Cable size parameter detection method and system based on point cloud feature registration

Abstract

This invention relates to the field of power cable testing technology, and discloses a method and system for detecting cable dimensional parameters based on point cloud feature registration. The method includes: acquiring multispectral image signals; obtaining the light trace energy matrix through physical simulation; constructing an error fingerprint database by combining topological mapping; performing semantic segmentation on the image signals to obtain a probability segmentation mask; calling the fingerprint database to obtain a correction matrix and extracting the original sub-pixel point cloud; performing confidence assessment based on environmental stress parameters to obtain a weighted confidence feature set; performing geometric center alignment to output the cable geometric center set, and then extracting the dimensions to obtain the cable cross-section measurement results. This invention achieves digital decoupling of stress-induced optomechanical coupling pseudo-drift, solves the technical problems of numerical jumps in insulation thickness and deviations in eccentricity calculation, and improves the confidence and measurement accuracy of cable cross-section configuration reconstruction under complex mechanical stress conditions.

Description

Technical Field

[0001] This invention relates to the field of power cable testing technology, and more specifically, to a method and system for detecting cable size parameters based on point cloud feature registration. Background Technology

[0002] With the deepening of urban power grid construction, high-precision detection of insulation thickness and eccentricity of cables laid in underground utility tunnels or substations is a crucial link in ensuring power transmission safety. Traditional machine vision measurement technology optimizes the measurement by adjusting macroscopic parameters such as camera resolution, exposure intensity, or edge extraction threshold. However, it faces measurement bottlenecks in areas where cables are bent at large angles or subjected to external compression. The system-generated detection reports show abnormal jumps in insulation thickness values, causing eccentricity calculation results to deviate from design standards. Existing technologies often simplify the edge of the cable insulation layer as a geometrically abrupt interface, relying on static geometric projection features for edge extraction. This ignores the fact that cables are under non-uniform mechanical stress in the laying environment, including bending stress and radial compressive stress. The rearrangement of molecular chain structure caused by mechanical stress produces a photoelastic effect, causing the refractive index of the cable insulation material to evolve from an isotropic scalar to an anisotropic tensor that varies with stress distribution. The existence of the refractive index gradient induces nonlinear path deflection of the probe light within the insulation layer and forms complex secondary reflections. This causes a nonlinear spatial displacement between the geometric edge perceived by the camera and the physical boundary, forming an optomechanical coupled edge pseudo-drift phenomenon. In the point cloud feature registration process, traditional algorithms cannot remove the false displacement information caused by the refractive index gradient, causing the registration calculation to fall into the local optimum trap formed by stress field interference, and thus cannot restore the high-confidence physical configuration of the cable cross section. Therefore, how to shift from static geometric analysis to dynamic multi-physics coupling prediction, and transform the optimization of macroscopic image parameters into precise decoupling of stress-induced pseudo displacement, thereby breaking through the measurement limitations brought about by optomechanical coupling, is a technical problem to be solved in this field.

[0003] In the prior art, Chinese patent CN107578047B discloses a method for detecting the eccentricity of power cables. This method acquires two-dimensional images of the cable cross-section, uses an edge detection algorithm to extract the contours of the insulation layer and conductive core, obtains the geometric center of the two contours through geometric fitting, and calculates the eccentricity based on the ratio of the center offset to the insulation thickness. This enables automated detection of parameters such as insulation thickness and eccentricity, effectively improving the efficiency and stability of offline detection of cable cross-section parameters. Chinese patent CN114821571B discloses a point cloud processing method for power cable identification and reconstruction. This method achieves accurate reconstruction of the cable contour through three-dimensional point cloud acquisition, noise reduction filtering, and feature registration, optimizes point cloud matching and spatial positioning accuracy, and can output three-dimensional dimensional parameters such as cable outer diameter and insulation layer thickness, adapting to the non-contact detection needs of power cables in three-dimensional scenes.

[0004] However, while the two existing technologies mentioned above have certain application value in automated detection of cable eccentricity and reconstruction of 3D point cloud contours, they fail to solve the core technical pain points of pseudo-drift of optomechanical coupling edges, visual distortion induced by refractive index gradient, and point cloud registration getting trapped in local optima under complex mechanical stress conditions. Specifically, the Chinese patent with authorization announcement number CN107578047B relies solely on the geometric projection features of static 2D images to extract edges, simplifying the insulation layer boundary to an ideal geometric abrupt interface. It completely ignores the photoelastic effect and refractive index anisotropy caused by bending and radial compression stress, failing to remove stress-induced false displacement information. This easily leads to insulation thickness jumps and eccentricity measurement distortions in areas with large-angle bending or compression. The Chinese patent with authorization announcement number CN114821571B focuses on optimizing the point cloud registration algorithm, but it does not establish a coupling mapping model between the stress field and optical signals, nor does it consider the modulation effect of stress on light propagation. Point cloud registration is still easily affected by stress field interference and trapped in local optima, failing to restore the high-confidence physical configuration of the cable cross-section. Both methods remain at the level of static geometric analysis and fail to achieve the transformation from macroscopic image parameter optimization to precise decoupling of stress-induced pseudo-displacement. They cannot meet the high-precision and high-robustness testing requirements of cable insulation parameters in complex stress scenarios such as underground pipe corridors and substations. Summary of the Invention

[0005] This invention is applicable to cable inspection scenarios in various laying environments such as underground pipe corridors or substations, and can meet the needs of precise cross-sectional measurement under complex mechanical stress conditions. It achieves digital decoupling of the pseudo-drift phenomenon at the optomechanical coupling edge induced by mechanical stress through the topological mapping relationship of multispectral image signals excited by multi-band pulsed light sources, light trace energy matrix, error fingerprint database, and probability segmentation masks. A weighted confidence feature set generated by combining the original sub-pixel point cloud with environmental stress parameters transforms the uncertainty of edge features into geometric weight coefficients for quantitative evaluation, accurately identifying spurious displacements caused by non-uniform refractive index distribution and suppressing geometric centroid drift. The cable geometric center set, combined with the outer diameter distribution sequence and center offset vector, achieves sub-pixel-level quantitative extraction of dimensional parameters, restoring the physical configuration of the cable cross-section and providing high-confidence data support for operation and maintenance. This solves the technical problems of abnormal insulation thickness jumps and measurement deviations from standards, reducing measurement errors and manual inspection costs. This invention comprehensively improves measurement accuracy, restoration confidence, and operation and maintenance economy.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A cable size parameter detection method based on point cloud feature registration includes:

[0008] Multispectral image signals characterizing the optical features of the cable cross-section under laying conditions are acquired. Physical simulation is performed on the multispectral image signals to obtain the light trace energy matrix. Topological mapping calculation is performed between the light trace energy matrix and the multispectral image signals to obtain the error fingerprint database.

[0009] Semantic segmentation processing is performed on the multispectral image signal to obtain a probability segmentation mask that represents the regional distribution of each component of the cable. The error fingerprint library is called to perform spatial position offset correction on the probability segmentation mask to obtain a correction matrix. The correction matrix is ​​then used to perform calculations to obtain the original sub-pixel point cloud. Environmental stress parameters are obtained to perform confidence evaluation on the original sub-pixel point cloud to obtain a weighted confidence feature set.

[0010] Geometric center alignment is performed using a weighted confidence feature set, outputting a cable geometric center set. Based on the cable geometric center set, dimensional parameters are extracted, and the cable cross-section measurement results are output.

[0011] Furthermore, the method for acquiring the multispectral image signal includes:

[0012] A multi-band pulse light source is used to sequentially emit pulses of multiple center wavelengths according to a preset band sequence to perform time-division illumination excitation on the cross section of the cable under test. An industrial camera is used to synchronously capture the reflected signal of the cross section of the cable under test at the sampling time when the corresponding center wavelength is triggered, and a series of single-band feature matrices are obtained.

[0013] By utilizing the physical principle that light of different frequencies has different refractive index dispersion characteristics in cable insulation materials, the single-band feature matrices corresponding to each center wavelength are stacked in dimensional order according to the band sequence to generate the multispectral image signal containing spatial pixel coordinates and spectral response dimensions.

[0014] Furthermore, the method for performing physical simulation on multispectral image signals includes:

[0015] A two-dimensional rectangular coordinate system is constructed with the geometric center of the cross-section of the cable to be tested as the origin, and is defined as the cross-sectional coordinate system;

[0016] In the cross-sectional coordinate system, with the origin as the starting point, the radial vector pointing to the spatial pixel coordinates is defined as the scanning ray, and the horizontal angle between the scanning ray and the positive direction of the horizontal axis of the cross-sectional coordinate system is defined as the polar azimuth angle.

[0017] Extend the preset axial step length along the direction perpendicular to the plane of the cross-section coordinate system to construct the three-dimensional simulation space of the cable with a cylindrical structure. Use the Cartesian meshing method to divide the three-dimensional simulation space of the cable into voxel units with a three-dimensional matrix structure, and define the vertices of the voxel units as spatial mesh nodes.

[0018] Obtain the mechanical stress vector acting on the surface of the cable under test, define the outer periphery of the cable's three-dimensional simulation space as the load boundary and apply the mechanical stress vector, calculate the stress tensor at each spatial grid node, and define the set of stress tensors corresponding to all spatial grid nodes in space as the stress field distribution.

[0019] Furthermore, the method for performing physical simulation on multispectral image signals also includes:

[0020] Obtain the initial refractive index and the trace of the stress tensor corresponding to each center wavelength. Perform an accumulation operation using the product of the initial refractive index and the preset Kronecker sign, the product of the stress tensor and the preset stress optical constant, and the product of the trace of the stress tensor and the preset stress optical constant to obtain the refractive index tensor corresponding to each spatial grid node.

[0021] Path tracing of virtual photons is performed within the three-dimensional simulation space of the cable. The virtual photons are probabilistic statistical sample units used to simulate the characteristics of optical energy propagation.

[0022] By using the refractive index tensor to perform micro-element-level vector perturbation correction on the motion direction of the virtual photon, the emission direction vector is obtained.

[0023] Furthermore, the obtained light trace energy matrix includes:

[0024] Using the plane where the photoelectric sensor of the industrial camera is located as a reference, a two-dimensional virtual detection plane is constructed and defined as the imaging plane;

[0025] The two-dimensional spatial position indices of all virtual photons intersecting the imaging plane are statistically analyzed and defined as the landing point coordinates;

[0026] The proportion of residual light intensity caused by medium absorption or scattering during the propagation of the virtual photon is calculated and defined as the residual energy weight.

[0027] The residual energy weights are accumulated according to the stated landing point coordinates to obtain the light trace energy matrix.

[0028] Furthermore, the obtained error fingerprint database includes:

[0029] The measured derivative matrix is ​​obtained by performing convolution operation on the single-band feature matrix, and the theoretical derivative matrix is ​​generated by performing convolution operation on the trace energy matrix.

[0030] Search along the scanning ray for two adjacent matrix elements in the measured derivative matrix and the theoretical derivative matrix that are located on the preset scanning ray path. When the product of the values ​​of the two adjacent matrix elements is less than zero, it is marked as a zero intersection point.

[0031] By collecting the zero-crossing points in the measured derivative matrix, we obtain a set of observation edge points consisting of multiple observation edge points. By collecting the zero-crossing points in the theoretical derivative matrix, we obtain a set of theoretical edge points consisting of multiple theoretical edge points.

[0032] For observed edge points and theoretical edge points at the same polar azimuth angle, the difference between the spatial pixel coordinates of the observed edge points and the coordinates of the theoretical edge points is calculated and defined as the pixel offset. The pixel offset is then encapsulated in a data structure to obtain an error fingerprint database.

[0033] Furthermore, the probability segmentation mask includes:

[0034] The multispectral image signal is input into the encoder of a pre-defined deep convolutional neural network to extract multiple feature maps, which are defined as multi-scale feature maps.

[0035] Global average pooling is performed on the spatial dimension of the multi-scale feature map to extract the description vector describing the band features, and a fully connected neural network is called to perform a nonlinear transformation on the description vector to obtain the channel weight coefficients.

[0036] The enhanced band feature representation is obtained by performing element-wise arithmetic product between the channel weight coefficients and the multi-scale feature map.

[0037] A densely connected branch is constructed to perform channel-by-channel tensor concatenation fusion on the enhanced band feature representation, determine the probability of each spatial pixel coordinate belonging to different regions of the cable, and output a probability segmentation mask.

[0038] Furthermore, obtaining the original sub-pixel point cloud includes:

[0039] Traverse each spatial pixel coordinate in the probability segmentation mask, retrieve the corresponding pixel offset from the error fingerprint database, perform vector subtraction operation on each spatial pixel coordinate and the corresponding pixel offset to obtain the corrected spatial coordinates, and map all the corrected spatial coordinates and their associated class probabilities to obtain the correction matrix.

[0040] Perform differentiation on the correction matrix to lock the zero-crossing region determined by the product of adjacent element values ​​being less than zero, extract the class probabilities arranged in this zero-crossing region, and define it as a probability profile;

[0041] On the probability profile, two adjacent pixels whose numerical distributions fall on either side of a preset edge threshold are retrieved and defined as step boundary point pairs. Linear interpolation is then performed using the corrected spatial coordinates and class probabilities corresponding to the step boundary point pairs to obtain sub-pixel coordinates. All sub-pixel coordinates under the polar azimuth angles are aggregated to obtain the original sub-pixel point cloud.

[0042] Furthermore, the confidence assessment includes:

[0043] The system acquires real-time ambient temperature, relative humidity, and radial compressive stress and combines them into an environmental state vector. It also acquires multiple historical environmental state vectors as a training sample set.

[0044] Calculate the spatial correlation between any two historical environmental state vectors in the training sample set, perform matrix arrangement on each spatial correlation value to obtain the original covariance matrix, and superimpose a preset system noise term on the main diagonal of the original covariance matrix to obtain the training set covariance matrix.

[0045] Calculate the similarity between the environmental state vector and each of the historical environmental state vectors in the training sample set, and encapsulate each similarity value into a vector to obtain the test covariance vector.

[0046] Perform matrix inversion on the training set covariance matrix to obtain the precision matrix. Perform inner product operation on the transpose row vector of the test covariance vector, the precision matrix, and the test covariance vector, and output a scalar value, which is defined as the projected energy term.

[0047] Furthermore, the obtained weighted confidence feature set includes:

[0048] Set the prior variance, calculate the difference between the prior variance and the projected energy term, and obtain the prediction variance for each sub-pixel coordinate;

[0049] The predicted variance is converted into geometric weight coefficients using a preset mapping function;

[0050] The geometric weight coefficients are encapsulated with each corresponding sub-pixel coordinate in the original sub-pixel point cloud to generate a weighted confidence feature set.

[0051] Furthermore, the geometric center set of the output cable includes:

[0052] Subtract the horizontal axis value of the weighted mean center from the horizontal axis value of each subpixel coordinate in the cross-sectional coordinate system, and subtract the vertical axis value of the weighted mean center from the vertical axis value of the subpixel coordinate to obtain the decentralized coordinate vector.

[0053] A weighted covariance matrix is ​​constructed using decentralized coordinate vectors and their corresponding geometric weight coefficients;

[0054] Calculate the weighted mean of the subpixel coordinates along the horizontal axis and the weighted mean of the vertical axis in the weighted confidence feature set, and combine them to obtain the center of the weighted mean.

[0055] A second-order matrix is ​​generated by performing an outer product operation between the decentralized coordinate vector and its transpose row vector, and then a weighted product operation is performed with the geometric weight coefficients. The weighted covariance matrix is ​​obtained by summing the results.

[0056] Perform eigenvalue decomposition on the weighted covariance matrix and extract the principal eigenvectors. Perform spatial rotation and alignment along the direction of the principal eigenvectors, and output the set of cable geometric centers consisting of the geometric centers of the insulation layer and the centers of the conductive cores.

[0057] Furthermore, the output cable cross-section measurement results include:

[0058] With the center of the conductive wire core as the pole, the angle between the scanning ray and the positive direction of the horizontal axis of the cross-sectional coordinate system is defined as the circumferential polar angle, and a radial search path is constructed.

[0059] Calculate the Euclidean distance between the sub-pixel coordinates of the pre-identified cable insulation layer region and the pole at each circumferential polar angle, and aggregate the outer diameter distribution sequence.

[0060] Calculate the average distance between the sub-pixel coordinates of the pre-identified conductive core region and the center of the conductive core, and define it as the equivalent radius of the core.

[0061] Subtract the equivalent radius of the wire core from the values ​​of each Euclidean distance in the outer diameter distribution sequence to obtain a thickness distribution set consisting of a series of radial thickness values. The arithmetic mean and minimum value of the thickness distribution set are defined as the average insulation thickness and the minimum insulation thickness, respectively.

[0062] The spatial displacement vector of the geometric center of the insulation layer relative to the center of the conductive core is calculated and defined as the center offset vector. The modulus of the center offset vector is divided by the average insulation thickness to obtain the eccentricity. The average insulation thickness, minimum insulation thickness and eccentricity are digitally encapsulated and the cable cross-section measurement results are output.

[0063] A cable size parameter detection system based on point cloud feature registration is used to implement the aforementioned cable size parameter detection method based on point cloud feature registration. The system includes:

[0064] The pseudo-drift decoupling module is used to acquire multispectral image signals that characterize the optical features of the cable cross-section under laying conditions, perform physical simulation on the multispectral image signals to obtain the light trace energy matrix, and perform topological mapping calculation on the light trace energy matrix and the multispectral image signals to obtain the error fingerprint library.

[0065] Feature weighting module: used to perform semantic segmentation processing on multispectral image signals, obtain a probability segmentation mask that represents the regional distribution of each component of the cable, call the error fingerprint library to perform spatial position offset correction on the probability segmentation mask, obtain a correction matrix, and use the correction matrix to perform calculations to obtain the original sub-pixel point cloud, obtain environmental stress parameters to perform confidence evaluation on the original sub-pixel point cloud, and obtain a weighted confidence feature set;

[0066] The cross-section calculation module is used to perform geometric center alignment using a weighted confidence feature set, output the cable geometric center set, extract dimensional parameters based on the cable geometric center set, and output the cable cross-section measurement results.

[0067] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0068] This invention achieves digital decoupling of the pseudo-drift phenomenon at the optomechanical coupling edge induced by mechanical stress through the topological mapping relationship of multispectral image signals, light trace energy matrix, and error fingerprint database. This solves the problem of measurement distortion caused by traditional visual algorithms ignoring the modulation effect of physical stress on optical signals. The weighted confidence feature set transforms the traditional qualitative judgment of edge positioning into a quantitative calculation of confidence assessment of the original sub-pixel point cloud using environmental stress parameters. It accurately identifies spurious displacements caused by non-uniform refractive index distribution through geometric weight coefficients, suppressing geometric centroid drift. The cable geometric center set, combined with the outer diameter distribution sequence and center offset vector, realizes sub-pixel-level quantitative extraction of dimensional parameters, which not only restores the physical configuration of the cable cross-section but also provides accurate data support for operation and maintenance. This solves the technical problems of numerical jumps in insulation thickness and deviations in eccentricity measurement from the standard, reducing measurement costs and improving the economy of operation and maintenance. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 A flowchart of a cable size parameter detection method based on point cloud feature registration provided in an embodiment of the present invention;

[0071] Figure 2 A schematic diagram of a cable detection system based on multi-band pulse excitation provided in an embodiment of the present invention;

[0072] Figure 3 This invention provides a schematic diagram of photon path deflection logic based on Monte Carlo tracing.

[0073] Figure 4 This is a schematic diagram of the spatial distribution of observed edge points and theoretical edge points provided in an embodiment of the present invention;

[0074] Figure 5 The physical logic diagram for performing spatial alignment and geometric parameter measurement based on a weighted confidence feature set provided in this embodiment of the invention;

[0075] Figure 6 This is a functional block diagram of a cable size parameter detection system based on point cloud feature registration provided in an embodiment of the present invention. Detailed Implementation

[0076] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0077] Example 1

[0078] Please see Figure 1 As shown, this embodiment provides a cable size parameter detection method based on point cloud feature registration, including:

[0079] Step S10: Obtain multispectral image signals that characterize the optical features of the cable cross-section in the laying state; perform physical simulation on the multispectral image signals to obtain the light trace energy matrix; perform topological mapping calculation on the light trace energy matrix and the multispectral image signals to obtain the error fingerprint database.

[0080] Further, step S10 includes:

[0081] Step S11: Obtain multispectral image signals that characterize the optical features of the cable cross-section in its laying state.

[0082] In the field of power cable operation and maintenance, cables, as the core physical carrier of power transmission systems, are often subjected to non-uniform mechanical stress due to their own weight, support preload, or bending torque in actual laying environments. When subjected to mechanical stress, the molecular chain structure within the cable insulation material undergoes microscopic rearrangement, resulting in a photoelastic effect. This causes the refractive index of the cable insulation material to evolve from an isotropic scalar to an anisotropic tensor that varies with stress distribution. However, traditional machine vision measurement techniques typically simplify the edge of the cable insulation layer as an ideal geometric abrupt interface, completely ignoring the nonlinear path deflection of the probe light as it penetrates a medium with a refractive index gradient. This leads to spurious geometric deviations in the edge features captured by industrial cameras due to refractive displacement, i.e., the phenomenon of optomechanically coupled edge pseudo-drift induced by mechanical stress. To eliminate the distortion in physical dimension measurement caused by model simplification, a reference frame is established that can accurately map the stress state of the cable insulation material to its optical response characteristics. The aim is to transform passive, monochromatic image acquisition into an active, multi-physics-sensing multispectral image signal extraction process, providing a high-fidelity digital input benchmark for removing the pseudo-drift phenomenon at the optomechanical coupling edge induced by mechanical stress.

[0083] To acquire multispectral image signals characterizing the optical features of the cable cross-section under installation, a multi-band pulsed light source is used to perform time-division illumination excitation on the cross-section of the cable under test. The multi-band pulsed light source refers to an LED array capable of sequentially emitting light with a set center wavelength according to a preset band sequence, including red, green, blue, and near-infrared bands. Since light of different frequencies exhibits different refractive index dispersion characteristics in the cable's insulation material, meaning that the same stress field produces differentiated deflection effects on probe light of different wavelengths, acquiring multidimensional spectral information spanning the visible to near-infrared bands significantly improves the algorithm's accuracy in identifying stress-induced refractive index gradients. Based on the refractive index dispersion characteristics, the center wavelengths of each band in the band sequence are differentiated. For example, the center wavelength of the red band is set to 625 nm, the green band to 525 nm, the blue band to 470 nm, and the near-infrared band to 850 nm.

[0084] The sampling time is defined as the instantaneous point at which each center wavelength in the corresponding band sequence is triggered within each scanning cycle. The scanning cycle refers to the time span required to complete the acquisition of all center wavelengths in the band sequence. The scanning cycle is set to balance the sensor readout rate of the industrial camera with the dynamic throughput of real-time on-site detection, aiming to ensure continuous and real-time quality monitoring at the cable laying site. For example, it is set to 100 milliseconds. The sampling time is set based on the physical trigger response time of the pulse light source in each band of the band sequence and the minimum exposure delay of the industrial camera, aiming to ensure that the image acquisition of each center wavelength does not interfere with each other on the time axis and has sufficient exposure energy. For example, within each scanning cycle, the sampling time is set to 0 milliseconds, 25 milliseconds, 50 milliseconds, and 75 milliseconds, respectively, thereby achieving the orderly extraction of four different center wavelength signals. The industrial camera synchronously captures the cable cross-sectional reflection signal at the corresponding center wavelength at each sampling time, obtaining a series of single-band feature matrices. The single-band feature matrix refers to a two-dimensional pixel array that stores and quantizes the detection light corresponding to the center wavelength at a specific sampling time, after reflection from the cable cross-section and capture by the photoelectric sensor of the industrial camera. Each matrix element in the single-band feature matrix represents the local reflection intensity of the cable cross-section under the excitation of that center wavelength. A series of single-band feature matrices are stacked dimensionally according to the band sequence to generate a three-dimensional data cube with high-dimensional information carrying capacity, defined as a multispectral image signal. The multispectral image signal includes spatial pixel coordinates and a spectral response dimension; the spatial pixel coordinates are two-dimensional geometric coordinates determined by the row and column indices of the industrial camera sensor, used to characterize the spatial distribution of the cable cross-section on the imaging plane; the spectral response dimension is a third-dimensional index determined by the band sequence, used to characterize the physical response differences of the same spatial pixel location under different frequency detection light.

[0085] To accurately describe the spatial relationships during multispectral image signal acquisition, a two-dimensional Cartesian coordinate system, defined as the cross-sectional coordinate system, is constructed with the geometric center of the cross-section of the cable under test as the origin. The X-axis of the cross-sectional coordinate system is defined as a straight line parallel to the ground plane, and the Y-axis is defined as a straight line perpendicular to the ground plane and passing through the origin. See also... Figure 2 This is a schematic diagram of a cable detection system based on multi-band pulse excitation provided in an embodiment of the present invention. Figure 2In the diagram, a dark gray circle at the center simulates the cross-section of the cable under test, while a black circle surrounding it represents the outer boundary of the cable insulation. Four light blue circles surrounding the dark gray circle represent multi-band pulsed light sources, and a black square in front of the dark gray circle's axis represents an industrial camera. At the edge of the dark gray circle, directional red vector arrows illustrate the mechanical stress field experienced by the cable under test. The cross-sectional coordinate system provides a unique spatial coordinate mapping for each pixel in the multispectral image signal, ensuring strict alignment between spectral information and physical geometric location.

[0086] Step S12: Perform physical simulation on the multispectral image signal to obtain the light trace energy matrix.

[0087] After acquiring multispectral image signals, a physical simulation based on the Monte Carlo algorithm is performed to quantify the nonlinear propagation behavior of probe light in a non-uniform insulating medium with a refractive index gradient. Because the mechanical stress field experienced by the cable during installation causes a continuously changing refractive index distribution within the insulation layer, this change in physical properties results in a spatial topological misalignment between the optical edge features recorded in the multispectral image signal and the actual physical geometric boundaries. To perceive and reconstruct this edge displacement logic caused by photoelastic effects, a simulation framework capable of accurately mapping environmental stress field characteristics to optical energy evolution is established. The goal is to output a light trace energy matrix that is stripped of external environmental noise and retains only the modulation law of the material's physical properties by simulating the trajectory of a massive number of virtual photons.

[0088] A three-dimensional simulation space for the cable is constructed, and spatial mesh nodes are defined. Specifically, using a cross-sectional coordinate system as the geometric reference, the outer diameter of the insulation layer and the radius of the conductor of the cable under test are obtained. The outer diameter of the insulation layer refers to the straight-line distance from the origin to the outermost edge of the insulation layer on the cross-section of the cable under test; the radius of the conductor refers to the straight-line distance from the origin to the outer edge of the conductive conductor on the cross-section of the cable under test. Based on the cross-sectional coordinate system, a cylindrical feature space is constructed by extending a preset axial step length along a direction perpendicular to the plane of the cross-sectional coordinate system, which is defined as the three-dimensional simulation space of the cable. The axial step length is set according to the optical depth of field range of an industrial camera, and is exemplarily set to 5 mm. The three-dimensional simulation space of the cable is divided into voxel units of a three-dimensional matrix structure using the Cartesian mesh method. The voxel unit refers to the smallest cubic volume element that constitutes the interior of the three-dimensional simulation space of the cable, used to carry the discretized values ​​of local physical properties. The vertices of the voxel units are defined as spatial mesh nodes. Each spatial mesh node has a unique coordinate vector. Specifically, the grid spacing of the spatial grid nodes must ensure sufficient resolution in the edge region of the cable insulation layer; for example, it is set to 10 micrometers. The grid spacing refers to the geometric distance between two adjacent spatial grid nodes, and its value determines the volume of the voxel unit.

[0089] To characterize the refractive index non-uniformity caused by stress, the stress field distribution is obtained and a refractive index tensor is assigned. Specifically, the mechanical stress vector acting on the surface of the cable under test is obtained. This mechanical stress vector refers to the magnitude and direction of the radial pressure acting on the surface of the cable insulation layer. Based on the elastic modulus and Poisson's ratio of the cable material under test, the stress state inside the three-dimensional simulation space of the cable is calculated using the finite element analysis method. Specifically, the outer peripheral surface of the three-dimensional simulation space of the cable is defined as the loaded boundary and a mechanical stress vector is applied, while the internal interface corresponding to the core radius is defined as a displacement-fixed constraint boundary. By iteratively solving the linear elastic static equilibrium differential equation, the stress tensor borne by each spatial grid node in the three-dimensional simulation space of the cable is calculated. The stress tensors of each spatial grid node are combined to generate a stress field distribution. The linear elastic static equilibrium differential equation refers to the fundamental mechanical equation that integrates the three-dimensional spatial static equilibrium conditions of the continuous medium, the geometric deformation compatibility conditions, and the generalized Hooke's law of the material, used to describe the internal displacement field and stress tensor distribution of an elastic body under stress. The stress tensor is a second-order tensor used to characterize the internal interaction force per unit area at the spatial grid node. In the matrix representation, the main diagonal elements represent normal stress, and the off-diagonal elements represent shear stress. The stress field distribution refers to the spatial set of the stress tensors corresponding to each spatial grid node in the three-dimensional simulation space of the cable. The stress field distribution is mapped to a refractive index field using the stress optics law. Specifically, each spatial grid node is assigned a specific refractive index tensor A. The refractive index tensor is used to describe the birefringence properties of the material under stress, and the formula for calculating the refractive index tensor is: ,in, Let B represent the stress tensor and B be the initial refractive index of the material in a stress-free state. Based on the material manual for the cable insulation material and the center wavelength settings of the current simulation band, the corresponding initial refractive index is set for each center wavelength in the band sequence. For example, the initial refractive index of the red band is set to 1.510, the initial refractive index of the green band is set to 1.515, the initial refractive index of the blue band is set to 1.522, and the initial refractive index of the near-infrared band is set to 1.505. The symbol is Kronecker, where X and Y are the values ​​of the horizontal and vertical axes in the cross-sectional coordinate system. When X and Y are equal... =1, otherwise =0. Let be the trace of the stress tensor, whose value is equal to the sum of the main diagonals of the stress tensor. and This is a preset stress optical constant, the value of which is determined based on the photoelastic experimental data of the material. For example, and Set to respectively and The calculation formula is based on Maxwell's law of stress optics, establishing a linear mapping relationship between the refractive index tensor and the stress tensor. Specifically, the reference optical properties of the material are determined by multiplying the initial refractive index by the Kronecker sign; the anisotropic birefringence effect caused by shear stress is quantified by multiplying the stress optical constant by the stress tensor; and the isotropic refractive index shift caused by volumetric strain is quantified by multiplying the stress optical constant by the trace of the stress tensor. This ensures that the changes in microscopic optical properties at spatial grid nodes can be accurately calculated under complex stress field distributions.

[0090] After each spatial grid node is assigned a refractive index tensor, virtual photon path tracing is performed within the three-dimensional simulation space of the cable. A virtual photon refers to the smallest probabilistic statistical sample unit used to simulate the characteristics of optical energy propagation during numerical simulation. The total number of virtual photons, N, is set to ensure that the statistical fluctuations of the simulation results are lower than the background noise threshold of the multispectral image signal. The background noise threshold refers to the maximum quantization limit of the background interference signal determined by the sensor's own physical characteristics and the inherent random level fluctuations of the back-end electronic readout circuitry when industrial cameras and other photoelectric sensors lack effective external light illumination or are in a reference operating state. For example, the total number of virtual photons, N, is set to... Let k represent the index of the virtual photon, ranging from 1 to N. The emission start point of the virtual photon is determined using the spatial coordinates of a multi-band pulsed light source, and the incident direction vector of each virtual photon is determined based on the lens aperture angle of the industrial camera. The incident direction vector refers to the initial unit direction vector of motion when the virtual photon enters the 3D simulation space of the cable. The exit direction vector is obtained by calculating the path deflection state of each virtual photon as it penetrates each voxel unit due to the non-uniform distribution of refractive index. The emission direction vector refers to the unit direction vector when a virtual photon leaves the current voxel unit and enters an adjacent voxel unit. Specifically, it is the k-th emission direction vector. For example, the calculation formula is as follows: ,in, This represents the incident direction vector of the k-th virtual photon. The scalar value representing the projection of the refractive index tensor corresponding to the current spatial grid node of the virtual photon onto the direction of the virtual photon's travel is obtained by performing a quadratic matrix operation between the incident direction vector and the refractive index tensor corresponding to the spatial grid node. It is a three-dimensional vector whose components are determined by the rate of change of the refractive index tensor between two adjacent spatial grid nodes. The calculation method involves extracting the difference in the refractive index tensor of adjacent spatial grid nodes along the three axes and dividing it by the grid spacing. S represents the tiny step size of the virtual photon's movement within the current voxel unit. Its value is set according to the grid spacing of the spatial grid nodes. This tiny step size must be strictly smaller than the grid spacing, for example, half the grid spacing. For instance, it is set to 5 micrometers. The calculation formula is based on the variational expression of Fermat's principle in discrete non-uniform media, using the refractive index tensor to perform micro-level vector perturbation correction on the virtual photon's motion direction. The physical essence lies in simulating the minimum time path evolution of the probe ray in a continuously changing refractive index field, ensuring that the virtual photon's trajectory accurately reflects the optical deflection induced by mechanical stress.

[0091] Using the plane of the industrial camera's photoelectric sensor as a reference, a two-dimensional virtual detection plane is constructed, defined as the imaging plane. This imaging plane records the landing coordinates of virtual photons after they exit the three-dimensional simulation space of the cable. The landing coordinates refer to the two-dimensional spatial position index when the virtual photon intersects the imaging plane. The residual energy weights of all virtual photons reaching the imaging plane are calculated. The residual energy weight refers to the proportion of residual light intensity due to medium absorption or scattering during the propagation of the virtual photon within the cable's three-dimensional simulation space; its initial value is 1. The residual energy weights of all virtual photons are accumulated according to the landing coordinates of the imaging plane to generate a light trace energy matrix. The light trace energy matrix is ​​a two-dimensional numerical array used to quantify the theoretical imaging intensity, storing the expected spatial distribution of light intensity considering the stress field distribution. See also... Figure 3 This is a schematic diagram of photon path deflection logic based on Monte Carlo tracing provided by an embodiment of the present invention. The grid lines in the background are used to simulate the computational grid in the three-dimensional simulation space of a cable; the continuous blue curve represents the trajectory of a virtual photon as it penetrates different voxel units; the varying shades of red in the grid background characterize the spatial gradient of the refractive index tensor determined by the stress field distribution. The depth of the red is positively correlated with the magnitude of the spatial gradient of the refractive index tensor, i.e., the darker the area, the larger the spatial gradient, and the lighter the area, the smaller the spatial gradient. When the blue curve passes through the red area, it undergoes a significant directional deflection due to the modulation of the refractive index tensor, simulating the pseudo-drift phenomenon at the optomechanical coupling edge induced by mechanical stress.

[0092] Step S13: Perform topological mapping calculation on the light trace energy matrix and the multispectral image signal to obtain the error fingerprint database.

[0093] After obtaining the light trace energy matrix, to achieve a precise transition from the observed optical pseudo-edges to the real physical boundaries, a spatial correlation operation is performed between the light trace energy matrix and the multispectral image signal. Since the cable insulation edge recorded in the multispectral image signal contains optomechanically coupled edge pseudo-drift induced by mechanical stress, and the light trace energy matrix quantifies the desired light intensity evolution characteristics caused by the non-uniform distribution of the material's refractive index in the simulation dimension, a spatial correction field capable of decoupling visual displacement and physical deformation is established through topological mapping calculations. The aim is to address the problem of macroscopic measurement data distortion caused by changes in material physical properties and output a structured error fingerprint database.

[0094] Specifically, the Laplacian operator is used to perform spatial convolution operations on the single-band feature matrix and the trace energy matrix in the multispectral image signal. The Laplacian operator is an isotropic linear operator used to characterize the second-order partial derivatives of the brightness spatial distribution of a two-dimensional discrete matrix. When performing discretization processing on the single-band feature matrix or the trace energy matrix, the Laplacian operator is implemented using a 3×3 convolution kernel matrix. For example, the convolution kernel matrix preferably adopts a discrete operator form with a center pixel weight of -4 and other pixel weights of 1. The convolution operation process involves using the convolution kernel matrix as a sliding window to perform weighted summation with each pixel value and its neighborhood elements in both the single-band feature matrix and the light trace energy matrix. Convolution is then performed on the single-band feature matrix to obtain the measured derivative matrix, and on the light trace energy matrix to generate the theoretical derivative matrix. The measured derivative matrix is ​​a numerical array storing the second-order derivative values ​​of the single-band feature matrix at each pixel coordinate. The size of its matrix elements reflects the rate of change of the light intensity gradient in the measured image signal, and is used to lock the apparent geometric boundaries containing stress interference. The theoretical derivative matrix is ​​a numerical array storing the second-order derivative values ​​of the light trace energy matrix at each pixel coordinate. The size of its matrix elements reflects the degree of abrupt change in energy distribution in the physical simulation model, and is used to lock the theoretical geometric boundaries that conform to the laws of physical propagation.

[0095] In the cross-sectional coordinate system, starting from the origin, the radial vector pointing to the spatial pixel coordinates is defined as the scanning ray, and the horizontal angle between the scanning ray and the positive X-axis is defined as the polar azimuth angle. The polar azimuth angle is cyclically increased within a preset angle step value from 0 to 360 degrees, thereby performing ray scanning within the cross-sectional coordinate system. The sequence of matrix elements covered by the scanning ray in the measured derivative matrix or theoretical derivative matrix is ​​defined as the ray path. The angle step value is set to balance the geometric resolution of the cable's circumferential edge with the retrieval efficiency of the error fingerprint database, aiming to ensure the capture of minute edge offsets caused by local non-uniform stress. For example, it is set to 1 degree. At each polar azimuth angle, the matrix elements traversed along the ray path are iterated. Specifically, for the measured derivative matrix, two adjacent matrix elements on the ray path are retrieved. If the signs of the two matrix elements switch (i.e., their arithmetic product is less than zero), a measured zero-intersection point is determined to exist between the spatial positions corresponding to these two matrix elements, defined as the observed edge point. All observation edge points locked at the pole azimuth angle are aggregated to obtain the observation edge point set. This observation edge point set refers to a sequence of measured pixel coordinates containing the pseudo-drift of the optomechanical coupling edge induced by mechanical stress. Simultaneously, the same ray scanning and sign-switching detection process is performed on the theoretical derivative matrix to lock theoretical zero-crossing points, which are defined as theoretical edge points, and these are aggregated to obtain the theoretical edge point set. This theoretical edge point set refers to a sequence of ideal pixel coordinates representing the true physical boundary of the cable cross-section under test.

[0096] After obtaining the set of observed edge points and the set of theoretical edge points, spatial deviation quantization calculation is performed. Specifically, within the cross-sectional coordinate system, each observed edge point in the observed edge point set is matched with a theoretical edge point in the theoretical edge point set that is at the same polar azimuth angle. The spatial coordinate difference D between the observed edge point and its corresponding theoretical edge point is calculated to obtain the pixel offset. The specific calculation formula is as follows: Where E1 and E2 represent the observed edge point and the theoretical edge point, respectively. and These represent the horizontal and vertical coordinates of the observed edge points in the cross-sectional coordinate system, respectively. and These represent the horizontal and vertical coordinates of the theoretical edge point in the cross-sectional coordinate system, respectively. The calculation formula is based on the vector distance criterion of Euclidean geometry. By quantifying the absolute value of the spatial displacement between the measured point and the theoretical point locked by the zero-crossing detection of the second derivative on the same ray path, the nonlinear visual displacement artifact caused by the refractive index tensor distribution inside the cable insulation layer is eliminated. The spatial pixel coordinates of each observed edge point in the cross-sectional coordinate system are used as the search key, and the calculated corresponding pixel offset is used as the storage attribute value. Data structure encapsulation is performed to obtain the error fingerprint database, which is a digital index table mapping the spatial pixel coordinates to the physical deviation. See also Figure 4 This diagram illustrates the spatial distribution of observed and theoretical edge points according to an embodiment of the present invention. The circle enclosed by the solid black line represents the set of theoretical edge points, reflecting the true physical geometric boundary of the cable under test. The curve enclosed by the dashed red line represents the set of observed edge points, reflecting the apparent distortion profile caused by stress. The solid blue line emanating from the origin is the scanning ray; two black dots are distributed on the scanning ray. The black dot at the beginning of the ray is the origin of the coordinate system, and the black dot at the intersection of the solid black lines is the theoretical edge point. The red dots on the scanning ray are the observed edge points. The spatial displacement deviation between the theoretical and observed edge points is quantified by the blue arrow connecting them.

[0097] Step S10 addresses the technical challenge of traditional visual measurements neglecting the non-uniform refractive index distribution of cable insulation under mechanical stress, leading to nonlinear spatial displacement between the camera-perceived geometric edges and the physical boundaries. This is achieved through multispectral image signals, a light trace energy matrix, and an error fingerprint database. Digital decoupling of the pseudo-drift phenomenon at the optomechanical coupling edge induced by mechanical stress is realized. Specifically, the multispectral image signal improves the accuracy of identifying stress-induced refractive index gradients; the light trace energy matrix quantifies the modulation law of the stress field on the evolution of optical energy; and the error fingerprint database establishes a correlation mapping between visual displacement and physical deformation.

[0098] Step S20: Perform semantic segmentation processing on the multispectral image signal to obtain a probability segmentation mask that represents the regional distribution of each component of the cable. Call the error fingerprint library to perform spatial position offset correction on the probability segmentation mask to obtain a correction matrix. Use the correction matrix to perform calculations to obtain the original sub-pixel point cloud. Obtain environmental stress parameters to perform confidence evaluation on the original sub-pixel point cloud to obtain a weighted confidence feature set.

[0099] Further, step S20 includes:

[0100] Step S21: Perform semantic segmentation processing on the multispectral image signal to obtain a probability segmentation mask that characterizes the regional distribution of each component of the cable.

[0101] After acquiring multispectral image signals, a deep convolutional neural network based on an integrated channel attention mechanism is executed to accurately extract feature regions of the cable cross-section from the complex optical background. Due to the high reflectivity of the cable under test during installation and the differences in scattering effects of different wavelengths of probe light in anisotropic insulating media, a single-dimensional, single-band feature matrix cannot provide stable edge discrimination criteria. Therefore, a semantic extraction framework capable of feature enhancement and spatial region segmentation of multispectral information is established. The aim is to suppress invalid information affected by environmental noise by adaptively assigning weights to the spectral channels, outputting a probabilistic segmentation mask with pixel-level semantic constraints.

[0102] A deep convolutional neural network with an integrated channel attention mechanism is invoked to perform semantic segmentation on multispectral image signals. Specifically, the multispectral image signal, composed of stacked single-band feature matrices corresponding to the red, green, blue, and near-infrared bands, is used as input data. The multispectral image signal is input to the encoder of the deep convolutional neural network in the form of a four-channel three-dimensional tensor. The encoder consists of multiple cascaded convolutional units, where each convolutional unit is the smallest computational module composed of convolutional layers, batch normalization layers, and linear rectified activation functions in sequence. It is used to extract local spatial geometric features of the image through a sliding window operation. The encoder reduces the spatial resolution of the multispectral image signal through layer-by-layer convolution and downsampling operations, while increasing the depth of the feature channels, generating multi-scale feature maps representing deep semantic information of the cable cross-section. A channel weight adaptive module is inserted into the output path of each level of multi-scale feature map extracted by the encoder. This channel weight adaptive module is an embedded attention computation unit used to dynamically adjust the influence of different spectral bands based on feature contribution. Specifically, using the global average pooling operator, each multi-scale feature map is compressed into a set of one-dimensional vectors in spatial dimension, obtaining a descriptive vector representing the global information of each spectral band. A two-layer fully connected neural network is then used to transform these descriptive vectors. The first fully connected layer performs dimensionality reduction on the vectors to extract the core correlations between bands and enhances the nonlinear expression using the ReLU activation function. The second fully connected layer restores the vectors to their original dimensions. Finally, the output values ​​are normalized to between 0 and 1 using the Sigmoid activation function to generate channel weight coefficients. The generated channel weight coefficients are then element-wise multiplied with the multi-scale feature map to obtain the enhanced band feature representation.

[0103] An improved U-Net++ architecture is used to perform multi-scale feature extraction and decoding of the enhanced band feature representation. The core of the U-Net++ architecture lies in the construction of densely connected branches. These densely connected branches refer to network paths composed of a series of nested and overlapping intermediate convolutional sub-nodes introduced between encoder nodes and their corresponding decoder nodes. The construction method involves multiple feature fusions and refinements of the enhanced band feature representation in each layer before it is passed to the decoder, thereby reducing the semantic distribution gap between the encoder and decoder. This structure effectively suppresses edge signal diffusion caused by mechanical stress, ensuring extremely high topology reconstruction accuracy at the cable insulation layer boundary. The end of the deep convolutional neural network performs pixel-by-pixel category classification calculations using the Softmax activation function. Specifically, the last convolutional layer of the deep convolutional neural network outputs a multi-channel feature matrix, with the number of channels corresponding to the preset number of categories. The number of categories is determined based on the cross-sectional physical structure of the cable under test and the types of features to be extracted for the measurement task. This aims to ensure the exclusion of texture interference from non-target areas during subsequent edge extraction by assigning independent category labels to different functional layers within the cable. For example, it is set to 3, corresponding to the cable insulation layer region, the conductive core region, and the mesh background region, respectively. The Softmax activation function is used to perform exponential normalization calculations on the values ​​of each pixel position in each channel to determine the category probability of the pixel belonging to the cable insulation layer region, the conductive core region, or the mesh background region, outputting a probability segmentation mask. The probability segmentation mask is a two-dimensional probability distribution matrix perfectly aligned with the multispectral image signal in the cross-sectional coordinate system, where each matrix element is the category probability output by the Softmax activation function.

[0104] Step S22: Call the error fingerprint library to perform spatial position offset correction on the probability segmentation mask to obtain the correction matrix, and use the correction matrix to perform calculations to obtain the original sub-pixel point cloud.

[0105] After obtaining the probability segmentation mask, to restore the apparent probability distribution affected by the nonlinear path deflection of the probe light to the true physical geometric distribution, a spatial position offset correction operation using an error fingerprint database is performed. Since the insulating layer edges recorded in the probability segmentation mask still contain pseudo-drift amounts induced by mechanical stress in the optomechanical coupling edges, a spatial reconstruction logic based on vector compensation is established. The aim is to reverse the displacement offset stored in the error fingerprint database onto the probability distribution, thereby removing the spatial distortion caused by the optical path deflection and outputting a set of original sub-pixel point clouds with sub-pixel accuracy.

[0106] Specifically, each spatial pixel coordinate in the probabilistic segmentation mask is traversed, and the current spatial pixel coordinate is used as an index key to retrieve the corresponding pixel offset from the error fingerprint database. A vector subtraction operation is performed on each spatial pixel coordinate and its corresponding pixel offset to obtain the corrected spatial coordinates. A numerical array generated by remapping all the corrected spatial coordinates and their associated class probabilities is defined as the correction matrix. Each element in the correction matrix contains a corrected spatial coordinate and its corresponding probability value, achieving topological alignment between semantic features and the actual physical contour of the cable in space. The probability value refers to the confidence weight value output by the deep convolutional neural network in the probabilistic segmentation mask, representing that the corresponding corrected spatial coordinate belongs to the cable insulation layer region.

[0107] The Laplacian operator is invoked again to perform a second-order differential operation on the correction matrix, locking the pixel region where the second-order derivative value in the correction matrix switches between positive and negative, defined as a zero-crossing region. A preset local neighborhood is extracted centered on the zero-crossing point within the zero-crossing region. The local neighborhood refers to a pixel array containing complete probability gradient evolution features. The local neighborhood is set based on balancing the completeness of local feature coverage with the computational efficiency of sub-pixel extraction, ensuring that the neighborhood window can completely cover the transition zone where the probability value evolves from a high level in the target region to a low level in the background region. A discrete probability value sequence is extracted within the local neighborhood along the polar azimuth direction of the cross-sectional coordinate system, defined as a probability profile. Two adjacent pixel points whose probability values ​​are located on either side of a preset edge threshold are retrieved on the probability profile, defined as a step boundary point pair. The preset edge threshold is used to identify the probability center position of the physical boundary of the insulating layer; for example, it is set to 0.5. When extracting the outer boundary of the cable insulation layer, the probability values ​​in the probability profile are specified as the category probabilities corresponding to the cable insulation layer region in the probability segmentation mask; when extracting the outer boundary of the conductive core, the probability values ​​in the probability profile are switched to the category probabilities corresponding to the conductive core region in the probability segmentation mask. Step boundary point pair retrieval is performed independently on the probability profiles corresponding to the two categories of probabilities to obtain step boundary point pairs for the outer boundary of the insulation layer and the outer boundary of the conductive core.

[0108] Linear interpolation is performed using the corrected spatial coordinates and corresponding probability values ​​of two pixels in the step boundary point pair. Specifically, based on the edge threshold, and according to the probability distance ratio between the two pixels in the step boundary point pair relative to the edge threshold, the class probabilities corresponding to the two adjacent pixels in the step boundary point pair are extracted. The absolute value of the difference between the edge threshold and the class probability of the first pixel among the two adjacent pixels is calculated to obtain the local probability deviation. The absolute value of the difference between the class probabilities of the two adjacent pixels is calculated to obtain the total probability step. The local probability deviation is divided by the total probability step, and the resulting ratio is the probability distance ratio. The spatial position corresponding to the probability value being exactly equal to the edge threshold is defined as the sub-pixel coordinates. The calculation process is based on a sub-pixel edge localization algorithm based on linear interpolation. Due to the discreteness of physical space sampling of industrial camera photoelectric sensors, traditional pixel-level edge localization is inevitably limited by grid resolution. By using the class probability of step boundary point pairs and a preset edge threshold to construct a proportional mapping relationship, the sub-pixel coordinates that break through the integer pixel grid limitation are calculated, thereby effectively removing the geometric measurement deviation caused by image digitization sampling. The sub-pixel coordinates extracted under all polar azimuth angles are aggregated to obtain the original sub-pixel point cloud.

[0109] Step S23: Obtain environmental stress parameters and perform confidence assessment on the original sub-pixel point cloud to obtain a weighted confidence feature set.

[0110] After obtaining the original sub-pixel point cloud, an environment-aware confidence assessment is performed to quantify the reliability of each sub-pixel coordinate in a dynamic installation environment. Since environmental fluctuations at the detection site alter the background noise distribution of the measurement system, a dynamic weighted evaluation framework based on statistical learning theory is established. The aim is to quantify the uncertainty of feature points by sensing changes in environmental parameters and output a weighted confidence feature set.

[0111] Specifically, environmental stress parameters are acquired in real time using external sensor units. These environmental stress parameters include ambient temperature, relative humidity, and radial compressive stress of the cable outer sheath obtained through pressure sensors. A Gaussian process regression method is used to model the mapping between these environmental stress parameters and the variance of the measurement points. Specifically, an environmental state vector under the current operating conditions is constructed, including values ​​for ambient temperature, relative humidity, and radial compressive stress. A training sample set consisting of multiple historical environmental state vectors is obtained. The number of training sample sets is determined based on the dynamic distribution range of the multidimensional environmental parameters of the cable under test at the testing site. This dynamic distribution range refers to the multidimensional feature space boundary defined by the maximum and minimum values ​​of ambient temperature, relative humidity, and radial compressive stress in historical monitoring data; for example, 150 sets are set. The Matrn covariance function with a known smoothing parameter of 5 / 2 in the Gaussian process regression algorithm is used to calculate the spatial correlation between each pair of elements within the training sample set. The calculated spatial correlation values ​​are then matrix-arranged according to the calculation order of the corresponding elements to obtain the original covariance matrix. The spatial correlation refers to the statistical dependence between any two environmental state vectors, used to characterize the cooperative influence of the proximity of the physical environment on the distribution of measurement deviation. The original covariance matrix is ​​a symmetric numerical array storing the pairwise spatial correlations between all elements within the training sample set. A system noise term is superimposed on the main diagonal of the original covariance matrix to obtain the training set covariance matrix. This system noise term characterizes the inherent random interference in the photoelectric conversion process of the industrial camera; for example, it is set as follows: The Matern5 / 2 kernel function is called again to calculate the similarity between the current environmental state vector and each element within the training sample set. This similarity refers to the degree of matching between the current sampled environment and the historical calibration environment in terms of statistical features. The calculated similarity values ​​are then indexed and vectorized according to the order of the training sample set to obtain the test covariance vector. The test covariance vector is a column vector representing the correlation between the current operating condition and historical operating conditions. Obtain the prior variance, which represents the maximum fluctuation intensity of the Matern5 / 2 kernel function feature space before obtaining observation information. For example, it is set to 1.0. Perform matrix inversion on the training set covariance matrix. Specifically, obtain the inverse matrix of the training set covariance matrix, which is defined as the precision matrix. Perform a transpose operation on the test covariance vector to obtain a row vector. Multiply this row vector by the precision matrix on the left to obtain an intermediate mapping vector. Perform an inner product operation on the obtained intermediate mapping vector and the column vector of the test covariance vector. Finally, output a scalar value, which is defined as the projected energy term. The projected energy term represents the proportion of energy of the current working condition information in the feature space defined by the historical working condition information.

[0112] The difference between the prior variance and the projected energy term is calculated, and the predicted variance of each sub-pixel coordinate under the current environmental condition is output. The smaller the predicted variance, the closer the current operating condition is to the calibration condition, and the higher the positioning accuracy. The predicted variance is converted into a geometric weighting coefficient F using a preset mapping function. The formula for calculating the geometric weighting coefficient is as follows: ,in, Let g represent the exponential function, and g represent the prediction variance. The preset noise tolerance threshold is set based on the following: using the background noise of the measurement system as a benchmark, the upper limit of the variance that the sub-pixel coordinates can reflect physical reality is defined. For example, it is set to 0.02. The smoothing factor is used to adjust the evolution rate of the geometric weight coefficients as the prediction variance increases, balancing the preservation of feature points and the anomaly suppression effect. For example, it is set to 0.005. The calculation formula is constructed based on the nonlinear saturation and exponential decay characteristics of the logistic mapping function. Within the confidence interval with low prediction variance, the geometric weight coefficients are nonlinearly made to approach the upper limit of 1, thereby maximizing the preservation of high-fidelity boundary features. In the discrete interval where the variance exceeds the limit, active weight weakening is performed, thereby reducing the impact of environmental interference points on subsequent geometric fitting. The geometric weight coefficients are logically encapsulated one-to-one with each sub-pixel coordinate in the original sub-pixel point cloud to generate a weighted confidence feature set. The weighted confidence feature set refers to a heterogeneous data set composed of each sub-pixel coordinate in the original sub-pixel point cloud and its associated geometric weight coefficients.

[0113] Step S20 addresses the interference of high reflectivity and environmental fluctuations at the cable laying site on the robustness of edge positioning by using a probabilistic segmentation mask, the original sub-pixel point cloud, and a weighted confidence feature set. This achieves feature point extraction with sub-pixel accuracy and physical reliability evaluation. Specifically, the probabilistic segmentation mask enables accurate division of heterogeneous regions of the cable; the original sub-pixel point cloud achieves spatial topological alignment between semantic features and the actual physical contour of the cable; and the weighted confidence feature set actively weakens the weights of low-confidence feature points.

[0114] Step S30: Perform geometric center alignment using the weighted confidence feature set, output the cable geometric center set, extract dimensional parameters based on the cable geometric center set, and output the cable cross-section measurement results.

[0115] Further, step S30 includes:

[0116] Step S31: Perform geometric center alignment using the weighted confidence feature set and output the cable geometric center set.

[0117] After obtaining the weighted confidence feature set, a spatial alignment operation based on weighted principal component analysis is performed to eliminate the pose offset and rotation deviation of the cable in the image sensor's field of view. Since feature points at different locations in the original sub-pixel point cloud are affected by environmental interference to varying degrees, a structured registration framework guided by geometric weight coefficients is established. The aim is to suppress geometric centroid drift caused by edge signal diffusion or reflective interference, and to output the physical geometric center of the cable cross-section.

[0118] The sub-pixel coordinates in the weighted confidence feature set are extracted to calculate the weighted mean center. Specifically: the horizontal axis value of each sub-pixel coordinate is multiplied by the corresponding geometric weight coefficient, and then divided by the sum of all geometric weight coefficients to obtain the horizontal axis weighted mean; the vertical axis weighted mean is calculated simultaneously; the coordinate point formed by the horizontal axis weighted mean and the vertical axis weighted mean is defined as the weighted mean center. Using the weighted mean center as a reference, the sub-pixel coordinates in the weighted confidence feature set are decentered. Specifically: the weighted confidence feature set is traversed, and the horizontal axis value of the weighted mean center is subtracted from the horizontal axis value of each sub-pixel coordinate in the cross-sectional coordinate system, and the vertical axis value of the sub-pixel coordinate is subtracted from the vertical axis value of the weighted mean center to obtain the decentered coordinate vector; the decentered coordinate vector and the corresponding geometric weight coefficient are used to construct a weighted covariance matrix. Specifically: The decentralized coordinate vector is multiplied by its transpose row vector to obtain a second-order matrix. Each element of this matrix is ​​then multiplied by its corresponding geometric weight coefficient to obtain a weighted second-order matrix. All weighted second-order matrices are summed to obtain a weighted covariance matrix. The characteristic equation of the weighted covariance matrix is ​​solved to obtain two eigenvalues ​​and two corresponding eigenvectors. The largest eigenvalue is extracted and defined as the maximum eigenvalue. The unit eigenvector corresponding to the maximum eigenvalue is also extracted and defined as the principal eigenvector. The geometric alignment direction is defined based on the direction of the principal eigenvector. This geometric alignment direction represents the physical principal axis along which the energy distribution of the cable cross-section is most concentrated on the imaging plane. The sub-pixel coordinates in the weighted confidence feature set are spatially rotated and translated along this geometric alignment direction to align it with the principal axis of the cross-sectional coordinate system. For the sub-pixel coordinates of the regions identified as cable insulation, the corresponding weighted mean center is calculated and defined as the geometric center of the insulation layer; for the sub-pixel coordinates of the regions identified as conductive cores, the corresponding weighted mean center is calculated and defined as the center of the conductive core. The geometric centers of the insulation layer and the conductive core are combined to obtain the set of cable geometric centers.

[0119] Step S32: Perform radial distance measurement based on the spatially aligned feature distribution, calculate and output the cable cross-section measurement results.

[0120] After determining the geometric center set of the cable, multidimensional geometric parameters are extracted and quantified to obtain key indicators reflecting the cable manufacturing quality. Since the original sub-pixel point cloud has undergone spatial location backtracking using an error fingerprint database and confidence filtering using geometric weight coefficients, a dimensional measurement model with sub-pixel accuracy can be established. The goal is to output physically significant cable cross-section measurement results by calculating the relative geometric relationships between the various functional layers of the cable.

[0121] Specifically, with the center of the conductive core as the pole and the positive X-axis of the cross-sectional coordinate system as the reference, the angle between the scanning ray and the reference is defined as the circumferential polar angle. The cable cross-section is divided into a full circumferential angular domain using angle step values, constructing a radial search path covering a 360-degree spatial range. At each discrete orientation determined by the circumferential polar angle, the sub-pixel coordinates of the region identified as the cable insulation layer are retrieved along the direction of the scanning ray, and the Euclidean distance between these sub-pixel coordinates and the center of the conductive core is calculated, defined as the outer diameter radial distance. The numerical sequence of outer diameter radial distances at all discrete orientations is defined as the outer diameter distribution sequence. Simultaneously, the sub-pixel coordinates of the outer boundary of the region identified as the conductive core are retrieved, and the average radial distance between these sub-pixel coordinates and the center of the conductive core is calculated, defined as the equivalent radius of the core. The equivalent radius of the core is subtracted from the value corresponding to each discrete orientation in the outer diameter distribution sequence to obtain a thickness distribution set consisting of a series of radial thickness values. Statistical analysis is performed on the thickness distribution set to calculate the arithmetic mean of the thickness values, defined as the average insulation thickness; the minimum value in the thickness distribution set is retrieved and defined as the minimum insulation thickness. In the cross-sectional coordinate system, the spatial displacement vector of the geometric center of the insulation layer relative to the center of the conductive core is calculated, defined as the center offset vector. The eccentricity is calculated by dividing the magnitude of the center offset vector by the average insulation thickness. The average insulation thickness, minimum insulation thickness, and eccentricity are digitally packaged and the cable cross-section measurement results are output. See also... Figure 5This is a physical logic diagram for performing spatial alignment and geometric parameter measurement based on a weighted confidence feature set, provided in an embodiment of the present invention. The light gray ring in the center represents the conductive core region, and the outer dashed ring represents the insulating layer region. Discrete dots scattered on the edge contour represent sub-pixel coordinates, with their corresponding geometric weight coefficients represented by color intensity. Black dots represent reliable edge features with high geometric weight coefficients, which dominate the contribution to spatial alignment and thickness calculation. Gray dots represent abnormal edge features with low geometric weight coefficients due to environmental noise interference or optomechanical coupling, which are actively weakened in subsequent geometric fitting. The diagram marks the conductive core center determined by weighted principal component analysis, i.e., O1 in the diagram, and the insulating layer geometric center O2. The red vector arrow connecting the two points is the center offset vector. The blue dashed line pointing from the conductive core center to the edge represents the sampling path in discrete orientations, used to quantify the radial distance of the outer diameter in that direction.

[0122] Step S30, through the cable geometric center set and cable cross-section measurement results, solves the technical problem that feature point clouds are easily trapped in local optima caused by stress field interference during registration, leading to jumps in insulation thickness values ​​and deviations in eccentricity calculations from the standard. This achieves high-fidelity reconstruction of the cable's physical cross-sectional configuration under non-uniform mechanical stress conditions. Specifically, the cable geometric center set, as a spatial reference composed of the insulation layer's geometric center and the conductor core's center, suppresses geometric center of gravity drift caused by edge signal dispersion. The cable cross-section measurement results, as digitally packaged data consisting of average insulation thickness, minimum insulation thickness, and eccentricity, enable sub-pixel-level quantitative evaluation of cable manufacturing quality and insulation performance.

[0123] Example 2

[0124] This embodiment, based on Embodiment 1, provides a cable size parameter detection system based on point cloud feature registration, such as... Figure 6 As shown, it includes:

[0125] The pseudo-drift decoupling module is used to acquire multispectral image signals that characterize the optical features of the cable cross-section under laying conditions, perform physical simulation on the multispectral image signals to obtain the light trace energy matrix, and perform topological mapping calculation on the light trace energy matrix and the multispectral image signals to obtain the error fingerprint library.

[0126] Feature weighting module: used to perform semantic segmentation processing on multispectral image signals, obtain a probability segmentation mask that represents the regional distribution of each component of the cable, call the error fingerprint library to perform spatial position offset correction on the probability segmentation mask, obtain a correction matrix, and use the correction matrix to perform calculations to obtain the original sub-pixel point cloud, obtain environmental stress parameters to perform confidence evaluation on the original sub-pixel point cloud, and obtain a weighted confidence feature set;

[0127] The cross-section calculation module is used to perform geometric center alignment using a weighted confidence feature set, output the cable geometric center set, extract dimensional parameters based on the cable geometric center set, and output the cable cross-section measurement results.

[0128] In the pseudo-drift decoupling module, the process involves acquiring a multispectral image signal characterizing the optical features of the cable cross-section under laying conditions, performing physical simulation on the multispectral image signal to obtain a light trace energy matrix, and performing a topological mapping calculation between the light trace energy matrix and the multispectral image signal to obtain an error fingerprint database, including:

[0129] Step S11: Obtain multispectral raw image signals characterizing the optical features of the cable cross-section in its laying state;

[0130] Step S12: Perform physical simulation on the original multispectral image signal to obtain the light trace energy distribution matrix;

[0131] Step S13: Perform topological mapping calculation on the light trace energy distribution matrix and the original multispectral image signal to obtain the error correction fingerprint database.

[0132] In the feature weighting module, semantic segmentation processing is performed on the multispectral image signal to obtain a probability segmentation mask representing the regional distribution of each component of the cable. An error fingerprint database is then used to perform spatial offset correction on the probability segmentation mask to obtain a correction matrix. This correction matrix is ​​then used to perform calculations to obtain the original sub-pixel point cloud. Environmental stress parameters are then obtained to perform confidence assessment on the original sub-pixel point cloud, resulting in a weighted confidence feature set, including:

[0133] Step S21: Perform semantic segmentation processing on the multispectral original image signal to obtain a probability segmentation mask that characterizes the regional distribution of each component of the cable.

[0134] Step S22: Call the error correction fingerprint library to perform spatial position offset correction on the probability segmentation mask to obtain the correction probability matrix, and use the correction probability matrix to perform calculations to obtain the original sub-pixel point cloud;

[0135] Step S23: Obtain environmental stress parameters and perform confidence assessment on the original sub-pixel point cloud to obtain a weighted confidence feature set.

[0136] The cross-section calculation module is used to perform geometric center alignment using a weighted confidence feature set, output the cable geometric center set, extract dimensional parameters based on the cable geometric center set, and output the cable cross-section measurement results, including:

[0137] Step S31: Perform geometric center alignment using the weighted confidence feature set and output the cable geometric center set;

[0138] Step S32: Extract dimensional parameters based on the cable geometric center set and output the cable cross-section measurement results.

[0139] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0140] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A cable size parameter detection method based on point cloud feature registration, characterized in that, The method includes: Multispectral image signals characterizing the optical features of the cable cross-section under laying conditions are acquired. Physical simulation is performed on the multispectral image signals to obtain the light trace energy matrix. Topological mapping calculation is performed between the light trace energy matrix and the multispectral image signals to obtain the error fingerprint database. The method for obtaining the light trace energy matrix includes: constructing a two-dimensional virtual detection plane, defined as the imaging plane, based on the plane where the photoelectric sensor of the industrial camera is located; The two-dimensional spatial position indices of all virtual photons intersecting the imaging plane are statistically analyzed and defined as the landing point coordinates; The proportion of residual light intensity caused by medium absorption or scattering during the propagation of the virtual photon is calculated and defined as the residual energy weight. The residual energy weights are summed according to the landing point coordinates to obtain the light trace energy matrix. The method for obtaining the error fingerprint database includes: performing convolution operations on the single-band feature matrix to obtain the measured derivative matrix, and performing convolution operations on the light trace energy matrix to generate the theoretical derivative matrix; Search along the scanning ray for two adjacent matrix elements in the measured derivative matrix and the theoretical derivative matrix that are located on the preset scanning ray path. When the product of the values ​​of the two adjacent matrix elements is less than zero, it is marked as a zero intersection point. By collecting the zero-crossing points in the measured derivative matrix, we obtain a set of observation edge points consisting of multiple observation edge points. By collecting the zero-crossing points in the theoretical derivative matrix, we obtain a set of theoretical edge points consisting of multiple theoretical edge points. For observed edge points and theoretical edge points at the same polar azimuth angle, the difference between the spatial pixel coordinates of the observed edge points and the coordinates of the theoretical edge points is calculated and defined as the pixel offset. The pixel offset is then encapsulated in a data structure to obtain an error fingerprint database. Semantic segmentation processing is performed on the multispectral image signal to obtain a probability segmentation mask that represents the regional distribution of each component of the cable. The error fingerprint library is called to perform spatial position offset correction on the probability segmentation mask to obtain a correction matrix. The correction matrix is ​​then used to perform calculations to obtain the original sub-pixel point cloud. Environmental stress parameters are obtained to perform confidence evaluation on the original sub-pixel point cloud to obtain a weighted confidence feature set. The method for obtaining the probability segmentation mask includes: inputting a multispectral image signal into the encoder of a preset deep convolutional neural network, extracting multiple feature maps, and defining them as multi-scale feature maps; Global average pooling is performed on the spatial dimension of the multi-scale feature map to extract the description vector describing the band features, and a fully connected neural network is called to perform a nonlinear transformation on the description vector to obtain the channel weight coefficients. The enhanced band feature representation is obtained by performing element-wise arithmetic product between the channel weight coefficients and the multi-scale feature map. Construct densely connected branches to perform channel-by-channel tensor concatenation fusion on the enhanced band feature representation, determine the probability of each spatial pixel coordinate belonging to different regions of the cable, and output a probability segmentation mask; Geometric center alignment is performed using a weighted confidence feature set, outputting a cable geometric center set. Based on the cable geometric center set, dimensional parameters are extracted, and the cable cross-section measurement results are output.

2. The cable size parameter detection method based on point cloud feature registration according to claim 1, characterized in that, The method for acquiring the multispectral image signal includes: A multi-band pulse light source is used to sequentially emit multiple pulses of a set center wavelength according to a preset band sequence to perform time-division illumination excitation on the cross section of the cable under test. An industrial camera is used to synchronously capture the reflected signal of the cross section of the cable under test at the sampling time when the corresponding center wavelength is triggered, and a series of single-band feature matrices are obtained. By utilizing the physical principle that light of different frequencies has different refractive index dispersion characteristics in cable insulation materials, the single-band feature matrices corresponding to each center wavelength are stacked in dimensional order according to the band sequence to generate the multispectral image signal containing spatial pixel coordinates and spectral response dimensions.

3. The cable size parameter detection method based on point cloud feature registration according to claim 1, characterized in that, The method for performing physical simulation on multispectral image signals includes: A two-dimensional rectangular coordinate system is constructed with the geometric center of the cross-section of the cable to be tested as the origin, and is defined as the cross-sectional coordinate system; In the cross-sectional coordinate system, with the origin as the starting point, the radial vector pointing to the spatial pixel coordinates is defined as the scanning ray, and the horizontal angle between the scanning ray and the positive direction of the horizontal axis of the cross-sectional coordinate system is defined as the polar azimuth angle. Extend the preset axial step length along the direction perpendicular to the plane of the cross-section coordinate system to construct the three-dimensional simulation space of the cable with a cylindrical structure. Use the Cartesian meshing method to divide the three-dimensional simulation space of the cable into voxel units with a three-dimensional matrix structure, and define the vertices of the voxel units as spatial mesh nodes. Obtain the mechanical stress vector acting on the surface of the cable under test, define the outer periphery of the cable's three-dimensional simulation space as the load boundary and apply the mechanical stress vector, calculate the stress tensor at each spatial grid node, and define the set of stress tensors corresponding to all spatial grid nodes in space as the stress field distribution.

4. The cable size parameter detection method based on point cloud feature registration according to claim 3, characterized in that, The method for performing physical simulation on multispectral image signals further includes: Obtain the initial refractive index and the trace of the stress tensor corresponding to each center wavelength. Perform an accumulation operation using the product of the initial refractive index and the preset Kronecker sign, the product of the stress tensor and the preset stress optical constant, and the product of the trace of the stress tensor and the preset stress optical constant to obtain the refractive index tensor corresponding to each spatial grid node. Path tracing of virtual photons is performed within the three-dimensional simulation space of the cable. The virtual photons are probabilistic statistical sample units used to simulate the characteristics of optical energy propagation. By using the refractive index tensor to perform micro-element-level vector perturbation correction on the motion direction of the virtual photon, the emission direction vector is obtained.

5. The cable size parameter detection method based on point cloud feature registration according to claim 4, characterized in that, The obtained original sub-pixel point cloud includes: Traverse each spatial pixel coordinate in the probability segmentation mask, retrieve the corresponding pixel offset from the error fingerprint database, perform vector subtraction operation on each spatial pixel coordinate and the corresponding pixel offset to obtain the corrected spatial coordinates, and map all the corrected spatial coordinates and their associated class probabilities to obtain the correction matrix. Perform differentiation on the correction matrix to lock the zero-crossing region determined by the product of adjacent element values ​​being less than zero, extract the class probabilities arranged in this zero-crossing region, and define it as a probability profile; On the probability profile, two adjacent pixels whose numerical distributions fall on either side of a preset edge threshold are retrieved and defined as step boundary point pairs. Linear interpolation is then performed using the corrected spatial coordinates and class probabilities corresponding to the step boundary point pairs to obtain sub-pixel coordinates. All sub-pixel coordinates under the polar azimuth angles are aggregated to obtain the original sub-pixel point cloud.

6. The cable size parameter detection method based on point cloud feature registration according to claim 1, characterized in that, The confidence assessment includes: The system acquires real-time ambient temperature, relative humidity, and radial compressive stress and combines them into an environmental state vector. It also acquires multiple historical environmental state vectors as a training sample set. Calculate the spatial correlation between any two historical environmental state vectors in the training sample set, perform matrix arrangement on each spatial correlation value to obtain the original covariance matrix, and superimpose a preset system noise term on the main diagonal of the original covariance matrix to obtain the training set covariance matrix. Calculate the similarity between the environmental state vector and each of the historical environmental state vectors in the training sample set, and encapsulate each similarity value into a vector to obtain the test covariance vector. Perform matrix inversion on the training set covariance matrix to obtain the precision matrix. Perform inner product operation on the transpose row vector of the test covariance vector, the precision matrix, and the test covariance vector, and output a scalar value, which is defined as the projected energy term.

7. The cable size parameter detection method based on point cloud feature registration according to claim 6, characterized in that, The obtained weighted confidence feature set includes: Set the prior variance, calculate the difference between the prior variance and the projected energy term, and obtain the prediction variance for each sub-pixel coordinate; The predicted variance is converted into geometric weight coefficients using a preset mapping function; The geometric weight coefficients are encapsulated with each corresponding sub-pixel coordinate in the original sub-pixel point cloud to generate a weighted confidence feature set.

8. The cable size parameter detection method based on point cloud feature registration according to claim 7, characterized in that, The geometric center set of the output cable includes: Subtract the horizontal axis value of the weighted mean center from the horizontal axis value of each subpixel coordinate in the cross-sectional coordinate system, and subtract the vertical axis value of the weighted mean center from the vertical axis value of the subpixel coordinate to obtain the decentralized coordinate vector. A weighted covariance matrix is ​​constructed using decentralized coordinate vectors and their corresponding geometric weight coefficients; Calculate the weighted mean of the subpixel coordinates along the horizontal axis and the weighted mean of the vertical axis in the weighted confidence feature set, and combine them to obtain the center of the weighted mean. A second-order matrix is ​​generated by performing an outer product operation between the decentralized coordinate vector and its transpose row vector, and then a weighted product operation is performed with the geometric weight coefficients. The weighted covariance matrix is ​​obtained by summing the results. Perform eigenvalue decomposition on the weighted covariance matrix and extract the principal eigenvectors. Perform spatial rotation and alignment along the direction of the principal eigenvectors, and output the set of cable geometric centers consisting of the geometric centers of the insulation layer and the centers of the conductive cores.

9. The cable size parameter detection method based on point cloud feature registration according to claim 8, characterized in that, The output cable cross-section measurement results include: With the center of the conductive wire core as the pole, the angle between the scanning ray and the positive direction of the horizontal axis of the cross-sectional coordinate system is defined as the circumferential polar angle, and a radial search path is constructed. Calculate the Euclidean distance between the sub-pixel coordinates of the pre-identified cable insulation layer region and the pole at each circumferential polar angle, and aggregate the outer diameter distribution sequence. Calculate the average distance between the sub-pixel coordinates of the pre-identified conductive core region and the center of the conductive core, and define it as the equivalent radius of the core. Subtract the equivalent radius of the wire core from the values ​​of each Euclidean distance in the outer diameter distribution sequence to obtain a thickness distribution set consisting of a series of radial thickness values. The arithmetic mean and minimum value of the thickness distribution set are defined as the average insulation thickness and the minimum insulation thickness, respectively. The spatial displacement vector of the geometric center of the insulation layer relative to the center of the conductive core is calculated and defined as the center offset vector. The modulus of the center offset vector is divided by the average insulation thickness to obtain the eccentricity. The average insulation thickness, minimum insulation thickness and eccentricity are digitally encapsulated and the cable cross-section measurement results are output.

10. A cable size parameter detection system based on point cloud feature registration, used to implement the cable size parameter detection method based on point cloud feature registration as described in any one of claims 1-9, characterized in that, The system includes: The pseudo-drift decoupling module is used to acquire multispectral image signals that characterize the optical features of the cable cross-section under laying conditions, perform physical simulation on the multispectral image signals to obtain the light trace energy matrix, and perform topological mapping calculation on the light trace energy matrix and the multispectral image signals to obtain the error fingerprint library. The method for obtaining the light trace energy matrix includes: constructing a two-dimensional virtual detection plane, defined as the imaging plane, based on the plane where the photoelectric sensor of the industrial camera is located; The two-dimensional spatial position indices of all virtual photons intersecting the imaging plane are statistically analyzed and defined as the landing point coordinates; The proportion of residual light intensity caused by medium absorption or scattering during the propagation of the virtual photon is calculated and defined as the residual energy weight. The residual energy weights are summed according to the landing point coordinates to obtain the light trace energy matrix. The method for obtaining the error fingerprint database includes: performing convolution operations on the single-band feature matrix to obtain the measured derivative matrix, and performing convolution operations on the light trace energy matrix to generate the theoretical derivative matrix; Search along the scanning ray for two adjacent matrix elements in the measured derivative matrix and the theoretical derivative matrix that are located on the preset scanning ray path. When the product of the values ​​of the two adjacent matrix elements is less than zero, it is marked as a zero intersection point. By collecting the zero-crossing points in the measured derivative matrix, we obtain a set of observation edge points consisting of multiple observation edge points. By collecting the zero-crossing points in the theoretical derivative matrix, we obtain a set of theoretical edge points consisting of multiple theoretical edge points. For observed edge points and theoretical edge points at the same polar azimuth angle, the difference between the spatial pixel coordinates of the observed edge points and the coordinates of the theoretical edge points is calculated and defined as the pixel offset. The pixel offset is then encapsulated in a data structure to obtain an error fingerprint database. Feature weighting module: used to perform semantic segmentation processing on multispectral image signals, obtain a probability segmentation mask that represents the regional distribution of each component of the cable, call the error fingerprint library to perform spatial position offset correction on the probability segmentation mask, obtain a correction matrix, and use the correction matrix to perform calculations to obtain the original sub-pixel point cloud, obtain environmental stress parameters to perform confidence evaluation on the original sub-pixel point cloud, and obtain a weighted confidence feature set; The method for obtaining the probability segmentation mask includes: inputting a multispectral image signal into the encoder of a preset deep convolutional neural network, extracting multiple feature maps, and defining them as multi-scale feature maps; Global average pooling is performed on the spatial dimension of the multi-scale feature map to extract the description vector describing the band features, and a fully connected neural network is called to perform a nonlinear transformation on the description vector to obtain the channel weight coefficients. The enhanced band feature representation is obtained by performing element-wise arithmetic product between the channel weight coefficients and the multi-scale feature map. Construct densely connected branches to perform channel-by-channel tensor concatenation fusion on the enhanced band feature representation, determine the probability of each spatial pixel coordinate belonging to different regions of the cable, and output a probability segmentation mask; The cross-section calculation module is used to perform geometric center alignment using a weighted confidence feature set, output the cable geometric center set, extract dimensional parameters based on the cable geometric center set, and output the cable cross-section measurement results.

Images