Isolator conducting arm deflection detection device based on visual detection

Abstract

This invention relates to the field of power equipment monitoring technology and discloses a vision-based detection device for the flexural deformation of a disconnector switch's conductive arm. The device includes a multi-modal data acquisition module that adjusts the polarization state based on illumination feedback and acquires images of the polarized light field and a three-dimensional temperature field; a thermal correction and normal vector calculation module that derives the local refractive index gradient using the temperature field to correct thermal distortion and calculates the surface normal vector field of the conductive arm; a dynamic mode shape extraction module that extracts the intrinsic natural frequencies based on optical flow; a three-dimensional shape reconstruction module that fuses the initial depth map and the normal vector field to generate a three-dimensional geometric mesh model; and a physical coupling quantization evaluation module that inputs the mesh model, temperature field, and natural frequencies into a physical information neural network, using the geometric fitting residual, the thermal stress coupled elastic beam partial differential equation, and the frequency as joint constraints to output flexural quantization parameters. This invention eliminates interference from specular light and thermal flow distortion and, combined with physical constraints, improves the accuracy of deformation detection.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically to a visual inspection-based device for detecting the flexural deformation of the conductive arm of a disconnector switch. Background Technology

[0002] The conductive arms of high-voltage disconnect switches in substations are prone to flexural deformation due to multiple environmental stresses during long-term energized operation. Obtaining their three-dimensional morphology and flexural parameters under operating conditions is a prerequisite for equipment operation and maintenance diagnosis. Traditional contact-based measurements require power outages and cannot reflect the true mechanical characteristics of the equipment under energized thermal stress. Non-contact visual reconstruction technology, when implemented in substations, is hampered by the dual interference of strong light and heat environments. On the one hand, the specular reflection of the smooth metal surface of the conductive arm can lead to local pixel overexposure and loss of texture features in the image. On the other hand, the Joule heat generated by the high current operation of the conductive arm creates a non-uniform spatial temperature gradient in the optical path, and the spatiotemporal changes in air refractive index cause deflection of the visual imaging optical path. This imaging distortion caused by self-heating radiation and high-frequency surface reflection interference fundamentally disrupt the basic geometric mapping relationship of visual inspection.

[0003] In the visual shape reconstruction stage, existing algorithms generally rely on macroscopic low-frequency feature matching of images to generate initial depth data, lacking effective means to extract the microscopic high-frequency topological structure of metals. Due to the difficulty in deeply fusing macroscopic spatial contours with microscopic shape features, the reconstructed geometric mesh elements lose crucial microscopic deformation details, making it difficult to construct a truly high-fidelity 3D model with physical details.

[0004] In the deformation quantification assessment stage, traditional visual algorithms can only output static spatial coordinate system parameters, completely detaching the dynamic mechanical stiffness response of the conductive arm under service conditions. Existing pure data-driven deep learning networks, when processing such deformation assessments, typically directly map optical images into displacement results, lacking objective physical constraints from mechanical partial differential equations and thermal stress boundary conditions in their internal computational nodes. Because physical laws are not introduced to limit the solution domain, the model ignores the degradation of the elastic modulus of metallic materials under heating conditions, leading to deformation calculation results that deviate from the actual physical characteristics of the material. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a vision-based detection device for the flexural deformation of the conductive arm of a disconnecting switch. This device solves the problem that existing non-contact vision detection methods are difficult to accurately reconstruct the microstructure of metals under strong light overexposure and thermal distortion interference from hot airflow, and lack mechanical and physical constraints, resulting in lower deformation measurement accuracy of the conductive arm of a high-voltage disconnecting switch under energized operation than the inaccurate mechanical assessment.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a vision-based detection device for the flexural deformation of a disconnector switch conductive arm, the device comprising: The multimodal data acquisition module is used to adjust the polarization phase of the light entering the data acquisition device based on the light intensity feedback of the current field of view, and simultaneously acquire the polarization light field image array and three-dimensional temperature field image of the conductive arm. The thermal correction and normal vector calculation module is used to derive the spatial local refractive index gradient based on the three-dimensional temperature field image to perform thermal distortion correction on the polarization field image array, and calculate the normal vector field of the conductive arm surface based on the corrected polarization field image array. The dynamic mode extraction module is used to acquire the time-series image sequence of the conductive arm and extract the intrinsic natural frequencies of the conductive arm in operation based on the optical flow method. The three-dimensional shape reconstruction module is used to fuse the initial depth map corresponding to the polarized light field image array with the normal vector field to generate a three-dimensional geometric mesh model of the conductive arm. The physical coupling quantization evaluation module is used to input the three-dimensional geometric mesh model, the three-dimensional temperature field image and the intrinsic natural frequency into the physical information neural network for processing. The physical information neural network uses the geometric fitting residual between the network prediction result and the three-dimensional geometric mesh model, the partial differential equation of elastic beam deflection including thermal stress coupling and the intrinsic natural frequency as joint physical constraints to output the deflection quantization parameters of the conductive arm.

[0007] Preferably, to address the issues of uneven surface air density and optical path deflection caused by heat generation of the conductive arm under high current operation, the thermal correction and normal calculation module performs distortion correction based on the physical mapping relationship between air refractive index and temperature. The calculation of its spatial local refractive index gradient follows the following principle formula: ; In the formula, Representing spatial coordinates The local refractive index at that location; Indicates the standard reference atmospheric refractive index; This represents the local absolute temperature measured from the three-dimensional temperature field image; The reference ambient temperature is obtained in real time by the device's built-in environmental sensor. This represents the air refractive index temperature coefficient, a constant obtained by fitting historical prior experimental data. By calculating the integral offset of this refractive index field along the light propagation path, reverse displacement compensation is achieved for the thermal distortion pixel coordinates in the polarized light field image array.

[0008] Preferably, when the thermal correction and normal calculation module calculates the normal vector field of the conductive arm surface, it extracts the image intensity under different polarization directions to calculate the Stokes parameters. Combining the polarization bidirectional reflection distribution function and the complex refractive index characteristics of the metal, it decouples the diffuse reflection component of the metal surface from the specular reflection specular highlight, thereby mapping the zenith angle and azimuth angle of the local normal vector.

[0009] Preferably, the 3D shape reconstruction module achieves deep fusion by constructing a cross-scale variational energy functional model. The initial visual depth map has macroscopic low-frequency accuracy, while the normal vector field based on polarization calculation possesses microscopic high-frequency details. Its variational energy functional model is expressed as: ; In the formula, Represents the total energy functional; This represents the three-dimensional depth distribution of the target to be optimized; The spatial gradient representing the target depth; This represents the desired image plane gradient field derived from the normal vector field; This represents the initial depth map; The weighted penalty coefficients, representing the balance between local and global constraints, are obtained based on 3D scanning calibration data from similar historical equipment. This energy functional model is minimized using a variational method to generate the final 3D geometric mesh model.

[0010] Preferably, the overall loss function of the physical information neural network is a data-driven loss. Partial differential equation constraint loss and dynamic frequency constraint loss Weighted combination.

[0011] The constraint loss of the partial differential equation The residual calculation terms include the partial differential equations for the deflection of a non-isothermal elastic beam. Since the conductive arm undergoes stiffness degradation under thermal stress, the network internally uses a three-dimensional temperature field image as the thermal stress boundary condition, and calculates the temperature-dependent spatial elastic modulus parameter distribution using the following formula:

[0012] In the formula, This represents the local dynamic elastic modulus affected by temperature. Indicates the metal material of the conductive arm at the reference temperature The standard elastic modulus is obtained based on the material's factory mechanical parameters. This represents the temperature softening coefficient of the metallic material, obtained based on historical tensile test data. The physical information neural network utilizes this dynamic elastic modulus. Substitute the values ​​into the partial differential equation to calculate the structural residuals, ensuring that the output flexural quantization parameters strictly conform to the laws of mechanics and physics.

[0013] Preferably, the multimodal data acquisition module includes an illumination evaluation unit, a polarization state adjustment unit, a light field acquisition unit, and an infrared acquisition unit; the polarization state adjustment unit is equipped with a phase delay component, which is used to dynamically change the delay angle according to the control signal fed back by the illumination intensity, so as to actively adapt to the high light interference environment.

[0014] Preferably, the dynamic mode shape extraction module continuously acquires the time-series image sequence at a preset frame rate, extracts the pixel-level velocity field according to the optical flow conservation equation and integrates it to obtain the time-series displacement field, and then performs a fast Fourier transform to extract the fundamental frequency response, determine the intrinsic natural frequency and the corresponding spatial mode shape vector, and uses this as a dynamic physical constraint reflecting the overall stiffness boundary of the conductive arm to be input into the neural network.

[0015] Preferably, the deflection quantification parameters output by the physical coupling quantification evaluation module include the maximum deflection, the relative deflection angle of the connection point, and the spatial extreme curvature; the device also includes a multi-level safety early warning module, which is used to compare the deflection quantification parameters with a preset deformation classification threshold database to trigger an early warning signal, and generate and output the deformation trend curve of the conductive arm by combining the pre-stored historical deflection quantification parameters.

[0016] This invention provides a vision-based detection device for detecting the flexural deformation of the conductive arm of an isolating switch. It offers the following advantages: 1. This invention utilizes a three-dimensional temperature field to correct the influence of thermal distortion on the polarized light field, eliminating coordinate offsets caused by thermal disturbances; combined with polarization state feedback adjustment and polarization parameter calculation, it removes specular reflection interference, outputs the microscopic normal vector field of the target surface, and realizes the underlying calibration and linkage of multimodal physical data.

[0017] 2. This invention constructs a cross-scale variational energy functional model, performs a minimization integral operation on the initial depth map obtained by visual matching and the microscopic topological details contained in the normal vector field, and fuses and outputs a three-dimensional geometric mesh model, thus solving the technical problem of losing local structural parameters in the shape reconstruction of a single visual algorithm.

[0018] 3. This invention uses the three-dimensional temperature field, the intrinsic natural frequency, and the three-dimensional geometric mesh model as joint constraints to input physical information into the neural network. By limiting the solution domain through geometric boundaries, non-isothermal partial differential equations, and frequency residuals, the solution results of the flexural state conform to the multi-physics coupling law and the thermal degradation law. Attached Figure Description

[0019] Figure 1 This is a block diagram of the disconnector switch conductive arm flexure deformation detection system of the present invention; Figure 2 This is a schematic diagram of the hardware deployment and spatial structure of the device of the present invention; Figure 3 This is a flowchart of the thermal distortion correction and normal calculation of the present invention; Figure 4 This is a schematic diagram illustrating the principle of three-dimensional topography reconstruction using cross-scale variational energy functionals in this invention. Figure 5 This is a multi-physics coupling topology diagram of the physical information neural network of the present invention; Figure 6 This is a schematic diagram of the multi-level early warning logic and deformation trend evolution curve of the present invention.

[0020] Explanation of reference numerals in the attached figures: 100, observation pan-tilt unit; 101, light field acquisition unit; 102, polarization state adjustment unit; 103, infrared acquisition unit; 104, ambient temperature and humidity sensor; 105, data processing host; Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see the appendix Figure 1 To be continued Figure 6 This invention provides a vision-based detection device for the flexural deformation of a disconnector switch conductive arm. This device is deployed at a substation site to perform non-contact three-dimensional morphology measurement and physical and mechanical condition assessment of the conductive arm of a high-current disconnector switch in operation.

[0023] The vision-based detection device for detecting the flexural deformation of the conductive arm of the disconnector switch may include: a multimodal data acquisition module, a thermal correction and normal calculation module, a dynamic mode extraction module, a three-dimensional morphology reconstruction module, and a physical coupling quantitative evaluation module. The above modules are connected in sequence through a high-speed data bus to realize data flow.

[0024] At the physical level, this device is mainly implemented using an observation gimbal 100 and a data processing host 105. The multimodal data acquisition module serves as the data input for this device, and its hardware is integrated and installed on the observation gimbal 100. It includes a light field acquisition unit 101, a polarization state adjustment unit 102, an infrared acquisition unit 103, and an environmental temperature and humidity sensor 104. The data processing host 105 internally houses a processor and storage medium for running the algorithm programs related to the thermal correction and normal calculation module, the dynamic mode shape extraction module, the three-dimensional topography reconstruction module, and the physical coupling quantitative evaluation module.

[0025] The multimodal data acquisition module is used to adjust the polarization phase of the light entering the data acquisition device based on the light intensity feedback of the current field of view, and simultaneously acquire the polarization field image array and three-dimensional temperature field image of the conductive arm.

[0026] The light field acquisition unit 101 employs a high frame rate microlens array polarization camera, with four nanowire gratings of different transmission axis directions covering its focal plane array. This allows for the acquisition of image arrays in four polarization directions (0°, 45°, 90°, and 135°) in a single exposure. The infrared acquisition unit 103 uses an uncooled focal plane microbolometer. The light field acquisition unit 101 and the infrared acquisition unit 103 are spatially fixed inside the observation gimbal 100 by a rigid robotic arm. Their optical axis centers remain parallel, and they undergo binocular extrinsic parameter matrix calibration using a stereo calibration plate before leaving the factory. This ensures they are in a spatially registered state, enabling them to synchronously trigger and capture the polarized light field image array and the three-dimensional temperature field image, respectively, guaranteeing a strict correspondence between the coordinates of the same spatial point in the optical and thermodynamic images.

[0027] Under strong sunlight or under the searchlights used for nighttime inspections at substations, the surface of the directly collected metal conductive arm has a severe specular high-gloss reflection area, resulting in pixel saturation and overexposure in the visual image, and loss of local texture information. The multimodal data acquisition module is equipped with an illumination evaluation unit and a polarization state adjustment unit 102. The illumination evaluation unit runs in the data processing host 105 and extracts the grayscale histogram of the image output by the light field acquisition unit 101 in real time to evaluate the illumination intensity distribution in the current field of view. When the proportion of pixels in the high grayscale range in the image is detected to exceed a set threshold, a control signal corresponding to the illumination intensity feedback is generated.

[0028] The polarization state adjustment unit 102, as a physical hardware execution terminal, is installed at the front end of the optical path of the light field acquisition unit 101. The polarization state adjustment unit 102 is equipped with a phase delay component, which is a birefringent nematic liquid crystal delay film. The polarization state adjustment unit 102 dynamically changes the applied voltage according to the control signal sent by the data processing host 105, and dynamically changes the delay angle of the phase delay component by utilizing the birefringence effect generated by the deflection of liquid crystal molecules. By actively adjusting the polarization state phase difference of the incident light, the high-light reflection beam in a specific vibration direction is effectively attenuated to enter the light field acquisition unit 101, thereby eliminating pixel overexposure interference caused by specular reflection at the optical hardware acquisition level and providing low-noise background light field data for microscopic normal calculation.

[0029] In this embodiment, the thermal correction and normal vector calculation module receives a three-dimensional temperature field image and a polarized light field image array synchronously transmitted by the multimodal data acquisition module. It then performs thermal distortion correction on the polarized light field image array and further calculates the normal vector field of the conductive arm surface. Under high current operation, the conductive arm generates Joule heating, causing the surface air to heat up and forming a spatial temperature gradient. This results in uneven distribution of air density and refractive index, causing physical deflection of the light propagation path.

[0030] In this invention, the thermal correction and normal calculation module transforms the three-dimensional temperature field image into a local refractive index gradient field based on the physical mapping relationship between air refractive index and temperature. The calculation of the physical mapping relationship of refractive index is based on the following formula:

[0031] In the formula, Representing spatial coordinates The local refractive index at that location; This represents the standard reference atmospheric refractive index constant, which is obtained from the standard atmospheric environmental parameters in the optical handbook. This represents the local absolute temperature extracted from the coordinates corresponding to the three-dimensional temperature field image; The reference ambient temperature is acquired in real time by the built-in environmental sensor. This represents the temperature coefficient of air refractive index, which is obtained by fitting historical empirical data from a handbook of air thermodynamic properties.

[0032] Based on the generated local refractive index gradient field, the thermal correction and normal calculation module establishes a refractive index compensation model. Under this model, the algorithm calculates the integral refraction offset along the line of sight based on the principle of ray tracing. Based on the calculated integral refraction offset, the thermal correction and normal calculation module performs reverse displacement mapping on the thermally distorted pixel coordinates in the polarized light field image array to complete the thermal distortion correction of the polarized light field image array and eliminate the physical interference of hot airflow disturbance on visual imaging.

[0033] After completing thermal distortion correction, the thermal correction and normal vector calculation module begins to calculate the normal vector field of the conductive arm surface. This module extracts the image intensity matrix under different polarization directions from the corrected polarized light field image array to calculate the Stokes parameters. The calculation logic of the Stokes parameters is as follows:

[0034]

[0035]

[0036] In the formula, Represents the total light intensity component; This represents the intensity difference components of linearly polarized light in the horizontal and vertical directions. This represents the intensity difference component of linearly polarized light in a direction orthogonal at 45 degrees. , , , These represent the pixel grayscale intensity values ​​of the corrected polarized light field image array in the polarization direction channels of 0 degrees, 45 degrees, 90 degrees, and 135 degrees, respectively. These intensity values ​​are directly extracted from the single exposure data matrix of the multimodal data acquisition module.

[0037] Based on the calculated Stokes parameters, the thermal correction and normal calculation module further calculates the linear polarization degree and polarization angle of each pixel. The derivation and calculation of the relevant parameters are completed according to the following formulas:

[0038]

[0039] In the formula, Indicates the degree of linear polarization; Indicates the polarization angle, when When the effective energy of a pixel is below a preset threshold, the pixel is marked as an inefficient light sampling point and will not participate in the subsequent normal vector calculation to prevent division by zero anomalies.

[0040] Combining the polarization bidirectional reflectance distribution function and the complex refractive index characteristics of metals, the thermal correction and normal calculation module deconstructs the light reflection model using the aforementioned polarization parameters. The reflected light from the conductive arm's metal surface includes a diffuse reflection component and a specular reflection highlight component. The algorithm uses a linear polarization degree threshold condition to separate and decouple the diffuse reflection component from the specular reflection highlight.

[0041] In the decoupled specular reflection light field region, there is a direct projection relationship between the polarization angle and the azimuth angle of the local normal vector of the metal surface. The algorithm establishes the mapping relationship between the polarization angle and the azimuth angle, and solves for the azimuth angle of the local surface normal vector; it also solves for the zenith angle of the local surface normal vector using the correspondence between the polarization extremum condition and the incident angle in the Fresnel equation. The thermal correction and normal vector calculation module integrates the zenith angle and azimuth angle parameters corresponding to all pixels in the image to construct a conductive arm surface normal vector field with high-frequency microscopic physical details.

[0042] In this embodiment, the dynamic mode shape extraction module continuously acquires a sequence of time-series images at a preset frame rate, using this as the initial data source for analyzing the operating state of the conductive arm. This module reads adjacent frame image matrices sequentially along the time axis and calculates the grayscale gradient of each pixel in both the spatial and temporal dimensions. In this invention, the algorithm establishes a pixel motion model based on the optical flow conservation equation, extracting the pixel-level velocity field on the surface of the conductive arm. The calculation logic of the optical flow conservation equation is as follows:

[0043] In the formula, and These represent the spatial pixel grayscale gradients of the image in the horizontal and vertical directions, respectively. The system obtains these spatial gradient parameters by performing a difference operation on adjacent pixels within the image matrix. This represents the temporal grayscale gradient of the corresponding pixel position between adjacent image frames; and These represent the horizontal and vertical velocity components of the pixel-level velocity field to be solved, respectively. The system solves this equation by combining multi-scale pyramid and optical flow smoothing constraint equations. Specifically, the system introduces a global smoothing constraint term based on the Horn-Schunck algorithm, which minimizes the Laplacian operator of the velocity field components. and The sum of squares and integrals are used to obtain the velocity vector matrix of the global pixels through alternating iterative calculations.

[0044] After obtaining the velocity components of each pixel using the above calculations, the dynamic mode extraction module performs discrete integration on the extracted pixel-level velocity field along the time axis. The integration operation accumulates the velocity vectors of each frame along a continuous time sequence to obtain the temporal displacement field during the operation of the conductive arm. The evolution of the temporal displacement field is calculated based on the following formula:

[0045] In the formula, Represents spatial pixel coordinates At the time point The cumulative displacement vector; This represents the velocity component output from the aforementioned calculation. and The synthesized two-dimensional velocity vector; This parameter represents the time interval between adjacent image frames and is directly obtained based on the hardware-preset sampling frame rate parameter of the data acquisition device. After integration calculation, the system generates a two-dimensional time-varying displacement matrix containing spatial coordinate sequences and time mapping information.

[0046] After acquiring the time-series displacement field data, the dynamic mode shape extraction module performs a Fast Fourier Transform (FFT) on the time-series displacement data column for each pixel in the spatial coordinate system. The displacement data array in the time domain is transformed to the frequency domain, generating the corresponding amplitude spectrum matrix. The calculation model of the FFT is as follows:

[0047] In the formula, Indicates the first in the frequency domain Complex spectral response parameters corresponding to discrete frequency points; This represents the total number of sampling frames within the time-series image sequence extraction window; Index labels representing discrete data in a time series; This represents the imaginary unit. Using this frequency domain transformation formula, the algorithm extracts disordered vibration data from the system environment and locates the absolute peak of the energy distribution matrix in the generated amplitude spectrum.

[0048] The dynamic mode shape extraction module extracts the frequency response components corresponding to the locations of the peak and extreme points of the amplitude spectrum energy, and uses these as the fundamental frequency response. Based on the extracted fundamental frequency response, this module determines the intrinsic natural frequencies of the conductive arm in its current operating state. Simultaneously, the algorithm extracts the spatial complex matrix under this fundamental frequency response value, analyzes the amplitude and phase information in the complex matrix, determines the relative proportions of spatial displacement of each pixel during natural vibration, and generates the corresponding spatial mode shape vector. The system outputs the intrinsic natural frequencies and spatial mode shape vectors, constructing dynamic characteristic parameters reflecting the stiffness state of the conductive arm's mechanical structure.

[0049] In this embodiment, the 3D topography reconstruction module receives the initial depth map corresponding to the polarized light field image array and the previously calculated normal vector field, performs cross-scale 3D topography data fusion operation, and the system uses the normal vector field to derive the desired image plane gradient field. For the normal vector mapped to any pixel in the normal vector field, the algorithm extracts the projection components of this vector along three independent spatial dimensions in a 3D Cartesian coordinate system; using the spatial algebraic ratio relationship between the projection components, the theoretical depth change rate in the horizontal and vertical directions of the image plane is calculated, thereby generating the desired image plane gradient field corresponding to the pixel coordinates of the entire image. The calculation logic of the vector elements inside the desired image plane gradient field is as follows:

[0050]

[0051] In the formula, , as well as Representing spatial pixel coordinates The numerical parameters of the orthogonal projection of the local normal vector at a point along the three axes of the Cartesian coordinate system; and Together they constitute the desired image plane gradient field. Furthermore, considering the steep edges in the depth direction on the surface of the conductive arm (e.g., bends caused by severe deformation), the projection component of the normal in the depth direction... It will approach zero, when When the absolute value of the gradient is lower than the preset numerical stability threshold, the algorithm extracts the local neighborhood normal vector of the pixel and performs weighted interpolation or smooth extrapolation to stably calculate the expected gradient components, thereby ensuring the stability of the numerical solution of subsequent Poisson reconstruction.

[0052] After generating the desired image plane gradient field, the 3D shape reconstruction module uses the normal vector field as a local normal constraint and the initial depth map as a global depth constraint to construct a cross-scale variational energy functional model. This variational energy functional model integrates the preservation of microscopic details and the fitting of macroscopic contours in visual 3D reconstruction into a unified integral operation model, with the specific equation form as follows:

[0053] In the formula, This indicates the total variational energy functional that requires a minimization solution; This represents the two-dimensional continuous integral space domain corresponding to the image photosensitive matrix; This represents the three-dimensional depth distribution matrix of the target to be solved; This represents the gradient matrix parameter obtained by performing spatial differentiation on the three-dimensional depth distribution matrix of the target in the two-dimensional image plane; This represents the initial depth map pre-calculated and output by the multimodal data acquisition module based on light field feature matching; This represents the penalty coefficient used to balance the weights of local and global constraints. This penalty coefficient parameter is obtained by fitting the comparison error of offline calibration data of standard three-dimensional laser scanning of the same type of conductive arm in history.

[0054] Based on the established mathematical model, the system minimizes the variational energy functional model through numerical analysis, achieving the fusion and integration of the local normal gradient and the global depth. The algorithm introduces Euler-Lagrange partial differential equations to transform the functional extremum operation into a discrete linear Poisson equation solution process. The system calls a preprocessing conjugate gradient algorithm to iteratively solve this discrete linear Poisson equation system multiple times. When the convergence threshold of the iterative operation is reached, the system outputs a two-dimensional depth distribution matrix that satisfies the functional minimum condition.

[0055] Based on the two-dimensional depth distribution matrix output by the above solution, the three-dimensional shape reconstruction module maps the two-dimensional planar pixel coordinate array and the calculated depth parameters to a three-dimensional Euclidean physical space. The algorithm connects adjacent discrete coordinate points in three-dimensional space according to the set spatial sampling interval parameter, and calls the Delaunay triangulation algorithm to construct a polygonal topological element structure based on the principle of maximizing the minimum angle. The system summarizes all topological elements to generate a three-dimensional geometric mesh model of the conductive arm with a complete structural shape, and uses this model as the reference input data for the network geometric boundary of the subsequent physical field coupling quantization calculation.

[0056] In this embodiment, the physical coupling quantization evaluation module is used to perform joint inference calculations of multiple physics fields. In this invention, the physical information neural network internally employs a multi-layer feedforward fully connected multilayer perceptron as a continuous function approximator network structure, for example, setting 4 to 8 hidden layers, each containing 50 to 100 neurons, and using the Tanh function as the nonlinear activation function. The input of the physical information neural network is set to the three-dimensional spatial physical coordinates of the conductive arm. Its output is set to the predicted deflection field function matrix at the corresponding coordinates. The three-dimensional geometric mesh model generated by the three-dimensional topography reconstruction module, the three-dimensional temperature field image synchronously captured by the multi-modal data acquisition module, and the intrinsic natural frequencies calculated by the dynamic mode extraction module are not directly used as inputs to network nodes. Instead, they are used as external physical prior data inputs to the internally deployed loss calculation module to construct a joint physical constraint mechanism.

[0057] The physical information neural network constructs a joint physical constraint mechanism based on the input data. The overall loss function of the physical information neural network is a weighted combination of data-driven loss, partial differential equation constraint loss, and dynamic frequency constraint loss. The system calculates the output state matrix of the aforementioned multi-layer feedforward fully connected neurons based on the network's forward propagation algorithm. This data-driven loss is used to constrain the geometric consistency between the predicted deflection field output by the network and the three-dimensional geometric mesh model at discrete sampling points, based on the geometric fitting residual. The calculation logic of the data-driven loss is as follows:

[0058] In the formula, Indicates the data-driven loss value; This represents the total number of discrete sampling points in three-dimensional space extracted from the surface of the three-dimensional geometric mesh model; The physical information neural network is for the first The predicted deflection value output from the coordinates of each spatial sampling point; This represents the corresponding spatial reference deflection value extracted based on the three-dimensional coordinate matrix data of the three-dimensional geometric mesh model.

[0059] The constraint loss in the partial differential equation includes the residual calculation term of the partial differential equation for the deflection of a non-isothermal elastic beam. A physical information neural network extracts the thermodynamic calibration values ​​of the corresponding coordinate pixels in the three-dimensional temperature field image, calculates the temperature-affected spatial elastic modulus parameter distribution matrix, and the conversion calculation formula for the spatial elastic modulus is as follows:

[0060] In the formula, This represents the local dynamic elastic modulus affected by temperature, generated by network derivation. Indicates the metal material of the conductive arm at the reference temperature The standard elastic modulus is obtained by analyzing the mechanical parameters of the material of the conductive arm being tested from the manufacturer's mechanical specifications document. This represents the temperature softening coefficient of metallic materials, which is obtained by fitting prior empirical data from historical standard tensile thermal stress tests of the same type of material.

[0061] After calculating and obtaining the local dynamic elastic modulus, the network uses the distribution matrix of this dynamic elastic modulus as a thermal stress boundary condition to calculate the residual of the partial differential equation for the deflection of the non-isothermal elastic beam. The calculation logic of the constraint loss of the partial differential equation is as follows:

[0062] In the formula, This represents the constraint loss value of a partial differential equation; This represents the total number of collocation points set within the computational grid domain inside the partial differential equation; This represents the moment of inertia constant of the conductor arm's cross section; This represents the distributed parameters of the external electrodynamic load at the designated location. The network uses automatic differentiation techniques to calculate... Substituting the higher-order derivatives relative to spatial coordinates into the above loss formula, i.e., the dynamic computation graph constructed based on the deep learning framework, the chain rule is used to process the network output. By performing precise analytical differentiation of the input coordinates, the grid discretization truncation error caused by the traditional finite difference method is avoided.

[0063] The dynamic frequency constraint loss is used to establish residual constraints between the acquired intrinsic natural frequencies and the theoretical frequencies derived internally by the physical information neural network based on dynamic stiffness. The network derives the theoretical frequency response characteristic value corresponding to the stiffness parameter based on the local dynamic elastic modulus parameter matrix, and then calculates the sum of squares of the differences between the derived characteristic value and the input intrinsic natural frequencies to generate the dynamic frequency constraint loss. The system calls the Adam optimizer to perform an initial global gradient descent search, and seamlessly switches to a finite-memory quasi-Newton optimizer for local exact optimization after the rate of change of the loss function converges. According to the chain rule, backpropagation is performed on the above three combined losses, and the network layer weight matrix parameters are iteratively updated until the overall loss function value decreases to the preset convergence limit.

[0064] After the physical information neural network converges, it outputs the final version of the predicted deflection field. The physical coupling quantization evaluation module performs inverse calculus calculation based on the distribution function of the predicted deflection field parameters. The algorithm extracts the extreme point information of the numerical distribution matrix of the deflection field to generate the maximum deflection parameter. The first spatial derivative of the predicted deflection field distribution function is calculated to obtain the relative deflection angle parameter of the connection point. The second spatial derivative of the distribution function is calculated to obtain the spatial extreme curvature parameter. The system integrates and extracts the above parameters and finally outputs the deflection quantization parameter.

[0065] In this embodiment, the system also includes a multi-level safety early warning module for performing final structural status diagnosis and maintenance prompts. The multi-level safety early warning module receives the deflection quantification parameters parsed and output by the physical coupling quantification evaluation module. The deflection quantification parameters, obtained from the aforementioned calculation steps, include maximum deflection, relative deflection angle of the connection point, and spatial extreme curvature. The multi-level safety early warning module internally deploys a preset deformation classification threshold database, which contains factory-set design tolerance limits for different models of conductive arms.

[0066] The multi-level safety early warning module extracts the deflection quantification parameters, performs dimensionless normalization on them, and calculates and constructs a comprehensive structural deformation damage index based on various parameters. The system assesses the overall deterioration degree of the conductive arm based on this damage index. The calculation model for the comprehensive structural deformation damage index is as follows:

[0067] In the formula, This represents the overall structural deformation damage index output by the system calculation; , and These correspond to the maximum deflection, the relative deflection angle of the connection point, and the spatial extreme curvature input from the physical coupling quantification evaluation module, respectively. , and These represent the maximum allowable deflection, the allowable angle, and the allowable curvature constant stored internally in the deformation classification threshold database, respectively. , and These represent the weighting coefficients for each deformation parameter. The weighting coefficients are pre-calibrated based on historical fatigue failure test records of metal structures.

[0068] The multi-level safety early warning module compares the calculated comprehensive structural deformation damage index with the set level ranges within the deformation classification threshold database. When the comprehensive structural deformation damage index exceeds the upper limit of the set normal service range, the system determines that the deflection quantification parameter exceeds the set range. Based on the specific numerical range into which the damage index falls, the system triggers corresponding graded early warning signals. In specific implementation, for example, the normal service range is set as follows: The warning range is The danger zone is When the index falls into the warning range, a level two warning is triggered to restrict load operation. When the index falls into the danger range, a level one warning is triggered to trigger emergency power outage maintenance. An alarm command data packet containing the spatial location code of the abnormal equipment and deformation parameters is sent to the station-end centralized control center through the underlying communication gateway.

[0069] After determining the current state, the multi-level safety warning module retrieves pre-stored historical deflection quantization parameters from the local storage medium. The system then concatenates the currently calculated deflection quantization parameters with the historical deflection quantization parameters according to the running time sequence, forming a deformation time evolution sequence matrix. The algorithm performs polynomial regression fitting calculations on this sequence matrix to construct a time evolution mapping function for the maximum deformation. The mathematical expression of the mapping function is as follows:

[0070] In the formula, Indicates time node Corresponding expected trend deformation; , and These represent the zero-order baseline coefficient, the first-order linear evolution coefficient, and the second-order nonlinear degradation coefficient, respectively, derived from regression fitting calculations. These three coefficients are obtained by solving the matrix equation based on the least squares method by minimizing the mean square error objective function between the historical discrete deflection quantification parameters and the trend variables at the corresponding time nodes.

[0071] Based on the aforementioned polynomial regression calculation model, the multi-level safety early warning module generates a continuous numerical evolution sequence, which is then rendered and output as a deformation trend curve of the conductive arm. The system synchronously pushes this deformation trend curve of the conductive arm and the aforementioned graded early warning signals to the human-machine interface, intuitively displaying the mechanical deformation trajectory of the conductive arm under long-term high-load operation and environmental stress coupling, thus completing the entire closed-loop detection and quantitative evaluation process.

Claims

1. A vision-based detection device for detecting the flexural deformation of a disconnector switch conductive arm, characterized in that, include: The multimodal data acquisition module is used to adjust the polarization state phase of the light entering the data acquisition device based on the light intensity feedback of the current field of view, and simultaneously acquire the polarization field image array and three-dimensional temperature field image of the conductive arm. The thermal correction and normal vector calculation module is used to derive the spatial local refractive index gradient based on the three-dimensional temperature field image to perform thermal distortion correction on the polarization field image array, and calculate the normal vector field of the conductive arm surface based on the corrected polarization field image array. The dynamic mode extraction module is used to acquire the time-series image sequence of the conductive arm and extract the intrinsic natural frequencies of the conductive arm in operation based on the optical flow method. The three-dimensional shape reconstruction module is used to fuse the initial depth map corresponding to the polarized light field image array with the normal vector field to generate a three-dimensional geometric mesh model of the conductive arm. The physical coupling quantization evaluation module is used to input the three-dimensional geometric mesh model, the three-dimensional temperature field image and the intrinsic natural frequency into the physical information neural network for processing. The physical information neural network uses the geometric fitting residual between the network prediction result and the three-dimensional geometric mesh model, the partial differential equation of elastic beam deflection including thermal stress coupling and the intrinsic natural frequency as joint physical constraints to output the deflection quantization parameters of the conductive arm.

2. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 1, characterized in that, The multimodal data acquisition module includes an illumination evaluation unit, a polarization state adjustment unit, a light field acquisition unit, and an infrared acquisition unit; The illumination evaluation unit is used to evaluate the illumination intensity distribution within the current field of view and generate a control signal corresponding to the illumination intensity feedback; the polarization state adjustment unit is equipped with a phase delay component, which is used to dynamically change the delay angle of the phase delay component according to the control signal to adjust the polarization state of the incident light; the light field acquisition unit and the infrared acquisition unit are in a spatially registered state, which is used to synchronously trigger and capture the polarized light field image array and the three-dimensional temperature field image respectively.

3. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 1, characterized in that, The thermal correction and normal calculation module is specifically used for: When performing thermal distortion correction, Based on the physical mapping relationship between air refractive index and temperature, the three-dimensional temperature field image is transformed into a local refractive index gradient field; a refractive index compensation model is established based on the local refractive index gradient field to calculate the integral refractive offset on the light propagation path, and the thermal distortion pixel coordinates in the polarized light field image array are corrected accordingly.

4. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 3, characterized in that, The thermal correction and normal vector calculation module is specifically used to calculate the normal vector field on the surface of the conductive arm when: The image intensities under different polarization directions in the corrected polarized light field image array are extracted to calculate the Stokes parameters; the degree of linear polarization and the polarization angle are calculated based on the Stokes parameters; the diffuse reflection component of the metal surface and the specular reflection specular highlight are decoupled by combining the polarization bidirectional reflection distribution function and the complex refractive index characteristics of the metal; the zenith angle and azimuth angle of the local normal vector are calculated through the mapping relationship between the extremum condition and the polarization angle, and then the normal vector field is constructed.

5. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 1, characterized in that, The dynamic mode extraction module, when extracting the intrinsic natural frequencies of the conductive arm in operation, is specifically used for: The time-series image sequence is continuously acquired at a preset frame rate. The pixel-level velocity field is extracted according to the optical flow conservation equation, and the time-series displacement field during the operation of the conductive arm is obtained by time integration. The time-series displacement field is subjected to fast Fourier transform to extract the fundamental frequency response in the amplitude spectrum and determine the intrinsic natural frequency and the corresponding spatial mode shape vector.

6. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 1, characterized in that, The 3D topography reconstruction module is specifically used to generate the 3D geometric mesh model of the conductive arm: The desired image plane gradient field is derived using the normal vector field. The normal vector field is used as a local normal constraint, and the initial depth map is used as a global depth constraint to construct a cross-scale variational energy functional model. Depth fusion and integration are achieved by minimizing the variational energy functional model to generate the three-dimensional geometric mesh model.

7. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 1, characterized in that, The overall loss function of the physical information neural network is composed of a weighted combination of data-driven loss, partial differential equation constraint loss, and dynamic frequency constraint loss.

8. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 7, characterized in that, The data-driven loss is used to constrain the predicted deflection field output by the network and the geometric consistency of the three-dimensional geometric mesh model at discrete sampling points based on the geometric fitting residual. The partial differential equation constraint loss includes the residual calculation term of the partial differential equation of non-isothermal elastic beam deflection, wherein the spatial elastic modulus parameter distribution affected by temperature is calculated based on the three-dimensional temperature field image, and is used as a thermal stress boundary condition to calculate the residual of the partial differential equation of non-isothermal elastic beam deflection. The dynamic frequency constraint loss is used to establish residual constraints between the intrinsic natural frequencies and the theoretical frequencies derived from the dynamic stiffness within the physical information neural network.

9. The vision-based detection device for detecting the flexural deformation of the conductive arm of a disconnector switch according to claim 8, characterized in that, The deflection quantization parameters output by the physical coupling quantization evaluation module include the maximum deflection, the relative deflection angle of the connection point, and the spatial extreme curvature. The deflection quantization parameters are obtained by parsing the predicted deflection field output by the physical information neural network.

10. The vision-based detection device for detecting the flexural deformation of a disconnector switch conductive arm according to claim 9, characterized in that, Also includes: The multi-level safety early warning module is used to compare the flexure quantification parameter with a preset deformation classification threshold database. When the flexure quantification parameter exceeds the set range, the corresponding graded early warning signal is triggered, and the deformation trend curve of the conductive arm is generated and output by combining the pre-stored historical flexure quantification parameters.

Images