Method for visually detecting action state of arrester tripper based on edge computing

Abstract

The present application relates to power transmission and distribution equipment monitoring technical field, disclose a lightning arrester tripping device action state visual detection method based on edge computing, including: using lateral edge density frame bottom end candidate domain, based on radial alignment score reconstruction bottom end coaxial center and develop as polar coordinate system.On this basis, extract the outward opening positive radial projection gradient to construct the release seam gradient field and the angular density distribution, then carry out the first order harmonic expansion calculation eccentric release seam first order harmonic dominant degree, and combine the main direction release seam stretch and the one-way evolution quantity to carry out the threshold-free continuous mapping, output the continuous state probability.The present application only encapsulates the most evidence moment image to return and adaptively update the non-action baseline, completely get rid of the heavy image reasoning burden, realize the high-sensitive detection and closed-loop early warning of the early hidden danger of equipment with very low communication load.

Description

Technical Field

[0001] This invention relates to the field of power transmission and distribution equipment monitoring technology, and more specifically, to a visual detection method for the action status of surge arrester trip units based on edge computing. Background Technology

[0002] With the advancement of power distribution network automation, remote video inspections of surge arrester trip units are becoming increasingly common. Existing remote monitoring and analysis largely rely on identifying obvious signs of tripping components falling off. However, in real outdoor conditions, the analysis process faces two significant challenges. First, after a deflagration, the flexible cable detached from the arrester will oscillate irregularly due to the explosive force, its own weight, and wind load. The densely layered insulating skirts on the surge arrester surface will form dense lateral blocking strips within the monitoring field of view. When the cable's oscillation trajectory intersects with the skirts, the image will alternately present a pseudo-static appearance of movement at similar times, easily misleading the detection logic. Second, under the combined influence of moisture intrusion and surface contamination, the bottom encapsulation area often experiences concealed deformation, with one side preferentially detaching. This causes the lead terminals to first slightly eccentrically open on one side before fully detaching. These subtle precursors are often mixed in with the refraction and reflection of external water films and mud stains. Due to the lack of large-scale changes in appearance, current detection methods are extremely insensitive to them. Furthermore, edge-side monitoring nodes suffer from insufficient computing power and communication bandwidth, making it impossible for equipment to retain ultra-high-definition continuous images for detailed investigation over extended periods. Therefore, traditional methods often misreport actions as inactive events due to brief visual obstructions, or miss the crucial window for early warning by blindly relying on excessively large displacement thresholds. Summary of the Invention

[0003] This invention provides a visual detection method for the operating status of surge arrester trip units based on edge computing, which solves the technical problems mentioned in the background art.

[0004] This invention provides a visual detection method for the operating status of surge arrester trip units based on edge computing, applied to edge computing nodes equipped with visual acquisition devices, including: The system acquires the input video image sequence captured by the vision acquisition device, extracts the current frame image and the previous frame image within the current buffer window, and maps the timestamp of the input video image sequence to normalized time. Extract the normalized x-coordinate and normalized y-coordinate of the current frame image, use the lateral edge distribution density of the arrester skirt to locate the arrester axis and the bottom candidate domain, and capture the bottom candidate domain image corresponding to the bottom candidate domain. Within the candidate domain at the bottom end, the coaxial center at the bottom end is reconstructed based on the radial alignment score, and polar coordinate normalization is performed based on the coaxial center at the bottom end to construct a polar coordinate domain. In the polar coordinate domain, the outwardly expanding positive radial projection gradient is extracted to construct the gap-releasing gradient field, and the angular density distribution is generated by radial integration. The first harmonic expansion of the angular density distribution is performed to calculate the dominance and direction of the first harmonic of the eccentric release slit. The expansion amount of the release gap is extracted along the main harmonic direction, and the unidirectional evolution amount is calculated by combining the change of the first harmonic dominance of the eccentric release gap with the normalization time. The first harmonic dominance of the eccentric release gap, the gap extension amount, and the unidirectional evolution amount are continuously mapped without thresholds to obtain the continuous score corresponding to each action state, and the action state at the edge end is output. Within the current cache window, select the moment with the strongest action evidence, obtain the bottom candidate domain image captured at that moment and perform edge backpropagation with the action evidence weight map generated based on the release gradient field, and adaptively update the non-action baseline map corresponding to the previous frame image based on the action state at the edge.

[0005] The beneficial effects of this invention include: It overcomes the limitations of traditional visual judgment, which heavily relies on large displacement appearances, by deeply exploring the evolution mechanism of the tripping micro-gap in a coaxial reference frame, quantifying and purifying the chaotic and difficult-to-distinguish unilateral opening signs into the first harmonic dominance of the eccentric release gap. Utilizing a threshold-free mapping detection system based on radial extension measurement and temporal accumulation, this invention successfully eliminates redundant interference from water film reflection and dense skirt textures at computationally limited edge environments, enabling the sensitive capture of easily overlooked pre-action signals in the early stages of action evolution. Furthermore, by employing a scoring logic based on continuous smooth mapping combined with temporal evolution evaluation indicators, this invention successfully eliminates the pseudo-static illusion caused by flexible cable swing and skirt interruption. By extracting and encapsulating simplified evidence only from the highest confidence segment and dynamically updating the non-action background benchmark locally, while retaining complete maintenance audit credentials, the burden on the wireless network is reduced to an extremely low level, achieving efficient capture and zero-missed monitoring of early arrester failure conditions. Attached Figure Description

[0006] Figure 1 This is a flowchart of the visual detection method for the operation status of surge arrester trip unit based on edge computing according to the present invention. Detailed Implementation

[0007] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0008] like Figure 1 As shown, the visual detection method for the operation status of surge arrester tripping devices based on edge computing is applied to edge computing nodes equipped with visual acquisition devices, including: The system acquires the input video image sequence captured by the vision acquisition device, extracts the current frame image and the previous frame image within the current buffer window, and maps the timestamp of the input video image sequence to normalized time. Extract the normalized x-coordinate and normalized y-coordinate of the current frame image, use the lateral edge distribution density of the arrester skirt to locate the arrester axis and the bottom candidate domain, and capture the bottom candidate domain image corresponding to the bottom candidate domain. Within the candidate domain at the bottom end, the coaxial center at the bottom end is reconstructed based on the radial alignment score, and polar coordinate normalization is performed based on the coaxial center at the bottom end to construct a polar coordinate domain. In the polar coordinate domain, the outwardly expanding positive radial projection gradient is extracted to construct the gap-releasing gradient field, and the angular density distribution is generated by radial integration. The first harmonic expansion of the angular density distribution is performed to calculate the dominance and direction of the first harmonic of the eccentric release slit. The expansion amount of the release gap is extracted along the main harmonic direction, and the unidirectional evolution amount is calculated by combining the change of the first harmonic dominance of the eccentric release gap with the normalization time. The first harmonic dominance of the eccentric release gap, the gap extension amount, and the unidirectional evolution amount are continuously mapped without thresholds to obtain the continuous score corresponding to each action state, and the action state at the edge end is output. Within the current cache window, select the moment with the strongest action evidence, obtain the bottom candidate domain image captured at that moment and perform edge backpropagation with the action evidence weight map generated based on the release gradient field, and adaptively update the non-action baseline map corresponding to the previous frame image based on the action state at the edge.

[0009] Preferably, the step of locating the arrester axis and bottom candidate region by utilizing the lateral edge distribution density of the arrester skirt includes: Obtain the normalized brightness of the current frame image. The horizontal edge density is obtained by calculating the vertical gradient magnitude of the normalized brightness and dividing it by the maximum gradient magnitude within the current image region. : The lateral edge density is integrated along the longitudinal direction and then normalized by the lateral integration to obtain the lateral edge quality distribution. : The abscissa of the arrester axis is calculated using the weighted integral of the normalized abscissa under the lateral edge mass distribution. The lateral discrete scale is calculated using the square root of the weighted variance. : Within the horizontal coordinate neighborhood of the arrester axis, the lateral edge density is integrated along the lateral direction and then normalized by longitudinal integration to obtain the longitudinal edge mass distribution. : The ordinate of the bottom candidate domain is obtained by dividing the integral of the square of the normalized ordinate under the longitudinal edge quality distribution by the integral of the normalized ordinate under the longitudinal edge quality distribution. : The bottom candidate region is defined based on the abscissa of the arrester axis, the lateral discrete scale, and the ordinate of the bottom candidate region. : in, These are the normalized x-axis and normalized y-axis, respectively; Normalized time; Normalized brightness; Horizontal edge density; The current image region; The mass distribution is at the lateral edge. The horizontal coordinate of the surge arrester axis; It is a horizontal discrete scale; The quality distribution is at the longitudinal edge. The ordinate of the bottom candidate domain; This is the bottom candidate domain; These are the horizontal and vertical spatial integral elements, respectively.

[0010] The current frame image is the original image frame acquired by the vision acquisition device at the current moment, which can be continuously acquired by an industrial camera installed in the observation direction of the surge arrester.

[0011] Normalized luminance is the luminance representation of each pixel in the current frame image after grayscale conversion, luminance normalization, and illumination correction.

[0012] The longitudinal gradient magnitude is the absolute measure of the strength of the change in normalized brightness along the vertical direction of the image, used to characterize the salience of the lateral edge of the surge arrester skirt.

[0013] The current image region is the spatial range that participates in the maximum gradient magnitude search and the lateral edge density calculation, and is used to limit the image region covered by a single calculation.

[0014] The maximum gradient magnitude is the maximum value among all vertical gradient magnitudes of all pixels in the current image region, and is used to normalize the density of horizontal edges.

[0015] The lateral edge density is the edge intensity distribution obtained by normalizing the longitudinal gradient magnitude of the normalized brightness after normalizing it to the maximum gradient magnitude.

[0016] The intermediate lateral integral is the unnormalized lateral cumulative quantity obtained by integrating the lateral edge density along the normalized ordinate direction.

[0017] The lateral edge mass distribution is the lateral mass distribution obtained by normalizing the lateral integral of the intermediate lateral quantity. It is used to characterize the relative concentration of lateral edge energy at different normalized abscissas.

[0018] Normalized x-coordinates are position coordinates obtained by normalizing the horizontal position of the current frame image according to the image width, and are used to unify the horizontal spatial scale of images with different resolutions.

[0019] The horizontal coordinate of the surge arrester axis is the weighted center position of the normalized horizontal coordinate under the lateral edge mass distribution, used to represent the lateral center of the surge arrester body.

[0020] Weighted variance is a second-order statistic formed under the lateral edge mass distribution, representing the degree of dispersion of the normalized abscissa relative to the abscissa of the arrester axis.

[0021] The lateral discrete scale is the lateral scale quantity obtained by taking the square root of the weighted variance, and is used to characterize the lateral width of the surge arrester body in the image.

[0022] The abscissa neighborhood of the surge arrester axis is a local lateral analysis region centered on the abscissa of the surge arrester axis and defined by a lateral discrete scale.

[0023] The intermediate longitudinal integral is the unnormalized longitudinal cumulative quantity obtained by integrating the lateral edge density along the lateral direction within the horizontal coordinate neighborhood of the arrester axis.

[0024] The longitudinal edge mass distribution is the longitudinal mass distribution obtained by normalizing the longitudinal integral of the middle longitudinal quantity. It is used to characterize the degree of concentration of edge energy along the normalized ordinate.

[0025] Normalized ordinates are position coordinates obtained by normalizing the vertical position of the current frame image according to the image height, and are used to unify the vertical spatial scale of images with different resolutions.

[0026] The ordinate of the bottom candidate region is a bottom position index determined based on the first and second order statistical results of the normalized ordinate under the longitudinal edge quality distribution, and is used to locate the starting longitudinal position of the bottom candidate region of the surge arrester.

[0027] The bottom candidate domain is the bottom spatial region defined by combining the horizontal coordinate of the arrester axis, the horizontal discrete scale, and the vertical coordinate of the bottom candidate domain.

[0028] A transverse spatial integral element is an integral placeholder used when integrating in the transverse direction, used to represent the traversal variable in the transverse integration process.

[0029] The longitudinal spatial integral element is an integral placeholder used when integrating in the longitudinal direction, used to represent the traversal variable in the longitudinal integration process.

[0030] In practice, for the generation of normalized brightness, the current frame image needs to be converted into a single-channel grayscale image first, and then the pixel values ​​are linearly mapped to the 0 to 1 range according to the camera bit depth. For brightness drift caused by backlight, shadow, water film reflection and stains, 3×3 median filtering is first performed to suppress isolated noise, then large-scale smoothing is performed to estimate the background illumination, and the original grayscale value is subtracted from the background illumination and then stretched back to the 0 to 1 range. For nighttime lighting scenes, percentile stretching can be used to compress the low percentile brightness to 0 and the high percentile brightness to 1, and the excess part is saturated and cropped.

[0031] In practice, for the calculation of the vertical gradient magnitude and the maximum gradient magnitude, a 3×3 vertical Sobel difference kernel or an equivalent central difference kernel is used to obtain the vertical gradient on the normalized brightness, and then the absolute value of the gradient result is taken to obtain the vertical gradient magnitude; the maximum gradient magnitude is obtained by searching pixel by pixel in the current image area; if the maximum gradient magnitude is close to 0, the denominator should be replaced with a very small positive number to prevent division by zero error in the horizontal edge density.

[0032] In practical implementation, for the discrete implementation of integration and normalization, the integral in the formula needs to be summed in the program according to the pixel grid. First, each pixel is mapped to the corresponding normalized horizontal and vertical coordinates, and then accumulated along the specified direction. When normalizing the intermediate horizontal and vertical integral quantities, the denominator should be set to the sum of the whole domain. When the denominator is less than the preset minimum positive number, it should directly revert to a uniform distribution or use the stable result of the previous frame.

[0033] In specific implementation, for the boundary processing of the neighborhood of the horizontal coordinate of the surge arrester axis, it is necessary to form a horizontal integration interval with the horizontal coordinate of the surge arrester axis as the center and the horizontal discrete scale as half width. Then, it is necessary to first determine whether the interval exceeds the normalized horizontal range of 0 to 1. If the left boundary is less than 0, it is truncated to 0. If the right boundary is greater than 1, it is truncated to 1. If the effective interval width is too narrow, at least one minimum pixel width neighborhood must be retained with the horizontal coordinate of the surge arrester axis as the center.

[0034] In practice, for the delineation, cropping, and pixel mapping of the bottom candidate region, it is necessary to first form the candidate region in the normalized coordinate system according to the formula, and then map the bottom candidate region back to pixel coordinates according to the current image width and height. The horizontal range is taken as two horizontal discrete scales to the left and right of the horizontal coordinate of the lightning arrester axis, and the vertical range is taken as the vertical coordinate of the bottom candidate region minus one horizontal discrete scale to the bottom edge of the image. If the width or height of the bottom candidate region after mapping is less than the minimum analysis size, it is expanded according to the minimum analysis size. If part of the bottom candidate region exceeds the image boundary, the excess part is cropped and the pixel index is corrected simultaneously.

[0035] Preferably, the step of reconstructing the coaxial center of the bottom end based on the radial alignment score within the bottom end candidate domain, and performing polar coordinate normalization based on the coaxial center of the bottom end to construct a polar coordinate domain includes: Candidate coaxial centers are set within the bottom candidate domain. Construct a polar coordinate radial unit vector pointing from the center of the candidate coaxial axis to each pixel. ; Calculate the gradient vector of the normalized brightness. The magnitude of the gradient vector is extracted as the absolute gradient. The surface positive component of the absolute gradient within the bottom candidate domain is calculated as the absolute gradient surface positive component. Calculate the maximum non-negativity of the dot product of the gradient vector and the polar coordinate radial unit vector. Divide the surface positive fraction of the maximum non-negativity in the bottom candidate domain by the absolute gradient surface positive fraction to obtain the radial alignment score of the candidate coaxial center. : The position that maximizes the radial alignment score is selected as the bottom coaxial center. : Using the distance from the pixel to the coaxial center of the bottom edge The reference radius is obtained by calculating a weighted average based on the absolute gradient. : Divide the distance by the reference radius to convert it into polar coordinate radius. The azimuth angle of the pixel relative to the coaxial center at the bottom is normalized to polar coordinates using a circular method. To construct the polar coordinate domain: in, It is a radial unit vector in polar coordinates; The absolute gradient; Radial alignment score; For the surface active differential element within the bottom candidate domain; The bottom end is coaxially centered; The coordinate components are the coaxial center at the bottom end; This is the distance from the pixel to the center of the coaxial axis. Used as a reference radius; The radius is in polar coordinates; The angle is in polar coordinates.

[0036] The candidate coaxial center is the central hypothetical point inside the bottom candidate domain used to participate in the radial alignment scoring search.

[0037] The x-coordinate of the candidate coaxial center is the horizontal coordinate component of the candidate coaxial center in the image coordinate system.

[0038] The ordinate of the candidate coaxial center is the vertical coordinate component of the candidate coaxial center in the image coordinate system.

[0039] The polar coordinate radial unit vector is a unit direction vector pointing from the center of the candidate coaxial direction to a pixel position within the bottom candidate domain. It is used to determine whether the gradient direction is consistent with the radial outward direction.

[0040] The gradient vector is a combination of changes in normalized brightness in both the horizontal and vertical directions, used to characterize the direction and intensity of local brightness changes.

[0041] The magnitude of the gradient vector is the length of the gradient vector, used to characterize the total intensity of local brightness changes.

[0042] The absolute gradient is the non-negative intensity quantity corresponding to the magnitude of the gradient vector, used to form a normalized benchmark for radial alignment scoring.

[0043] The absolute gradient surface positive fraction is the total intensity obtained by accumulating the absolute gradient within the bottom candidate domain.

[0044] The surface integral element is a two-dimensional integral placeholder used for area integration within the bottom candidate domain, used to represent the pixel-by-pixel surface accumulation process.

[0045] The radial alignment score is the ratio of the intensity of the gradient vector and the polar coordinate radial unit vector projected onto the bottom candidate domain to the total intensity of the absolute gradient. It is used to evaluate whether the candidate coaxial center is consistent with the true bottom center.

[0046] The bottom coaxial center is the true center estimate position that maximizes the radial alignment score.

[0047] The x-coordinate component of the bottom coaxial center is the horizontal coordinate component of the bottom coaxial center.

[0048] The ordinate component of the bottom coaxial center is the longitudinal coordinate component of the bottom coaxial center.

[0049] The distance from a pixel to the coaxial center is the distance between any pixel in the bottom candidate domain and the determined bottom coaxial center.

[0050] The reference radius is the baseline radius obtained by weighting the distance from the pixel to the coaxial center according to the absolute gradient, and is used to establish the scale-independent polar coordinate radius.

[0051] The azimuth angle is the circumferential angle position of a pixel relative to the direction of its bottom coaxial center, used to characterize the directional distribution of pixels on a circle.

[0052] The polar radius is the radius normalized relative to a reference radius, representing the distance from a pixel to the coaxial center. It is used to standardize the radial scale for different device sizes and shooting distances.

[0053] Polar coordinate angle is an angular position quantity formed by normalizing the azimuth angle by one revolution.

[0054] In practice, for setting the candidate coaxial center, the bottom candidate domain needs to be mapped to a pixel coordinate region first, and then the outer ring pixels that are too close to the boundary of the bottom candidate domain are removed to avoid radial vector distortion caused by the candidate center being too close to the boundary. In the remaining region, a candidate grid is established with a step size of 1 or 2 pixels, and the radial alignment score is calculated for each candidate point. When the resolution is high, a coarse search can be performed with a large step size first, and then a fine search can be performed with a small step size in the high-scoring region.

[0055] In practice, for the discrete acquisition of gradient vector and absolute gradient, gradient components in two directions need to be calculated on the normalized brightness using horizontal and vertical difference kernels respectively, and then the two components are combined to form gradient vector; the magnitude of gradient vector is absolute gradient; to reduce noise disturbance to gradient direction, mild Gaussian smoothing can be performed on normalized brightness before difference.

[0056] In practice, for the discrete integration, extremum search, and parallel extremum processing of radial alignment scoring, it is necessary to perform a dot product between the gradient vector of each pixel in the bottom candidate domain and the radial unit vector pointing from the candidate coaxial center to that pixel, retain only the non-negative part, and accumulate it pixel by pixel to obtain the numerator; accumulate the absolute gradient pixel by pixel to obtain the denominator; divide the two to obtain the radial alignment score; if multiple candidate centers obtain the same highest score, the candidate point closest to the geometric center of the bottom candidate domain is selected first; if they are still the same, the candidate point with the smallest distance from the bottom coaxial center of the previous frame is selected first.

[0057] In practice, for the effective pixel set of the reference radius and zero denominator protection, it is necessary to prioritize the use of all pixels in the bottom candidate domain to participate in the weighted average. Pixels with too small absolute gradients can be given very low weights or ignored directly. If the denominator is too small, it means that the current frame lacks reliable edges, and it should revert to the geometric mean radius of the bottom candidate domain or use the reference radius of the previous frame.

[0058] In practice, for the sampling resolution, interpolation method and angle normalization constructed in the polar coordinate domain, it is necessary to first take the bottom coaxial center as the pole and map each pixel in the bottom candidate domain to the polar coordinate radius and polar coordinate angle; the polar coordinate angle should uniformly fall into the range of 0 to 1, the starting point of the circle is fixed in the positive horizontal direction of the image, and the process should be consistent throughout the entire process by rotating counterclockwise or clockwise; for resampling from the Cartesian grid to the polar coordinate grid, bilinear interpolation should be used first.

[0059] Preferably, the step of extracting the outwardly projected positive radial gradient in the polar coordinate domain to construct the gap-releasing gradient field and generating an angular density distribution by radial integration includes: In the polar coordinate domain, calculate the gradient vector and the polar coordinate radial unit vector. The maximum non-negativity of the dot product is used to construct the gap-releasing gradient field by dividing it by the sum of the integrals of the maximum non-negativity in the polar coordinate domain. : Calculate the absolute value of the difference between the polar radius and the numerical value, multiply the absolute value by the release gradient field, and take a double integral in the polar coordinate domain to obtain the adaptive radial spread. : The angular density distribution is generated by radially integrating the release gradient field within a radial interval bounded by the adaptive radial expansion amount and a numerical value of 1. : in, It is a radial unit vector in polar coordinates; Polar coordinates; For the release gradient field; For adaptive radial expansion; It exhibits an angular density distribution.

[0060] The polar coordinate domain is the spatial range of the bottom region with the coaxial center at the bottom as the pole, and expressed by polar coordinate radius and polar coordinate angle.

[0061] The polar coordinate radial unit vector is a radial unit direction quantity that points from the coaxial center at the bottom to a point in the polar coordinate domain. It is used to represent the reference direction of outward expansion.

[0062] The release gradient field is a gradient field formed by normalizing the positive projection of the gradient vector in the radial unit vector direction of polar coordinates. It is used to highlight possible unilateral release regions.

[0063] The adaptive radial extension is the degree of deviation of the polar coordinate radius from the reference radius position. It is a ring width index obtained by weighting under the release gradient field and is used to adaptively determine the radial integration interval.

[0064] The angular density distribution is the angular distribution quantity obtained by radially integrating the release gradient field within an adaptive ring around the reference radius. It is used to characterize the degree of concentration of release evidence in the circumferential direction.

[0065] In practice, for the alignment calculation of the gradient vector and the radial unit vector in the polar coordinate domain, it is necessary to first calculate the gradient vector in the original image coordinate system to obtain the horizontal gradient component and the vertical gradient component of each pixel. When entering the polar coordinate domain, there is no need to recalculate the gradient. Only the radial unit vector is recalculated for the same pixel position based on the bottom coaxial center. Then, the original gradient vector and the radial unit vector are multiplied by a dot product. When the dot product result is negative, it is truncated to 0.

[0066] In practice, to protect the normalized denominator of the release gradient field, it is necessary to first calculate the non-negative value of the gradient vector multiplied by the radial unit vector in the polar coordinate domain pixel by pixel, and then sum all the non-negative values ​​as the denominator. If the denominator is too small, it means that the current frame lacks reliable external expansion evidence. In this case, the release gradient field should be set to a zero field, or the stable result of the previous frame should be used.

[0067] In practical implementation, for the upper and lower limit constraints of the adaptive radial expansion, the average radius deviation needs to be calculated according to the formula and then restricted to the effective radius range. It must not exceed the remaining radius range between the outer boundary of the polar coordinate domain and the reference radius position. If the calculation result is too large, it should be truncated to the upper limit of the allowable range. If the result is too small, a minimum annular thickness should be maintained to ensure the computability of the angular integral.

[0068] In practice, for the radial interval truncation around the value 1, the radial integration interval needs to be centered on the reference radius position, with the left and right boundaries set as 1 minus the adaptive radial expansion and 1 plus the adaptive radial expansion, respectively; if the left boundary is less than the minimum radius of the polar coordinate domain, it is truncation to the minimum radius; if the right boundary is greater than the maximum radius of the polar coordinate domain, it is truncation to the maximum radius.

[0069] In practice, for the discrete integration step size and annular zone realization of the angular density distribution, it is necessary to first determine the number of angular samples and the number of radial samples, and then accumulate the release gradient field radially within the effective annular zone around the reference radius to form a one-dimensional angular value at each angular position; the number of angular samples can be set to 72, 90 or 180 discrete angles according to the image resolution and the size of the bottom region.

[0070] Preferably, the step of performing first-order harmonic expansion on the angular density distribution and calculating the dominance and direction of the first-order harmonic of the eccentric release slit includes: The first-order cosine coefficients are obtained by multiplying the angular density distribution with a circular cosine function based on the polar coordinate angle and performing discrete integration. : The first-order sine coefficients are obtained by multiplying the angular density distribution with a circumferential sine function based on the polar coordinate angle and performing discrete integration. : The square root of the sum of the squares of the first-order cosine coefficient and the squares of the first-order sine coefficient is taken to obtain the first-order harmonic dominance of the eccentric release slit. : Calculate the arctangent phase angle of the quotient of the first-order sine coefficient and the first-order cosine coefficient, and divide the arctangent phase angle by the total radians of the circle to obtain the principal harmonic direction. : in, These are the first-order cosine coefficients and the first-order sine coefficients, respectively; The first harmonic dominance of the eccentric release gap; The main harmonic direction.

[0071] The circular cosine function is a cosine basis function with polar coordinate angle as the independent variable, used to extract the first-order periodic component of a circle.

[0072] The first-order cosine coefficients are the first-order cosine component coefficients obtained by multiplying the angular density distribution with the circular cosine function and then performing discrete integration.

[0073] The circumferential sine function is a sine basis function with polar coordinate angle as the independent variable, used to extract the first-order periodic component of a circle.

[0074] The first-order sine coefficients are the first-order sine component coefficients obtained by multiplying the angular density distribution with the circumferential sine function and then performing discrete integration.

[0075] The first harmonic dominance of the eccentric release gap is the first harmonic amplitude obtained by synthesizing the first-order cosine coefficient and the first-order sine coefficient. It is used to measure whether the gap release evidence shows obvious unilateral eccentricity.

[0076] The arctangent phase angle is the first harmonic direction angle determined by the first-order sine coefficient and the first-order cosine coefficient.

[0077] The total radian of a circle is the total angle corresponding to a complete revolution, used to normalize the phase angle to a uniform directional scale.

[0078] The principal harmonic direction is the main release direction index obtained by normalizing the arctangent phase angle according to the total radian of the circle. It is used to indicate the direction position where the eccentric release is most significant.

[0079] In practice, for discrete sampling of angular density distribution, the polar coordinate angle needs to be uniformly divided into a fixed number of angular sampling units, and then the annular pixels in each sampling unit are radially integrated to form a discrete angular sequence; the beginning and end of the sequence must be connected to ensure circumferential closure.

[0080] In practice, for the discrete summation and periodic closure processing of the first-order cosine coefficients and the first-order sine coefficients, it is necessary to multiply the angular density value of each discrete angle unit by the corresponding cosine base value and sine base value, and then perform full circumference summation; the first angle unit and the last angle unit are considered to be adjacent, and the circumferential sequence must not be treated as an open sequence.

[0081] In practical implementation, regarding the range and normalization of the first harmonic dominance of the eccentric release gap, it is necessary to ensure that the total amount of the release gap gradient field has been normalized in the polar coordinate domain, and that the angular density distribution originates from the radial accumulation of the ring zone near the reference radius, so that the first harmonic dominance of the eccentric release gap falls within the range of 0 to 1. In the program implementation, the first harmonic dominance of the eccentric release gap can be limited to 0 to 1 to prevent slight out-of-bounds errors caused by discrete errors from entering subsequent continuous mappings.

[0082] In specific implementation, when determining the main harmonic direction in low-energy or double-zero scenarios, if the first-order cosine coefficient and the first-order sine coefficient are both very close to 0, it indicates that the current angular distribution is approximately uniform or there is no effective evidence of gap release. In this case, the main harmonic direction is not considered a reliable direction, and the main harmonic direction of the previous frame can be directly used, or it can be marked as an invalid direction and the direction-related calculation is paused.

[0083] Preferably, the step of extracting the release gap extension along the dominant harmonic direction and calculating the unidirectional evolution amount by combining the cumulative change of the first harmonic dominance of the eccentric release gap with the normalized time includes: Construct the maximum non-negativity value of the circular cosine function with the difference between the polar coordinate angle and the principal harmonic direction as the variable, and use it as the cosine weighting function. : Multiplying the release gradient field by the cosine weighted function and integrating along polar coordinates, then dividing by the sum of the angular integrals of the cosine weighted function, yields the radial distribution of the principal harmonic directions. : Obtain the maximum polar coordinate radius corresponding to the boundary of the bottom candidate domain. Within the range from the numerical value to the maximum polar radius, the dimensionless first moment radially distributed in the principal harmonic direction is calculated to obtain the release tension. : Extract the differential increment of the first harmonic dominance of the eccentric release gap with respect to the normalized time, take the absolute value of the differential increment as the absolute differential increment, and take the differential increment that is greater than zero as the positive differential increment. Dividing the time integral of the positive differential increment by the time integral of the absolute differential increment yields the unidirectional evolution quantity. : in, It is a cosine weighted function; Radial distribution in the main harmonic direction; The maximum polar radius; This is the amount of seam stretch; It is a unidirectional evolution quantity; It is a time integral infinitesimal element.

[0084] The cosine-weighted function is a non-negative weighted function constructed using the difference between the polar coordinate angle and the principal harmonic direction as a variable. It is used to highlight the gap release evidence near the principal harmonic direction.

[0085] The difference between the polar coordinate angle and the principal harmonic direction is a circumferential difference that describes the degree of deviation of any angular position relative to the principal harmonic direction.

[0086] The radial distribution in the principal harmonic direction is the radial distribution of the release gradient field along the angular direction after being integrated by the cosine weighted function. It is used to characterize the radial extension of the release energy in the principal harmonic direction.

[0087] The bottom candidate domain boundary is the outer contour interface of the bottom candidate domain in space, used to determine the maximum radius range that can be reached during polar coordinate analysis.

[0088] The maximum polar radius is the maximum effective radius obtained by mapping the bottom candidate domain boundary to the polar coordinate domain, and is used to limit the upper bound of the integral of the release gap extension.

[0089] The dimensionless first moment is a normalized first-order statistic formed by the radial offset of the principal harmonics distributed radially in the interval outside the reference radius.

[0090] The release extension is a dimensionless first moment result radially distributed in the principal harmonic direction outside the reference radius, used to represent the extent to which the release extends outward along the principal direction.

[0091] The differential increment is the instantaneous increase or decrease in the dominance of the first harmonic of the eccentric release slit as a function of normalized time.

[0092] The absolute differential increment is the total change obtained by taking the absolute value of the differential increment, and is used to statistically analyze the total intensity of the change in dominance.

[0093] The positive differential increment is the part of the differential increment that is greater than 0, and it is used to characterize the positive growth of dominance over time.

[0094] The unidirectional evolution quantity is the proportion of the time integral of the positive differential increment to the time integral of the absolute differential increment, and is used to measure whether the first harmonic dominance of the eccentric release slit shows a continuous unidirectional increase along time.

[0095] A time integral infinitesimal is an integral placeholder used when integrating with normalized time, used to represent the traversal variable in the time series integration process.

[0096] In practice, for the circular processing of angle differences in the cosine weighted function, the difference between the polar coordinate angle and the principal harmonic direction needs to be calculated as the shortest difference on the circle to ensure that the angles on both sides of the starting point of the circle are still considered to be adjacent to each other. After the calculation is completed, the difference is fed into the cosine function and the negative values ​​are truncated.

[0097] In practice, for the discrete integral and denominator protection of the radial distribution of the main harmonic direction, it is necessary to first collect all angle samples at each fixed radius position, then multiply the release gradient field and cosine weighted function angle by angle and accumulate them, and finally divide by the sum of the angular weights; if the sum of the angular weights is too small, it means that the current main direction is unreliable and should be reverted to zero distribution or the stable distribution of the previous frame should be used.

[0098] In practice, to obtain the maximum polar coordinate radius, it is necessary to emit rays from the coaxial center at the bottom to all angular directions, calculate the radius of the farthest intersection point between each ray and the boundary of the bottom candidate domain, and then select the maximum value within the effective angular region participating in the principal harmonic direction analysis. If a discrete polar coordinate grid is used, the maximum reachable radius can be directly read from the effective radius index after the bottom candidate domain is mapped.

[0099] In specific implementation, for the discrete realization of the dimensionless first moment of the release extension and the handling of abnormal boundaries, it is necessary to accumulate the radial distribution of the main harmonic direction point by point according to the radius offset outside the reference radius position, that is, within the effective interval from 1 to the maximum polar coordinate radius. Then, divide by the length of the reachable extension interval and the total amount of the radial distribution of the main harmonic direction to obtain the dimensionless first moment. If the maximum polar coordinate radius is not greater than 1, it means that there is no effective space outside. At this time, the release extension should be set to 0.

[0100] In specific implementation, for the time-series calculation of differential increment, absolute differential increment, positive differential increment, and unidirectional evolution, it is necessary to use the eccentric slit first harmonic dominance sequence sorted by time within the current buffer window as the basis. The discrete differential increment is obtained by dividing the difference of the eccentric slit first harmonic dominance of adjacent frames by the adjacent normalized time difference. The positive part is accumulated to obtain the total positive change, and the absolute value of all differences is accumulated to obtain the total change. The unidirectional evolution is then formed by the ratio of the two. When the sampling time interval is not uniform, the interval normalized by the real timestamp must be used, and the frame number must not be used directly.

[0101] Preferably, the step of performing a threshold-free continuous mapping of the first harmonic dominance of the eccentric release gap, the gap extension, and the unidirectional evolution to obtain a continuous score corresponding to each action state, and outputting the edge-end action state, includes: The inactive score is obtained by multiplying the difference between the numerical value and the first harmonic dominance of the eccentric release slot by the difference between the numerical value and the unidirectional evolution. : Multiply the first harmonic dominance of the eccentric release gap, the unidirectional evolution, and the difference between the numerical value and the release gap extension to obtain the quasi-action score. : Multiplying the first harmonic dominance of the eccentric release gap by the gap extension amount yields the action score. : Perform natural exponential normalization on the inaction score, the quasi-action score, and the action score to obtain the corresponding inaction state probability, quasi-action state probability, and action state probability, respectively. : The category corresponding to the extreme value probability is output as the edge action state. : in, The scores are respectively: no action score, near action score, and action score; For the corresponding state probabilities, and Index for status categories; This represents the action state at the edge.

[0102] The inactive score is a continuous score of inactive state obtained by mapping the first harmonic dominance of the eccentric release slit and the unidirectional evolution quantity. It is used to characterize the degree to which the equipment remains inactive.

[0103] The quasi-action score is a continuous score of quasi-action obtained by mapping the first harmonic dominance of the eccentric release gap, the unidirectional evolution, and the gap extension. It is used to characterize the degree to which the equipment has eccentric opening precursors but has not yet fully activated.

[0104] The action score is a continuous action score obtained by mapping the first harmonic dominance of the eccentric release slot and the release slot extension, and is used to characterize the degree to which the equipment has formed a clear action appearance.

[0105] State probability is the class probability obtained by normalizing the inaction score, quasi-action score, quasi-action score, and action score using the natural exponent.

[0106] The state category index is the category number corresponding to the state probability, where 0 corresponds to the inactive state, 1 corresponds to the quasi-active state, and 2 corresponds to the active state.

[0107] The state category index is the summation number in the denominator of the natural index normalization denominator used to iterate through all category scores.

[0108] The edge action state is the category output result corresponding to the extreme probability in the state probability.

[0109] In specific implementation, to address the range constraints of the first harmonic dominance, venting extension, and unidirectional evolution of the eccentric venting before entering continuous mapping, it is necessary to perform range correction on the first harmonic dominance, venting extension, and unidirectional evolution of the eccentric venting before calculating the non-action score, quasi-action score, and action score, so that they stably fall within the 0 to 1 range; values ​​less than 0 due to discrete errors are truncated to 0, and values ​​slightly greater than 1 are truncated to 1.

[0110] In practice, for the natural exponent normalization and stabilization of state probabilities, it is necessary to first find the maximum score among the no-action score, quasi-action score, and action score, then subtract the maximum score from each score and perform the natural exponent calculation, and finally normalize the denominator to avoid numerical overflow when the score difference is too large.

[0111] In practice, when the edge action states have extreme probabilities, if the probabilities of two or three states are completely equal or the difference is less than the minimum discriminant, the higher risk state should be output according to the risk priority principle, with the priority being that the action has been taken higher than the quasi-action, and the quasi-action has been taken higher than the no-action. If the system emphasizes conservative feedback, the quasi-action can also be output and feedback review can be triggered when they are in parallel.

[0112] Preferably, the step of selecting the moment with the strongest action evidence within the current cache window, obtaining the bottom candidate domain image captured at that moment, performing edge backpropagation with the action evidence weight map generated based on the gap gradient field, and adaptively updating the non-action baseline map corresponding to the previous frame image based on the edge action state includes: Within the current cache window, find the position where the sum of the probabilities of the quasi-action state and the action state reaches its maximum, and mark it as the moment with the strongest evidence of action. : The release gradient field at the moment with the strongest action evidence is normalized to its maximum value to generate the action evidence weight map. : The evidence tightness is obtained by dividing the surface area distribution of the action evidence weight map within the bottom candidate region by the global surface area distribution of the bottom candidate region. : The edge backpropagation is achieved by encapsulating the action evidence weight map, the evidence compactness, and the bottom candidate region image captured at that moment; the probability of the inactive state at that moment is extracted as the inactive confidence parameter. : Save the inactive baseline map from the previous frame. Multiplying the difference between the numerical value and the inactive confidence parameter by the value minus the value of the inactive confidence parameter, and adding the product of the normalized brightness and the inactive confidence parameter at that moment, yields the updated inactive baseline map. : in, The moment with the most evidence of action; For action evidence weighting; This represents the spatial extent of the bottom candidate domain in polar coordinates. For the tightness of evidence; For parameters indicating no action; This is the baseline image of the previous frame that was not in action. This is the updated baseline map of inactive data.

[0113] The current cache window is the time window used by edge nodes to save consecutive frames, perform temporal accumulation, search for the moment with the most action evidence, and complete the edge backhaul selection. It is preferably 12 to 32 frames. This range can ensure the temporal stability of dominance and evolution, while avoiding the introduction of historical trailing caused by cable swing and wind load disturbance due to excessively long windows, and at the same time reducing edge-side storage occupation and response latency.

[0114] The sum of the quasi-action state probability and the action state probability is a comprehensive probability measure used to assess the strength of action evidence at a given moment, and is used to select the moment with the strongest action evidence from the current cache window.

[0115] The moment with the strongest evidence of action is the position within the current cache window where the sum of the probabilities of the quasi-action state and the already-action state reaches its maximum, which is used to determine the most suitable evidence frame for transmission.

[0116] The action evidence weight map is a weight distribution map formed by normalizing the release gradient field corresponding to the moment with the strongest action evidence. It is used to intuitively mark the spatial location and strength of action evidence in the bottom candidate domain image.

[0117] The spatial range corresponding to the bottom candidate domain in the polar coordinate system is the effective integration region formed after the bottom candidate domain is mapped to the polar coordinate domain, which is used to calculate the tightness of evidence.

[0118] Evidence compactness is the average weight level of the action evidence weight map in the spatial range corresponding to the bottom candidate domain in the polar coordinate system, used to characterize whether action evidence is concentrated and compact.

[0119] The bottom candidate domain image is a local image extracted from the frame corresponding to the moment with the most action evidence according to the bottom candidate domain.

[0120] The inaction confidence parameter is the probability of the inaction state corresponding to the moment with the strongest evidence of action, and is used to adjust the weight of the inaction baseline map to update the current normalized brightness.

[0121] The previous frame is the video image that is immediately preceding the current frame in time.

[0122] The inactive baseline map is the inactive background baseline map saved in the previous round, used to continuously record the stable appearance of the device in an inactive state.

[0123] The updated inactive baseline map is a new baseline map obtained by fusing the inactive baseline map with the current normalized brightness weighted based on the inactive confidence parameter.

[0124] In practice, a fixed-length sliding time window is required to be used to determine the length of the current buffer window, the sliding method, and the granularity of the search for the moment with the most action evidence. The window slides forward by one frame for each new image received, and all temporal features within the window are updated synchronously. The moment with the most action evidence is searched frame by frame within the window without skipping frames.

[0125] In practice, for the protection of the maximum value normalized to zero denominator of the action evidence weight map, it is necessary to first search for the global maximum value in the release gradient field corresponding to the moment with the most action evidence, and then divide the A value of each pixel by the maximum value to obtain the action evidence weight map. If the maximum value is close to 0, it means that there is no reliable release evidence at present. At this time, the action evidence weight map should be set to a zero map, or the stable weight map of the previous moment should be used.

[0126] In practice, for the discretization of the polar coordinate area integral for evidence compactness, it is necessary to sum the action evidence weight map pixel by pixel within the spatial range corresponding to the bottom candidate domain in the polar coordinate system, and then divide it by the total number of effective pixels within the corresponding spatial range in the polar coordinate system to obtain the average weight. If non-uniform polar coordinate sampling is used, the area weight at different radius positions should also be corrected.

[0127] In specific implementation, regarding the encapsulation fields, encoding methods, and triggering mechanisms for edge backhaul, edge backhaul should at least encapsulate the bottom candidate domain image, action evidence weight map, evidence compactness, edge action state, and state probability at the corresponding moment corresponding to the moment with the strongest action evidence. The bottom candidate domain image can be encoded using lossless compression or high-quality lossy compression, and the action evidence weight map can be quantized into an 8-bit grayscale image or a heatmap index. The backhaul triggering time is based on the completion of a complete judgment in the current cache window.

[0128] In practice, for the initialization, continuous updating, and abnormal freezing of the inactive baseline map, when the system is first started, the normalized brightness should be averaged on multiple consecutive stable inactive frames to form an initial inactive baseline map. During subsequent operation, the inactive baseline map should be continuously updated with weighted values ​​according to the inactive confidence parameter. If acquisition anomalies such as lens occlusion, severe defocusing, raindrop attachment, or large-area overexposure occur, the update should be paused to avoid writing abnormal appearances into the baseline.

[0129] It should be noted that the input and output parameters in the calculation formulas of this application are all dimensionless calculations performed through normalization processing. The formulas are all derived from software simulations based on a large amount of collected data, and the preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0130] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A visual detection method for the operating status of surge arrester trip units based on edge computing, applied to edge computing nodes equipped with visual acquisition devices, characterized in that, include: The system acquires the input video image sequence captured by the vision acquisition device, extracts the current frame image and the previous frame image within the current buffer window, and maps the timestamp of the input video image sequence to normalized time. Extract the normalized x-coordinate and normalized y-coordinate of the current frame image, use the lateral edge distribution density of the arrester skirt to locate the arrester axis and the bottom candidate domain, and capture the bottom candidate domain image corresponding to the bottom candidate domain. Within the candidate domain at the bottom end, the coaxial center at the bottom end is reconstructed based on the radial alignment score, and polar coordinate normalization is performed based on the coaxial center at the bottom end to construct a polar coordinate domain. In the polar coordinate domain, the outwardly expanding positive radial projection gradient is extracted to construct the gap-releasing gradient field, and the angular density distribution is generated by radial integration. The first harmonic expansion of the angular density distribution is performed to calculate the dominance and direction of the first harmonic of the eccentric release slit. The expansion amount of the release gap is extracted along the main harmonic direction, and the unidirectional evolution amount is calculated by combining the change of the first harmonic dominance of the eccentric release gap with the normalization time. The first harmonic dominance of the eccentric release gap, the gap extension amount, and the unidirectional evolution amount are continuously mapped without thresholds to obtain the continuous score corresponding to each action state, and the action state at the edge end is output. Within the current cache window, select the moment with the strongest action evidence, obtain the bottom candidate domain image captured at that moment and perform edge backpropagation with the action evidence weight map generated based on the release gradient field, and adaptively update the non-action baseline map corresponding to the previous frame image based on the action state at the edge.

2. The visual detection method for the operation status of a surge arrester trip unit based on edge computing according to claim 1, characterized in that, The method of locating the arrester axis and bottom candidate region by utilizing the lateral edge distribution density of the arrester skirt includes: Obtain the normalized brightness of the current frame image, calculate the vertical gradient magnitude of the normalized brightness and divide it by the maximum gradient magnitude within the current image region to obtain the horizontal edge density; The lateral edge density is integrated along the longitudinal direction and then normalized by the lateral integration to obtain the lateral edge quality distribution. The abscissa of the arrester axis is calculated using the weighted integral of the normalized abscissa under the lateral edge mass distribution, and the lateral discrete scale is calculated using the square root of the weighted variance. The lateral edge density is integrated along the lateral direction and normalized by integrating along the longitudinal direction within the horizontal coordinate neighborhood of the arrester axis to obtain the longitudinal edge mass distribution. The ordinate of the bottom candidate domain is obtained by dividing the integral of the square of the normalized ordinate under the longitudinal edge quality distribution by the integral of the normalized ordinate under the longitudinal edge quality distribution. The bottom candidate region is defined based on the abscissa of the arrester axis, the lateral discrete scale, and the ordinate of the bottom candidate region.

3. The visual detection method for the operation status of a surge arrester trip unit based on edge computing according to claim 2, characterized in that, Within the candidate domain at the bottom end, the coaxial center of the bottom end is reconstructed based on the radial alignment score, and polar coordinate normalization is performed based on the coaxial center of the bottom end to construct a polar coordinate domain, including: Within the bottom candidate domain, a candidate coaxial center is set, and a polar coordinate radial unit vector is constructed pointing from the candidate coaxial center to each pixel. Calculate the gradient vector of the normalized brightness, extract the magnitude of the gradient vector as the absolute gradient, and calculate the surface positive component of the absolute gradient in the bottom candidate domain as the absolute gradient surface positive component. Calculate the maximum non-negative value of the dot product of the gradient vector and the polar coordinate radial unit vector, divide the surface positive fraction of the maximum non-negative value in the bottom candidate domain by the absolute gradient surface positive fraction, and obtain the radial alignment score of the candidate coaxial center. The position that maximizes the radial alignment score is selected as the bottom coaxial center; The reference radius is obtained by calculating a weighted average of the distances from the pixels to the coaxial center at the bottom using the absolute gradient. The distance is divided by the reference radius to convert it into a polar coordinate radius, and the azimuth angle of the pixel relative to the bottom coaxial center is normalized by the circle to convert it into a polar coordinate angle, so as to construct the polar coordinate domain.

4. The visual detection method for the action status of a surge arrester trip unit based on edge computing according to claim 3, characterized in that, The step of extracting the outwardly expanding positive radial projection gradient in the polar coordinate domain to construct the gap-releasing gradient field, and generating an angular density distribution by radial integration, includes: In the polar coordinate domain, the maximum non-negative value of the dot product of the gradient vector and the polar coordinate radial unit vector is calculated, and then divided by the sum of the integrals of the maximum non-negative value in the polar coordinate domain to construct the release gradient field; Calculate the absolute value of the difference between the polar coordinate radius and the numerical value, multiply the absolute value by the release gradient field, and take a double integral in the polar coordinate domain to obtain the adaptive radial expansion amount; The angular density distribution is generated by radially integrating the release gradient field within a radial interval bounded by the adaptive radial expansion amount and around a numerical value of 1.

5. The visual detection method for the operation status of a surge arrester trip unit based on edge computing according to claim 4, characterized in that, The step of performing first-order harmonic expansion on the angular density distribution and calculating the dominance and direction of the first-order harmonic of the eccentric release slit includes: The first-order cosine coefficients are obtained by multiplying the angular density distribution with the circular cosine function based on the polar coordinate angle and performing discrete integration. The first-order sine coefficients are obtained by multiplying the angular density distribution with a circumferential sine function based on the polar coordinate angle and performing discrete integration. The square root of the sum of the squares of the first-order cosine coefficient and the squares of the first-order sine coefficient is taken to obtain the first-order harmonic dominance of the eccentric release slit. Calculate the arctangent phase angle of the quotient of the first-order sine coefficient and the first-order cosine coefficient, and divide the arctangent phase angle by the total radians of the circle to obtain the principal harmonic direction.

6. The visual detection method for the operation status of a surge arrester trip unit based on edge computing according to claim 5, characterized in that, The step of extracting the release gap extension along the dominant harmonic direction and calculating the unidirectional evolution by combining the cumulative change of the first harmonic dominance of the eccentric release gap with the normalized time includes: Construct the maximum non-negativity value of the circular cosine function with the difference between the polar coordinate angle and the principal harmonic direction as the variable, and use it as the cosine weighting function; Multiply the release gradient field with the cosine weighted function and integrate along the polar coordinate angle, then divide by the sum of the angles of the cosine weighted function to obtain the radial distribution of the principal harmonic direction; Obtain the maximum polar radius corresponding to the bottom candidate domain boundary, and calculate the dimensionless first moment of the radial distribution of the principal harmonic direction within the range of value one to the maximum polar radius to obtain the release gap extension amount. Extract the differential increment of the first harmonic dominance of the eccentric release gap with respect to the normalized time, take the absolute value of the differential increment as the absolute differential increment, and take the differential increment that is greater than zero as the positive differential increment. The unidirectional evolution quantity is obtained by dividing the time integral of the positive differential increment by the time integral of the absolute differential increment.

7. The visual detection method for the operation status of a surge arrester trip unit based on edge computing according to claim 6, characterized in that, The process involves performing a threshold-free continuous mapping between the first-order harmonic dominance of the eccentric release slot, the release slot extension, and the unidirectional evolution, to obtain continuous scores corresponding to each action state, and outputting the edge-end action states, including: The score for no action is obtained by multiplying the difference between the numerical value and the first harmonic dominance of the eccentric release gap by the difference between the numerical value and the unidirectional evolution. The quasi-action score is obtained by multiplying the first harmonic dominance of the eccentric release gap, the unidirectional evolution amount, and the difference between the numerical value and the release gap extension amount. Multiply the first harmonic dominance of the eccentric release gap by the gap extension to obtain the action score; Perform natural exponential normalization on the inaction score, the quasi-action score, and the action score to obtain the corresponding inaction state probability, quasi-action state probability, and action state probability, respectively. The category corresponding to the extreme value probability is output as the edge action state.

8. The visual detection method for the action status of a surge arrester trip unit based on edge computing according to claim 7, characterized in that, The step of selecting the moment with the strongest action evidence within the current cache window, acquiring the bottom candidate domain image captured at that moment, performing edge backpropagation with the action evidence weight map generated based on the gap gradient field, and adaptively updating the non-action baseline map corresponding to the previous frame image based on the edge action state includes: Within the current cache window, find the position where the sum of the probability of the quasi-action state and the probability of the already-action state reaches its maximum, and mark it as the moment with the strongest evidence of action; The release gradient field at the moment with the strongest action evidence is normalized to its maximum value to generate the action evidence weight map. The evidence compactness is obtained by dividing the surface positive score of the action evidence weight map within the bottom candidate domain by the global surface positive score of the bottom candidate domain. The edge backpropagation is achieved by encapsulating the action evidence weight map, the evidence compactness, and the bottom candidate domain image captured at that moment. Extract the probability of the inactive state at that moment as the inactive confidence parameter; The updated inactive baseline map is obtained by multiplying the inactive baseline map saved in the previous frame by the difference between the value one and the inactive confidence parameter, and adding the product of the normalized brightness and the inactive confidence parameter at this moment.

Images