio, spot concentration and short-term growth of the dual-band image are calculated, which is the enhanced feature identification in the generation stage. The time-constrained weighted clustering model is trained using pollution flashover test samples. The input features of the time-constrained weighted clustering model include the stage identification enhancement features. The four continuous intervals corresponding to the normalized time axis formed after normalization according to the time position are used as the initial cluster centers. After iterative training, the physical stages corresponding to the pollution flashover test samples are obtained. The physical stages include: initial stage, partial discharge stage, arc propagation stage and flashover breakdown stage. For each frame of the sample to be tested, the comprehensive discrimination score of the four candidate stages is calculated, and the candidate stage with the largest comprehensive discrimination score is selected as the preliminary identification stage of the current frame. The preliminary identification result is constrained and judged. When the candidate stage satisfies the constraint judgment at the same time, the current candidate stage is taken as the identification result of the pollution flashover stage at the current moment. Based on the test samples, a stable phase sequence under multiple frames is determined, thereby generating a pollution flashover risk index, and a graded early warning output is performed based on the pollution flashover risk index.

[0009] Furthermore, including: The calculation of the dual-band area response ratio, spot concentration, and short-term growth of the dual-band image based on the standardized dual-band six-dimensional feature vector includes: The standardized six-dimensional feature vector is represented as: ; in, This represents the sum of pixel areas of all discharge spot regions in the standardized solar-blind ultraviolet image. The pixel area of ​​the single largest discharge spot region in the standardized solar-blind ultraviolet image. This represents the number of connected regions of independent discharge spots in the standardized solar-blind ultraviolet image. This represents the sum of the pixel areas of all discharge spot regions in the standardized UVA image. The pixel area of ​​the single largest discharge spot region in the standardized UVA image. The number of connected regions of independent discharge spots in the standardized UVA image; The dual-band area response ratio is used to reflect the enhancement degree of the UVA channel relative to the solar-blind ultraviolet channel, and is expressed as: ; in, ε To prevent constants with a denominator of zero; t Indicates the current frame number or the current time. The concentration of the light spot reflects the trend of discharge concentrating from discrete light spots to a continuous arc, including the concentration of light spots in the solar-blind ultraviolet band. and UVA band light spot concentration , respectively represented as: ; The short-term growth amount reflects the direction of change in discharge characteristics between adjacent frames, and is expressed as follows: .

[0010] Furthermore, including: The step of using four consecutive intervals corresponding to the normalized time axis formed according to time position as the initial cluster centers includes: Each group of flashover test samples is recorded as a sample sequence, with the complete frame sequence from the start of pressurization to the end of flashover. For the s There are 3 sample sequences, and the total number of frames in the sample sequence is T. s Then the first t The time position of the frame is normalized as follows: ; in, Used to characterize the relative evolution position of the frame throughout the entire flashover process, with a size range of 0 to 1. The smaller the value, the closer the frame is to the early stage of flashover development. The larger the value, the closer the frame is to the later stage of the flashover development.

[0011] Furthermore, including: The physical stages corresponding to the flashover test samples obtained after the iterative training include: Cluster center initialization: The initial cluster centers are determined by four consecutive intervals of the normalized time axis obtained by normalizing according to the time position; Iterative optimization process: During the sample allocation process, for each frame, the combined cost allocated to each cluster center is calculated, and the cluster with the smallest combined cost is selected as the new stage label for that frame, thereby minimizing the overall objective function in the iteration. Model parameter calculation: Repeatedly execute sample allocation, cluster center update, stage average normalized time position update and stage transition probability matrix update until the stopping condition is met. The stage transition probability matrix is ​​used to describe the possibility of mutual transfer between the four clusters during the pollution flashover discharge process. The stage average normalized time position is used to determine the time sequence of the four clusters. After clustering is completed, the stage feature center, stage feature radius, stage average normalized time position and stage transition probability matrix are calculated for each cluster. The stage feature center of each cluster is labeled, and the physical stage of the cluster is determined by combining the transition relationship in the stage transition probability matrix. Output Model: After completing the stage calibration, an offline temporal constrained weighted clustering model is obtained, which includes stage feature centers, feature weights, stage feature radii, stage transition probability matrices, stage average time positions, and physical stage labels.

[0012] Furthermore, including: The step of defining the stage of each cluster center and determining the physical stage of the cluster based on the transition relationships in the stage transition probability matrix includes: The clusters are sorted from early to late according to the average occurrence position of the stage feature centers on the normalized time axis, i.e. from low to high. Compare the total spot area, maximum spot area, number of spots, dual-band area response ratio, and spot concentration corresponding to the stage feature centers of each cluster; The cluster with the earliest average normalized time position and the lowest ranking in terms of total spot area of ​​the solar-blind ultraviolet channel, maximum spot area of ​​the solar-blind ultraviolet channel, total spot area of ​​the UVA channel, and maximum spot area of ​​the UVA channel among the four cluster centers, along with a UVA channel response lower than that of the solar-blind ultraviolet channel, is designated as the initial stage. The number of light spots exceeds the level corresponding to the initial stage, the total light spot area increases but the proportion of the largest light spot area does not reach a high value, the discharge is distributed in multiple discrete points and the cluster center transferred from the initial stage is identified as the partial discharge stage. The total and maximum spot areas of the UVA channel exceed the levels corresponding to the partial discharge stage, the dual-band area response ratio increases, the spot concentration increases, and the clustering center of the spot development from multi-point dispersion to local connectivity is identified as the arc expansion stage. The cluster center with the latest average normalized time, the highest total spot area and the largest spot area, the strong continuous discharge region of the UVA channel, the highest spot concentration and no subsequent stable low-stage transfer is identified as the flashover breakdown stage.

[0013] Furthermore, including: The stage average normalized time position is used to determine the temporal order of the four stages, with the earliest being the initial stage and the latest being the flashover breakdown stage. For each stage, the normalized time positions of all frames belonging to that stage are collected, and the corresponding arithmetic mean is calculated. Therefore, the stage average normalized time position is expressed as: ; in, For the first j The stage average normalized time position of each cluster For the first s The first flashover sample sequence t Normalized time position of the frame. The stage label is j The number of training sample frames, Let t be the stage label of the s-th flashover sample sequence.

[0014] Furthermore, including: The selection of the cluster with the lowest combination cost as the new stage label for the frame includes: During the sample allocation process, for the first s The first of the flashover sample sequences t Frame, the current frame is assigned to a candidate cluster. j The combined cost is expressed as: ; in, Indicates the first s In the flashover sample sequence, the th t Frames are assigned to candidate clusters j The combined cost; Weighted feature distance; Penalty for reverse transfer; Penalty for skipping stages; Penalty for deviation from normalized time position; These are the weight coefficients for the reverse transition penalty, the cross-stage jump penalty, and the normalized temporal position deviation penalty, respectively. The stage label of the previous frame is... ; After all training sample frames have been assigned stage labels, the cluster centers and stage mean normalized time positions of each cluster are updated according to the new stage labels. j The cluster centers have been updated to: ; No. j The stage-averaged normalized time position of each cluster is updated as follows: ; in, Assigned to the current number j The number of training sample frames for each cluster. For the first s The first of the flashover sample sequences t The comprehensive feature vector corresponding to the frame, and the candidate cluster number is j .

[0015] Furthermore, including: The repeated execution of sample allocation, cluster center update, stage average normalized time position update, and stage transition probability matrix update until the stopping condition is met includes: The stopping condition is that the maximum change in cluster centers in two adjacent iterations is less than a preset convergence threshold, or the number of iterations reaches a preset maximum number of iterations. The maximum change in the cluster centers is expressed as: ; in, For the first r After the nth iteration j Cluster centers, For the first r After -1 iterations, the th j Cluster centers, The maximum value of the change in the four cluster centers in two consecutive iterations, when Stop iteration when This is the preset convergence threshold.

[0016] Furthermore, including: The calculation of the comprehensive discrimination score for each frame of the sample to be tested, involving four candidate stages, includes: During online recognition, the weighted distance between the feature vector corresponding to the sample to be tested and the feature center of each stage is calculated, and the distance similarity is calculated based on the feature radius of each stage. Calculate the distance interval between the nearest stage center and the second nearest stage center to obtain the distance interval confidence. The confidence level of the stage feature radius of the candidate stage corresponding to the current frame is calculated based on the stage feature radius. By combining the transition probability from the previous stable stage to the candidate stage and the consistency between the current short-term growth rate and the typical growth direction of the candidate stage, a comprehensive discrimination score for the candidate stage is formed.

[0017] Furthermore, including: The comprehensive discrimination score for the candidate stage is expressed as: ; In the formula, Indicates the first t Time of the first j The comprehensive discrimination score for each candidate stage; Indicates distance similarity; Indicates the confidence level of the distance interval; Indicates the confidence level of the stage feature radius; Indicates the transition from the previous stable phase to the [missing information]. j The stage transition probability of each candidate stage; This indicates consistency in short-term growth trends; to Let be the weight coefficients of each discriminant factor, and let their sum be 1; During online recognition, the weighted distance between the feature vector to be tested and the feature centers at each stage is first calculated. For the first stage... j The weighted distance between each candidate stage is expressed as: ; in, Indicates the first t The comprehensive feature vector Z(t) of the sample under test at time t and the first time j Each stage of characteristic center Weighted distance between them; K The dimension of the comprehensive feature vector; Indicates the first k Dimensional feature weights; Let Z(t) be the eigenvector of the composite feature vector at time t. k 3D eigenvalues; Indicates the first j Each stage of characteristic center The k Dimensional eigenvalues. The smaller the value, the better the sample being tested is compared to the first... j The more similar the characteristics of the candidate stages; Further calculation based on weighted distance j The distance similarity of each candidate stage is expressed as: ; in, Indicates the first j Distance similarity between candidate stages; For the firstj The stage feature radius of each stage; ε is a constant to prevent the denominator from being zero. This formula converts the weighted distance into a similarity value in the range [0,1]. The smaller, The larger.

[0018] Furthermore, including: The distance interval between the nearest stage center and the second nearest stage center is calculated to obtain the distance interval confidence score, which is expressed as follows: For the j The candidate phase, first in addition to the first... j Among the remaining stages other than the candidate stages, determine the stage with the smallest weighted distance to the sample to be tested, and record this smallest distance as . Then the first j The distance interval confidence of each candidate stage is expressed as: ; in, Indicates the first j The confidence level of the distance interval between candidate stages, if If the weighted distance is less than the weighted distance corresponding to other stages, then Take a positive value; if If it is not less than the minimum weighted distance in other stages, then Setting it to 0, the confidence level of this distance interval is used to determine the test sample relative to the first... j Does each candidate stage possess a distinguishable distance advantage?

[0019] Furthermore, including: The step of calculating the confidence level of the stage feature radius corresponding to the candidate stage of the current frame based on the stage feature radius includes: Calculate the first based on the stage characteristic radius. j The confidence level of the stage feature radius of each candidate stage is expressed as: ; in, Indicates the first j The credibility of the stage feature radius of each candidate stage, if Located at the stage characteristic radius Within the range, then If it is a positive value; Exceeding the stage feature radius range, then Take 0.

[0020] Furthermore, including: The combination of the transition probability from the previous stable stage to the candidate stage and the consistency between the current short-term growth rate and the typical growth direction of the candidate stage forms a comprehensive discrimination score for the candidate stage, including: The previous stable stage is the label of the stable stage output after the most recent adaptive timing smoothing, denoted as . For the first j The stage transition probability of each candidate stage is expressed as: ; in, Indicates the transition from the previous stable phase Transfer to the j The probability of each candidate stage is provided by the stage transition probability matrix obtained from the offline modeling stage. If no previous stable stage exists at the current time, then... Set all four candidate stages to the same value so that the transition probability term does not affect the identification result of the first stage; Short-term growth trend consistency is used to determine whether the current characteristic change direction is consistent with the typical growth direction of the candidate stage. The current short-term growth amount is represented as follows: For the first j For each candidate stage, the average growth direction is calculated based on the difference in the comprehensive feature vectors of adjacent frames belonging to that stage in the offline training samples, denoted as . , No. j The consistency of the short-term growth trends of the candidate stages is represented as follows: ; in, Indicates the first j The short-term growth trend of the candidate stage is consistent, and the value range is [0,1]. When the current short-term growth direction is consistent with the short-term growth trend of the candidate stage, the short-term growth trend is consistent with the short-term growth trend of the candidate stage. j When the typical growth direction of each candidate stage is consistent, Increase; when the two increase in opposite directions. Decrease.

[0021] Furthermore, including: The constraint judgment on the preliminary identification results includes: During the candidate phase The result of the pollution flashover discharge stage identification at the current moment is determined when the following conditions are met simultaneously: First, the sample to be tested is in the candidate stage. Within the radius of the stage characteristic; second, Candidate stage It has the advantage of distance interval compared to other stages; Third, if a previous stable phase exists ,but Candidate stage There is a stage transition relationship supported by offline training samples between the previous stable stage and the previous stable stage.

[0022] Secondly, the present invention also provides a system for identifying and classifying the flashover stage of polluted insulators, the system comprising: The preprocessing module is used to simultaneously acquire dual-band ultraviolet images or video frame sequences of polluted insulator discharge using a solar-blind ultraviolet imaging channel and a UVA imaging channel, forming a dual-band image sequence to be identified, and preprocessing the dual-band image sequence to obtain a standardized dual-band six-dimensional feature vector, including: the total pixel area of ​​all discharge spot regions in the solar-blind ultraviolet image, the pixel area of ​​the single largest discharge spot region in the solar-blind ultraviolet image, the number of connected regions of independent discharge spots in the solar-blind ultraviolet image, the total pixel area of ​​all discharge spot regions in the UVA image, the pixel area of ​​the single largest discharge spot region in the UVA image, and the number of connected regions of independent discharge spots in the UVA image; The enhanced feature determination module is used to calculate the dual-band area response ratio, spot concentration and short-term growth of the dual-band image based on the standardized dual-band six-dimensional feature vector, that is, to identify enhanced features during the generation stage. The model training module is used to train the time-constrained weighted clustering model using pollution flashover test samples. The input features of the time-constrained weighted clustering model include the stage identification enhancement features. The four continuous intervals corresponding to the normalized time axis formed after normalization according to the time position are used as the initial cluster centers. After iterative training, the physical stages corresponding to the pollution flashover test samples are obtained. The physical stages include: initial stage, partial discharge stage, arc propagation stage and flashover breakdown stage. The physical stage determination module is used to calculate the comprehensive discrimination score of four candidate stages for each frame of the sample to be tested, and select the candidate stage with the largest comprehensive discrimination score as the preliminary identification stage of the current frame. The preliminary identification result is constrained and judged. When the candidate stage satisfies the constraint judgment at the same time, the current candidate stage is taken as the pollution flashover stage identification result at the current moment. The graded early warning module is used to determine the stable phase sequence under multiple frames based on the test sample, thereby generating a pollution flashover risk index, and outputting graded early warnings based on the pollution flashover risk index.

[0023] Thirdly, the present invention provides an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the machine-readable instructions are executed by the processor to perform the steps of the first aspect.

[0024] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps implemented in the first aspect described above.

[0025] Beneficial effects: Compared with the prior art, the present invention has the following advantages: After extracting basic features such as the area and number of dual-band light spots, this invention further introduces dual-band area response ratio, light spot concentration, short-term growth, and normalized time position. Through feature weights, stage transition constraints, stage feature radii, stable stage identification confidence, and risk indices, a discharge stage model is established offline based on historical samples. This makes stage modeling, online discrimination, and early warning output interconnected and subject to progressive constraints. Compared with single solar-blind ultraviolet imaging or single UVA imaging, this invention can simultaneously utilize the high sensitivity of the solar-blind ultraviolet channel to early weak discharges and the morphological characterization advantage of the UVA channel for mid-to-late-stage arc expansion, thus more completely characterizing the development process of pollution flashover discharge from weak to strong and from discrete to connected.

[0026] By using time-constrained weighted clustering and stage transition probability matrix, this invention can suppress non-physical stage bounces caused by noise, local flicker, or short-term arc extinction. By jointly judging stage feature radius, distance interval confidence, and short-term growth trend consistency, it can reduce misjudgments at the boundaries of adjacent stages. Through adaptive time-series smoothing and risk index grading mechanisms, it can improve the stability of early warning output while maintaining the real-time nature of early warning, providing a more reliable decision-making basis for continuous observation, key inspections, live-line cleaning, planned power outage maintenance, and test protection. Attached Figure Description

[0027] Figure 1 This is a flowchart of the method for identifying and classifying flashover stages of polluted insulators according to an embodiment of the present invention; Figure 2 This is a flowchart of the dual-band ultraviolet image preprocessing and feature extraction process described in an embodiment of the present invention; Figure 3 This is a flowchart illustrating the offline modeling and online comprehensive identification process for temporal constraints as described in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the typical characteristics of the four stages of flashover discharge as described in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the discharge stage risk index early warning determination process as described in an embodiment of the present invention. Figure 6 These are example images of visible light images, solar-blind ultraviolet images, and UVA images described in embodiments of the present invention; Figure 7 This is a schematic diagram illustrating the stage identification and early warning results described in an embodiment of the present invention; Figure 8 The diagram shows the actual application effect described in the embodiment of the present invention. Detailed Implementation

[0028] 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, and 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.

[0029] like Figure 1 As shown, the present invention includes the following technical solution: 1. Simultaneous acquisition of dual-band ultraviolet images: The dual-band ultraviolet images or video frame sequences of polluted insulator discharge are acquired simultaneously using a solar-blind ultraviolet imaging channel and a UVA imaging channel.

[0030] The preferred imaging method for the solar-blind ultraviolet channel is a solar-blind ultraviolet filter + a fully transparent ultraviolet lens + an ICCD camera. The preferred imaging method for the UVA imaging channel is a fully transparent ultraviolet lens combined with a CMOS camera. The two channels simultaneously acquire dual-band ultraviolet images of the same discharge process, forming a dual-band image sequence to be identified.

[0031] In this embodiment, the working wavelength of the solar-blind ultraviolet imaging channel is 240–280 nm, and the working wavelength of the UVA imaging channel is 320–400 nm.

[0032] 2. Dual-band image preprocessing: like Figure 2 As shown, the preprocessing of the dual-band image sequence specifically includes: (1) Perform timestamp matching on the two-channel images to obtain dual-band image frame pairs corresponding to the same or adjacent acquisition times; (2) Spatial registration is performed on the dual-band images, and the region of interest (ROI) is unified; when the resolution, field of view, or pixel scale of the two channels are inconsistent, the ROI area is normalized:

[0033] In the formula, These represent the normalized spot area, spot area, and ROI area in a single frame of solar-blind ultraviolet image, respectively. This represents the normalized total spot area in a single frame of UVA image, including the spot area and the ROI area.

[0034] (3) Convert the image to grayscale; (4) Threshold segmentation is performed on the image to obtain a binary image of the discharge region; In this embodiment, a fixed threshold segmentation is used for solar-blind ultraviolet images, and an Otsu threshold segmentation is used for UVA images. Morphological optimization is performed through opening and closing operations. The structural element used is a rectangular structural element, and the fixed threshold for the solar-blind ultraviolet images is preferably 0.8.

[0035] (5) Remove noise and correct the light spot boundary through morphological operations.

[0036] Preferably, the morphological processing includes one opening operation and one closing operation, and the structuring element is a 3×3 pixel rectangle structuring element.

[0037] 3. Dual-band image feature extraction: Connectivity analysis was performed on the preprocessed binary image to extract features from the normalized solar-blind ultraviolet and UVA images, including the total spot area. That is, the sum of the normalized pixel areas of all discharge spot regions in the image; the maximum spot area. , which is the normalized pixel area of ​​a single largest discharge spot region in the image; the number of spots N is the number of connected regions of independent discharge spots in the image. Then, a dual-band six-dimensional feature vector is constructed: ; Where SB (Sun Blind) represents the sun-blind ultraviolet band, UVA represents the UVA band, and t represents the t-th frame or the t-th time.

[0038] To eliminate differences in the dimensions of different features, the feature vectors are standardized. The standardized parameters used in the online recognition stage are consistent with those used in the offline modeling stage.

[0039] Z-score standardization is preferred. Let the first... k The mean and standard deviation of the 3D features in the training samples are respectively and Then, the standardized result of this feature at time t is: ; In the formula, To prevent constants with a denominator of zero.

[0040] Thus, the standardized feature vector is obtained: ; Based on the standardized basic eigenvectors, the dual-band area response ratio R, spot concentration C, and short-duration growth are further calculated. Formation stage identification of comprehensive feature vector :

[0041] in, Used to reflect the enhancement of the UVA channel relative to the solar-blind ultraviolet channel. and It is used to reflect the trend of discharge concentrating from discrete light spots to a continuous arc. This is used to reflect the direction of change in discharge characteristics between adjacent frames. It should be noted that the six-dimensional basic feature vector... Used for basic quantitative representation, comprehensive feature vector Used for subsequent stage modeling and online recognition.

[0042] It should be noted that the dual-band area response ratio, spot concentration, and short-term growth rate supplement the basic image features from three perspectives: dual-band relative response, spatial connectivity, and temporal variation trend.

[0043] The dual-band area response ratio borrows from the idea of ​​band ratio analysis, but instead of directly using the gray values, radiation intensity, or reflectivity of the two bands as a ratio, it uses the area of ​​the discharge spot obtained by segmentation in the solar-blind ultraviolet channel and the UVA channel as the discharge area response of each channel to characterize the relative response relationship between the two ultraviolet bands at the same time.

[0044] The spot concentration is derived from connected component analysis and the concept of maximum region proportion. It quantifies the dominance of the main discharge region over the overall discharge region by using the ratio of the area of ​​the largest discharge connected region to the sum of the areas of all discharge connected regions. For example, if there are 5 separate discharge connected regions in a frame of an image, with areas of 10, 12, 8, 9, and 11 respectively, then the total discharge region area is 50, the area of ​​the largest connected region is 12, and the spot concentration is 0.24, indicating that the discharge in this frame is still mainly distributed in a multi-point discrete manner. If the total discharge region area in another frame of an image is 80, and the area of ​​the largest connected region is 60, then the spot concentration is 0.75, indicating that the discharge region has been mainly concentrated in a main discharge region, and the discharge pattern tends to be a continuous arc or a main discharge channel.

[0045] The short-term growth rate, based on the first-order difference concept of time series, is used to characterize the short-term variation trend of features such as discharge area, number of light spots, and concentration between adjacent frames. This feature can reflect the state of discharge enhancement, weakening, or stabilization, reducing the possibility of misjudgment caused by relying solely on static features of a single frame.

[0046] Therefore, the six-dimensional basic feature vector is used for the basic quantitative representation of the discharge image, while the dual-band area response ratio, spot concentration, and short-term growth are used to enhance the stage discrimination capability. Together, they constitute the stage recognition comprehensive feature vector for subsequent stage modeling and online identification.

[0047] 4. Establishment of an offline flashover discharge stage model: like Figure 3As shown, a pollution flashover discharge stage model is constructed based on pre-collected historical pollution flashover sample data. This invention introduces dual-band feature weights, a normalized time axis, adjacent stage transition relationships, and stage feature radii during the stage modeling process, forming a dual-band time-constrained weighted clustering model to address the problems of blurred boundaries between pollution flashover development stages, overlapping features of adjacent stages, and the lack of physical order in conventional clustering labels. The specific process is as follows: (1) Collect flashover test samples of different insulator types, different pollution levels, and different humidity conditions, and record the complete frame sequence of each group of flashover test samples from the start of pressurization to the end of flashover as a sample sequence; for the s-th sample sequence, let the total number of frames of the sample sequence be . Then the time position of the t-th frame is normalized as follows: ; in, Used to characterize the relative evolution position of the frame throughout the entire flashover process, with a size range of 0 to 1. The smaller the value, the closer the frame is to the early stage of flashover development. The larger the value, the closer the frame is to the later stage of the flashover development.

[0048] In this embodiment, the normalized time position is used for cluster center initialization, stage physical calibration, and stage transition relationship establishment in the subsequent offline modeling stage. The online identification stage does not depend on the complete duration of the test sample from the start of pressurization to the end of flashover, but rather constrains the stage evolution order through the previous stable stage, stage transition probability, and short-term growth trend.

[0049] (2) Perform the same preprocessing and feature extraction as above on the sample sequence to form a dual-band basic feature matrix and perform standardization processing; on the basis of the dual-band basic features, further calculate the dual-band area response ratio, spot concentration and short-term growth as stage identification enhancement features. Through the above enhancement features, the complementary relationship between the solar blind ultraviolet channel being sensitive to early weak discharge and the UVA channel being sensitive to mid-to-late arc expansion is incorporated into the stage identification.

[0050] Preferably, the dual-band area response ratio Concentration of sun-blind ultraviolet spots UVA spot concentration The short-term growth is the difference between the normalized feature vector of the current frame and the normalized feature vector of the previous frame. Through these enhanced features, the complementary relationship between the solar-blind ultraviolet channel's sensitivity to early weak discharge and the UVA channel's sensitivity to mid-to-late arc expansion is incorporated into the stage identification process.

[0051] (3) Adaptively determine feature weights based on the contribution of each feature dimension in the training samples to stage differentiation. Specifically, the entire flashover process of each training sample is divided into several continuous time slices according to the normalized time position, and the mean and variance of each feature dimension in each normalized time slice are calculated respectively. For the first... k The first feature is defined as follows: the variance between its average values ​​within each normalized time slice is used as the inter-slice variation, characterizing the difference in variation of the feature across different time slices; the average variance within each normalized time slice is used as the intra-slice volatility, characterizing the degree of volatility of the feature within the same time slice. The second feature is determined based on the ratio of inter-slice variation to intra-slice volatility. k The stage discrimination of each feature is calculated, and the stage discrimination of each feature is normalized and used as the corresponding feature weight.

[0052] (4) Using time-constrained weighting K The -means algorithm performs phased modeling on the training samples. Cluster centers are initially selected based on four consecutive intervals along the normalized time axis.

[0053] Specifically, the normalized time axis is divided into four continuous intervals. The mean of the comprehensive feature vector of the training samples falling into each continuous interval is calculated and used as four initial cluster centers. The four initial cluster centers are numbered 1, 2, 3, and 4 in order of their normalized time position from earliest to latest.

[0054] During the sample allocation process, for the first s The first of the flashover sample sequences t Frame, let its comprehensive feature vector be Normalized time position is The previous frame stage label is The candidate cluster number is j , No. j The cluster centers are , No. j The stage average normalized time position of each cluster is The current frame is then assigned to a candidate cluster. j The combined cost is expressed as: ; in, Indicates the first s In the flashover sample sequence, the th t Frames are assigned to candidate clusters j The combined cost; Weighted feature distance; Penalty for reverse transfer; Penalty for skipping stages; Penalty for deviation from normalized time position; The weight coefficients for the reverse transition penalty, the cross-stage jump penalty, and the normalized time position deviation penalty are determined during the offline training phase.

[0055] Specifically, the training samples are first weighted without incorporating reverse transition penalties, cross-stage jump penalties, or normalized time position deviation penalties. K -means clustering is performed, and the average weighted feature distance from the training sample frames to their respective cluster centers is calculated, with this average weighted feature distance recorded as the baseline. Then, the reverse transition penalty coefficient α is set to twice this baseline, the cross-stage jump penalty coefficient β is set to once this baseline, and the normalized time position deviation penalty coefficient γ is set to 0.5 times this baseline. Finally, the determined... Execution timing constraint weighted K -means clustering modeling. Through the above method, reverse jumps are strongly constrained, cross-stage jumps are subject to conventional constraints, and normalized time position deviations are used as auxiliary constraints in stage division, thereby improving the temporal stability of stage labels without changing the dominant role of feature clustering.

[0056] The weighted feature distance is: ; in, K To determine the dimension of the comprehensive feature vector, Indicates the first k Adaptive weights for dimensional features For the comprehensive feature vector The k 3D eigenvalues For the first j Cluster centers The k 3D eigenvalues. Through this weighted feature distance, the features with higher stage discrimination in step (3) contribute more to the cluster distance calculation.

[0057] The reverse transfer penalty is expressed as follows: ; When the candidate cluster number j Smaller than the previous frame stage label When, it indicates that the current frame has undergone a reverse transition from the subsequent stage to the preceding stage relative to the previous frame, and the penalty term takes a positive value; when When that happens, the penalty item is set to 0.

[0058] The cross-stage jump penalty is expressed as follows: ; When the candidate cluster number j Stage label from the previous frame When the absolute value of the difference between adjacent frames is greater than 1, it indicates that the adjacent frame stage labels cross at least one intermediate stage, and the penalty term takes a positive value; when When that happens, the penalty item is set to 0.

[0059] The normalized time position deviation penalty is expressed as follows: ; in, For the s-th flashover sample sequence, the th... t Normalized time position of the frame. For the first j The stage mean normalized time position of each cluster. This penalty term is used to constrain the normalized time position of the sample frame to match the stage mean normalized time position of the candidate cluster.

[0060] For each frame, calculate the combined cost when assigning it to one of the four candidate clusters, and select the candidate cluster with the smallest combined cost as the new stage label for that frame. After all training sample frames have been assigned stage labels, update the cluster center and stage mean normalized time position of each cluster based on the new stage labels. j The cluster centers have been updated to: ; No. j The stage-averaged normalized time position of each cluster is updated as follows: ; in, Assigned to the current number j The number of training sample frames for each cluster.

[0061] Simultaneously, the stage transition probability matrix is ​​calculated based on the stage labels of adjacent frames in the same flashover sample sequence. The i-th cluster moves towards the i-th... j The transition probability of each cluster is expressed as: ; in, For all training sample sequences, two adjacent frames are transferred from cluster i to cluster i. j Number of times, For clusters i Total number of transfers from the starting point.

[0062] Repeat the above sample allocation, cluster center update, stage average normalized time position update, and stage transition probability matrix update until the stopping condition is met. The stopping condition is that the maximum change in cluster centers between two adjacent iterations is less than a preset convergence threshold, or the number of iterations reaches a preset maximum number of iterations.

[0063] The maximum change in the cluster centers is expressed as: ; in, For the first r After the nth iteration j Cluster centers, For the first r After -1 iterations, the th j Cluster centers, This represents the maximum change in the four cluster centers between two consecutive iterations. Stop iteration when This is a preset convergence threshold. Since the comprehensive feature vector has already been standardized, Pick to The maximum number of iterations is between 100 and 300.

[0064] Through the aforementioned combined cost function, the reverse transition penalty, cross-stage jump penalty, and normalized time position deviation penalty are directly integrated into the sample allocation process of each iteration. For any candidate cluster, if it causes a reverse transition or cross-stage jump in the current frame relative to the previous frame, or if its average normalized time position deviates from the normalized time position of the current frame, the corresponding combined cost increases; the candidate cluster with the minimum combined cost is determined as the new stage label for the current frame. Thus, while maintaining clustering based on sample feature similarity, the iterative process ensures that the stage labels of the same flashover sample sequence satisfy the overall trend of gradual evolution from the initial stage, partial discharge stage, arc expansion stage to flashover breakdown stage. After the iteration is completed, four clusters with time-order constraints are obtained, and then the four clusters are mapped to the corresponding physical stages according to the stage labeling rules.

[0065] (5) The optimal number of clusters is set to 4. After iterative optimization, the clustering effectiveness is verified by the sum of squared errors (SSE), silhouette coefficient, CH index, and DB index. Simultaneously, based on the stage transition probability matrix and the physical evolution of flashover discharge from the initial stage of weak discharge, partial discharge, arc expansion to flashover breakdown, the number of clusters is checked for consistency. k When the clustering effectiveness index is good and the stage transition matrix shows sequential transition characteristics, the final clustering result is determined to be a four-stage model.

[0066] In this embodiment, the number of clusters is set to be... The total number of training sample frames is N, and the th... j The cluster centers are The comprehensive feature vector is The mean of the combined feature vector of all training samples is When the four-stage model is finally adopted, The sum of squared errors (SSE) measures the compactness of clustering results and is expressed as: ; in, For the s-th flashover sample sequence, the th... t Frame stage labels, For the first j Cluster centers, To train the comprehensive feature vector of the samples, For the first j The mean vector of each cluster. The SSE is the Euclidean distance. The smaller the SSE, the more closely the points in each cluster are clustered around its center, and the better the clustering effect.

[0067] The silhouette coefficient measures both the compactness and separation of clusters, with a value ranging from [−1, 1]. A larger value indicates better clustering performance. For any sample frame... , is represented as: ; in, For sample frames The average distance to other points within the same cluster. For sample frames The minimum average distance to all points in other clusters.

[0068] The CH index evaluates clustering performance by the ratio of inter-cluster deviation to intra-cluster deviation, and is expressed as: ; in, The cluster-level deviation matrix is... for traces, Represented as: ; in, The inter-cluster deviation matrix, for traces, Represented as: ; The stage label is j The number of training sample frames, Let N be the mean of the combined feature vectors of all training samples, and N be the total number of training sample frames. This represents the number of clusters. A larger CH index indicates a greater inter-cluster dispersion compared to intra-cluster dispersion, resulting in better clustering performance.

[0069] The DB index is based on the ratio of the average intra-cluster distance to the inter-cluster center distance. A smaller value indicates a better clustering effect, expressed as: ; in, For all satisfied training sample frames to the first j Cluster centers average distance, For all satisfied Training sample frames to the l-th cluster center average distance, For the first j The Euclidean distance between the first cluster center and the l-th cluster center.

[0070] (6) After completing the clustering, calculate the stage feature center, stage feature radius, stage average normalized time position and stage transition probability matrix for each cluster.

[0071] The stage feature center is the cluster center vector of each cluster after clustering is completed. j The stage feature centers of each cluster are represented as follows: ; in, j =1,2,3,4, which correspond to four clusters with time order; K The dimension of the comprehensive feature vector; For the first j The cluster centers at the in k The value taken on the dimensional feature. The _th j The cluster centers are currently labeled as j The comprehensive feature vector of all training sample frames is calculated and expressed as: ; in, For the first s The first flashover sample sequence t Frame stage labels, The current stage is tagged as j The number of training sample frames, This is the comprehensive feature vector corresponding to the training sample frame.

[0072] The stage feature radius is preferably a quantile value of the weighted distance from the training sample to the stage feature center, which is used to determine whether the sample to be tested is within the credible range of the stage during online identification.

[0073] In this embodiment, the stage transition probability matrix is ​​a 4×4 matrix, and the elements of matrix P are... This indicates that the current frame belongs to a cluster between two adjacent frames. i At that time, the next frame moves to the cluster. j The probability of is calculated using the following formula:

[0074] in, This indicates that in the entire training sample sequence, two adjacent frames are separated by stage labels. i Transfer to stage label j Number of times, Indicates the stage label i Total number of transfers from the starting point.

[0075] The calculation process of this matrix is ​​as follows: for each training sample sequence, weighted by time constraints. K -means obtains the stage label for each frame. It iterates through the adjacent frames of each sample sequence, counting the stages... i To the stage j The number of transitions. For each i Divide the number of transitions by the number of clusters. i The total number of occurrences is obtained This matrix is ​​saved as part of the offline model for use during online recognition.

[0076] In this embodiment, the stage average normalized time position can preliminarily determine the temporal order of the four stages, with the earliest being the initial stage and the latest being the flashover breakdown stage. Preferably, in one approach, for each stage, the normalized time positions of all frames belonging to that stage are collected, and their arithmetic mean is calculated, expressed as: ; in, For the first j The stage average normalized time position of each cluster For the first s The first flashover sample sequence t Normalized time position of the frame. The stage label is j The number of training sample frames. Of course, during the iteration process, no physical stages are formed. At this time, cluster calculation with time order can be used, that is, the four stages are modified to correspond to the four clusters.

[0077] (7) When calibrating each cluster center in stages, the calibration is carried out according to the combination rules of time location, feature intensity, spot shape and transfer relationship.

[0078] First, according to the j The stage mean normalized time position of each cluster The four clusters were sorted in ascending order. The sorted clusters correspond to the early, pre-mid-stage, mid-late stage, and late stage of the flashover process, respectively. Because... The current stage is labeled as j The normalized time position of all training sample frames is calculated, and therefore it can reflect the actual occurrence time of the cluster in the entire process of pollution flashover.

[0079] Secondly, compare the cluster centers. The corresponding dual-band basic and enhanced features include the total spot area of ​​the solar-blind ultraviolet channel, the maximum spot area of ​​the solar-blind ultraviolet channel, the number of spots in the solar-blind ultraviolet channel, the total spot area of ​​the UVA channel, the maximum spot area of ​​the UVA channel, the number of spots in the UVA channel, the dual-band area response ratio, the concentration of solar-blind ultraviolet spots, and the concentration of UVA spots. For each of the above features, the relative ranking among the four cluster centers is used as the basis for stage labeling.

[0080] Specifically, in this embodiment, the values ​​of a certain feature in the four cluster centers are sorted from smallest to largest. The earlier the value is sorted, the lower the value is sorted, indicating that the feature is at a low level in the four stages, and the later the value is sorted, indicating that the feature is at a high level in the four stages.

[0081] In this embodiment, the spot concentration is characterized by the ratio of the maximum spot area to the total spot area. A higher spot concentration indicates a lower proportion of the maximum spot area in the total spot area, with the discharge region mainly consisting of multiple scattered spots; a lower spot concentration indicates a higher proportion of the maximum spot area in the total spot area, with the discharge region mainly consisting of locally connected or continuous arcs. Therefore, the multi-point discrete distribution is determined by both the number of spots and the spot concentration: when the number of spots in a cluster center increases relative to the initial stage, and the spot concentration does not reach the highest ranking among the four cluster centers, the discharge pattern corresponding to that cluster center is determined to be a multi-point discrete distribution; when the spot concentration reaches a higher ranking among the four cluster centers, and the maximum spot area and total spot area of ​​the UVA channel increase simultaneously, the discharge pattern is determined to evolve from multi-point discrete to locally connected or continuous arcs.

[0082] Secondly, combining the stage transition probability matrix The transition relationship between the preceding and following stages is confirmed. If a cluster mainly transitions from clusters with earlier time positions and mainly transitions to clusters with later time positions, then the cluster conforms to the sequential evolution relationship of the flashover stage. If a cluster does not have stable low-stage transitions afterward and its mean normalized time position is the latest, then the cluster conforms to the late termination characteristics of the flashover breakdown stage.

[0083] like Figure 4 As shown, specifically, the average normalized time position The cluster with the earliest ranking, and whose total spot area, maximum spot area, total spot area, and maximum spot area of ​​the UVA channel are all low among the four cluster centers, while the UVA channel response is lower than that of the cluster center with the solar-blind UV channel response, is designated as the initial stage. This stage corresponds to the early weak discharge state of pollution flashover, with few and small discharge spots. The weak discharge signal is mainly captured by the solar-blind UV channel, and the UVA channel response is not yet obvious.

[0084] Mean Normalized Time Position Following the initial stage, the number of light spots increases relative to the initial stage, and the total light spot area also rises. However, the proportion of the largest light spot area to the total light spot area does not reach the highest ranking among the four cluster centers, and this cluster mainly consists of cluster centers transferred from the initial stage, thus identifying it as the partial discharge stage. In this stage, multiple partial discharge points begin to appear, and the discharge area is distributed discretely in multiple points, without yet forming a stable continuous arc.

[0085] Mean Normalized Time Position Following the partial discharge stage, the total and maximum UVA channel spot areas further increase compared to the partial discharge stage. The dual-band area response ratio and spot concentration also increase, and the clustering centers where the discharge morphology evolves from multi-point discreteness to localized connectivity are identified as the arc expansion stage. In this stage, the UVA channel's characterization of arc channel area growth and discharge channel connectivity is enhanced, and the proportion of the maximum spot area increases, indicating that the discharge expands from scattered points to a locally continuous arc.

[0086] Mean Normalized Time Position The cluster with the latest ranking, exhibiting the highest total spot area of ​​the solar-blind ultraviolet channel, the highest spot area of ​​the solar-blind ultraviolet channel, the highest spot area of ​​the UVA channel, and the highest spot concentration among the four cluster centers, and lacking any stable low-stage transition cluster centers in the stage transition probability matrix, is designated as the flashover breakdown stage. This stage corresponds to a strong continuous discharge state, where the discharge region expands from a local arc to the main discharge channel, representing the highest risk level.

[0087] After completing the stage calibration, an offline pollution flashover stage identification model is obtained, which includes stage feature center, feature weight, stage feature radius, stage transition probability matrix, stage average time position, and physical stage label.

[0088] 5. Online identification of flashover stage: Repeat the above steps on the dual-band ultraviolet image of the polluted insulator to be tested to obtain the dual-band basic features and stage identification enhancement features of the sample to be tested, and process them using the standardized parameters and feature weights obtained in the offline modeling stage.

[0089] During online recognition, the weighted distance between the feature vector to be tested and the feature centers of each stage is first calculated. The weighted distance is defined in the same way as the weighted feature distance in the offline modeling stage.

[0090] For the j The weighted distance between each candidate stage is expressed as: ; in, Indicates the first t The comprehensive feature vector Z(t) of the sample under test at time t and the first time j Each stage of characteristic center Weighted distance between them; K The dimension of the comprehensive feature vector; Indicates the first k Dimensional feature weights; Let Z(t) be the eigenvector of the composite feature vector at time t. k 3D eigenvalues; Indicates the first j Each stage of characteristic center The k Dimensional eigenvalues. The smaller the value, the better the sample being tested is compared to the first... j The more similar the characteristics of the candidate stages, the better.

[0091] Further calculation based on weighted distance j The distance similarity of each candidate stage is expressed as: ; in, Indicates the first j Distance similarity between candidate stages; For the first j The stage feature radius for each stage; ε is a constant to prevent the denominator from being zero. This formula converts the weighted distance into a similarity value in the range [0,1], and The smaller, The larger.

[0092] Secondly, calculate the distance interval confidence score. For the first... j The candidate phase, first in addition to the first... j Among the remaining stages other than the candidate stages, determine the stage with the smallest weighted distance to the sample to be tested, and record this smallest distance as . Then the first j The distance interval confidence of each candidate stage is expressed as: ; in, Indicates the first j The confidence level of the distance interval between candidate stages. If If the weighted distance is less than the weighted distance corresponding to other stages, then Take a positive value; if If it is not less than the minimum weighted distance in other stages, then Set it to 0. Using this distance interval confidence level, we can determine the relative confidence level of the test sample compared to the first... j Does each candidate stage possess a distinguishable distance advantage?

[0093] Next, calculate the first based on the stage characteristic radius. j The confidence level of the stage feature radius of each candidate stage is expressed as: ; in, Indicates the first j The credibility of the stage feature radius of each candidate stage. If Located at the stage characteristic radius Within the range, then If it is a positive value; Exceeding the stage feature radius range, then Take 0.

[0094] The previous stable stage is the label of the stable stage output after the most recent adaptive timing smoothing, denoted as . For the first j The stage transition probability of each candidate stage is expressed as: ;in, Indicates the transition from the previous stable phase Transfer to the j The probability of each candidate stage is provided by the stage transition probability matrix obtained from the offline modeling stage. When there is no previous stable stage at the current time, [the probability is...]. All four candidate stages are set to the same value so that the transition probability term does not affect the identification result of the first stage.

[0095] Short-term growth trend consistency is used to determine whether the current characteristic change direction is consistent with the typical growth direction of the candidate stage. The current short-term growth amount is expressed as: ΔZ( t )=Z( t )−Z( t -1), for the th j For each candidate stage, the average growth direction is calculated based on the difference in the comprehensive feature vectors of adjacent frames belonging to that stage in the offline training samples, denoted as . . No.j The consistency of the short-term growth trends of the candidate stages is represented as follows: ;in, Indicates the first j The short-term growth trend of the candidate stage is consistent, with a value range of [0,1]. When the current short-term growth direction is consistent with the short-term growth trend of the candidate stage... j When the typical growth direction of each candidate stage is consistent, Increase; when the two increase in opposite directions. Decrease.

[0096] By combining distance similarity, distance interval confidence, stage feature radius confidence, stage transition probability, and short-term growth trend consistency, the first... j The comprehensive discrimination score for each candidate stage is expressed as: ; In the formula, Represents the time t at time t. j The comprehensive discrimination score for each candidate stage; Indicates distance similarity; Indicates the confidence level of the distance interval; Indicates the confidence level of the stage feature radius; Indicates the transition from the previous stable phase to the [missing information]. j The stage transition probability of each candidate stage; This indicates consistency in short-term growth trends; to Let be the weight coefficients of each discriminant factor, and let them sum to 1.

[0097] In this embodiment, to We set the values ​​to 0.25, 0.20, 0.20, 0.20, and 0.15 respectively, making distance similarity the primary discrimination factor, while also combining distance interval, stage feature radius, stage transition probability, and short-term growth trend for joint discrimination.

[0098] For each frame, the comprehensive discrimination score of the four candidate stages is calculated, and the candidate stage with the highest comprehensive discrimination score is selected as the preliminary recognition stage for the current frame. y t To avoid misjudgments at stage boundaries or under noise interference, further constraints are applied to the preliminary identification results. When candidate stages... y t The current flashover discharge stage identification result is taken when the following conditions are met simultaneously: First, the sample to be tested is located in the candidate stage y. t Within the radius of the stage characteristic; second, That is, candidate stage y t It has the advantage of distance interval compared to other stages; Third, if a previous stable phase exists ,but That is, candidate stage y t There is a stage transition relationship supported by offline training samples between the previous stable stage and the previous stable stage.

[0099] If the score difference between the stage corresponding to the highest comprehensive discrimination score and the stage corresponding to the second highest comprehensive discrimination score is less than 0.05, or if there is no transition relationship supported by offline training samples between the candidate stage and the previous stable stage, then the preliminary identification result of a single frame is not directly used. Instead, the previous stable stage, the adjacent transition stage, or the state to be confirmed is output by combining the identification results of consecutive frames. Therefore, the final identification result of the current frame is jointly determined by the weighted distance, distance similarity, distance interval confidence, stage feature radius confidence, stage transition probability, and short-term growth trend consistency, rather than solely by the distance corresponding to the nearest stage center.

[0100] 6. Adaptive timing smoothing and hierarchical early warning output: like Figure 5 As shown, in order to reduce misjudgments caused by single-frame recognition fluctuations, this invention adaptively determines the smoothing window length based on recognition confidence and stage label fluctuation, and uses confidence-weighted voting to obtain a stable stage sequence.

[0101] Specifically, the stage identification result obtained by online identification of the t-th frame is denoted as y. t The maximum value among the comprehensive discrimination scores of the four candidate stages in frame t is taken as the single-frame recognition confidence score of that frame. ; in, For time t, the first... j The comprehensive discrimination score for each candidate stage. Because The result is obtained by weighting distance similarity, distance interval confidence, stage feature radius confidence, stage transition probability, and short-term growth trend consistency, and each discriminant factor is normalized to the range [0,1]. Therefore, Used to characterize the reliability of the identification results at the t-th frame stage.

[0102] The label volatility is obtained by counting the number of label changes in the most recent L frames of the sample under test. ; in, The tag volatility within the most recent L frames. This is an indicator function; it takes the value 1 if the stage labels of two adjacent frames are different, and 0 otherwise. The larger the value, the more frequent the stage label changes within the most recent L frames. The average recognition confidence score for the most recent L frames is then calculated. ; in, The average recognition confidence score of the most recent L frames is used to characterize the overall credibility of the recognition results of the most recent consecutive frames. This average recognition confidence score is derived from the comprehensive discrimination score in the online recognition stage.

[0103] according to and The smooth window length is adaptively determined. Specifically, when and When the most recent frame identification result is deemed reliable and the stage label is stable, the smoothing window length is set to 3 frames to improve the response speed to stage changes; when or At that time, the smoothing window length is set to 5 frames to balance real-time performance and stability; when or If the most recent frame recognition result is deemed unstable, the smoothing window length is set to 7 frames; if two consecutive window judgments both satisfy the condition... and The smoothing window length is set to 9 frames to enhance the suppression of noise fluctuations and transient flicker. Therefore, the window length adaptively switches between 3, 5, 7, and 9 frames.

[0104] When voting within the window, different weights are assigned to the recognition results of each frame based on temporal distance and recognition confidence. For the current frame within the smoothing window... m The voting weight of a frame is represented as follows: ,in, W is the current smoothing window length. For the first m Frame-by-frame recognition confidence. t−m For the first m The time distance between the current frame and the previous frame. Therefore, the closer the frame is to the current moment and the higher the identification confidence, the greater its contribution to the window voting.

[0105] For the candidate stage j Its weighted voting score within the current window is expressed as: ; in, Candidate stage j The weighted voting score. The candidate stage with the highest weighted voting score is selected as the smoothing stage result for the current frame, denoted as... , This is the stable stage label for frame t after adaptive temporal smoothing. It is derived from the outputs at consecutive time points. A stable stage sequence is formed. The stable stage sequence refers to the stage label sequence obtained after single-frame online recognition, smooth window adaptive determination, confidence-weighted voting, and stage transition confirmation.

[0106] When there is evidence of a transition to a higher-risk stage, a shorter confirmation time is used. Specifically, if a candidate stage is higher than the current stable stage, and the high-risk candidate stage has the highest weighted voting score for two consecutive frames, then the stable stage is updated to the high-risk candidate stage. This setting is used to promptly raise the warning level when flashover progresses from partial discharge to arc propagation, and from arc propagation to flashover breakdown.

[0107] When a stage declines towards a low-risk stage, a longer holding period is used. Specifically, if a candidate stage is lower than the current stable stage, the stable stage is only updated to the low-risk candidate stage if it has been the stage with the highest weighted voting score for 5 consecutive frames; otherwise, the current stable stage remains unchanged. This setting is used to avoid frequent drops in warning levels during the critical development stage of pollution flashover due to transient extinction, local flicker, or image noise.

[0108] The pollution flashover risk index is further calculated based on the stable phase sequence, and a graded early warning output is generated accordingly. The pollution flashover risk index is derived from the normalized stable phase levels. Duration of the current stage The probability of transitioning to a higher-risk stage The maximum spot area growth rate of the UVA channel within the stable phase sequence Dual-band area response ratio growth rate Confidence level for identifying stable phases Weighted average yields: ; In this embodiment, the stability phase level The sequence number is directly taken from the current stable stage, such as initial=1, partial discharge=2, arc propagation=3, flashover breakdown=4, and the duration of the current stage. To count the number of frames or the duration of consecutive identical stage labels in a stable stage sequence, based on the stable stage sequence. To query the current stable stage from the stage transition probability matrix obtained from offline modeling. The probability of transitioning to a higher stage. The maximum spot area growth rate of the UVA channel. The rate of increase of the dual-band area response ratio is calculated based on the change in the maximum spot area of ​​the UVA channel during the current stable phase. The confidence level for identifying the stable phase is calculated based on the change in the dual-band area response ratio during the current stable phase duration. It is obtained by weighting the single-frame recognition confidence of each frame within the current smoothing window: Where W is the current smoothing window length, For the first m The voting weight of a frame within the current smoothing window. For the first m The confidence level of single-frame recognition. Therefore, It represents the overall confidence level of the current stable phase sequence within the smooth window, rather than the confidence level of the single-frame distance interval.

[0109] In this embodiment, to We used values ​​of 0.30, 0.15, 0.20, 0.15, 0.10, and 0.10 respectively to make the stable phase level and the probability of transitioning to a higher risk phase the main risk factors. We also combined the current phase duration, the growth rate of the maximum spot area of ​​the UVA channel, the growth rate of the dual-band area response ratio, and the confidence level of stable phase identification to conduct risk assessment.

[0110] By introducing duration and growth rate factors into the risk index, the early warning output is no longer a simple mapping of stage labels, but can reflect the development speed and stage deterioration trend of pollution flashover.

[0111] The tiered early warning strategy includes: (1) When the equipment is in the initial stage and the pollution flashover risk index is lower than the first threshold, output low-level concern information and record the early weak discharge occurrence time and the response intensity of the solar blind ultraviolet channel. (2) When the equipment enters the partial discharge stage, or when the number of light spots, total light spot area or risk index continues to rise even though it is in the initial stage, output general abnormal information and suggest strengthening inspection or continuous observation; (3) When the equipment moves from the partial discharge stage to the arc expansion stage and remains in the arc expansion stage, or when the probability of transition from the partial discharge stage to the arc expansion stage and the growth rate of the UVA channel both exceed the preset threshold, a high-level warning message is output to prompt the equipment to arrange for troubleshooting, cleaning or maintenance as soon as possible. (4) When the equipment enters the flashover breakdown stage, or when the stable stage has not fully entered the flashover breakdown stage but the risk index exceeds the highest threshold and the strong discharge characteristics continue to exist, an alarm message is output for test protection, emergency handling or linkage shutdown.

[0112] This invention organically combines pollution flashover stage identification with graded early warning through the above-mentioned dual-band time-series constraint modeling, confidence joint discrimination, adaptive time-series smoothing and risk index early warning mechanism, providing a stable basis for differentiated operation and maintenance strategies such as continuous observation, key inspection, live cleaning, and planned power outage maintenance.

[0113] 7. Model update mechanism: When significant changes occur in imaging system parameters, insulator type, pollution level range, operating environment, or sample distribution, the offline stage model can be retrained or calibrated based on newly added representative samples. Retraining or calibration includes adjusting standardized parameters, feature weights, stage feature centers, stage feature radii, stage transition probability matrices, and risk index thresholds. Except for model updates, re-clustering of each test sample is unnecessary during normal application.

[0114] This embodiment uses an artificial simulated contaminated insulator wet discharge test platform to verify the method of the present invention. The test platform includes a 50Hz power frequency test transformer, a protective resistor, a voltage divider measurement unit, an artificial fog chamber, test insulators, and a dual-band ultraviolet imaging system. The test insulators include different types of porcelain insulators and silicone rubber composite insulators. The artificial contamination on the insulator surface is prepared using sodium chloride and kaolin, with the preferred mass ratio of sodium chloride to kaolin being 1:5. Light, medium, and heavy contamination states are created by adjusting the equivalent salt density ESDD. The test uses uniform wetting and slow voltage increase to induce discharge on the contaminated insulator surface, with the preferred voltage increase rate being 1... k V / s, to obtain the complete development process from initial weak discharge, partial discharge to arc propagation and final flashover breakdown.

[0115] (1) Simultaneous acquisition of dual-band ultraviolet images: like Figure 6 Image (a) is a visible light image. Figure 6 Image (b) is a solar-blind ultraviolet image. Figure 6 (c) in the figure represents the UVA image. The dual-band ultraviolet imaging system in this embodiment includes a solar-blind ultraviolet imaging channel and a UVA imaging channel. The solar-blind ultraviolet imaging channel has a resolution of 640×480px, a frame rate of 25FPS, and an operating wavelength of 240–280nm; the UVA imaging channel has a resolution of 1280×1024px, a frame rate of 50FPS, and an operating wavelength of 320–400nm.

[0116] Before the experiment, the two channels were time-aligned using timestamp matching, and spatial registration was completed using insulator contour features. For cases where the UVA channel frame rate was higher than the solar-blind UV channel frame rate, the UVA frame corresponding to the solar-blind UV frame was selected according to the nearest timestamp principle, resulting in a one-to-one corresponding dual-band image frame sequence. To ensure consistency in subsequent analysis, the complete process from the start of pressurization to the end of flashover was selected as the sample sequence.

[0117] (2) Dual-band image preprocessing and feature extraction: The acquired dual-band images were processed frame by frame. First, ROI unification and area normalization were performed on both channels to avoid incomparable area features due to differences in resolution and field of view between the two channels. Solar-blind ultraviolet images were segmented using a fixed threshold of 0.8, while UVA images were segmented using the Otsu thresholding method. After segmentation, one opening operation and one closing operation were performed, with the structuring element being a 3×3 pixel rectangle. After these processes, binary images of the discharge region with clearer boundaries and less background noise were obtained, providing a foundation for subsequent feature extraction.

[0118] Subsequently, connected component analysis was performed on each frame of the image to extract the total spot area, the maximum spot area, and the number of spots, constructing a dual-band six-dimensional basic feature sequence, which was then normalized using Z-score. Based on this, the dual-band area response ratio, solar-blind ultraviolet spot concentration, UVA spot concentration, and short-term growth between adjacent frames were calculated to form a comprehensive feature sequence for stage identification.

[0119] (3) Offline stage model establishment: Representative flashover samples from different insulator types, pollution levels, and test cycles were selected as the training set. Consistent preprocessing and feature extraction were performed on the training set images to obtain the training sample feature matrix, and the Z-score normalization method described above was applied. For each complete flashover sample sequence, the time axis was normalized using the total number of frames from the start of voltage application to the end of the flashover, yielding the normalized time position for each frame.

[0120] Based on the six-dimensional dual-band fundamental features, the dual-band area response ratio, solar-blind ultraviolet spot concentration, UVA spot concentration, and short-term growth between adjacent frames are further calculated. Feature weights are determined according to the degree of difference between different normalized time slices and the degree of fluctuation within the same time slice for each feature dimension in the training samples, so that features that contribute more to stage differentiation have higher weights in the modeling.

[0121] This embodiment uses time-constrained weighted average. K The -means algorithm is used to establish the stage model. First, cluster centers are initialized according to four consecutive intervals of the normalized time axis. Then, weighted feature distances are calculated during sample allocation, and reverse transition penalties and cross-stage jump penalties are introduced to suppress non-physical bounces of stage labels within the same flashover process. Finally, the stage feature centers, stage average time positions, and stage transition probability matrices are iteratively updated. After comprehensive evaluation using SSE, silhouette coefficient, CH index, and DB index, the model is selected. k =4 is used as the stage division number.

[0122] After clustering, the stage feature center, stage feature radius, average normalized time position, and stage transition probability matrix of each cluster are calculated. Stages are calibrated based on time position, feature intensity, spot morphology, and transition relationships: the earliest appearing cluster with dominant small, discrete solar-blind ultraviolet spots is calibrated as the initial stage; clusters with increasing spot numbers and total area but not yet forming a large-area continuous arc are calibrated as the partial discharge stage; clusters with significantly increased UVA channel area and maximum spot area, and increased spot concentration are calibrated as the arc expansion stage; and clusters with the latest appearance and obvious strong continuous discharge characteristics are calibrated as the flashover breakdown stage. This completes the offline establishment of the flashover discharge stage model for polluted insulators.

[0123] (4) Online stage identification Another set of pollution flashover test data for polluted insulators that was not involved in the modeling was selected as the test sample. The test sample images underwent the same preprocessing and feature extraction as in the training phase, and were standardized and weighted using the mean, standard deviation, and feature weights obtained in the training phase.

[0124] Then, the weighted distance between the feature vector of the test sample at each time step and the feature centers of the four stages is calculated. Simultaneously, the confidence level of the distance interval between the nearest and second-nearest stages, whether the test sample falls within the stage feature radius, the transition probability from the previous stable stage to the candidate stage, and the consistency of the short-term growth trend are calculated. Only when the candidate stage simultaneously satisfies the distance, confidence, and transition constraints is that candidate stage output as the identification result for the current time step; when single-frame evidence is insufficient, the previous stable stage, the adjacent transition stage, or the state to be confirmed is output.

[0125] (5) Adaptive time-series smoothing and hierarchical early warning Considering that single-frame images are susceptible to transient interference, noise fluctuations, and local flicker, which can easily lead to short-term stage jumps, adaptive temporal smoothing is applied to the recognition results of consecutive frames. This embodiment statistically analyzes the stage label volatility and calculates the average similarity confidence within the most recent consecutive frames. When the label volatility is low and the confidence is high, a shorter window is used to improve response speed; when the label volatility increases or the confidence decreases, the window length is automatically increased to improve the stability of the stage sequence. Within the window, a stable stage sequence is obtained using a voting method weighted by time distance and recognition confidence.

[0126] Based on this, a pollution flashover risk index is calculated according to the stability level, the duration of the current stage, the probability of transitioning to a higher risk stage, the growth rate of the maximum spot area of ​​the UVA channel, the growth rate of the dual-band area response ratio, and the identification confidence level. A graded early warning information is then output based on the risk index. Specifically, when the stable stage is in its initial phase and the risk index is low, a low-level concern information is output; when a local discharge stage occurs or the risk index continues to rise, a general abnormality information is output, prompting increased inspections or continuous monitoring; when the arc expansion stage begins, or when the probability of transitioning from the partial discharge stage to the arc expansion stage and the UVA channel growth rate both exceed a preset threshold, a high-level early warning information is output, prompting prompt arrangements for troubleshooting, cleaning, or maintenance; when the flashover breakdown stage begins, or when the risk index exceeds the highest threshold and strong discharge characteristics persist, an alarm information is output for experimental protection or emergency response.

[0127] Implementation results: (1) The solar-blind ultraviolet channel responds earlier to weak discharge signals in the initial stage of flashover discharge, making it more suitable for early discharge identification; the UVA channel more clearly characterizes the discharge channel morphology, area growth, and strong discharge development process in the stages of partial discharge enhancement and arc expansion, making it more suitable for dynamic state assessment. Compared with single-band imaging, dual-band joint analysis can more completely characterize the evolution process of surface discharge of polluted insulators from weak to strong and from discrete to continuous, especially showing more obvious complementary advantages in distinguishing between early glow discharge and later arc discharge.

[0128] (2) Clustering effectiveness analysis was performed on the training samples using different numbers of clusters. The results showed that the SSE curve was... k A clear inflection point appeared near =4, and the DB index was... k A value of 4 is obtained, and the silhouette coefficient and CH index also perform well around this cluster number. Further analysis of the stage transition probability matrix reveals that... k When the value is 4, the stages exhibit a sequential evolution relationship of initial stage, partial discharge stage, arc propagation stage, and flashover breakdown stage. This indicates that the established four-stage model not only has statistical separability but also conforms to the physical development law of pollution flashover discharge, such as... Figure 7 As shown, the identification result is a flashover breakdown stage warning level; the current alarm characteristics are: full-screen high-brightness coverage of the solar blind ultraviolet channel; and the UVA channel forms a strong main discharge channel.

[0129] (3) To verify the effectiveness of the stage identification results, 100 sets of dual-band synchronous image frames were randomly selected, and the stage identification results were compared with those obtained through manual interpretation. The results show that after using dual-band temporal constraint weighted clustering, stage radius confidence, and transition relationship joint discrimination, the overall accuracy of stage identification can reach 95%, such as... Figure 8 As shown. Figure 8This is a confusion matrix between the model's predicted category and the human interpretation category, where the horizontal axis represents the model's predicted category and the vertical axis represents the human interpretation category. Samples on the diagonal of the matrix represent results consistent with the model's prediction and the human interpretation, while samples off-diagonal represent misidentifications. Figure 8 It can be seen that among the 100 randomly selected dual-band synchronous image frames, 95 samples fell on the diagonal of the confusion matrix, resulting in an overall recognition accuracy of 95.0%. Specifically, samples from the initial stage, arc expansion stage, and flashover breakdown stage mainly fell on the corresponding diagonal positions; some samples from the partial discharge stage were identified as belonging to the initial stage, indicating that misidentification mainly occurred between adjacent stages in the early stages of flashover. This phenomenon is consistent with the actual characteristics of both the initial and partial discharge stages: weak light spots, small area, discrete morphology, and overlapping transitions. Overall, Figure 8 This indicates that the method described in this embodiment can reliably identify different stages of flashover discharge, and the stage identification results are in good agreement with the results of manual interpretation.

[0130] (4) Through adaptive timing smoothing and confidence-weighted voting, the smoothing intensity can be automatically adjusted according to the current signal fluctuation: maintaining a faster response when the identification confidence is high, and enhancing noise suppression capability when the tag fluctuation is obvious. Especially when the flashover transitions from the partial discharge stage to the arc extension stage, combined with the UVA channel growth rate and stage transition probability in the risk index, it can form a continuous and reliable high-level early warning output earlier, while avoiding frequent rises and falls in the warning level due to single-frame noise.

[0131] In summary, the proposed method for identifying and classifying flashover stages of polluted insulators based on dual-band ultraviolet images can leverage the high sensitivity of the solar-blind ultraviolet channel to early weak discharges and the excellent characterization ability of the UVA channel for the mid-to-late arc expansion process. Through time-constrained weighted modeling, joint stage confidence discrimination, adaptive time-series smoothing, and risk index early warning mechanisms, it can stably divide the flashover development process into stages and classify risks.

[0132] This embodiment provides a system for identifying and classifying the flashover stage of polluted insulators, the system comprising: The preprocessing module is used to simultaneously acquire dual-band ultraviolet images or video frame sequences of polluted insulator discharge using a solar-blind ultraviolet imaging channel and a UVA imaging channel, forming a dual-band image sequence to be identified, and preprocessing the dual-band image sequence to obtain a standardized dual-band six-dimensional feature vector, including: the total pixel area of ​​all discharge spot regions in the solar-blind ultraviolet image, the pixel area of ​​the single largest discharge spot region in the solar-blind ultraviolet image, the number of connected regions of independent discharge spots in the solar-blind ultraviolet image, the total pixel area of ​​all discharge spot regions in the UVA image, the pixel area of ​​the single largest discharge spot region in the UVA image, and the number of connected regions of independent discharge spots in the UVA image; The enhanced feature determination module is used to calculate the dual-band area response ratio, spot concentration and short-term growth of the dual-band image based on the standardized dual-band six-dimensional feature vector, that is, to identify enhanced features during the generation stage. The model training module is used to train the time-constrained weighted clustering model using pollution flashover test samples. The input features of the time-constrained weighted clustering model include the stage identification enhancement features. The four continuous intervals corresponding to the normalized time axis formed after normalization according to the time position are used as the initial cluster centers. After iterative training, the physical stages corresponding to the pollution flashover test samples are obtained. The physical stages include: initial stage, partial discharge stage, arc propagation stage and flashover breakdown stage. The physical stage determination module is used to obtain the comprehensive discrimination score of different candidate physical stages corresponding to the current frame in the sample based on the characteristics of the sample to be tested, and to determine the physical stage of pollution flashover based on the comprehensive discrimination score; The graded early warning module is used to determine the stable phase sequence under multiple frames based on the test sample, thereby generating a pollution flashover risk index, and outputting graded early warnings based on the pollution flashover risk index.

[0133] The other technical features of the pollution flashover stage identification and graded early warning system for polluted insulators described in this embodiment are similar to the corresponding pollution flashover stage identification and graded early warning method for polluted insulators, and will not be repeated here.

[0134] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. "A plurality of" means two or more, unless otherwise explicitly specified.

[0135] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0136] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0137] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0138] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0139] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0140] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0141] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0142] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0143] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for identifying and classifying the flashover stage of polluted insulators, characterized in that, The method includes: A dual-band ultraviolet image or video frame sequence of polluted insulator discharge is simultaneously acquired using a solar-blind ultraviolet imaging channel and a UVA imaging channel to form a dual-band image sequence to be identified. The dual-band image sequence is then preprocessed to obtain a standardized dual-band six-dimensional feature vector, including: the total pixel area of ​​all discharge spot regions in the solar-blind ultraviolet image, the pixel area of ​​the single largest discharge spot region in the solar-blind ultraviolet image, the number of connected regions of independent discharge spots in the solar-blind ultraviolet image, the total pixel area of ​​all discharge spot regions in the UVA image, the pixel area of ​​the single largest discharge spot region in the UVA image, and the number of connected regions of independent discharge spots in the UVA image. Based on the standardized dual-band six-dimensional feature vector, the dual-band area response ratio, spot concentration and short-term growth of the dual-band image are calculated, which is the enhanced feature identification in the generation stage. The time-constrained weighted clustering model is trained using pollution flashover test samples. The input features of the time-constrained weighted clustering model include the stage identification enhancement features. The four continuous intervals corresponding to the normalized time axis formed after normalization according to the time position are used as the initial cluster centers. After iterative training, the physical stages corresponding to the pollution flashover test samples are obtained. The physical stages include: initial stage, partial discharge stage, arc propagation stage and flashover breakdown stage. For each frame of the sample to be tested, the comprehensive discrimination score of the four candidate stages is calculated, and the candidate stage with the largest comprehensive discrimination score is selected as the preliminary identification stage of the current frame. The preliminary identification result is constrained and judged. When the candidate stage satisfies the constraint judgment at the same time, the current candidate stage is taken as the identification result of the pollution flashover stage at the current moment. Based on the test samples, a stable phase sequence under multiple frames is determined, thereby generating a pollution flashover risk index, and a graded early warning output is performed based on the pollution flashover risk index.

2. The method for identifying and classifying flashover stages of polluted insulators according to claim 1, characterized in that, The calculation of the dual-band area response ratio, spot concentration, and short-term growth of the dual-band image based on the standardized dual-band six-dimensional feature vector includes: The standardized six-dimensional feature vector is represented as: ； in, This represents the sum of pixel areas of all discharge spot regions in the standardized solar-blind ultraviolet image. The pixel area of ​​the single largest discharge spot region in the standardized solar-blind ultraviolet image. This represents the number of connected regions of independent discharge spots in the standardized solar-blind ultraviolet image. This represents the sum of the pixel areas of all discharge spot regions in the standardized UVA image. The pixel area of ​​the single largest discharge spot region in the standardized UVA image. The number of connected regions of independent discharge spots in the standardized UVA image; The dual-band area response ratio is used to reflect the enhancement degree of the UVA channel relative to the solar-blind ultraviolet channel, and is expressed as: ； in, ε To prevent constants with a denominator of zero; t Indicates the current frame number or the current time. The concentration of the light spot reflects the trend of discharge concentrating from discrete light spots to a continuous arc, including the concentration of light spots in the solar-blind ultraviolet band. and UVA band light spot concentration , respectively represented as: ； The short-term growth amount reflects the direction of change in discharge characteristics between adjacent frames, and is expressed as follows: 。 3. The method for identifying and classifying flashover stages of polluted insulators according to claim 1, characterized in that, The step of using four consecutive intervals corresponding to the normalized time axis formed according to time position as the initial cluster centers includes: Each group of flashover test samples is recorded as a sample sequence, with the complete frame sequence from the start of pressurization to the end of flashover. For the s There are 3 sample sequences, and the total number of frames in the sample sequence is 1. Then the first t The time position of the frame is normalized as follows: ； in, Used to characterize the relative evolution position of the frame throughout the entire flashover process, with a size range of 0 to 1. The smaller the value, the closer the frame is to the early stage of flashover development. The larger the value, the closer the frame is to the later stage of the flashover development.

4. The method for identifying and classifying flashover stages of polluted insulators according to claim 3, characterized in that, The physical stages corresponding to the flashover test samples obtained after the iterative training include: Cluster center initialization: The initial cluster centers are determined by four consecutive intervals of the normalized time axis obtained by normalizing according to the time position; Iterative optimization process: During the sample allocation process, for each frame, the combined cost allocated to each cluster center is calculated, and the cluster with the smallest combined cost is selected as the new stage label for that frame, thereby minimizing the overall objective function in the iteration. Model parameter calculation: Repeatedly execute sample allocation, cluster center update, stage average normalized time position update and stage transition probability matrix update until the stopping condition is met. The stage transition probability matrix is ​​used to describe the possibility of mutual transfer between the four clusters during the pollution flashover discharge process. The stage average normalized time position is used to determine the time sequence of the four clusters. After clustering is completed, the stage feature center, stage feature radius, stage average normalized time position and stage transition probability matrix are calculated for each cluster. The stage feature centers of each cluster are labeled, and the physical stage of each cluster is determined by combining the transition relationship in the stage transition probability matrix. Output Model: After completing the stage calibration, an offline temporal constrained weighted clustering model is obtained, which includes stage feature centers, feature weights, stage feature radii, stage transition probability matrices, stage average time positions, and physical stage labels.

5. The method for identifying and classifying flashover stages of polluted insulators according to claim 4, characterized in that, The step of defining the stage of each cluster center and determining the physical stage of each cluster by combining the transition relationships in the stage transition probability matrix includes: The clusters are sorted from early to late according to the average occurrence position of the stage feature centers on the normalized time axis. That is, the four clusters after sorting correspond to the early, early-middle, middle-late and late stages of the pollution flashover process, respectively. Compare the total spot area, maximum spot area, number of spots, dual-band area response ratio, and spot concentration corresponding to the stage feature centers of each cluster; The cluster with the earliest average normalized time position and the lowest total spot area, maximum spot area, and UVA channel spot area among the four cluster centers, along with the lowest UVA channel spot area, was identified as the initial stage. Furthermore, the cluster center with the lowest UVA channel response compared to the solar-blind ultraviolet channel response was also identified as the initial stage. The number of light spots exceeds the level corresponding to the initial stage, the total light spot area increases but the proportion of the largest light spot area does not reach the highest value, the discharge is distributed in multiple discrete points and the cluster centers transferred from the initial stage are identified as the partial discharge stage. The total and maximum spot areas of the UVA channel exceed the levels corresponding to the partial discharge stage, the dual-band area response ratio increases, the spot concentration increases, and the clustering center of the spot development from multi-point dispersion to local connectivity is identified as the arc expansion stage. The cluster center with the latest average normalized time, the highest total spot area and the largest spot area, the strong continuous discharge region of the UVA channel, the highest spot concentration and no subsequent stable low-stage transfer is identified as the flashover breakdown stage.

6. The method for identifying and classifying flashover stages of polluted insulators according to claim 5, characterized in that, The stage average normalized time position is used to determine the temporal order of the four stages. For each stage, the normalized time positions of all frames belonging to that stage are collected, and the corresponding arithmetic mean is calculated. Therefore, the stage average normalized time position is expressed as: ； in, For the first j The stage average normalized time position of each cluster For the first s The first flashover sample sequence t Normalized time position of the frame. The candidate stage is labeled as j The number of training sample frames, For the first s The first flashover sample sequence t Frame stage label.

7. The method for identifying and classifying flashover stages of polluted insulators according to claim 4, characterized in that, The selection of the cluster with the lowest combination cost as the new stage label for the frame includes: During the sample allocation process, for the first s The first of the flashover sample sequences t Frame, the current frame is assigned to a candidate cluster. j The combined cost is expressed as: ； in, Indicates the first s In the flashover sample sequence, the th t Frames are assigned to candidate clusters j The combined cost; Weighted feature distance; Penalty for reverse transfer; Penalty for skipping stages; Penalty for deviation from normalized time position; These are the weight coefficients for the reverse transition penalty, the cross-stage jump penalty, and the normalized temporal position deviation penalty, respectively. The stage label of the previous frame is... ; After all training sample frames have been assigned stage labels, the cluster centers and stage mean normalized time positions of each cluster are updated according to the new stage labels. j The cluster centers have been updated to: ； No. j The stage-averaged normalized time position of each cluster is updated as follows: ； in, Assigned to the current number j The number of training sample frames for each cluster. For the first s The first of the flashover sample sequences t The comprehensive feature vector corresponding to the frame, and the candidate cluster number is j .

8. The method for identifying and classifying flashover stages of polluted insulators according to claim 7, characterized in that, The repeated execution of sample allocation, cluster center update, stage average normalized time position update, and stage transition probability matrix update until the stopping condition is met includes: The stopping condition is that the maximum change in cluster centers in two adjacent iterations is less than a preset convergence threshold, or the number of iterations reaches a preset maximum number of iterations. The maximum change in the cluster centers is expressed as: ； in, For the first r After the nth iteration j Cluster centers, For the first r After -1 iterations, the th j Cluster centers, The maximum value of the change in the four cluster centers in two consecutive iterations, when Stop iteration when This is the preset convergence threshold.

9. The method for identifying and classifying flashover stages of polluted insulators according to claim 4, characterized in that, The calculation of the comprehensive discrimination score for each frame of the sample to be tested, involving four candidate stages, includes: During online recognition, the weighted distance between the feature vector corresponding to the sample to be tested and the feature center of each stage is calculated, and the distance similarity is calculated based on the feature radius of each stage. Calculate the distance interval between the nearest stage center and the second nearest stage center to obtain the distance interval confidence. The confidence level of the stage feature radius of the candidate stage corresponding to the current frame is calculated based on the stage feature radius. By combining the transition probability from the previous stable stage to the candidate stage and the consistency between the current short-term growth rate and the typical growth direction of the candidate stage, a comprehensive discrimination score for the candidate stage is formed.

10. The method for identifying and classifying flashover stages of polluted insulators according to claim 9, characterized in that, The comprehensive discrimination score for the candidate stage is expressed as: ； In the formula, Indicates the first t Time of the first j The comprehensive discrimination score for each candidate stage; Indicates distance similarity; Indicates the confidence level of the distance interval; Indicates the confidence level of the stage feature radius; Indicates the transition from the previous stable phase to the [missing information]. j The stage transition probability of each candidate stage; This indicates consistency in short-term growth trends; to Let be the weight coefficients of each discriminant factor, and let their sum be 1; During online recognition, the weighted distance between the feature vector to be tested and the feature centers at each stage is first calculated. For the first stage... j The weighted distance between each candidate stage is expressed as: ； in, Indicates the first t The comprehensive feature vector Z(t) of the sample under test at time t and the first time j Candidate stage feature centers Weighted distance between them; K The dimension of the comprehensive feature vector; Indicates the first k Dimensional feature weights; Let Z(t) be the eigenvector of the composite feature vector at time t. k 3D eigenvalues; Indicates the first j Candidate stage feature centers The k Dimensional eigenvalues. The smaller the value, the better the sample being tested is compared to the first... j The more similar the characteristics of the candidate stages; Further calculation based on weighted distance j The distance similarity of each candidate stage is expressed as: ； in, Indicates the first j Distance similarity between candidate stages; For the first j The stage feature radius of each stage; ε is a constant to prevent the denominator from being zero. This formula converts the weighted distance into a similarity value in the range [0,1]. The smaller, The larger.

11. The method for identifying and classifying flashover stages of polluted insulators according to claim 10, characterized in that, The distance interval between the nearest stage center and the second nearest stage center is calculated to obtain the distance interval confidence score, which is expressed as follows: For the j The candidate phase, first in addition to the first... j Among the remaining stages other than the candidate stages, determine the stage with the smallest weighted distance to the sample to be tested, and record this smallest distance as . Then the first j The distance interval confidence of each candidate stage is expressed as: ； in, Indicates the first j The confidence level of the distance interval between candidate stages, if If the weighted distance is less than the weighted distance corresponding to other stages, then Take a positive value; if If it is not less than the minimum weighted distance in other stages, then Setting it to 0, the confidence level of this distance interval is used to determine the test sample relative to the first... j Does each candidate stage possess a distinguishable distance advantage? 12. The method for identifying and classifying flashover stages of polluted insulators according to claim 11, characterized in that, The step of calculating the confidence level of the stage feature radius corresponding to the candidate stage of the current frame based on the stage feature radius includes: Calculate the first based on the stage characteristic radius. j The confidence level of the stage feature radius of each candidate stage is expressed as: ； in, Indicates the first j The credibility of the stage feature radius of each candidate stage, if Located at the stage characteristic radius Within the range, then If it is a positive value; Exceeding the stage feature radius range, then Take 0.

13. The method for identifying and classifying flashover stages of polluted insulators according to claim 12, characterized in that, The combination of the transition probability from the previous stable stage to the candidate stage and the consistency between the current short-term growth rate and the typical growth direction of the candidate stage forms a comprehensive discrimination score for the candidate stage, including: The previous stable stage is the label of the stable stage output after the most recent adaptive timing smoothing, denoted as . For the first j The stage transition probability of each candidate stage is expressed as: ； in, Indicates the transition from the previous stable phase Transfer to the j The probability of each candidate stage is provided by the stage transition probability matrix obtained from the offline modeling stage. If no previous stable stage exists at the current time, then... Set all four candidate stages to the same value so that the transition probability term does not affect the identification result of the first stage; Short-term growth trend consistency is used to determine whether the current characteristic change direction is consistent with the typical growth direction of the candidate stage. The current short-term growth amount is represented as follows: For the first j For each candidate stage, the average growth direction is calculated based on the difference in the comprehensive feature vectors of adjacent frames belonging to that stage in the offline training samples, denoted as . , No. j The consistency of the short-term growth trends of the candidate stages is represented as follows: ； in, Indicates the first j The short-term growth trend of the candidate stage is consistent, and the value range is [0,1]. When the current short-term growth direction is consistent with the short-term growth trend of the candidate stage, the short-term growth trend is consistent with the short-term growth trend of the candidate stage. j When the typical growth direction of each candidate stage is consistent, Increase; when the two increase in opposite directions. Decrease.

14. The method for identifying and classifying flashover stages of polluted insulators according to claim 13, characterized in that, The constraint judgment on the preliminary identification results includes: During the candidate phase The result of the pollution flashover discharge stage identification at the current moment is determined when the following conditions are met simultaneously: First, the sample to be tested is in the candidate stage. Within the radius of the stage characteristic; second, Candidate stage It has the advantage of distance interval compared to other stages; Third, if a previous stable phase exists ,but Candidate stage There is a stage transition relationship supported by offline training samples between the previous stable stage and the previous stable stage.

15. A system for identifying and classifying flashover stages in polluted insulators, characterized in that, The system includes: The preprocessing module is used to simultaneously acquire dual-band ultraviolet images or video frame sequences of polluted insulator discharge using a solar-blind ultraviolet imaging channel and a UVA imaging channel, forming a dual-band image sequence to be identified, and preprocessing the dual-band image sequence to obtain a standardized dual-band six-dimensional feature vector, including: the total pixel area of ​​all discharge spot regions in the solar-blind ultraviolet image, the pixel area of ​​the single largest discharge spot region in the solar-blind ultraviolet image, the number of connected regions of independent discharge spots in the solar-blind ultraviolet image, the total pixel area of ​​all discharge spot regions in the UVA image, the pixel area of ​​the single largest discharge spot region in the UVA image, and the number of connected regions of independent discharge spots in the UVA image; The enhanced feature determination module is used to calculate the dual-band area response ratio, spot concentration and short-term growth of the dual-band image based on the standardized dual-band six-dimensional feature vector, that is, to identify enhanced features during the generation stage. The model training module is used to train the time-constrained weighted clustering model using pollution flashover test samples. The input features of the time-constrained weighted clustering model include the stage identification enhancement features. The four continuous intervals corresponding to the normalized time axis formed after normalization according to the time position are used as the initial cluster centers. After iterative training, the physical stages corresponding to the pollution flashover test samples are obtained. The physical stages include: initial stage, partial discharge stage, arc propagation stage and flashover breakdown stage. The physical stage determination module is used to calculate the comprehensive discrimination score of four candidate stages for each frame of the sample to be tested, and select the candidate stage with the largest comprehensive discrimination score as the preliminary identification stage of the current frame. The preliminary identification result is constrained and judged. When the candidate stage satisfies the constraint judgment at the same time, the current candidate stage is taken as the pollution flashover stage identification result at the current moment. The graded early warning module is used to determine the stable phase sequence under multiple frames based on the test sample, thereby generating a pollution flashover risk index, and outputting graded early warnings based on the pollution flashover risk index.

16. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the method for identifying and classifying flashover stages of polluted insulators as described in any one of claims 1 to 14.

17. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the method for identifying and classifying the flashover stage of polluted insulators as described in any one of claims 1 to 14.

Images