GIS (Geographic Information System) equipment X-ray detection method for realizing automatic compensation of spatial pose

Abstract

The invention provides a GIS equipment X-ray detection method for realizing space pose automatic compensation, and belongs to the technical field of GIS equipment detection.Point cloud data of GIS equipment are acquired through laser radar multi-view scanning, noise is eliminated by adopting self-adaptive multi-scale statistical filtering, and the spatial pose automatic compensation is realized. Non-rigid point cloud registration is realized by using a space registration optimization model combining a deep expansion network and sparse coding, the position of an internal conductor is corrected based on an electromagnetic field inverse problem inversion algorithm, and the space position and the ray angle of an X-ray machine are automatically adjusted through an attitude compensation adjustment function according to a pose deviation vector. The technical problem that the detection pose cannot be automatically compensated due to equipment deformation and positioning errors during X-ray detection of GIS equipment is solved.

Description

Technical Field

[0001] This invention belongs to the field of GIS equipment inspection technology, and more specifically, relates to an X-ray inspection method for GIS equipment that realizes automatic spatial pose compensation. Background Technology

[0002] X-ray inspection technology for GIS equipment is used to inspect the contact status and defects of conductor contacts inside gas-insulated switchgear. Traditional methods rely on manual measurement and experience to determine the X-ray machine's inspection position. A laser rangefinder measures the equipment's casing dimensions, calculates the internal contact positions based on design drawings, and the X-ray machine angle is manually adjusted for scanning. However, in existing technologies, GIS equipment undergoes non-rigid deformation due to thermal stress and mechanical loads during long-term operation, leading to discrepancies between design drawings and actual geometry. Furthermore, the laser rangefinder can only acquire discrete point coordinates on the casing surface, failing to construct a complete three-dimensional digital model, resulting in significant errors in estimating the internal contact positions. Traditional methods require repeated adjustments to the X-ray machine position for trial scans, relying on the inspector's experience to determine if the X-ray angle is aligned with the contacts, resulting in low inspection efficiency and a tendency to miss critical areas. In other words, existing technologies suffer from a technical problem where equipment deformation and positioning errors during GIS equipment X-ray inspection cannot be automatically compensated for. Summary of the Invention

[0003] In view of this, the present invention provides a method for X-ray inspection of GIS equipment that realizes automatic spatial pose compensation, which can solve the technical problem in the prior art that the detection pose cannot be automatically compensated due to equipment deformation and positioning error during X-ray inspection of GIS equipment.

[0004] This invention is implemented as follows: A method for X-ray detection of GIS equipment with automatic spatial pose compensation is provided, comprising the following steps: Multi-view scanning of the GIS equipment using a lidar to acquire an initial point cloud dataset; simultaneously, recording the spatial pose information of the lidar using an RTK differential positioning system; adaptive multi-scale statistical filtering of the initial point cloud dataset to eliminate noise points caused by electromagnetic interference and metal reflection, obtaining a preprocessed point cloud dataset; inputting the preprocessed point cloud dataset and spatial pose information into a spatial registration optimization model, which outputs a complete digital model of the GIS equipment containing the spatial coordinates of the conductor contact; and calculating X-rays based on the complete digital model of the GIS equipment. The material thickness distribution along the penetration path is used to determine the initial exposure parameter set based on the material thickness distribution using an exposure parameter prediction function. External electromagnetic field distribution data of the GIS equipment is collected, and an electromagnetic field inverse problem inversion algorithm is used to obtain the probability distribution estimate of the internal conductor position based on the external electromagnetic field distribution data. The spatial coordinates of the conductor contact in the complete GIS equipment digital model are corrected based on the probability distribution estimate to obtain the corrected contact coordinates. Based on the corrected contact coordinates and the current pose of the X-ray machine, a pose deviation vector is calculated. The spatial position and ray angle of the X-ray machine are automatically adjusted based on the pose deviation vector using an attitude compensation adjustment function. After automatic pose compensation, X-ray imaging detection is performed based on the initial exposure parameter set.

[0005] The RTK differential positioning system provides centimeter-level precision three-dimensional coordinates for a mobile station installed on a lidar and X-ray machine by receiving differential signals from a base station and satellites.

[0006] Specifically, the adaptive multi-scale statistical filtering process involves calculating the local density value and curvature change value of each point in the preprocessed point cloud dataset, dividing the point cloud data into high-density regions and low-density regions based on the local density values, performing statistical analysis on the high-density regions using a small-scale neighborhood, and performing statistical analysis on the low-density regions using a large-scale neighborhood.

[0007] The adaptive multi-scale statistical filtering process further includes calculating the average distance and standard deviation of each point to its neighboring points. When the average distance of a point exceeds a dynamic threshold, it is determined to be a noise point. The dynamic threshold is adjusted according to the curvature change value of the region where the point is located.

[0008] The spatial registration optimization model includes an input layer, a first feature extraction module, a sparse coding layer, a deep unfolded network layer, a second feature extraction module, a Riemannian manifold registration module, a local affine transformation estimation module, a global optimization layer, and an output layer.

[0009] The deep unfolded network layer unfolds the iterative process of the fast iterative shrinking threshold algorithm FISTA into 8 learnable network layers, each containing a soft threshold operator and a gradient descent update module.

[0010] The Riemannian manifold registration module embeds high-level semantic information into the Riemannian manifold space and calculates the point-pair similarity matrix using geodesic distance, which is calculated using Dijkstra's shortest path algorithm.

[0011] The local affine transformation estimation module constrains the transformation consistency of adjacent regions through graph Laplacian regularization and outputs the transformation parameters of each point cloud block. The global optimization layer uses the Log-Euclidean framework to iteratively solve the optimal transformation field on the Lie group manifold.

[0012] The spatial registration optimization model is trained end-to-end using the Adam optimizer. The loss function is a weighted sum of registration error loss and sparse regularization loss. The registration error loss is the mean square error between the predicted coordinates and the label data.

[0013] The calculation of the exposure parameter prediction function includes extracting the thickness and density values ​​of all material layers along the X-ray penetration path, querying the corresponding linear attenuation coefficient according to the material type, calculating the total equivalent thickness of the penetration path, and querying the initial tube voltage and initial exposure time values ​​based on the total equivalent thickness.

[0014] The electromagnetic field inverse problem inversion algorithm establishes a mathematical model of the electromagnetic field forward problem based on Maxwell's equations, uses the finite element method to discretize the internal space of GIS equipment, and calculates the electromagnetic field distribution in the external space by solving a sparse linear equation system.

[0015] Optionally, the electromagnetic field inverse problem inversion algorithm also introduces the Tikhonov regularization method to add a smoothness constraint term to the objective function. The regularization parameter is determined by the L-curve method, and the regularized optimization problem is solved iteratively by the conjugate gradient method.

[0016] Optionally, the electromagnetic field inverse problem inversion algorithm also introduces the Bayesian inference framework, treating the conductor position as a random variable, and obtaining the posterior probability distribution of the position parameters through the Markov chain Monte Carlo sampling method to form a probability distribution estimate.

[0017] The calculation of the attitude compensation adjustment function includes extracting the corrected contact coordinates as the target detection point coordinates, calculating the three-dimensional vector difference between the target detection point coordinates and the ray source coordinates as the ideal ray direction vector, and calculating the cosine of the angle between the ideal ray direction vector and the actual ray direction vector as the direction deviation metric.

[0018] The attitude compensation adjustment function determines the compensation adjustment strategy based on the range of the comprehensive deviation index value, including keeping the current position of the X-ray machine unchanged, adjusting only the pitch and azimuth angles of the X-ray machine, first translating the X-ray machine and then adjusting the angles, or replanning the spatial position of the X-ray machine.

[0019] Before using LiDAR to perform multi-view scanning of GIS equipment, the process also includes multi-view scanning path planning based on graph theory shortest path. The surface of the GIS equipment is discretized into candidate scanning view nodes, a directed graph model is constructed, and an improved Dijkstra algorithm is used to search for the shortest path covering all key components.

[0020] This invention establishes a complete digital model of GIS equipment through multi-view scanning with lidar, and obtains the probability distribution estimate of the internal conductor position by combining electromagnetic field inverse problem inversion. A non-rigid point cloud registration algorithm based on Riemannian manifolds is used to handle equipment deformation, achieving automatic compensation and adjustment of the X-ray machine's pose. The spatial registration optimization model employs a combination of deep unfolded networks and sparse coding, mapping the iterative optimization process in point cloud registration to learnable network layers. End-to-end training adaptively extracts the geometric features of the GIS equipment, overcoming the deformation sensitivity of traditional methods. The electromagnetic field inverse problem inversion algorithm infers the actual position of the internal conductor by measuring the external electromagnetic field distribution, cross-validating with the point cloud modeling results to improve positioning reliability. The pose compensation adjustment function automatically adjusts the spatial position and ray angle of the X-ray machine based on the pose deviation vector, avoiding repeated trial scans. In summary, this invention solves the technical problem mentioned in the background art where the detection pose cannot be automatically compensated due to equipment deformation and positioning errors during X-ray detection of GIS equipment. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 This is a schematic diagram of the registration optimization model.

[0023] Figure 3 This is a posterior probability distribution diagram for the Metropolis-Hastings algorithm, which is used to invert the electromagnetic field inverse problem.

[0024] Figure 4 This is a diagram showing the results of X-ray imaging detection of the probe insertion depth measurement. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] like Figure 1The diagram shows a flowchart of an X-ray inspection method for GIS equipment that achieves automatic spatial pose compensation, provided by the present invention. This method includes the following steps: S01. Use LiDAR to scan the GIS equipment from multiple perspectives to obtain the initial point cloud dataset, and at the same time record the spatial pose information of the LiDAR through the RTK differential positioning system. S02. Perform adaptive multi-scale statistical filtering on the initial point cloud dataset to eliminate noise points caused by electromagnetic interference and metal reflection, and obtain a preprocessed point cloud dataset. S03. Input the preprocessed point cloud dataset and spatial pose information into the spatial registration optimization model. The spatial registration optimization model outputs a complete GIS equipment digital model containing the spatial coordinates of the conductor contact. S04. Calculate the material thickness distribution along the X-ray penetration path based on the complete GIS equipment digital model, and determine the initial exposure parameter set based on the material thickness distribution using the exposure parameter prediction function; S05. Collect the external electromagnetic field distribution data of the GIS equipment, and obtain the probability distribution estimate of the internal conductor position based on the external electromagnetic field distribution data through the electromagnetic field inverse problem inversion algorithm. Correct the spatial coordinates of the conductor contact in the complete GIS equipment digital model based on the probability distribution estimate to obtain the corrected contact coordinates. S06. Based on the corrected contact coordinates and the current pose of the X-ray machine, calculate the pose deviation vector. Then, through the pose compensation adjustment function, automatically adjust the spatial position and ray angle of the X-ray machine according to the pose deviation vector. After completing the automatic pose compensation, perform X-ray imaging detection according to the initial exposure parameter set.

[0027] The RTK differential positioning system receives differential signals from a base station and satellites to provide centimeter-level precision three-dimensional coordinates for mobile stations installed on lidar and X-ray machines. By calibrating and establishing the relative positions of the RTK antenna phase center, lidar scanning center, GIS equipment flange reference point, X-ray machine focal point, and detector center, the spatial positions of all equipment are unified under the same geodetic coordinate system, achieving a high-precision absolute spatial positioning reference.

[0028] like Figure 2As shown, the adaptive multi-scale statistical filtering process specifically includes the following steps: calculating the local density value and curvature change value of each point in the preprocessed point cloud dataset; dividing the point cloud data into high-density regions and low-density regions based on the local density values; performing statistical analysis on high-density regions using small-scale neighborhood and on low-density regions using large-scale neighborhood; calculating the average distance and standard deviation from each point to its neighboring points, and classifying a point as a noise point when its average distance exceeds a dynamic threshold; adjusting the dynamic threshold based on the curvature change value of the region where the point is located, increasing the dynamic threshold for edge feature points with curvature change values ​​greater than the curvature threshold, and decreasing the dynamic threshold for flat region points with curvature change values ​​less than the curvature threshold; deleting all data identified as noise points and retaining the valid point cloud to form the preprocessed point cloud dataset. The curvature threshold is set to 0.15. .

[0029] The specific structure of the spatial registration optimization model is as follows: the input layer receives the preprocessed point cloud dataset and spatial pose information; the first feature extraction module uses the PointNet++ architecture to extract local geometric features from the preprocessed point cloud dataset; the sparse coding layer maps the local geometric features to an overcomplete dictionary space for sparse representation; the deep unfolded network layer unfolds the iterative process of the fast iterative shrinking thresholding algorithm FISTA into 8 learnable network layers, each containing a soft thresholding operator and a gradient descent update module; the second feature extraction module performs deep convolution on the sparse features output by the deep unfolded network layer to extract high-level semantic information; the Riemannian manifold registration module embeds the high-level semantic information into the Riemannian manifold space and calculates the point pair similarity matrix using geodesic distance; the local affine transformation estimation module constrains the transformation consistency of adjacent regions through graph Laplacian regularization and outputs the transformation parameters of each point cloud block; the global optimization layer uses the Log-Euclidean framework to iteratively solve the optimal transformation field on the Lie group manifold; and the output layer generates a complete GIS equipment digital model containing the spatial coordinates of the conductor contact based on the optimal transformation field. The overcomplete dictionary contains 512 atoms and has a sparsity constraint parameter of 0.05.

[0030] The steps for establishing the training dataset of the spatial registration optimization model specifically include: collecting laser point cloud data from 30 different models of GIS equipment as original samples; manually labeling the conductor contacts, springs, and insulators of each original sample with precise three-dimensional coordinates as label data; applying random translation, rotation, and non-rigid deformation to the original samples to simulate equipment installation errors and thermal deformation, generating deformation samples; adding Gaussian noise of different intensities and outliers to the deformation samples to simulate on-site interference, generating 200 training samples from each original sample; dividing the training samples into a training set and a validation set at an 8:2 ratio; performing data augmentation processing on the training set, including random cropping, scaling transformation, and mirror flipping, finally obtaining a complete training dataset containing 48,000 samples.

[0031] The specific steps for training the spatial registration optimization model include: initializing the overcomplete dictionary as a random orthogonal matrix, and initializing the soft threshold parameter of the deep unfolded network layer to 0.01; performing end-to-end training using the Adam optimizer with an initial learning rate of 0.001; defining the loss function as a weighted sum of registration error loss and sparse regularization loss, where the registration error loss uses the mean square error between the predicted coordinates and the label data, and the sparse regularization loss uses the L1 norm to constrain the sparsity coefficients; each training batch contains 16 samples, with 300 iterations; evaluating the registration accuracy on the validation set every 50 iterations during training, and reducing the learning rate to 0.1 times the original value when the registration error on the validation set no longer decreases for 10 consecutive iterations; and converging the training after the learning rate has been reduced 3 times, saving the model parameters with the smallest registration error as the final model.

[0032] The principle behind the combination of dictionary learning and deep unfolding network based on sparse coding in the spatial registration optimization model is as follows: Traditional sparse coding solves for the optimal sparse representation through iterative algorithms, but the number of iterations and convergence speed are difficult to control. The deep unfolding network maps each iteration of the Fast Iterative Shrinking Thresholding Algorithm (FISTA) to a network layer, using a soft thresholding operator as the activation function and the gradient descent step size as a learnable parameter. Dictionary learning and network training are performed simultaneously; atoms from the overcomplete dictionary are updated through backpropagation to adapt to the geometric features of the GIS point cloud. The unfolded network has a fixed number of 8 layers, ensuring both the quality of the sparse representation and enabling fast forward inference. Each network layer adaptively adjusts the shrinking threshold through learning, exhibiting robustness to preprocessed point cloud datasets with varying noise levels. This combination makes the sparse representation solution process end-to-end trainable, maintaining the interpretability of traditional optimization algorithms while significantly improving computational efficiency. Sparse coding can capture the inherent structural features of the preprocessed point cloud dataset, projecting high-dimensional point clouds into a low-dimensional sparse space, effectively compressing redundant information and highlighting key geometric features. The dictionary learning process automatically discovers typical local structural patterns of GIS equipment. Deeply unfolded network layers enhance the spatial registration optimization model's adaptability to deformation and noise through multi-layer feature abstraction. This combined approach significantly improves the accuracy and speed of point cloud registration, enabling the spatial registration optimization model to handle large-scale point cloud data containing hundreds of millions of points. Simultaneously, it exhibits strong robustness against electromagnetic interference and on-site noise from metal reflections, providing a reliable spatial reference for subsequent X-ray inspection, avoiding repeated scanning due to positioning errors, and improving on-site inspection efficiency.

[0033] The calculation steps of the exposure parameter prediction function are as follows: extract the thickness and density values ​​of all material layers along the X-ray penetration path; query the corresponding linear attenuation coefficient from the material attenuation coefficient database according to the material type; calculate the total equivalent thickness of the penetration path, expressed by the following formula: the thickness value of each material layer along the path divided by the reference aluminum thickness, multiplied by the ratio of the linear attenuation coefficient of the material to the linear attenuation coefficient of the aluminum, and summed for all layers to obtain the total equivalent thickness; query the initial tube voltage and initial exposure time values ​​from the energy-thickness database based on the total equivalent thickness; calculate the distance from the geometric center point of the penetration path to the detector; correct the initial exposure time value based on the distance value, the corrected exposure time value being equal to the initial exposure time value multiplied by the distance value divided by the standard distance, and output the corrected exposure time value and the initial tube voltage value as the initial exposure parameter set. The reference aluminum thickness is 10 mm, and the standard distance is 1000 mm.

[0034] The principle of the electromagnetic field inverse problem inversion algorithm is as follows: During operation, the internal conductors of the GIS equipment generate a time-varying electromagnetic field through alternating current. This time-varying electromagnetic field diffracts and transmits through the gaps in the metal casing and the insulating medium, forming a measurable electromagnetic radiation distribution outside the GIS equipment. A mathematical model of the electromagnetic field forward problem is established based on Maxwell's equations, which describe the differential relationships between electric field strength, magnetic field strength, current density, and charge density. The finite element method is used to discretize the internal space of the GIS equipment into tetrahedral mesh elements, and the electromagnetic field components are expanded using basis functions on each tetrahedral mesh element. Given the position and current distribution of the internal conductors as boundary conditions and excitation sources, the electromagnetic field distribution in the external space is calculated by solving a sparse linear equation system. An electromagnetic field feature library corresponding to different contact positions and contact states is established, with each configuration corresponding to a set of external field distribution data. The inverse problem solution process involves deploying multiple electromagnetic field sensor arrays outside the GIS equipment to collect measured electromagnetic field strength and phase data, forming external electromagnetic field distribution data. The external electromagnetic field distribution data is matched with simulated data in the electromagnetic field feature library. Due to the ill-posedness of the electromagnetic field equations, direct solution will produce unstable solutions. A Tikhonov regularization method is introduced, adding a smoothness constraint term to the objective function. The regularization parameter is determined using the L-curve method. The conjugate gradient method is used to iteratively solve the regularized optimization problem to obtain the optimal estimate of the internal conductor's position. Considering the influence of measurement noise and model error, a Bayesian inference framework is introduced, treating the conductor position as a random variable. The posterior probability distribution is calculated based on the prior probability distribution and the likelihood function. The posterior probability distribution of the position parameters is obtained through Markov chain Monte Carlo sampling, forming probability distribution estimates, which provide confidence intervals.

[0035] The specific implementation of the electromagnetic field inverse problem inversion algorithm in the scheme is as follows: In step S05, 16 triaxial electromagnetic field sensors are arranged axially and radially on the surface of the GIS equipment casing, with a spacing of 500mm between the triaxial electromagnetic field sensors. The triaxial electromagnetic field sensors collect electromagnetic field signals with a frequency range of 50Hz to 1000Hz, a sampling rate of 10kHz, and a collection time of 10s. The collected signals are subjected to a Fast Fourier Transform to extract the amplitude and phase information of the dominant frequency component, forming a 64-dimensional measured electromagnetic field feature vector. The 10 most similar sets of simulated data are retrieved from a pre-established electromagnetic field feature library, and the similarity is measured by the cosine distance of the 64-dimensional measured electromagnetic field feature vector. Using the conductor positions corresponding to the 10 sets of simulated data as the initial solution space, a Tikhonov regularized objective function is constructed, with the regularization parameter set to 0.001. The conjugate gradient method is used to iteratively solve the optimization problem, with the iteration termination condition being that the change in the objective function is less than 0.0001 or the number of iterations exceeds 100. After obtaining the optimal conductor position estimate, a Gaussian distribution with a mean and a standard deviation of 5 mm is used as the prior probability distribution. The Metropolis-Hastings algorithm is used to perform 10,000 Markov chain samplings to calculate the posterior probability distribution of the conductor position. The mean of the posterior probability distribution is extracted as the probability distribution estimate, and the standard deviation is extracted as the positioning uncertainty index. When the positioning uncertainty index is greater than 10 mm, the number of triaxial electromagnetic field sensors is increased or the positions of the triaxial electromagnetic field sensors are adjusted to re-acquire external electromagnetic field distribution data.

[0036] The technical advantages of the electromagnetic field inverse problem inversion algorithm are as follows: It overcomes the limitations of traditional X-ray detection, which relies on prior knowledge for blind scanning. By using non-invasive electromagnetic field measurements, it infers the actual position of conductors inside the GIS, providing a physical basis for the detection scheme. External electromagnetic field distribution data is sensitive to changes in conductor position; insufficient probe insertion depth or offset can significantly alter the external field distribution characteristics. The electromagnetic field inverse problem inversion algorithm can capture these minute changes. The finite element method ensures the accuracy of the physical model, the Tikhonov regularization method effectively suppresses solution oscillations caused by measurement noise, and the Bayesian inference framework quantifies the positioning uncertainty index, providing a basis for decision-making. The probability distribution estimate obtained by the electromagnetic field inverse problem inversion algorithm guides the X-ray machine to prioritize scanning high-probability areas, avoiding full-coverage scanning of the entire GIS equipment and significantly reducing detection time. Simultaneously, the probability distribution estimate verifies the accuracy of the laser point cloud modeling; the two methods mutually reinforce each other, improving overall positioning reliability. For operating GIS equipment, continuous monitoring of external electromagnetic field distribution data and real-time tracking of internal structural changes by the electromagnetic field inverse problem inversion algorithm enable timely detection of anomalies, providing early warnings for preventative maintenance and avoiding power outage losses due to sudden failures.

[0037] The principle of the non-rigid point cloud registration algorithm based on Riemannian manifolds in the Riemannian manifold registration module is as follows: During long-term operation, GIS equipment is affected by thermal expansion and contraction, mechanical stress, and installation errors, causing non-rigid deformation of its geometry. Traditional ICP algorithms that assume rigid body transformation cannot accurately register the deformed point cloud. A Riemannian manifold is a mathematical space that is locally similar to Euclidean space but has a curved geometry overall, suitable for describing the intrinsic geometric properties of surfaces. The preprocessed point cloud dataset is embedded into the Riemannian manifold space, and the similarity between point pairs is measured using geodesic distance on the manifold. A geodesic is the shortest path between two points on the Riemannian manifold, and its length reflects the true distance between points under the intrinsic geometry of the surface. Compared to the straight-line distance in Euclidean space, geodesic distance can more accurately characterize the proximity relationship of points on the surface and is invariant to local deformation. Rotational transformations in three-dimensional space form a special orthogonal group SO3, which is a Lie group manifold rather than a nonlinear space. A Log-Euclidean framework is employed to map the rotation matrix to the Lie algebra tangent space. Linear operations are then performed in the Lie algebra tangent space before mapping back to the Lie group manifold, avoiding singularity issues caused by direct optimization on the Lie group manifold. A local affine transformation model divides the preprocessed point cloud dataset into multiple overlapping regions, each described by an affine matrix representing its rotation, scaling, and translation transformations. The transformation parameters of adjacent overlapping regions are constrained for consistency using graph Laplacian regularization, which penalizes transformation differences between adjacent overlapping regions, ensuring the smoothness of the overall deformation field. The optimization objective function includes a registration error term and a graph Laplacian regularization constraint term. An alternating optimization strategy is used: first, the transformation parameters are fixed while updating the point correspondences; then, the point correspondences are fixed again while updating the transformation parameters, iterating until convergence.

[0038] The specific implementation of the non-rigid point cloud registration algorithm based on Riemannian manifolds in the spatial registration optimization model is as follows: The Riemannian manifold registration module receives the high-level semantic information output by the second feature extraction module and uses this high-level semantic information as the coordinate representation of the Riemannian manifold space. A manifold embedding is constructed using the feature functions of the Laplace-Beltrami operator, and the local covariance matrix of the preprocessed point cloud dataset is calculated. The first 10 principal components are extracted as manifold coordinates. A k-nearest neighbor graph is constructed in the Riemannian manifold space, with k set to 20. The geodesic distance between any two points is calculated using the Dijkstra shortest path algorithm. A point-to-point similarity matrix is ​​established based on the geodesic distance. The similarity values ​​of the point-to-point similarity matrix are calculated using a Gaussian kernel function, with the kernel width parameter adaptively adjusted to 0.5 times the median of the geodesic distance. The local affine transformation estimation module divides the preprocessed point cloud dataset into 64 cubic regions according to spatial location. Each cubic region has a side length of 200 mm, and adjacent cubic regions overlap by 50 mm. Each cube region's affine matrix contains 12 parameters, estimated using weighted least squares, with weights inversely proportional to the distance from a point to the cube region's center. The weight coefficient for the graph Laplacian regularization constraint is set to 0.1, and adjacent cube regions are defined by sharing a boundary. Within the Log-Euclidean framework, the rotation matrix is ​​mapped to an antisymmetric matrix space via matrix logarithm. The optimization process performs gradient descent in the antisymmetric matrix space, and the updated antisymmetric matrix is ​​mapped back to the rotation matrix via matrix exponentiation. The alternating optimization strategy iterates 20 times. In each iteration, nearest neighbor search is used to update the point correspondences, and the conjugate gradient method is used to solve the regularized least squares problem when updating the transformation parameters.

[0039] The technical advantages of the Riemannian-based non-rigid point cloud registration algorithm are as follows: It fundamentally solves the problem that traditional rigid registration cannot handle GIS equipment deformation, adapting to the deformation patterns of GIS equipment under thermal stress and mechanical loads. Riemannian geometry provides a theoretically rigorous mathematical framework; the intrinsic properties of geodesic distances make the registration results insensitive to coordinate system selection and overall rotation and translation, improving the robustness of the Riemannian-based non-rigid point cloud registration algorithm. The Log-Euclidean framework ensures the numerical stability of the rotation matrix optimization process, avoiding the gimbaling problem of traditional Euler angle representation and the normalization constraints of quaternion representation. The local affine transformation model controls model complexity while ensuring registration accuracy, reducing the number of affine matrix parameters and lowering the risk of overfitting compared to global non-rigid transformations. Graph Laplacian regularization constraints utilize prior knowledge of the spatial continuity of the preprocessed point cloud dataset to constrain the smoothness of the transformation field, preventing severe local distortions and ensuring physically reasonable registration results. The non-rigid point cloud registration algorithm based on Riemannian manifolds can accurately register point clouds of GIS equipment that have been in operation for many years, eliminating registration errors caused by deformation, enabling reliable fusion of data from multiple scans, and supporting long-term monitoring of GIS equipment deformation. Improved registration accuracy directly enhances the quality of the complete GIS equipment digital model, provides accurate spatial references for X-ray inspection, reduces detection omissions due to positioning errors, increases defect detection rates, and ensures the safe and stable operation of the power system.

[0040] The specific calculation steps of the attitude compensation adjustment function are as follows: extract the corrected contact coordinates as the target detection point coordinates, and extract the current coordinates of the X-ray machine focal point as the X-ray source coordinates; calculate the three-dimensional vector difference between the target detection point coordinates and the X-ray source coordinates as the ideal X-ray direction vector; extract the current coordinates of the X-ray machine detector center, and calculate the three-dimensional vector difference between the current coordinates of the X-ray machine detector center and the X-ray source coordinates as the actual X-ray direction vector; calculate the cosine of the angle between the ideal X-ray direction vector and the actual X-ray direction vector as the direction deviation metric; calculate the absolute value of the difference between the magnitude of the ideal X-ray direction vector and the magnitude of the actual X-ray direction vector as the distance deviation value; subtract 1 from the direction deviation metric, take the absolute value, and divide by 2 to obtain the normalized direction deviation value; divide the distance deviation value by the magnitude of the ideal X-ray direction vector to obtain the normalized distance deviation value; calculate the normalized direction deviation value. The weighted sum of the normalized distance deviation value and the average deviation value is used as the comprehensive deviation index value, with weighting coefficients of 0.6 and 0.4, respectively. The compensation adjustment strategy is determined according to the interval to which the comprehensive deviation index value belongs. When the comprehensive deviation index value a∈[0, 0.02), the current pose of the X-ray machine is kept unchanged and the detection is performed directly. When a∈[0.02, 0.05), only the pitch and azimuth angles of the X-ray machine are adjusted to make the actual ray direction vector coincide with the ideal ray direction vector. When a∈[0.05, 0.15), the X-ray machine is first translated to move the ray source coordinates along the ideal ray direction vector to compensate for 50% of the distance deviation value, and then the pitch and azimuth angles of the X-ray machine are adjusted. When a≥0.15, the spatial position of the X-ray machine is replanned, and the ray source coordinates are moved to a position 1000mm away from the target detection point coordinates and the angle between the ray direction vector and the ideal ray direction vector is less than 5 degrees.

[0041] The algorithm principle based on graph theory shortest path in the multi-view scanning path planning is as follows: The surface geometry of GIS equipment contains numerous occlusions and blind spots, making it impossible to acquire complete point cloud data from a single viewpoint. Therefore, scanning from multiple views and optimizing the scanning order are necessary. The surface of the GIS equipment is discretized into a uniformly distributed set of sampling points, each representing a potential scanning viewpoint. From this scanning viewpoint, the LiDAR covers a local area of ​​the GIS equipment. A directed graph model is constructed, with sampling points serving as nodes. The reachability and scanning quality between nodes constitute directed edges. The weight of the directed edges comprehensively considers the scanner's movement distance, point cloud acquisition quality, and viewpoint switching time. The point cloud acquisition quality score is calculated based on the angle between the laser incident angle and the surface normal vector. The closer the laser incident angle is to perpendicularity, the stronger the echo intensity, the higher the point cloud accuracy, and the higher the point cloud acquisition quality score. Topological constraints are established considering the occlusion relationships of the GIS equipment. If a key component cannot be observed from a certain node's viewpoint, that node is not associated with the detection task of the corresponding key component. An improved Dijkstra's algorithm is employed to find the shortest path covering all critical components. Starting from the initial node, the algorithm progressively expands to the nearest unvisited node until all critical components are covered by at least one node. A strategy from the Traveling Salesman Problem (TSP) is introduced to optimize the multi-view sequence, treating selected nodes as mandatory cities and minimizing the total path length to all nodes. A greedy strategy is used to construct the initial solution, selecting the next node from the current node with the shortest travel distance and covering the most new critical components. Dynamic programming is used to optimize local sub-paths; for a node subsequence of length 5, all permutations are enumerated, and the permutation with the shortest total travel distance replaces the original node subsequence.

[0042] The specific implementation of the multi-view scanning path planning algorithm based on graph theory shortest path in the scheme is as follows: Before step S01, the outer surface of the GIS equipment is extracted according to the CAD model or preliminary point cloud data of the GIS equipment. A sampling point is placed every 200mm along the normal of the outer surface of the GIS equipment, generating a total of 300 candidate scanning viewpoint nodes. The visible area of ​​each candidate scanning viewpoint node is calculated, and the ray casting method is used to determine whether the laser emitted from the candidate scanning viewpoint node is blocked by other parts of the GIS equipment. The positions of the conductor contact, the moving contact of the disconnect switch, and the spring are marked as 20 key components, and each key component is required to be covered by at least 3 different viewpoint nodes. The edge weight from candidate scanning viewpoint node i to candidate scanning viewpoint node j is calculated. The edge weight formula is as follows: the Euclidean distance between the coordinates of candidate scanning viewpoint node i and the coordinates of candidate scanning viewpoint node j is divided by the maximum moving speed of the scanner, plus the view switching time, then the average acquisition quality score of candidate scanning viewpoint node j for the key component is subtracted by the maximum quality score, and multiplied by the quality weight coefficient, which is set to 5. The point cloud acquisition quality score is calculated based on the laser incident angle. The point cloud acquisition quality score is equal to the square of the cosine of the laser incident angle. When the laser incident angle is greater than 60 degrees, the point cloud acquisition quality score is set to zero. A directed graph model containing 300 candidate scanning viewpoint nodes is constructed. An improved Dijkstra's algorithm is used to search for the shortest path, with the starting node set as the candidate scanning viewpoint node corresponding to the initial position of the scanner. During the search, a set of covered key components is maintained. The path search is completed when the set of covered key components contains all 20 key components. The Traveling Salesman Problem (TSP) is applied to optimize the candidate scanning viewpoint node sequence obtained by the search. A 2-opt local search algorithm is used, where the positions of any two candidate scanning viewpoint nodes in the candidate scanning viewpoint node sequence are swapped. If the total path length is reduced after the swap, the swap is accepted. This process is iterated 100 times until there is no room for improvement.

[0043] The technical advantages of the graph theory-based shortest path multi-view scanning path planning algorithm are as follows: This algorithm transforms the 3D spatial scanning path planning problem into a graph theory optimization problem, achieving efficient solutions using mature graph algorithms. By discretizing the GIS equipment surface into a finite set of candidate scanning viewpoint nodes, continuous spatial search is transformed into discrete combinatorial optimization, significantly reducing computational complexity. Edge weight design comprehensively considers travel distance, point cloud acquisition quality scores, and time costs, ensuring the optimization objective aligns with actual detection needs. Visibility analysis and occlusion constraints guarantee that selected candidate scanning viewpoint nodes can effectively acquire data, avoiding wasted time on ineffective scanning. The improved Dijkstra's algorithm ensures finding the globally optimal path, and the Traveling Salesman Problem (TSP) optimization further reduces the total scanner travel distance, significantly improving on-site operational efficiency. The combination of greedy strategies and dynamic programming balances computational efficiency and optimization quality, with dynamic programming handling local subproblems to prevent getting trapped in local optima. The graph theory-based shortest path multi-view scanning path planning algorithm adapts to the geometry of GIS equipment, automatically generating scanning schemes with complete coverage and optimal paths, reducing the subjectivity and blindness of manual planning. The reduction in scanning time decreases the impact of on-site operations on power system operation, improves the economy and feasibility of inspection, promotes the application of X-ray inspection technology in operating GIS equipment, and enhances the level of intelligent operation and maintenance of power equipment.

[0044] The initial point cloud dataset is a set of original three-dimensional coordinate points obtained by LiDAR scanning a GIS device from multiple perspectives. Each point contains spatial coordinates and echo intensity information. The pose deviation vector is a six-dimensional vector describing the translational and rotational differences between the current pose of the X-ray machine and the ideal detection pose. The pose deviation vector contains three translational components and three rotational components. The energy-thickness database is a pre-established parameter table relating the equivalent thickness of the material to the X-ray tube voltage and exposure time. The energy-thickness database is calibrated through numerous experiments to determine the exposure parameters required to obtain clear images at different thicknesses. The local density value is the ratio of the number of points in the neighborhood of each point in the preprocessed point cloud dataset to the neighborhood volume. The material thickness distribution is the spatial distribution information of the thickness values ​​and material types of each material layer along the X-ray penetration path.

[0045] Optionally, the present invention also provides a GIS equipment X-ray inspection system with automatic spatial pose compensation implemented by a computer. The computer is provided with a readable storage medium, which stores program instructions. When the program instructions are run in the computer, they execute the above-described method for GIS equipment X-ray inspection with automatic spatial pose compensation.

[0046] The specific implementation methods of the above steps are described in detail below.

[0047] The specific implementation of step S01 is as follows: First, LiDARs are placed around the GIS equipment in sequence according to the candidate scanning view node sequence generated by the multi-view scanning path planning algorithm based on graph theory shortest path. The LiDARs emit high-frequency laser pulses to the surface of the GIS equipment using the pulse time-of-flight ranging principle. After the laser pulses hit the surface of the GIS equipment, they are reflected back to the LiDAR receiver. The distance value of the LiDAR to the surface point is obtained by calculating the time difference from the emission to the reception of the laser pulse. Combined with the azimuth and elevation angles recorded by the precision angle encoder built into the LiDAR, the three-dimensional coordinates of the surface point in the LiDAR's own coordinate system are calculated. Multiple scans are accumulated to form an initial point cloud dataset containing hundreds of millions of points. Simultaneously, a mobile station receiver of the RTK differential positioning system is installed on the lidar and X-ray machine. The mobile station receiver of the RTK differential positioning system receives differential correction signals and satellite navigation signals sent by the base station. The three-dimensional coordinates of the mobile station receiver are calculated in real time through carrier phase differential technology, and the positioning accuracy reaches the centimeter level. The position and attitude of the lidar at each candidate scanning view node are recorded as spatial pose information. The spatial pose information includes two parts: position coordinates and rotation matrix, which provides an absolute spatial reference for subsequent point cloud registration and avoids registration errors caused by inconsistencies in coordinate systems.

[0048] The specific implementation of step S02 is as follows: First, each point in the initial point cloud dataset is traversed. A k-nearest neighbor search algorithm is used to find neighboring points within a spherical neighborhood centered on the given point. The neighborhood radius is adaptively determined based on the overall density of the point cloud. The average distance between the given point and all neighboring points is calculated as the local density value. Simultaneously, the principal curvature and secondary curvature are fitted to the surface formed by the neighboring points, and the difference between the two curvatures is used as the curvature change value. The curvature change value reflects the geometric complexity of the region where the point is located. Then, the initial point cloud dataset is divided into high-density and low-density regions based on the median of the local density values. Statistical analysis is performed using 5 neighboring points in the high-density region and 20 neighboring points in the low-density region. This division strategy balances computational efficiency and filtering accuracy. Next, the average distance and standard deviation to neighboring points are calculated for each point. The dynamic threshold for identifying noise points is dynamically adjusted based on the curvature change value. When the curvature change value is greater than the curvature threshold of 0.15... Points identified as edge features are identified by increasing their dynamic threshold to the average distance plus three standard deviations to prevent accidental deletion. Points with curvature changes less than the curvature threshold are identified as points in flat areas by decreasing their dynamic threshold to the average distance plus 1.5 standard deviations to enhance noise filtering in flat areas. Finally, all points with average distances exceeding the dynamic threshold are deleted, and the remaining valid point cloud is used to form a preprocessed point cloud dataset. This adaptive multi-scale statistical filtering effectively eliminates abnormal points caused by strong electromagnetic interference and metal reflections at the substation site, reducing the proportion of noise points from 15% to 25% to below 2%, providing high-quality input data for subsequent modeling.

[0049] The specific implementation of step S03 is as follows: The preprocessed point cloud dataset and spatial pose information are input into the input layer of the spatial registration optimization model. The input layer normalizes the preprocessed point cloud dataset so that all coordinate values ​​are within the range of 0 to 1. The first feature extraction module adopts a hierarchical point cloud feature learning method based on the PointNet++ architecture. Geometric features are abstracted layer by layer through a three-level structure: sampling layer, grouping layer, and feature extraction layer. The sampling layer uses the farthest point sampling algorithm to select representative points to reduce the amount of data. The grouping layer constructs a local neighborhood around each representative point. The feature extraction layer uses a multilayer perceptron to extract the geometric feature vectors of the local neighborhoods. The sparse coding layer receives the local geometric feature vectors and linearly combines them with the 512 atoms of an overcomplete dictionary to find sparse coefficients that minimize the number of non-zero elements and the reconstruction error. This process is implemented through a deep unfolded network layer. The deep unfolded network layer unfolds the eight iterations of the fast iterative shrinking thresholding algorithm FISTA into eight learnable network layers. Each layer first performs gradient descent to update the sparse coefficients, and then applies a soft thresholding operator to shrink the sparse coefficients. The threshold parameters of the soft thresholding operator are learned through backpropagation. This unfolding strategy makes the iterative solution process of sparse coding differentiable and trainable. The second feature extraction module performs deep convolution on the sparse features output by the deep unfolded network layer to extract high-level semantic information for subsequent registration. The Riemannian manifold registration module embeds the high-level semantic information into the Riemannian manifold space using the feature function of the Laplace-Beltrami operator. In this space, the Dijkstra shortest path algorithm is used to calculate the geodesic distance between any two points. A point-to-point similarity matrix is ​​established based on the geodesic distance. The geodesic distance is more suitable for the intrinsic geometric characteristics of the curved surfaces of GIS equipment than Euclidean distance. The local affine transformation estimation module divides the preprocessed point cloud dataset into 64 cubic regions with sides of 200mm and an overlap of 50mm. For each cubic region, it estimates an affine matrix containing 12 parameters, describing rotation, scaling, and translation transformations. Graph Laplacian regularization constrains the consistency of the affine matrices between adjacent cubic regions, with a weight coefficient set to 0.1 to prevent discontinuities caused by excessive differences in transformation between adjacent regions. The global optimization layer, within the Log-Euclidean framework, maps the rotation matrix to the Lie algebra tangent space for optimization, avoiding the singularity problem of direct optimization on the Lie group manifold. Iteratively solving for the optimal transformation field is achieved using the conjugate gradient method. The output layer transforms and aligns the preprocessed point cloud dataset based on the optimal transformation field, extracting the aligned conductor contact position coordinates as the conductor contact spatial coordinates. This generates a complete GIS equipment digital model containing the conductor contact spatial coordinates, accurately reflecting the 3D geometry of the GIS equipment and the spatial positions of its key internal components.

[0050] The specific implementation of step S04 is as follows: Based on the spatial coordinates of the conductor contact and the focal coordinates of the X-ray machine in the complete GIS equipment digital model, a ray tracing algorithm is used to calculate the X-ray penetration path from the focal point to the conductor contact. The thickness values ​​and material types of all material layers along the penetration path are extracted. Material types include aluminum alloy shell, copper conductor, epoxy resin insulator, and sulfur hexafluoride gas. The linear attenuation coefficient corresponding to each material type is queried from the material attenuation coefficient database. The exposure parameter prediction function first normalizes the thickness value of each material layer by dividing it by the reference aluminum thickness of 10mm, then multiplies it by the ratio of the linear attenuation coefficient of the material to the linear attenuation coefficient of the aluminum material. The calculation results of all material layers are summed to obtain the total equivalent thickness. This total equivalent thickness unifies the X-ray attenuation capabilities of different materials to the aluminum material equivalent thickness standard. Based on the total equivalent thickness, the initial tube voltage value and initial exposure time value are queried from the energy-thickness database through linear interpolation. The energy-thickness database stores a large number of experimentally calibrated exposure parameter combinations. Next, the distance from the geometric center of the penetration path to the detector is calculated. Based on the physical law that X-ray intensity is inversely proportional to the square of the distance, the initial exposure time is multiplied by the distance and divided by the square of the standard distance of 1000 mm to obtain the corrected exposure time. This correction compensates for the influence of distance variation on X-ray intensity. The corrected exposure time and initial tube voltage are output as the initial exposure parameter set, providing accurate exposure control parameters for the X-ray imaging detection in step S06, avoiding image blurring due to underexposure or image saturation due to overexposure.

[0051] The specific implementation of step S05 is as follows: 16 triaxial electromagnetic field sensors are uniformly arranged along the axial and radial directions on the surface of the GIS equipment casing. The triaxial electromagnetic field sensors are spaced 500mm apart to form an array. This array arrangement optimizes the sampling coverage of the electromagnetic field spatial distribution. The triaxial electromagnetic field sensors collect electromagnetic field signals with a frequency range of 50Hz to 1000Hz. The sampling rate is set to 10kHz to satisfy the Nyquist sampling theorem, and the acquisition time is 10s to obtain sufficient statistical samples. The collected electromagnetic field signals are subjected to a Fast Fourier Transform to extract the amplitude and phase information of the dominant frequency components. Each triaxial electromagnetic field sensor extracts 4 dominant frequency components to form a 64-dimensional measured electromagnetic field feature vector. The feature vector contains the frequency domain characteristics of the external electromagnetic field distribution data. The cosine distance between the 64-dimensional measured electromagnetic field feature vector and the feature vectors of all simulated data is calculated from a pre-established electromagnetic field feature library. The cosine distance measures the cosine value of the angle between two vectors, reflecting the degree of similarity. The 10 sets of simulated data with the smallest cosine distance are selected as candidates. The electromagnetic field inverse problem inversion algorithm uses the conductor positions corresponding to the 10 sets of simulated data as the initial solution space, constructs a Tikhonov regularized objective function, which includes a data fitting term and a smoothness regularization term. The regularization parameter is set to 0.001 to balance the fitting accuracy and the stability of the solution. The conjugate gradient method is used to iteratively solve the objective function. The conjugate gradient method utilizes conjugate directions to accelerate convergence. The iteration termination condition is that the change in the objective function is less than 0.0001 or the number of iterations exceeds 100, obtaining the optimal conductor position estimate. A Bayesian inference framework is introduced, using a Gaussian distribution with a mean and standard deviation of 5 mm as the prior probability distribution for the optimal conductor position estimate. The Metropolis-Hastings algorithm performs 10,000 Markov chain samplings. This algorithm explores the posterior probability distribution through an acceptance-rejection mechanism to calculate the posterior probability distribution of the conductor position. The mean of the posterior probability distribution is extracted as the probability distribution estimate, and the standard deviation of the posterior probability distribution is extracted as the positioning uncertainty index. When the positioning uncertainty index is greater than 10 mm, the positioning reliability is deemed insufficient, and the number of triaxial electromagnetic field sensors is increased or their arrangement is adjusted to re-collect external electromagnetic field distribution data. The spatial coordinates of the conductor contact in the complete GIS equipment digital model are corrected based on the probability distribution estimate, and the corrected coordinates are recorded as the corrected contact coordinates. This correction combines the advantages of both laser point cloud modeling and electromagnetic field inversion methods to improve the positioning accuracy and reliability of the conductor contact.

[0052] The specific implementation of step S06 is as follows: The attitude compensation adjustment function first extracts the corrected contact coordinates as the target detection point coordinates, extracts the current coordinates of the X-ray machine focal point as the X-ray source coordinates, and calculates the three-dimensional vector difference between the target detection point coordinates and the X-ray source coordinates as the ideal X-ray direction vector, which points in the direction where the X-rays should irradiate. The current coordinates of the X-ray machine detector center are extracted, and the three-dimensional vector difference between the current coordinates of the X-ray machine detector center and the X-ray source coordinates is calculated as the actual X-ray direction vector, which reflects the current actual orientation of the X-ray machine. The dot product of the ideal X-ray direction vector and the actual X-ray direction vector is divided by the product of the magnitudes of the two vectors to obtain the cosine of the included angle as a direction deviation metric. A direction deviation metric close to 1 indicates a consistent direction, while a deviation from 1 indicates an angular deviation. The absolute value of the difference between the magnitude of the ideal X-ray direction vector and the magnitude of the actual X-ray direction vector is calculated as the distance deviation value, which reflects the deviation in distance between the X-ray machine and the target detection point. The normalized direction deviation value is obtained by subtracting 1 from the direction deviation metric, taking the absolute value, and then dividing by 2. The normalized distance deviation value is obtained by dividing the distance deviation value by the magnitude of the ideal ray direction vector. This normalization process unifies deviations of different dimensions to the range of 0 to 1. The sum of the normalized direction deviation value multiplied by a weighting coefficient of 0.6 and the normalized distance deviation value multiplied by a weighting coefficient of 0.4 is calculated as the comprehensive deviation index value. The weighting coefficient allocation reflects the more significant impact of direction deviation on detection quality. Different compensation adjustment strategies are implemented according to the interval to which the comprehensive deviation index value belongs. When the comprehensive deviation index value a∈[0, 0.02), the X-ray machine pose is determined to meet the detection requirements, and the current X-ray machine pose remains unchanged, directly performing X-ray imaging detection based on the initial exposure parameter set. When a∈[0.02, 0.05), the direction deviation is determined to be small but requires adjustment. The pitch and azimuth drive motors of the X-ray machine are controlled to rotate, making the actual ray direction vector coincide with the ideal ray direction vector. The rotation angle is calculated based on the cross product and dot product of the two vectors. When a∈[0.05, 0.15), a significant deviation is identified, requiring comprehensive adjustment. First, the three-axis translation drive mechanism of the X-ray machine is controlled to move the X-ray source coordinates along the ideal X-ray direction vector to compensate for 50% of the distance deviation. Then, the pitch and azimuth angles of the X-ray machine are adjusted. This step-by-step adjustment strategy improves the compensation accuracy. When a≥0.15, the deviation is identified as too large, requiring replanning. The X-ray machine is controlled to move to a new position 1000mm away from the target detection point coordinates, with the angle between the X-ray direction vector and the ideal X-ray direction vector less than 5 degrees. This new position is obtained through inverse kinematics. After automatic pose compensation, X-ray imaging detection is performed according to the initial exposure parameter set to obtain a clear X-ray image of the conductor contacts inside the GIS equipment, which is used to determine the contact insertion depth, alignment, and mechanical defects.

[0053] It should be noted that the key technical ideas of this invention include a point cloud registration method combining dictionary learning based on sparse coding and deep unfolded networks, an internal structure-assisted localization method based on electromagnetic field inverse problem inversion, and a non-rigid point cloud registration method based on Riemannian manifolds. The point cloud registration method combining dictionary learning based on sparse coding and deep unfolded networks has the advantage over traditional iterative nearest-point algorithms in that traditional methods assume the point cloud undergoes rigid transformation and are sensitive to initial positions, easily getting trapped in local optima. In contrast, this method projects high-dimensional point clouds into a low-dimensional sparse space using sparse coding to extract intrinsic structural features. The deep unfolded network expands the iterative optimization algorithm into learnable network layers for end-to-end training, enabling the registration process to possess both the interpretability of traditional optimization algorithms and the powerful expressive capabilities of deep learning. It exhibits strong robustness to noise and deformation, significantly improving registration accuracy and speed. The internal structure-assisted positioning method based on electromagnetic field inversion has the advantage over traditional laser point cloud modeling. Laser radar can only scan the external surface and cannot directly observe the position of the internal conductor. In contrast, the proposed method utilizes the external electromagnetic field distribution generated by the internal current during the operation of GIS equipment. By solving the inverse problem of Maxwell's equations, the position of the internal conductor is inferred, providing a positioning verification method independent of point clouds. The two methods corroborate each other, significantly improving positioning reliability. At the same time, the probability distribution estimate obtained by electromagnetic field inversion guides the X-ray machine to prioritize scanning high-probability areas, avoiding full-coverage scanning and greatly reducing detection time. The advantage of the non-rigid point cloud registration method based on Riemannian manifolds over traditional rigid registration lies in the fact that GIS equipment undergoes non-rigid deformation due to thermal stress and mechanical loads during long-term operation. Traditional rigid registration cannot handle the registration errors caused by deformation. However, the proposed method utilizes the geodesic distance of Riemannian manifold geometry to measure the similarity of point pairs. The intrinsic property of geodesic distance is invariant to local deformation. Combined with a local affine transformation model and graph Laplace regularization constraints, the non-rigid transformation field is stably solved within the Log-Euclidean framework. Accurate registration of the deformed point cloud supports long-term monitoring of equipment deformation. The improved registration accuracy directly improves the quality of the complete GIS equipment digital model, providing a precise spatial reference for X-ray detection. The synergistic effect of the three key technical approaches lies in the fact that a point cloud registration method combining dictionary learning and deep unfolding networks based on sparse coding, and a non-rigid point cloud registration method based on Riemannian manifolds, jointly construct a high-precision and complete digital model of GIS equipment. An internal structure-assisted positioning method based on electromagnetic field inversion corrects the spatial coordinates of conductor contacts in the model. The combination of these three approaches achieves multi-source information fusion of laser point clouds, electromagnetic field inversion, and non-rigid registration, significantly improving the accuracy and reliability of positioning key components inside GIS equipment. This provides an accurate spatial reference for automatic X-ray machine pose compensation. The entire technical system breaks through the limitations of traditional X-ray inspection relying on manual experience and blind scanning, realizing intelligent and accurate inspection based on digital models, greatly improving inspection efficiency and defect detection rate, and ensuring the safe and stable operation of the power system.

[0054] Specifically, the principle of this invention is as follows: This invention can solve the technical problem of automatic compensation for detected pose. Its principle lies in combining laser point cloud spatial registration and electromagnetic field physical inversion to establish a mapping relationship from external measurement to internal structural positioning. The Riemannian manifold registration module in the spatial registration optimization model utilizes the intrinsic properties of manifold geometry to describe the non-rigid deformation of GIS equipment. The similarity of geodesic distance measurement point pairs is unaffected by overall rotation and translation. The local affine transformation model ensures the smooth continuity of the deformation field through graph Laplace regularization constraints, making the registration results conform to physical constraints. The deep unfolded network unfolds the iterative solution process of sparse coding into a fixed-layer neural network. Soft threshold parameters and dictionary atoms are learned through backpropagation, maintaining the interpretability of the optimization algorithm while improving computational efficiency, and adapting to noise generated by electromagnetic interference and metal reflections in the field. The electromagnetic field inverse problem inversion algorithm is based on the physical model of Maxwell's equations. Through Tikhonov regularization and a Bayesian inference framework, the ill-posed problem is transformed into a constrained optimization problem, obtaining the posterior probability distribution of the internal conductor's position. This quantifies the positioning uncertainty and provides a confidence assessment for pose compensation. The pose compensation adjustment function adjusts the X-ray machine's pose in stages according to the comprehensive deviation index value. Small deviations only adjust the angle, while large deviations require replanning the spatial position, achieving automated closed-loop control that meets the accuracy and efficiency requirements of actual detection.

[0055] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0056] The specific implementation of step S01 is as follows: When using LiDAR to perform multi-view scanning of the GIS equipment, the scanning view sequence is first determined according to a multi-view scanning path planning algorithm based on graph theory and shortest path. A sampling point is placed every 200mm along the normal direction of the outer surface of the GIS equipment, generating a total of 300 candidate scanning view nodes. The visible area of ​​each candidate scanning view node is calculated, and the positions of the conductor contact, the moving contact of the isolating switch, and the spring are marked as 20 key components. The Euclidean distance calculation formula is expressed as follows: ; In the formula, Candidate scanning view nodes To candidate scan view node Euclidean distance, in mm; Candidate scanning view nodes The three-dimensional coordinates are in mm; Candidate scanning view nodes The three-dimensional coordinates are given in mm. Calculate the candidate scan viewpoint nodes. To candidate scan view node The edge weights are expressed by the following formula: ; In the formula, Candidate scanning view nodes To candidate scan view node The edge weights, in units of seconds; This is the maximum scanner movement speed, in mm / s, with a default value of 200 mm / s. This is the time for the viewpoint to switch, in seconds (s), with a default value of 5 seconds. Candidate scanning view nodes The average acquisition quality score for key components is a dimensionless quantity. This is the maximum quality score; the default value is 1. This is the quality weighting coefficient, measured in seconds (s), with a default value of 5 seconds. The point cloud acquisition quality score is calculated based on the laser incident angle, as expressed in the following formula: ; In the formula, The laser incident angle is expressed in rad. When greater than 1.047 rad Zeroing. The lidar moves sequentially to each viewpoint position according to the optimized candidate scanning viewpoint node sequence, performs a 3D scan at each viewpoint position, and collects point cloud data to form an initial point cloud dataset. The RTK differential positioning system provides the rover station mounted on the lidar with centimeter-level accuracy in 3D coordinates by receiving differential signals from the base station and satellites, and records the spatial pose information of the lidar at each scanning viewpoint.

[0057] The specific implementation of step S02 is as follows: when performing adaptive multi-scale statistical filtering on the initial point cloud dataset, the local density value and curvature change value of each point in the preprocessed point cloud dataset are first calculated. The formula for calculating the local density value is as follows: ; In the formula, For point Local density values, in units of 1 / ; For point The number of points within its neighborhood is a dimensionless quantity; The volume of the neighborhood is expressed in units of 1. , defaults to The formula for calculating the curvature change is as follows: ; In the formula, For point The change in curvature, expressed in rad / m; Points The minimum, median, and maximum eigenvalues ​​of the neighborhood covariance matrix, in rad / m. The covariance matrix passes through points... The spatial coordinates of its neighboring points are calculated. The point cloud data is divided into high-density and low-density regions based on local density values, with the median local density value of all points serving as the dividing threshold. Statistical analysis is performed on high-density regions using a small-scale neighborhood with a radius of 10 mm, and on low-density regions using a large-scale neighborhood with a radius of 30 mm. The formula for calculating the average distance is as follows: ; In the formula, For point The average distance to its neighboring points is a dimensionless quantity; For point The three-dimensional coordinate vector, in mm; For point Within the neighborhood The three-dimensional coordinate vector of each point is in mm. Point With point The Euclidean distance between them is calculated using the following formula: The unit is mm, where For point The three coordinate components, For point The three coordinate components. The formula for calculating the standard deviation is as follows: ; In the formula, For point The standard deviation of the neighborhood distance is a dimensionless quantity. The formula for calculating the dynamic threshold is as follows: ; In the formula, For point The dynamic threshold, in mm; This is the standard deviation factor, which defaults to 2; Let be the curvature adjustment factor, which is a dimensionless quantity. The formula for calculating the curvature adjustment factor is as follows: ,when ; ,when ; In the formula, The curvature threshold is set to 0.15 rad / m. This represents the average distance from a point to its neighboring points. Exceeding the dynamic threshold If a point is identified as a noise point, all data identified as noise points are deleted, and the valid point cloud is retained to form a preprocessed point cloud dataset.

[0058] The specific implementation method of step S03 is the same as described above, and will not be repeated in detail here.

[0059] The specific implementation of step S04 is as follows: When calculating the material thickness distribution along the X-ray penetration path based on the complete GIS equipment digital model, the thickness and density values ​​of all material layers along the X-ray penetration path are extracted, and the corresponding linear attenuation coefficient is retrieved from the material attenuation coefficient database according to the material type. The formula for calculating the total equivalent thickness is expressed as follows: ; In the formula, The total equivalent thickness is a dimensionless quantity. The number of material layers; For the first The thickness of the layer material, in mm; The reference thickness for aluminum is 10mm. For the first The linear attenuation coefficient of the layer material, in units of ; The linear attenuation coefficient for aluminum is 0.43. The initial tube voltage value is retrieved from the energy-thickness database based on the total equivalent thickness. and initial exposure time value ,in The unit is kV. The unit is seconds (s). Calculate the distance from the geometric center of the penetration path to the detector. The unit is mm. The formula for calculating the corrected exposure time value is as follows: ; In the formula, This is the corrected exposure time value, in seconds. The standard distance is 1000mm. Output and As the initial exposure parameter set.

[0060] The specific implementation of step S05 is as follows: When collecting external electromagnetic field distribution data of the GIS equipment, 16 triaxial electromagnetic field sensors are arranged axially and radially on the surface of the GIS equipment casing, with a spacing of 500mm between the sensors. The triaxial electromagnetic field sensors collect electromagnetic field signals with a frequency range of 50Hz to 1000Hz, a sampling rate of 10kHz, and a collection time of 10s. The collected signals are subjected to a Fast Fourier Transform to extract the amplitude and phase information of the dominant frequency component, forming a 64-dimensional measured electromagnetic field feature vector. The 10 most similar sets of simulated data are retrieved from a pre-established electromagnetic field feature library, with cosine distance used as the similarity metric. The Tikhonov regularization objective function is expressed as follows: ; In the formula, This is the vector of position parameters for the internal conductors, in mm. This represents the number of electromagnetic field sensors, with a value of 16. For the first The measured electromagnetic field intensity vector of each sensor, in V / m; For a given conductor position Time The simulated electromagnetic field intensity vector of each sensor, in V / m; This is the regularization parameter, with a value of 0.001; This is an estimate of the initial conductor position, in mm. The Euclidean norm of the vector is used. The conjugate gradient method is employed iteratively to solve the optimization problem, with the iteration terminating when the change in the objective function is less than 0.0001 or the number of iterations exceeds 100. After obtaining the optimal conductor position estimate, a Gaussian distribution with a mean of 5 mm and a standard deviation of 5 mm is used as the prior probability distribution. The Metropolis-Hastings algorithm is used for 10,000 Markov chain samplings to calculate the posterior probability distribution of the conductor position, and the mean of the posterior probability distribution is extracted as the probability distribution estimate. The spatial coordinates of the conductor contacts in the complete GIS equipment digital model are corrected based on the probability distribution estimate to obtain the corrected contact coordinates.

[0061] The specific implementation of step S06 is as follows: when calculating the pose deviation vector based on the corrected contact coordinates and the current pose of the X-ray machine, the corrected contact coordinates are extracted as the target detection point coordinates. Extract the current coordinates of the X-ray machine focal spot as the X-ray source coordinates. The formula for calculating the ideal ray direction vector is as follows: ; In the formula, The vector represents the direction of the ideal ray, in mm. The coordinates of the target detection point, i.e., the corrected contact point coordinates, are three-dimensional coordinate vectors in mm. The coordinates of the X-ray source, i.e., the current coordinates of the X-ray machine's focal point, are a three-dimensional coordinate vector in mm. Extract the current coordinates of the X-ray machine detector center. The unit is mm. The formula for calculating the actual ray direction vector is as follows: ; In the formula, This is the actual ray direction vector, in mm; Here are the current coordinates of the X-ray detector center, a three-dimensional coordinate vector in mm. The formula for calculating the orientation deviation metric is as follows: ; In the formula, The angle between the ideal ray direction vector and the actual ray direction vector is expressed in rad. The dot product of vectors is represented by the formula: The unit is ; The vector magnitude is represented by the formula: The unit is mm; The vector magnitude is represented by the formula: The unit is mm; These are the three components of the ideal ray direction vector, in mm; These are the three components of the actual ray direction vector, in mm. The formula for calculating the distance deviation is as follows: ; In the formula, This represents the distance deviation value, in mm. The formula for calculating the normalized direction deviation value is as follows: ; In the formula, This is the normalized directional deviation value, a dimensionless quantity. The formula for calculating the normalized distance deviation value is as follows: ; In the formula, This is the normalized distance deviation value, a dimensionless quantity. The formula for calculating the comprehensive deviation index value is as follows: ; In the formula, The comprehensive deviation index value is a dimensionless quantity. A compensation adjustment strategy is determined based on the range to which the comprehensive deviation index value belongs. The spatial position and X-ray angle of the X-ray machine are automatically adjusted according to the pose deviation vector using the attitude compensation adjustment function. After automatic pose compensation is completed, X-ray imaging detection is performed based on the initial exposure parameter set.

[0062] To better understand and implement this invention, a specific application scenario of the invention is provided below as Example 2: To verify the effectiveness of the invention, technicians set up a test environment and selected a 220kV GIS device that had been operating in a substation for 6 years as the test object. This device recently exhibited abnormal partial discharge signals, requiring X-ray inspection to confirm the contact status of the internal conductor contacts. Technicians first deployed an RTK differential positioning system base station around the GIS device, and installed a rover station on the lidar and X-ray machine. Through calibration, the relative spatial positions of each device were established, unifying the coordinates of all devices to the same geodetic coordinate system, achieving a positioning accuracy of 15mm for the RTK system. Based on a multi-view scanning path planning algorithm using graph theory's shortest path, technicians extracted the outer surface of the GIS device's CAD model and placed sampling points every 200mm along the surface normal, generating a total of 286 candidate scanning viewpoint nodes. Eighteen key components, including conductor contacts, isolating switch moving contacts, and springs, were marked, requiring each key component to be covered by at least three different viewpoints. When calculating the edge weights between nodes, the quality weight coefficient was set to 5. The point cloud acquisition quality score was calculated based on the square of the cosine of the laser incident angle, with the score set to zero when the incident angle was greater than 60 degrees. An improved Dijkstra algorithm was used to search for the shortest path, with the starting node set to the node corresponding to the initial position of the LiDAR. After the search, a scan sequence containing 42 nodes was obtained. The technicians then optimized this sequence using the 2-opt local search algorithm. After 100 iterations, the total path length was reduced from 128m to 96m, and the total scanner movement time was shortened from 38 minutes to 28 minutes.

[0063] Technicians performed multi-view scanning according to the optimized path plan. The LiDAR collected point cloud data at each viewpoint, while the RTK system recorded the LiDAR's 3D coordinates and attitude angles, obtaining a total of [data missing]. The initial point cloud dataset contained numerous noisy and outlier points due to strong electromagnetic interference from GIS equipment and specular reflection from the metal casing. Technicians performed adaptive multi-scale statistical filtering on the dataset. First, they calculated the local density and curvature variation for each point, dividing the point cloud into high-density and low-density regions based on the local density values. Statistical analysis was performed on high-density regions using a small-scale neighborhood with a radius of 50 mm, and on low-density regions using a large-scale neighborhood with a radius of 150 mm. The average distance and standard deviation from each point to its neighbors were calculated. A curvature threshold of 0.15 rad / m was set. For edge feature points with curvature variation values ​​greater than the threshold, the dynamic threshold was increased by 20%, while for points in flat areas with curvature variation values ​​less than the threshold, the dynamic threshold was decreased by 30%. After filtering, points with high curvature variation values ​​were removed. One noise point, retained A preprocessed point cloud dataset is formed from the valid points.

[0064] Technicians input preprocessed point cloud datasets and spatial pose information into a spatial registration optimization model. This model uses the PointNet++ architecture to extract local geometric features. A sparse coding layer maps features to an overcomplete dictionary space with 512 atoms and a sparsity constraint parameter of 0.05. A deep unfolded network layer expands the FISTA algorithm iteration process into eight learnable network layers, each containing a soft threshold operator and a gradient descent update module. The soft threshold parameter is initialized to 0.01. The Riemannian manifold registration module embeds high-level semantic information into the manifold space, using the feature function of the Laplace-Beltrami operator to construct the manifold embedding, extracting the top 10 principal components as manifold coordinates. Technicians construct a k-nearest neighbor graph in the Riemannian manifold space, setting k to 20, and use Dijkstra's shortest path algorithm to calculate geodesic distances. The point-to-point similarity matrix is ​​calculated using a Gaussian kernel function, with the kernel width adaptively adjusted to 0.5 times the median of the geodesic distance. The local affine transformation estimation module divides the preprocessed point cloud dataset into 64 cubic regions with sides of 200mm and an overlap of 50mm between adjacent regions. The affine matrix of each region contains 12 parameters, which are estimated using the weighted least squares method. The weight coefficient of the graph Laplacian regularization constraint term is set to 0.1. The alternating optimization strategy converges after 20 iterations. The spatial registration optimization model outputs a complete GIS equipment digital model, and the spatial coordinate positioning accuracy of the conductor contact reaches 8mm.

[0065] Based on the complete GIS digital model of the equipment, technicians calculated the material thickness distribution along the X-ray penetration path. The penetration path sequentially passes through an aluminum alloy shell (12mm thick), an SF6 gas layer (85mm thick), a copper conductor (18mm thick), an SF6 gas layer (85mm thick), and then back to the aluminum alloy shell (12mm thick). Technicians then retrieved the linear attenuation coefficient from a material attenuation coefficient database based on the material type; the attenuation coefficient for aluminum alloy was 0.228. The copper attenuation coefficient is 0.596. The SF6 gas attenuation coefficient is 0.015. The total equivalent thickness of the penetration path was calculated. The reference thickness of the aluminum material was 10mm, the equivalent thickness of the aluminum alloy layer was 24mm, the equivalent thickness of the copper layer was 47.2mm, and the equivalent thickness of the SF6 gas layer was 5.1mm, for a total equivalent thickness of 76.3mm. Based on this total equivalent thickness, technicians retrieved the initial tube voltage and initial exposure time values ​​from the energy-thickness database, obtaining an initial tube voltage of 320kV and an initial exposure time of 8.5s. The distance from the geometric center point of the penetration path to the detector was calculated to be 1150mm. The standard distance was 1000mm. The exposure time was corrected using the inverse square law of distance, resulting in a corrected exposure time of 11.3s.

[0066] like Figure 3 As shown, technicians arranged 16 triaxial electromagnetic field sensors along the axial and radial directions on the surface of the GIS equipment casing, with a sensor spacing of 500 mm. The sensors acquired electromagnetic field signals with a frequency range of 50 Hz to 1000 Hz, a sampling rate of 10 kHz, and an acquisition time of 10 seconds. The acquired signals were subjected to a Fast Fourier Transform (FFT) to extract the amplitude and phase information of the dominant frequency component, forming a 64-dimensional measured electromagnetic field feature vector. The 10 most similar sets of simulated data were retrieved from a pre-established electromagnetic field feature library, with similarity measured by cosine distance, and a minimum cosine distance of 0.023. Using the conductor positions corresponding to these 10 sets of simulated data as the initial solution space, a Tikhonov regularized objective function was constructed, with the regularization parameter set to 0.001. The conjugate gradient method was used to iteratively solve the optimization problem. After 56 iterations, the change in the objective function was less than 0.0001, yielding the optimal conductor position estimate. Technicians used a Gaussian distribution with a mean and a standard deviation of 5 mm as the prior probability distribution, and performed 10,000 Markov chain samplings using the Metropolis-Hastings algorithm to calculate the posterior probability distribution of the conductor's position. The mean of the posterior probability distribution was used as the probability distribution estimate, and the standard deviation of 6.8 mm was used as the positioning uncertainty index, which is less than the 10 mm threshold and meets the accuracy requirements. Based on the probability distribution estimate, technicians corrected the spatial coordinates of the conductor contacts in the complete GIS equipment digital model. After correction, the contact coordinates were offset by 12 mm in the X-axis direction, 8 mm in the Y-axis direction, and 5 mm in the Z-axis direction relative to the initial coordinates.

[0067] Technicians calculated the pose deviation vector based on the corrected contact coordinates and the current pose of the X-ray machine. The corrected contact coordinates were extracted as the target detection point coordinates, and the current focal coordinates of the X-ray machine were used as the X-ray source coordinates. The ideal and actual X-ray direction vectors were calculated. The cosine of the angle between the ideal and actual X-ray direction vectors was 0.926, the direction deviation metric was 0.926, the distance deviation was 68 mm, and the magnitude of the ideal X-ray direction vector was 1035 mm. The normalized direction deviation was 0.037, the normalized distance deviation was 0.066, and the weighting coefficients were 0.6 and 0.4, respectively, resulting in a comprehensive deviation index of 0.048. Based on the attitude compensation adjustment function, the comprehensive deviation index value was within the interval [0.02, 0.05). Technicians only adjusted the pitch and azimuth angles of the X-ray machine to make the actual X-ray direction vector coincide with the ideal X-ray direction vector, adjusting the pitch angle to 15.6 degrees and the azimuth angle to 8.3 degrees. After automatic pose compensation, technicians performed X-ray imaging based on the corrected exposure time of 11.3s and the initial tube voltage of 320kV to obtain a clear image of the conductor contact status. Figure 4As shown in the image, the contact insertion depth is insufficient. The actual insertion depth is 42mm, while the designed insertion depth is 55mm, resulting in a deviation of 13mm. This confirms the cause of the abnormal partial discharge signal.

[0068] Technicians verified the positioning accuracy of the electromagnetic field inverse problem inversion algorithm by comparing data, as shown in Table 1.

[0069] Table 1. Comparison of Localization Results from Electromagnetic Field Inverse Problem Inversion Algorithm

[0070] Technicians analyzed the registration effect of a non-rigid point cloud registration algorithm based on Riemannian manifolds on a deformable GIS device. After six years of operation, the device was affected by thermal stress, resulting in an 8mm elongation of the outer shell in the axial direction and a 3mm expansion in the radial direction. The traditional ICP rigid registration algorithm assumes that the point cloud as a whole undergoes a rigid transformation, which cannot adapt to local non-rigid deformations. The root mean square of the residual after registration reached 26mm, leading to a contact positioning error exceeding 20mm. The Riemannian manifold registration module utilizes geodesic distance to measure the similarity of point pairs. A geodesic is the shortest path on the manifold, and its length reflects the true distance of points under the intrinsic geometry of the surface, exhibiting invariance to local deformations. The local affine transformation model divides the point cloud into 64 overlapping regions. Each region independently estimates rotation, scaling, and translation parameters. Adjacent regions are constrained for transformation consistency through graph Laplacian regularization, ensuring the smoothness of the overall deformation field. The Log-Euclidean framework maps rotation matrices to the Lie algebra tangent space for linear operations, avoiding the gimbaling problem of traditional Euler angle representation and the normalization constraints of quaternion representation, thus ensuring numerical stability. After registration, the root mean square residual is reduced to 5.2 mm, and the contact positioning error is controlled within 8 mm, significantly improving the quality of the complete GIS equipment digital model.

[0071] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0072] Table 2. Variable Explanation Table (Part 1)

[0073] Table 3. Variable Explanation Table (Part Two)

[0074] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for X-ray inspection of GIS equipment to achieve automatic spatial pose compensation, characterized in that, The process includes the following steps: using LiDAR to scan the GIS equipment from multiple perspectives to obtain an initial point cloud dataset, while simultaneously recording the spatial pose information of the LiDAR using an RTK differential positioning system; performing adaptive multi-scale statistical filtering on the initial point cloud dataset to eliminate noise points caused by electromagnetic interference and metal reflection, thereby obtaining a preprocessed point cloud dataset; inputting the preprocessed point cloud dataset and spatial pose information into a spatial registration optimization model, which outputs a complete digital model of the GIS equipment containing the spatial coordinates of the conductor contacts; The material thickness distribution along the X-ray penetration path is calculated based on the complete GIS equipment digital model. An initial exposure parameter set is determined based on the material thickness distribution using an exposure parameter prediction function. External electromagnetic field distribution data is collected from the GIS equipment. An electromagnetic field inverse problem algorithm is used to obtain an estimate of the probability distribution of the internal conductor's position based on the external electromagnetic field distribution data. The spatial coordinates of the conductor contact in the complete GIS equipment digital model are corrected based on the probability distribution estimate, resulting in the corrected contact coordinates. Based on the corrected contact coordinates and the current pose of the X-ray machine, a pose deviation vector is calculated. An attitude compensation adjustment function automatically adjusts the spatial position and ray angle of the X-ray machine based on the pose deviation vector. After automatic pose compensation, X-ray imaging detection is performed based on the initial exposure parameter set.

2. The method according to claim 1, characterized in that, The RTK differential positioning system provides centimeter-level accuracy three-dimensional coordinates for a mobile station installed on a lidar and X-ray machine by receiving differential signals from a base station and satellites.

3. The method according to claim 2, characterized in that, The adaptive multi-scale statistical filtering process specifically involves calculating the local density value and curvature change value of each point in the preprocessed point cloud dataset, dividing the point cloud data into high-density regions and low-density regions based on the local density values, performing statistical analysis on the high-density regions using a small-scale neighborhood, and performing statistical analysis on the low-density regions using a large-scale neighborhood.

4. The method according to claim 3, characterized in that, The adaptive multi-scale statistical filtering process also includes calculating the average distance and standard deviation of each point to its neighboring points. When the average distance of a point exceeds a dynamic threshold, it is determined to be a noise point. The dynamic threshold is adjusted according to the curvature change value of the region where the point is located.

5. The method according to claim 4, characterized in that, The spatial registration optimization model includes an input layer, a first feature extraction module, a sparse coding layer, a deep unfolded network layer, a second feature extraction module, a Riemannian manifold registration module, a local affine transformation estimation module, a global optimization layer, and an output layer.

6. The method according to claim 5, characterized in that, The deep unfolded network layer unfolds the iterative process of the fast iterative shrinking threshold algorithm FISTA into 8 learnable network layers, each containing a soft thresholding operator and a gradient descent update module.

7. The method according to claim 6, characterized in that, The Riemannian manifold registration module embeds high-level semantic information into the Riemannian manifold space and calculates the point-pair similarity matrix using geodesic distance, which is calculated using Dijkstra's shortest path algorithm.

8. The method according to claim 7, characterized in that, The local affine transformation estimation module constrains the transformation consistency of adjacent regions through graph Laplacian regularization and outputs the transformation parameters of each point cloud block. The global optimization layer uses the Log-Euclidean framework to iteratively solve the optimal transformation field on the Lie group manifold.

9. The method according to claim 8, characterized in that, The spatial registration optimization model is trained end-to-end using the Adam optimizer. The loss function is a weighted sum of registration error loss and sparse regularization loss. The registration error loss is the mean square error between the predicted coordinates and the label data.

10. The method according to claim 9, characterized in that, The calculation of the exposure parameter prediction function includes extracting the thickness and density values ​​of all material layers along the X-ray penetration path, querying the corresponding linear attenuation coefficient according to the material type, calculating the total equivalent thickness of the penetration path, and querying the initial tube voltage and initial exposure time values ​​based on the total equivalent thickness.

Images