Multi-mode online detection method and system for transformer oil

Abstract

The invention discloses a multi-mode online detection method and system for transformer oil, and relates to the technical field of online monitoring of transformer oil, and the method comprises the steps: deducing a dielectric constant based on a Maxwell equation and an electric displacement invariant principle in combination with a real-time voltage; segmenting, marking and sharpening a real-time visual image by combining a curvature prior smoothing approximation algorithm with a segmentation mark sharpening method; performing denoising enhancement on the real-time thermodynamic image by adopting a filtering mapping enhancement method to generate a real-time enhanced thermodynamic image, and deducing the temperature based on region screening and colorimetric mapping; deriving an acid value based on the gray value of the second target visual image in combination with a fluorescence standard curve, obtaining the actual height of an oil column in a third target visual image through an improved edge detection and scale scaling method, and calculating the change height of the oil column to reflect the viscosity; an aging index is generated through analysis of a logic regression device, and whether to adjust a monitoring interval or send an early warning prompt is dynamically decided based on an early warning decision mechanism, so that online self-adaptive detection and early warning of the transformer oil are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of on-line monitoring of transformer oil, and particularly relates to a multi-modal on-line detection method and system for transformer oil. Background Art

[0002] Power transmission and transformation equipment is an important part of the power system, and its operating conditions are directly related to the safe and economic operation of the entire power grid. As the core component in power transmission and transformation equipment, the transformer acts as a hub for power transmission and distribution. The state of the transformer oil inside has a crucial impact on the service life and failure risk of the transformer.

[0003] The existing Chinese patent application with the publication number CN103512962A proposes a simulation system and method for on-line detection of transformer oil gas chromatography. The system includes an oil sample configuration unit, an on-line test unit, and a comparative analysis unit. The oil sample configuration unit completes the configuration of the oil sample. The on-line test unit includes a circulating oil circuit, a circulating pump, a gas injection pump, an oil tank, and an oil circuit temperature control device, and the on-line test unit completes the on-line test of the oil sample. The comparative analysis unit includes measurement error analysis, cross-sensitivity test, minimum detection period test, measurement repeatability test, data transmission test, and data analysis function check.

[0004] Most of the existing technologies for on-line detection of transformer oil are based on separating the gases in the transformer oil and indirectly judging the state of the transformer oil through chromatographic detection. However, parameters such as the acid value and dielectric constant of the transformer oil need to rely on complex experimental equipment or manual operations to be carried out. The detection period is long, time-consuming and laborious, and it cannot monitor the state of the transformer oil in real time to timely discover potential faults. Summary of the Invention

[0005] In view of the deficiencies of the existing technology, the present invention proposes a multi-modal on-line detection method and system for transformer oil.

[0006] The technical solution adopted to achieve the object of the present invention is as follows:

[0007] A multi-modal on-line detection method for transformer oil includes the following specific steps:

[0008] Obtain a real-time visual image at the current node t A real-time thermal image And a real-time voltage U t ;

[0009] Based on Maxwell's equations and the principle of invariant electric displacement, combined with the real-time voltage U t Derive the dielectric constant ε at the current node t t ;

[0010] Process the real-time visual image using the curvature prior smoothing approximation algorithm Construct a non-convex high-order model using curvature information to reduce the staircase effect, construct an adaptive variational problem and perform smooth approximation through an optimization algorithm to generate a real-time smooth visual image

[0011] Process the real-time smooth visual image using the segmentation marker sharpening method Perform segmentation, marking, and sharpening to generate a first target visual image Second target visual image And a third target visual image

[0012] Remove the noise in the real-time thermal image through the filtering mapping enhancement method And convert it into an optimized grayscale image And generate a real-time enhanced thermal image through local histogram equalization Adopt region screening and colorimetric mapping to obtain the temperature T of the current node t t ;

[0013] Based on the second target visual image Obtain the fluorescence intensity from the grayscale values in it, and combine with the fluorescence standard curve derivation to obtain the acid value pH of the current node t t ;

[0014] Convert the oil column pixel height in the third target visual image Through the improved edge detection and scale stretching method Into the corresponding actual height l of the oil column t o And calculate the change height Δl of the oil column of the current node t t o ;

[0015] Input the dielectric constant ε of the current node t t , temperature T t , acid value pH t And the change height Δl of the oil column t o Into the logistic regressor, and perform dimensionality increase, dimensionality reduction, and non-linear mapping in sequence to generate the aging index ψ of the current node t t , and dynamically decide whether to adjust the monitoring interval Δt or send a warning prompt based on the warning decision mechanism

[0016] Furthermore, the specific derivation process of the dielectric constant ε of the current node t is as follows: t The specific derivation process is as follows:

[0017] The normal charge and normal voltage of the capacitor are Q0 and U0 respectively. According to Maxwell's equations, the normal electric displacement D0 is the product of the air permittivity ε0 and the normal electric field strength E0. At the same time, the normal electric displacement D0 is proportional to the normal charge Q0 and inversely proportional to the electrode distance r. From this, the normal electric field strength E0 is deduced;

[0018] Since the normal electric field strength E0 of the capacitor is the differential of the normal voltage U0 with respect to the electrode distance r, the first equation of the normal voltage U0 and the air permittivity ε0 can be deduced in reverse;

[0019] The real-time charge and real-time voltage of the capacitor at the current node t are Q(t) and U(t) respectively. Based on the principle of invariant electric displacement, the real-time electric displacement D(t) is equal to the normal electric displacement D0, and according to Maxwell's equations, the real-time charge Q(t) is equal to the normal charge Q0;

[0020] Since the capacitor has not changed, the second equation generated by replacing the real-time voltage U(t) and the real-time permittivity ε(t) at the current node t with the normal voltage U0 and the air permittivity ε0 in the first equation still holds;

[0021] Dividing the second equation by the first equation, the ratio of the real-time voltage U(t) to the normal voltage U0 is always equal to the ratio of the real-time permittivity ε(t) to the air permittivity ε0, and thus the real-time permittivity ε(t) at the current node t can be deduced.

[0022] Furthermore, the real-time visual image is processed by the curvature prior smoothing approximation algorithm to generate a real-time smoothed visual image including the following specific steps:

[0023] For the real-time visual image Integrate all the level curves on the Lipschitz continuous boundary domain Ω to obtain the TV regularization term The specific formula is as follows:

[0024]

[0025] where and are the grayscale image and pixel points of the real-time visual image respectively, and is the first-order gradient image corresponding to the grayscale image;

[0026] Introduce the curvature curve model, and fit and calculate the curvature regularization term through the divergence of the unit normal vector in the grayscale image The specific formula is as follows:

[0027]

[0028] where α is a fitting coefficient, is a grayscale image of the unit second-order gradient image,

[0029] Construct an adaptive variational problem for the weight matrix in the improved adaptive TV model. The adaptive variational problem is as follows:

[0030]

[0031] where λ and η are fixed regularization coefficients, b is an approximation coefficient, and are the unit second-order gradient, weight matrix, and first-order gradient at pixel point respectively. The weight matrix at pixel point is obtained based on the Gaussian kernel ;

[0032] Solve the adaptive variational problem using an optimization algorithm to obtain the optimal approximation coefficient b * and calculate the real-time smoothed visual image

[0033] Furthermore, use the segmentation marker sharpening method to process the real-time smoothed visual image to generate the first target visual image The second target visual image and the third target visual image including the following specific steps:

[0034] Mine the real-time smoothed visual image through the feature extraction layer based on the inverted residual structure The feature extraction layer based on the inverted residual structure realizes the dimensionality increase, feature mining, and dimensionality reduction of the real-time smoothed visual image through convolution, depthwise separable convolution, and again convolution, and generates the feature map f through ReLU and residual connection; t v ;

[0035] Capture long-range dependencies along the horizontal and vertical directions through the coordinate attention mechanism and retain accurate position information to generate the direction-aware attention map f t v,d and the position-aware attention map f t v,p and synergistically enhance the feature map f t v to generate the attention feature map f t v,a ;

[0036] Divide the real-time smoothed visual image through a multi-scale feature fusion layer into several patches, and generate corresponding weight vectors for each patch through secondary extraction and upsampling. Concatenate all the weight vectors according to the positions of the patches in the real-time smoothed visual image to generate a weight map

[0037] Linearly modulate and concatenate the attention feature map f t v,a and the weight map through a concatenated prediction layer, and then restore them through convolution and bicubic interpolation in sequence to generate a real-time visual marker image

[0038] Select a single marker, and reset the pixel values of all pixel points in the real-time visual marker image that do not carry the single marker to the boundary pixel value with the largest difference from the average pixel value of the pixel points carrying the single marker to generate a first target visual image A second target visual image and a third target visual image

[0039] Furthermore, process the real-time thermal image through a filtering mapping enhancement method to generate a real-time enhanced thermal image and obtain the temperature T of the current node t through region screening and colorimetric mapping t , including the following specific steps:

[0040] Use median filtering to filter out the salt-and-pepper noise and Gaussian noise in the real-time thermal image to generate a real-time optimized thermal image

[0041] Obtain an optimized grayscale image and divide it into N local grayscale regions, where N is the total number of local grayscale regions;

[0042] Sum up the probabilities of each grayscale value appearing in each local grayscale region to obtain a cumulative distribution function;

[0043] Obtain an optimized grayscale image Calculate the total number of grayscale values in the optimized grayscale image, multiply the total number of grayscale values minus 1 by the cumulative distribution function of each local grayscale region respectively and round down to generate a real-time enhanced thermal image

[0044] Retain the pixel points in the real-time enhanced thermal image that overlap with the non-boundary pixel points in the first target visual image ;

[0045] Perform colorimetric mapping according to the color temperature mapping function between pixel values and temperature in the colorimetric bar, and obtain a real-time enhanced thermal image for all the remaining pixel points in it, and take the average value as the temperature T of the current node t t 。

[0046] Furthermore, obtain the actual height l of the oil column at the current node t through an improved edge detection and scale stretching method t o and the change height Δl of the oil column t o , including the following specific steps:

[0047] Process the third target visual image using improved recursive filtering to solve the problem that Gaussian filtering has poor processing effect on salt-and-pepper noise and is prone to blurring edges, and generate an optimized third target visual image

[0048] Calculate the gradient magnitude and gradient direction of the optimized third target visual image using the Sobel operator, and perform bilinear interpolation on 4 pixel points including the target pixel point in the gradient direction to achieve sub-pixel positioning of the edge; Use the double-threshold determination method to determine the oil column pixel edge in the optimized third target visual image

[0049] and obtain the oil column pixel height of the current node t through rectangular approximation and the infiltration tube pixel height Take the ratio of the actual height of the infiltration tube to the infiltration tube pixel height as the scaling factor; Multiply the oil column pixel height

[0050] by the scaling factor to obtain the actual height l of the oil column at the current node t Subtract it from the initial oil column height t o to obtain the change height Δl of the oil column at the current node t t o 。

[0051] Furthermore, the improved recursive filtering includes the following specific steps:

[0052] Set the initial sliding window width w0, the maximum sliding window width w max and the judgment threshold ξ δ Define the pixel standard deviation of the third target visual imagewithin the sliding window as the local contrast;

[0053] Perform local contrast judgment before filtering the target pixel, and calculate the local contrast within the current sliding window And determine whether it is less than the judgment threshold ξ δ ;

[0054] If it is less than the judgment threshold ξ δ , update the pixel value of the target pixel to the pixel median within the current sliding window, and move the sliding window to the next pixel;

[0055] If it is greater than or equal to the judgment threshold ξ δ , perform sliding window judgment, and determine whether the current sliding window width w is less than the maximum sliding window width w max ;

[0056] If it is less than the maximum sliding window width w max , expand the sliding window width to w + 2, and jump to execute local contrast judgment; w is the current sliding window width;

[0057] If it is equal to the maximum sliding window width w max , update the pixel value of the target pixel to the pixel median of 4 adjacent pixels;

[0058] Repeat the above steps until all pixels in the third target visual image are filtered.

[0059] Furthermore, based on the early warning decision-making mechanism, dynamically decide whether to adjust the monitoring interval Δt or send an early warning prompt, including the following specific steps:

[0060] Obtain the aging index ψ of the transformer oil at the current node t t , and judge the aging index ψ t whether it is greater than or equal to the first aging threshold ψ1;

[0061] If it is greater than or equal to the first aging threshold ψ1, send an early warning prompt;

[0062] If it is less than the first aging threshold ψ1, further judge the aging index ψ t whether it is less than the second aging threshold ψ2;

[0063] If it is less than the second aging threshold ψ2, maintain the monitoring interval Δt as the initial monitoring interval Δ0 to generate the next node t + Δt;

[0064] If it is greater than or equal to the second aging threshold ψ2 and less than the first aging threshold ψ1, obtain the quantization layer number y where the aging index ψ t is located t , and adjust the monitoring interval Δt to the initial monitoring interval Δ0 minus the quantization layer number y tMultiply by the minimum adjustment interval Δ min , to generate the next node t + Δt.

[0065] A multi-modal on-line detection system for transformer oil is used to implement a multi-modal on-line detection method for transformer oil, including a system clock module, a data acquisition module, a multi-scale analysis module, and a feedback adjustment module;

[0066] The system clock module sets nodes based on the monitoring interval Δt and sends acquisition signals to the data acquisition module at each node;

[0067] The data acquisition module receives the acquisition signals, obtains real-time visual images, real-time thermal images, and real-time voltages, and feeds them back to the multi-scale analysis module;

[0068] The multi-scale analysis module derives the dielectric constant based on Maxwell's equations and the principle of invariant electric displacement, generates the first target visual image, the second target visual image, and the third target visual image based on the curvature prior smoothing approximation algorithm combined with the segmentation marker sharpening method, processes the real-time thermal image through the filter mapping enhancement method and combines region screening and colorimetric mapping to obtain the temperature, derives the acid value based on the second target visual image combined with the fluorescence standard curve, processes the third target visual image through the improved edge detection and scale stretching method to obtain the change height of the oil column, and sends the dielectric constant, temperature, acid value, and change height of the oil column to the feedback adjustment module;

[0069] The feedback adjustment module analyzes the dielectric constant, temperature, acid value, and change height of the oil column through a logistic regression analyzer to generate an aging index, and dynamically decides whether to adjust the monitoring interval Δt or send a warning prompt based on the warning decision mechanism.

[0070] Compared with the prior art, the present invention has the following remarkable advantages:

[0071] 1. By introducing methods such as infrared temperature measurement, acid value measurement with fluorescent indicators, capillary infiltration tube infiltration, and capacitor induction, and assisting algorithms or principles such as the filter mapping enhancement method, fluorescence standard curve, improved edge detection and scale stretching method, and Maxwell's equations, the original manual detection with cumbersome processes and complex equipment is changed into automated visual processing, infrared measurement, and electrical signal analysis, realizing regular automated on-line detection of the state of transformer oil. At the same time, the update of the detection method also reduces the application of precision equipment and intelligent equipment, reducing costs;

[0072] 2. Design a prior smoothing approximation algorithm for the design curve to attenuate abnormal high-frequency components in real-time visual images to reduce the staircase effect, and locally couple to generate a real-time smoothed visual image, which is assisted by the segmentation marker sharpening method. The segmentation marker sharpening method identifies the first target, the second target, and the third target in the real-time smoothed visual image through image segmentation technology and marks them to generate a real-time visual marker image, and generates the first target visual image, the second target visual image, and the third target visual image according to the marker sharpening for subsequent precise detection;

[0073] 3. Design a joint analysis and adjustment scheme for the logistic regression and the early warning decision mechanism, comprehensively consider the dielectric constant, temperature, acid value, and the change height of the oil column to estimate the aging index of the transformer oil. The early warning decision mechanism selects to send an early warning prompt or adaptively adjust the monitoring interval based on the range of the aging index to detect transformer oil problems earlier, realizing the effective early warning of the transformer oil and improving the operation safety of the transformer. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flowchart of a multi-modal on-line detection method for transformer oil in the present invention;

[0075] Figure 2 is a flowchart of the segmentation marker sharpening method in the present invention;

[0076] Figure 3 is a flowchart of the improved recursive filter in the present invention;

[0077] Figure 4 is a schematic diagram of a multi-modal on-line detection system for transformer oil in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0078] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0079] Embodiment 1

[0080] As Figure 1 shown, a specific embodiment of the present invention discloses a multi-modal on-line detection method for transformer oil, including the following specific steps:

[0081] At the current node t, actively collect real-time visual images through a multi-source visual detection device and passively sense through a variable capacitance sensor real-time thermal images and real-time voltage U t , and the real-time visual image and the real-time thermal image have the same dimension;

[0082] Combined with Maxwell's equations and the principle of invariant electric displacement, combined with the real-time voltage U tDetermine the dielectric constant ε of the transformer oil at the current node t based on the ratio to the normal voltage U0 t ;

[0083] Process the real-time visual image using the curve first-order prior smoothing approximation algorithm Utilize the curvature information to construct a non-convex high-order model to attenuate the abnormal high-frequency components in the real-time visual image in order to reduce the staircase effect, transform the local coupling into an adaptive variational problem and perform smoothing approximation through an optimization algorithm to generate a real-time smoothed visual image

[0084] Process the real-time smoothed visual image using the segmentation marker sharpening method Identify the pixel points belonging to the first target, second target, and third target in the real-time smoothed visual image through image segmentation technology and assign corresponding markers to generate a real-time visual marker image Generate the first target visual image based on the marker sharpening The second target visual image and the third target visual image wherein, the first target is the transformer oil, the second target is the fluorescent indicator, and the third target is the capillary infiltration tube;

[0085] Process the real-time thermal image using the filtering mapping enhancement method Remove the noise in the real-time thermal image through filtering and convert it into an optimized grayscale image Through the optimized grayscale image Perform local histogram equalization on the cumulative grayscale distribution function of the local area in it to generate a real-time enhanced thermal image Combine the first target visual image Perform region screening and obtain the temperature T of the transformer oil at the current node t through colorimetric mapping t ;

[0086] Since the gray value of each pixel point in the second target visual image represents the brightness of the pixel point, therefore, take the average gray value of all pixel points in the area where the fluorescent indicator is located in the second target visual image as the fluorescence intensity, and deduce and obtain the acid value pH of the transformer oil at the current node t based on the fluorescence standard curve t , wherein, the fluorescent indicator can be made of a pH-sensitive fluorescent dye based on amino reversible protonation, the fluorescence intensity of the fluorescent indicator will change with the change of the acid value of the contact substance, and the fluorescence standard curve can be obtained by pre-measuring transformer oils with multiple different acid values using the fluorescent indicator and fitting. In this embodiment, the fluorescent indicator selects boron dipyrromethene dyes;

[0087] Calibrating the visual image of the third target by an improved edge detection and scale stretching method the pixel height of the oil column and the pixel height of the dip tube Combined with the actual height of the dip tube Calculate the stretching scale to obtain the actual height l of the oil column at the current node t t o And calculate the change height Δl of the oil column at the current node t t o ;

[0088] Take the dielectric constant ε t 、temperature T t 、acid value pH t and the change height Δl of the oil column t o at the current node t as the input of a pre-trained logistic regression. The logistic regression first increases the dimension of the input through a fully connected layer and a batch normalization layer, then reduces the dimension through 2 fully connected layers in sequence, and finally realizes the nonlinear mapping from the dielectric constant ε t 、temperature T t 、acid value pH t and the change height Δl of the oil column t o at the current node t to the aging index ψ t of the transformer oil, and dynamically make decisions on whether to adjust the monitoring interval Δt or send a warning prompt based on the warning decision mechanism. Among them, the pre-training of the logistic regression requires collecting the temperature, dielectric constant, acid value and viscosity of a large number of transformer oil samples, and at the same time recording the aging index of these transformer oil samples. The aging index is in the range of 0 to 1 and can be obtained through various means such as chemical analysis, physical testing or expert evaluation.

[0089] Furthermore, the variable capacitance sensor can sense the capacitance change of the capacitor. The specific derivation process of the dielectric constant ε t of the transformer oil at the current node t is as follows:

[0090] When there is no transformer oil between the inner electrode and the outer electrode of the capacitor, the medium is air, and the normal charge amount and normal voltage of the capacitor are Q0 and U0 respectively;

[0091] According to Maxwell's equations, the normal electric displacement D0 of the capacitor when the medium is air is the product of the air dielectric constant ε0 and the normal electric field strength E0, that is, D0 = ε0·E0. At the same time, the normal electric displacement D0 is proportional to the normal charge amount Q0 and inversely proportional to the electrode distance r. The specific formula is D0 = Q0 / (2π·r). From this, the normal electric field strength E0 = Q0 / (2π·ε0·r) can be deduced;

[0092] Since the electric field strength E of the capacitor is the differential of the voltage U with respect to the electrode distance r, that is, the normal electric field strength E0, the normal voltage U0, and the electrode distance r satisfy The first equation of the normal voltage U0 and the air permittivity ε0 can be reversely deduced, and the first equation is as follows:

[0093]

[0094] where r1 and r2 are the radii of the inner electrode and the outer electrode respectively;

[0095] When the medium between the inner electrode and the outer electrode of the capacitor is transformer oil, the medium is transformer oil. The real-time charge quantity and real-time voltage of the capacitor at the current node t are Q(t) and U(t) respectively. Based on the principle of invariant electric displacement, the real-time electric displacement D(t) of the current node t is equal to the normal electric displacement D0, and according to Maxwell's equations, the real-time charge quantity Q(t) is also equal to the normal charge quantity Q0;

[0096] Since the electrode distance r, the inner electrode radius r1, and the outer electrode radius r2 of the capacitor remain unchanged, the second equation generated by replacing the real-time voltage U(t) and the real-time permittivity ε(t) at the current node t with the normal voltage U0 and the air permittivity ε0 in the first equation still holds;

[0097] Dividing the second equation by the first equation shows that the ratio of the real-time voltage U(t) of the capacitor at the current node t to the normal voltage U0 is always equal to the ratio of the real-time permittivity ε(t) at the current node t to the air permittivity ε0. Then the real-time permittivity ε(t) at the current node t = [U(t)·ε0] / U0.

[0098] Furthermore, the real-time visual image is processed by the curvature prior smoothing approximation algorithm to generate a real-time smoothed visual image including the following specific steps:

[0099] Since the real-time visual image is a bounded image on the Lipschitz continuous boundary domain Ω, integrating all the level curves of the real-time visual image obtains the TV regularization term , and the specific formula is as follows:

[0100]

[0101] where and are the grayscale image and pixel points of the real-time visual image respectively, and is the first-order gradient image corresponding to the grayscale image;

[0102] Since the TV model is susceptible to the step effect, the curvature curve model is introduced. t v The curvature regularization term is calculated by fitting the divergence of the unit normal vector in The specific formula is as follows:

[0103]

[0104] Among them, α is the fitting coefficient, Grayscale The unit second-order gradient map of

[0105] To reduce the effects of noise and blur, fully preserve the real-time visual image The edges, structures and other valuable details of the image are obtained by improving the weight matrix in the adaptive TV model to construct an adaptive variational problem. Smooth near real-time visual images The adaptive variational problem is as follows:

[0106]

[0107] Among them, λ and η are fixed regularization coefficients, b is the approximation coefficient, and Pixels The unit second-order gradient, weight matrix and first-order gradient at pixel The weight matrix at Based on Gaussian kernel Get, the specific formula is as follows:

[0108]

[0109] Among them, γ is the proportionality coefficient, and Grayscale In pixels The positive gradient on the horizontal axis i and the vertical axis j at point i, Gaussian kernel The specific calculation formula is as follows:

[0110]

[0111] Where δ is the standard deviation of the Gaussian kernel, and Pixels Pixel values ​​on the horizontal axis i and the vertical axis j;

[0112] Use optimization algorithm to solve adaptive variational problem and obtain the optimal approximation coefficient b * , optimal fitting coefficient α *, the optimal proportionality coefficient γ * and the optimal standard deviation δ * to obtain a real-time smoothed visual image In this embodiment, the optimization algorithm adopts the alternating direction multiplier method.

[0113] As Figure 2 shown, further, the real-time smoothed visual image is processed by the segmentation marker sharpening method to generate the first target visual image The second target visual image and the third target visual image include the following specific steps:

[0114] The features of the real-time smoothed visual image are mined through the feature extraction layer based on the inverted residual structure corresponding feature map f t v , the traditional feature extraction layer often realizes through the five-step strategy of dimensionality reduction, convolution, dimensionality increase, ReLU and residual connection. However, the feature extraction layer based on the inverted residual structure first increases the number of channels of the real-time smoothed visual image through a 1×1 convolution to achieve dimensionality increase, and then performs feature mining through a 3×3 depthwise separable convolution. The depthwise separable convolution can reduce the number of parameters and improve the operation speed. Finally, the number of channels is reduced again through a 1×1 convolution, and then the feature map f is generated through ReLU and residual connection ; t v ;

[0115] The attention degree for different positions of the feature map f is adjusted through the coordinate attention mechanism to imitate the human visual system. Different from the traditional attention mechanism that only focuses on the feature information of the encoding channels, the coordinate attention mechanism captures long-range dependencies along the horizontal direction and retains accurate position information along the vertical direction, respectively generating the direction-aware attention map f t v and the position-aware attention map f t v,d and they are used together to enhance the feature map f t v,p to generate the attention feature map t v ;

[0116] The context aggregation ability of the feature information in the real-time smoothed visual image is enhanced through the multi-scale feature fusion layer. The multi-scale feature fusion layer combines the real-time smoothed visual image in Divided into patches of the same size. For each patch, after 3×3 convolution, batch normalization, and Sigmoid, it is further extracted through 1×1 convolution, batch normalization, and Sigmoid. To avoid dimensionality reduction during the convolution process, bicubic interpolation is used for upsampling to generate the weight vector corresponding to each patch, and the weight vectors of all patches are concatenated according to the positions of the patches in the real-time smoothed visual image to generate a weight map

[0117] The attention feature map f t v,a and the weight map are linearly modulated respectively to be converted to the same dimension and concatenated. Then, after a 3×3 convolution for dimensionality reduction and restored to the same dimension as the real-time smoothed visual image through bicubic interpolation to generate a real-time visual marker image Real-time visual marker image The pixel points belonging to the first target, the second target, and the third target in the real-time visual marker image are assigned markers 1, 2, and 3 respectively;

[0118] Select a single marker from the 3 markers in turn, and reset the pixel values of all pixel points in the real-time visual marker image that do not carry the single marker to the boundary pixel value with the largest difference from the average pixel value of the pixel points carrying the single marker to achieve the sharpening of the real-time visual marker image The boundary pixel values include the upper limit pixel value 255 and the lower limit pixel value 0. Taking marker 1 as an example, assuming that the average pixel value of all pixel points carrying marker 1 is 197, then the pixel values of the remaining pixel points not carrying marker 1 should be reset to the lower limit pixel value 0, and the first target visual image The second target visual image and the third target visual image

[0119] Furthermore, the real-time thermal image is processed by the filtering mapping enhancement method to generate a real-time enhanced thermal image and the temperature T of the current node t is obtained through region screening and colorimetric mapping t , including the following specific steps:

[0120] Median filtering is used to filter out the salt-and-pepper noise and Gaussian noise in the real-time thermal image to reduce the influence of noise on the clarity of the real-time thermal image and generate a real-time optimized thermal image

[0121] Obtain the real-time optimized thermal image The corresponding optimized grayscale image and divided into N local grayscale regions, where N is the total number of local grayscale regions;

[0122] Sum the probabilities of each grayscale value in each local grayscale region to obtain the cumulative distribution function. Take the cumulative distribution function of the nth local grayscale region as an example, n = 1, 2, …, N, and the specific formula is as follows:

[0123]

[0124] where, u max represents the maximum grayscale value, and p n (u) represents the probability that the grayscale value is u in the nth local grayscale region;

[0125] Obtain the total number of grayscale values in the optimized grayscale image Subtract 1 from the total number of grayscale values, multiply each by the cumulative distribution function of each local grayscale region, and round down to obtain the enhanced thermal pixels corresponding to the local grayscale region to achieve local histogram equalization and generate a real-time enhanced thermal image and the real-time optimized thermal image Compared with the real-time optimized thermal image the real-time enhanced thermal image

[0126] has stronger infrared contrast, which is beneficial to the identification of the temperature of transformer oil; Since the first target visual image has been sharpened, it can be used as a region selection box for the real-time enhanced thermal image Retain the pixel points in the real-time enhanced thermal image that overlap with the non-boundary pixel values in the first target visual image

[0127] Perform colorimetric mapping according to the color temperature mapping function of the pixel value and temperature in the color bar to obtain the temperature corresponding to all the retained pixel points in the real-time enhanced thermal image and take the mean value as the temperature T of the current node t t , and the color temperature mapping function is a linear function, and the specific formula is as follows:

[0128] T = β1·x + β2;

[0129] where, x and T are the pixel value and the corresponding temperature in the color bar respectively, and β1 and β2 are the slope and intercept of the color temperature mapping function respectively, and the specific calculation formulas are as follows:

[0130]

[0131] β2 = T max-β1·x max ;

[0132] where x max and T max are respectively the maximum pixel value and the corresponding maximum temperature in the color bar, and x min and T min are respectively the minimum pixel value and the corresponding minimum temperature in the color bar. Based on the color temperature mapping function, only by substituting the pixel values of the retained pixel points in the real-time enhanced thermal image can the temperature corresponding to the retained pixel points be obtained.

[0133] Furthermore, the actual height l t o of the oil column and the changing height Δl t o of the current node t are obtained by the improved edge detection and scale stretching method, including the following specific steps:

[0134] Use the improved recursive filtering to replace the traditional Gaussian filtering to process the third target visual image to solve the problem that Gaussian filtering has poor processing effect on salt-and-pepper noise and is easy to blur edges, and generate the third target optimized visual image

[0135] Since the edges in the third target optimized visual image usually show a drastic change in gray intensity, use the Sobel operator to calculate the gradient magnitude and gradient direction of the third target optimized visual image . Among them, the gradient magnitude represents the degree of change in gray intensity, and the gradient direction represents the direction of change;

[0136] Since non-maximum suppression only compares the gradient magnitude of the target pixel point with the gradient magnitudes of two adjacent pixel points in the gradient direction, it is easy to cause random error and false edge problems when the edge and non-pixel points are in a non-coincident state. Use bilinear interpolation to replace the traditional non-maximum suppression process, and interpolate the 4 pixel points including the target pixel point in the gradient direction to accurately obtain the edge position within the grid area surrounded by the 4 pixel points, realizing sub-pixel level positioning of the edge;

[0137] Since the third target visual image before the improved recursive filtering is generated by sharpening, the existing double-threshold decision method can be directly used to determine the oil column pixel edge in the third target optimized visual image ;

[0138] Use a rectangular approximation to select the oil column pixels and capillary infiltration tube pixels, and obtain the oil column pixel height and the infiltration tube pixel height of the current node t Take the ratio of the actual height of the infiltration tube to the pixel height of the infiltration tube as the scaling factor;

[0139] Multiply the pixel height of the oil column by the scaling factor to obtain the actual height l of the oil column at the current node t t o , and take the difference from the initial oil column height corresponding to the brand-new transformer oil measured in advance to obtain the height change of the oil column at the current node t Since the viscosity change of the transformer oil will affect the height of the oil column in the capillary infiltration tube, therefore, the height change of the oil column Δl t o can indirectly indicate the viscosity change of the transformer oil from the brand-new transformer oil to the current node t.

[0140] As Figure 3 shown, furthermore, the improved recursive filtering introduces local contrast on the basis of the traditional recursive median filtering, and adjusts the size of the filtering window in different regions of the third target visual image to better adapt to the local features in the third target visual image . The improved recursive filtering includes the following specific steps:

[0141] Set the initial sliding window width to w0, w0 is an odd number, and the maximum sliding window width is set to w max , the judgment threshold is set to ξ δ , and define the standard deviation of the pixels of the third target visual image within the sliding window as the local contrast;

[0142] Before filtering any target pixel point in the third target visual image , perform local contrast judgment, calculate the local contrast within the current sliding window and judge whether it is less than the judgment threshold ξ δ ;

[0143] If the local contrast within the current sliding window is less than the judgment threshold ξ δ , update the pixel value of the target pixel point to the pixel median within the current sliding window, and move the sliding window to the next pixel point;

[0144] If the local contrast within the current sliding window is greater than or equal to the judgment threshold ξ δ , perform sliding window judgment, and judge whether the current sliding window width w is less than the maximum sliding window width w max ;

[0145] If the current sliding window width w is less than the maximum sliding window width w max , expand the sliding window width to w + 2, and jump to execute the local comparison decision;

[0146] If the current sliding window width w is equal to the maximum sliding window width w max , update the pixel value of the target pixel point to the pixel median of 4 adjacent pixel points;

[0147] Repeat the above steps until all pixel points in the third target visual image are filtered.

[0148] Furthermore, based on the early warning decision-making mechanism, dynamically decide whether to adjust the monitoring interval Δt or send an early warning prompt, including the following specific steps:

[0149] Obtain the aging index ψ of the transformer oil at the current node t t , and judge whether the aging index ψ t is greater than or equal to the first aging threshold ψ1. In this embodiment, the first aging threshold ψ1 is set to 0.8;

[0150] If the aging index ψ t is greater than or equal to the first aging threshold ψ1, it is determined that the transformer oil is severely aged, and an early warning prompt is sent to the staff;

[0151] If the aging index ψ t is less than the first aging threshold ψ1, further judge whether the aging index ψ t is less than the second aging threshold ψ2. In this embodiment, the second aging threshold ψ2 is set to 0.4;

[0152] If the aging index ψ t is less than the second aging threshold ψ2, it is determined that the transformer oil is relatively new, and the monitoring interval Δt is maintained at the initial monitoring interval Δ0 to generate the next node t + Δt;

[0153] If the aging index ψ t is greater than or equal to the second aging threshold ψ2 and less than the first aging threshold ψ1, obtain the quantization layer number y t where the aging index ψ is located t , adjust the monitoring interval Δt to the initial monitoring interval Δ0 minus the quantization layer number y t multiplied by the minimum adjustment interval Δ min , to generate the next node t + Δt, where the quantization layer number y t is divided with the second aging threshold ψ2 as the starting value, the first aging threshold ψ1 as the ending value, and 0.1 as the quantization step size. When the quantization layer number is y t , it represents the aging index ψ tLocated in the range [ψ2 + 0.1·(y t - 1), ψ2 + 0.1·y t ), in this embodiment, the minimum adjustment interval Δ min is 0.2 times of the initial monitoring interval Δ0.

[0154] Embodiment 2

[0155] As Figure 4 shown, a multi-modal on-line detection system for transformer oil is used to implement a multi-modal on-line detection method for transformer oil, including a system clock module, a data acquisition module, a multi-scale analysis module and a feedback regulation module;

[0156] The system clock module sets nodes based on the monitoring interval Δt and sends acquisition signals to the data acquisition module at each node;

[0157] The data acquisition module receives the acquisition signal, releases the fluorescent indicator, opens the oil sampling channel so that the transformer oil passes through the fluorescent indicator, the capillary infiltration tube and the capacitor in sequence, and obtains real-time visual images, real-time thermal images and real-time voltage through a multi-source vision detection device and a variable capacitance sensor and feeds them back to the multi-scale analysis module, and recovers the fluorescent indicator. Among them, the fluorescent indicator is placed in the recovery container at non-acquisition moments, and a substance that can assist the fluorescent indicator to quickly recover its normal fluorescence intensity is arranged in the recovery container, so as to realize the reuse of the fluorescent indicator. The multi-source vision detection device is composed of an infrared thermal imager and a high-definition camera;

[0158] The multi-scale analysis module deduces the dielectric constant by combining the real-time voltage with Maxwell's equations and the principle of invariant electric displacement, processes the real-time visual image based on the curvature prior smoothing approximation algorithm and generates the first target visual image, the second target visual image and the third target visual image by using the segmentation marking sharpening method, processes the real-time thermal image by the filtering mapping enhancement method and obtains the temperature by combining region screening and colorimetric mapping, deduces the fluorescence intensity based on the second target visual image, deduces and obtains the acid value by combining the fluorescence standard curve, and processes the third target visual image by the improved edge detection and scale stretching method to obtain the height change of the oil column, and sends the dielectric constant, temperature, acid value and height change of the oil column to the feedback regulation module;

[0159] The feedback regulation module analyzes the dielectric constant, temperature, acid value and height change of the oil column through a logistic regressor to generate an aging index, and dynamically decides whether to adjust the monitoring interval Δt or send a warning prompt based on the warning decision mechanism.

[0160] The present invention discloses a multi-modal online detection method and system for transformer oil, which derives the dielectric constant based on Maxwell's equations and the principle of invariant electric displacement in combination with real-time voltage; adopts a curvature prior smoothing approximation algorithm in combination with a segmentation marking sharpening method to segment, mark, and sharpen real-time visual images; adopts a filtering mapping enhancement method to denoise and enhance real-time thermal images to generate real-time enhanced thermal images, and derives the temperature based on region screening and colorimetric mapping; derives the acid value based on the gray value of the second target visual image in combination with a fluorescence standard curve, and obtains the actual height of the oil column in the third target visual image and calculates the change height of the oil column through an improved edge detection and scale stretching method to reflect the viscosity; analyzes and generates an aging index through a logistic regressor and dynamically makes a decision whether to adjust the monitoring interval or send a warning prompt based on an early warning decision mechanism, realizing online adaptive detection and early warning of transformer oil.

[0161] The above is only the preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions falling within the idea of the present invention belong to the protection scope of the present invention. It should be noted that for those of ordinary skill in the art in this technical field, several improvements and refinements made without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A transformer oil multi-modal online detection method, characterized in that: The specific steps include: Acquire real-time visual images, real-time thermal images, and real-time voltage at the current node, and derive the dielectric constant of the current node based on Maxwell equations and the principle of electric displacement invariance; The non-convex high-order model and adaptive variational problem are constructed by using the Qu-based priori smoothing approximation algorithm, and the real-time smooth visual image is obtained by solving the problem through the optimization algorithm. The first target visual image, the second target visual image and the third target visual image are generated by using the segmentation mark sharpening method. The noise of the real-time thermal image is removed by the filter mapping enhancement method and converted into an optimized grayscale image. The real-time enhanced thermal image is generated by local histogram equalization. The temperature of the current node is obtained by combining regional screening and colorimetric mapping. Based on the gray value of the second target visual image and combined with the fluorescence standard curve, the acid value of the current node is deduced, and the pixel height of the oil column in the third target visual image is converted into the actual height of the oil column through the improved edge detection and scale expansion method, and the change height of the oil column at the current node is calculated; The dielectric constant, temperature, acid value and oil column change height of the current node are analyzed through a logistic regression model and mapped to generate the aging index of the current node. Based on the early warning decision-making mechanism, a dynamic decision is made whether to adjust the monitoring interval or send an early warning prompt.

2. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The acquisition of real-time smooth visual images comprises the following specific steps: The TV regularization term is obtained by integrating all horizontal curves of the real-time visual image on the Lipschitz continuous boundary domain; The curvature curve model is introduced, and the curvature regularization term is calculated by fitting the divergence of the unit normal vector in the grayscale image; The weight matrix in the adaptive TV model is improved based on the Gaussian kernel, and the regularization coefficient is introduced to construct the adaptive variational problem; An optimization algorithm is used to solve the adaptive variational problem, obtain the optimal approximation coefficients and calculate the real-time smooth visual image.

3. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The method of using the segmentation mark sharpening method to generate the first target visual image, the second target visual image and the third target visual image comprises the following specific steps: The feature extraction layer based on the inverted residual structure mines real-time smooth visual images, and realizes dimension increase, feature mining and dimension reduction of real-time smooth visual images through convolution, depthwise separable convolution and convolution again, and generates feature maps through ReLU and residual connection; Through the coordinate attention mechanism, long-range dependencies are captured in the horizontal and vertical directions and the position information is retained, a direction-aware attention map and a position-aware attention map are generated, and the feature map is enhanced collaboratively to generate an attention feature map; The multi-scale feature fusion layer divides the real-time smooth visual image into multiple patches. Each patch generates a corresponding weight vector through secondary extraction and upsampling, and then generates a weight map based on the position of the patch in the real-time smooth visual image. The attention feature map and the weight map are linearly modulated and spliced ​​through the splicing prediction layer, and then restored through convolution and bicubic interpolation to generate a real-time visual markup image; A single marker is selected, and the pixel values ​​of all pixels in the real-time visual marker image that do not carry the single marker are reset to the boundary pixel values ​​that have the largest difference from the average pixel values ​​of the pixels carrying the single marker, thereby generating a first target visual image, a second target visual image, and a third target visual image.

4. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The generation of real-time enhanced thermal images and the acquisition of the temperature of the current node by combining area screening and colorimetric mapping include the following specific steps: Median filtering is used to filter out noise in real-time thermal images and generate real-time optimized thermal images; Obtain an optimized grayscale image and divide it into multiple local grayscale regions, and accumulate the probability of occurrence of each grayscale value in each local grayscale region to obtain a cumulative distribution function; The total number of grayscale values ​​in the optimized grayscale image is obtained, the total number of grayscale values ​​is subtracted by 1, and then multiplied by the cumulative distribution function of each local grayscale area and rounded down to generate a real-time enhanced thermal image; Retaining the pixel points in the real-time enhanced thermal image that overlap with the pixel points of the non-boundary pixel values ​​in the first target visual image; Colorimetric mapping is performed according to the color temperature mapping function of the pixel value and temperature in the colorimetric bar, the temperatures corresponding to all retained pixels in the real-time enhanced thermal image are obtained, and the average value is used as the temperature of the current node.

5. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The calculation of the oil column change height at the current node includes the following specific steps: The third target visual image is processed by improved recursive filtering to generate a third target optimized visual image; The Sobel operator is used to calculate the gradient magnitude and gradient direction of the third objective optimized visual image, and bilinear interpolation is used to interpolate the four pixel points including the target pixel point in the gradient direction; The double threshold judgment method is used to determine the oil column pixel edge in the third objective optimization visual image, and the oil column pixel height and the infiltration tube pixel height of the current node are obtained through rectangular approximation, and the ratio of the actual infiltration tube height to the infiltration tube pixel height is used as the scaling scale. The actual height of the oil column at the current node is obtained by multiplying the pixel height of the oil column by the telescopic scale, and the difference between the actual height of the oil column and the initial height of the oil column is used to obtain the changed height of the oil column at the current node.

6. A transformer oil multi-modal online detection method as claimed in claim 5, characterized in that: The improved recursive filtering comprises the following specific steps: Set the initial sliding window width, the maximum sliding window width and the judgment threshold, and define the pixel standard deviation of the third target visual image within the sliding window as the local contrast; Before filtering the target pixel, perform local contrast judgment, calculate the local contrast in the current sliding window and determine whether it is less than the judgment threshold; If it is less than the judgment threshold, the pixel value of the target pixel is updated to the median of the pixels in the current sliding window, and the sliding window is moved to the next pixel; If it is greater than or equal to the judgment threshold, the sliding window judgment is performed to determine whether the current sliding window width is less than the maximum sliding window width; If it is less than the maximum sliding window width, the sliding window width is expanded and the local comparison judgment is executed; If it is equal to the maximum sliding window width, update the pixel value of the target pixel to the median of the four adjacent pixels; Repeat the above steps until all pixels in the third target visual image are filtered.

7. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The specific derivation process of the dielectric constant of the current node is as follows: Based on Maxwell's equations, the normal electric displacement is the product of the air dielectric constant and the normal field strength, and is proportional to the normal charge and inversely proportional to the electrode distance, and the normal field strength is derived; Since the normal field strength is the differential of the normal voltage with respect to the electrode distance, the first equation of the normal voltage and the air dielectric constant is derived; Based on the principle of electric displacement invariance combined with Maxwell's equations, the real-time electric displacement is equal to the normal electric displacement, and the real-time charge is equal to the normal charge; The real-time voltage and the real-time dielectric constant of the current node are replaced with the normal voltage and the air dielectric constant in the first equation to generate the second equation; Dividing the second equation by the first equation yields that the ratio of the real-time voltage to the normal voltage is always equal to the ratio of the real-time dielectric constant to the air dielectric constant, and the real-time dielectric constant of the current node is derived.

8. A transformer oil multi-modal online detection method as claimed in claim 1, characterized in that: The dynamic decision-making based on the early warning decision mechanism whether to adjust the monitoring interval or send an early warning prompt includes the following specific steps: Obtaining an aging index of the transformer oil at the current node, and determining whether the aging index is greater than or equal to a first aging threshold; If it is greater than or equal to the first aging threshold, a warning prompt is issued; If it is less than the first aging threshold, determining whether the aging index is less than the second aging threshold; If it is less than the second aging threshold, the monitoring interval is maintained as the initial monitoring interval and the next node is generated; If it is greater than or equal to the second aging threshold and less than the first aging threshold, obtain the quantization layer number where the aging index is located, adjust the monitoring interval to the initial monitoring interval minus the quantization layer number multiplied by the minimum adjustment interval, and generate the next node.

9. A transformer oil multi-modal online detection system, characterized in that: It includes system clock module, data acquisition module, multi-scale analysis module and feedback regulation module; The system clock module sets nodes based on monitoring intervals and sends acquisition signals at each node; The data acquisition module receives the acquisition signal, obtains the real-time visual image, real-time thermal image and real-time voltage and feeds them back to the multi-scale analysis module; The multi-scale analysis module derives the dielectric constant based on Maxwell's equations and the principle of electric displacement invariance, generates the first target visual image, the second target visual image and the third target visual image based on the Qu first a priori smoothing approximation algorithm combined with the segmentation mark sharpening method, processes the real-time thermal image through the filter mapping enhancement method and obtains the temperature by combining regional screening and colorimetric mapping, obtains the acid value based on the second target visual image combined with the fluorescence standard curve, and processes the third target visual image through the improved edge detection and scale expansion method to obtain the height change of the oil column; The feedback regulation module analyzes the dielectric constant, temperature, acid value and oil column height change through a logistic regressor to generate an aging index, and dynamically decides whether to adjust the monitoring interval or send an early warning prompt based on the early warning decision mechanism.

10. A transformer oil multi-mode online detection system as claimed in claim 9, characterized in that: The data acquisition module receives the acquisition signal, lowers the fluorescent indicator, opens the oil extraction channel to allow the transformer oil to pass through the fluorescent indicator, the capillary infiltration tube and the capacitor in sequence, obtains real-time visual images, real-time thermal images and real-time voltage through multi-source visual detection equipment and variable capacitance sensors, and recovers the fluorescent indicator.

Images