Multi-modal sensor fusion-based inspection robot intelligent detection method and system

Abstract

The application relates to the technical field of intelligent detection, and discloses a multi-modal sensor fusion inspection robot intelligent detection method and system. The method comprises the following steps: collecting an original multi-modal image sequence of a target surface through a multi-spectral sensor, and obtaining a clear surface image through adaptive filtering; determining potential defect boundary information based on boundary gray gradient, contrast intensity and optical flow tracking technology; extracting three-dimensional point cloud data and constructing defect geometric morphology when the defect boundary is abnormal, and obtaining defect spatial distribution information; evaluating defect deformation characteristics, depth height and expansion rate in combination with curvature mapping and laser triangulation technology, and determining quantitative physical parameters; and on this basis, generating an alarm, sorting maintenance priorities and calculating future evolution probability, and obtaining a final risk assessment result. The application can realize intelligent detection, dynamic evaluation and risk warning of surface defects of industrial equipment, and significantly improves inspection efficiency.

Description

Technical Field

[0001] This application relates to the field of intelligent detection technology, and in particular to an intelligent detection method and system for inspection robots based on multimodal sensor fusion. Background Technology

[0002] With the development of intelligent operation and maintenance technology for industrial equipment, inspection methods have gradually evolved from manual inspection to automated and intelligent inspection. Early inspection techniques mainly relied on manual visual inspection and simple instrument measurements. Although these methods were low in implementation cost, they suffered from low detection efficiency, strong subjectivity, difficulty in continuous monitoring, and difficulty in detecting minute early defects. Subsequently, machine vision inspection technology based on visible light imaging was applied to equipment surface condition recognition. Through image acquisition and image processing, it can automatically detect surface defects such as cracks, corrosion, and wear, which has improved inspection efficiency and recognition accuracy to a certain extent.

[0003] As application scenarios become increasingly complex, single visible light image detection is gradually revealing its limitations. For example, under conditions of complex lighting, surface reflection, significant material differences, or weak defect features, single-modal images struggle to reliably extract effective defect features, easily leading to missed detections and false detections. To improve detection capabilities, various sensing technologies, such as multispectral imaging, infrared detection, laser ranging, and 3D point cloud reconstruction, are gradually being introduced into the field of inspection robots. Inspection systems are beginning to evolve from single-sensor systems to multi-sensor, multimodal fusion systems. By fusing surface texture information, spectral response information, and spatial depth information obtained from different sensors, the state of the target surface can be more comprehensively characterized, providing a richer data foundation for complex defect identification.

[0004] Building upon this foundation, existing technologies have further evolved from static defect identification to dynamic defect assessment. This means focusing not only on the presence of defects on the equipment surface but also on the morphological evolution, depth changes, expansion rate, and potential risks of those defects. Related technologies typically combine edge detection, feature matching, optical flow analysis, 3D modeling, and trend prediction to locate, measure, and provide early warnings of defects. However, existing intelligent inspection solutions for inspection robots still have the following shortcomings in practical applications: Firstly, multimodal images and depth data are easily affected by noise, perspective changes, and environmental disturbances during acquisition, leading to unstable boundary feature extraction and difficulty in accurately identifying potential defects. Secondly, existing solutions lack sufficient correlation analysis between the 3D geometric morphology of defects, depth degradation trends, and expansion rates, making it difficult to form a complete technical chain from defect detection and evolution analysis to risk assessment and maintenance decisions.

[0005] Especially in long-term industrial equipment operation scenarios, surface defects often exhibit a continuous evolution from minor anomalies to significant damage. Relying solely on single image features or single detection results often fails to reflect the true development trend of defects in a timely manner, and it is also difficult to accurately assess their impact on equipment safety and maintenance priorities. Therefore, how to accurately extract potential defect boundary information by integrating image, depth, and spatial distribution data collected by multimodal sensors in the operation scenario of inspection robots, and further realize defect geometry construction, deformation feature assessment, quantitative physical parameter calculation, and future evolution risk prediction, has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] This application provides an intelligent detection method and system for inspection robots based on multimodal sensor fusion, in order to solve the problems in the prior art where inspection robots have insufficient detection accuracy for equipment surface defects in complex industrial scenarios, weak dynamic evolution recognition capabilities, difficulty in accurately quantifying the three-dimensional morphology and depth changes of defects, and difficulty in conducting risk assessment and maintenance decisions based on multi-source detection results.

[0007] In a first aspect, this application provides an intelligent detection method for inspection robots based on multimodal sensor fusion, the method comprising: S1. Acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image; S2. Based on the analysis of the boundary gray-scale gradient and contrast intensity between pixels in the clear surface image, optical flow tracing technology is used to lock the dynamically changing area and determine the boundary information of potential defects. S3. If the boundary of a potential defect is irregular and the dynamic change of the boundary exceeds the preset change threshold, then extract the three-dimensional point cloud data of the corresponding area, combine the boundary depth information to construct the geometric shape of the defect, and obtain the spatial distribution information of the defect. S4. Based on the spatial distribution information of defects, the curvature mapping technique is used to evaluate the deformation characteristics of the defect region, and the geometric deformation and depth change of the defect region are analyzed to obtain the deformation characteristic evaluation results. S5. Based on the deformation characteristic assessment results, for defect areas whose evolution influence range exceeds the preset range, laser triangulation technology is used to obtain depth and height data, and the defect expansion rate is calculated by combining the defect evolution time span to determine the quantitative physical parameters of the defect. S6. If the quantified physical parameters exceed the preset safety threshold, an alarm signal is generated, and the defect evolution data is integrated to obtain the maintenance priority ranking. S7. Based on quantitative physical parameters and maintenance priority ranking, the evolution rate and depth deterioration data of defects are fused using probabilistic update technology to calculate the future evolution probability of defects and obtain the final risk assessment result.

[0008] Secondly, this application provides an intelligent inspection system for inspection robots based on multimodal sensor fusion, the system comprising: The image acquisition module is used to acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image; The boundary analysis module is used to analyze the boundary gray-scale gradient and contrast intensity between pixels based on a clear surface image, and uses optical flow tracing technology to lock dynamically changing areas and determine the boundary information of potential defects. The point cloud construction module is used to extract the 3D point cloud data of the corresponding area when the boundary of a potential defect is irregular and the dynamic change of the boundary exceeds a preset change threshold. It then combines the boundary depth information to construct the geometric shape of the defect and obtain the spatial distribution information of the defect. The deformation assessment module is used to assess the deformation characteristics of the defect region based on the spatial distribution information of the defect, using curvature mapping technology, and to analyze the geometric deformation and depth changes of the defect region to obtain the deformation characteristic assessment results. The parameter quantification module is used to obtain depth and height data by using laser triangulation technology for defect areas whose evolution influence range exceeds the preset range, based on the deformation characteristic evaluation results. Combined with the evolution time span of the defect, the defect expansion rate is calculated to determine the quantitative physical parameters of the defect. The early warning and sorting module is used to generate alarm signals when the quantified physical parameters exceed the preset safety threshold, integrate defect evolution data, and obtain maintenance priority sorting. The probability prediction module is used to calculate the future evolution probability of defects by combining the evolution rate and depth deterioration data of defects based on the quantified physical parameters and maintenance priorities, and obtain the final risk assessment result.

[0009] The beneficial effects are at least as follows: By acquiring raw multimodal image sequences of the target surface using a multispectral sensor and combining them with adaptive filtering to suppress boundary region noise, the image quality and defect boundary discernibility under complex lighting and surface conditions are improved, thereby enhancing the inspection robot's detection accuracy and stability for minute surface defects. Based on boundary gray-scale gradient, contrast intensity, and optical flow tracing technologies, the dynamic change regions of potential defects are identified and tracked. This allows the robot to reflect the dynamic evolution state of defects based on static feature recognition, improving the ability to identify early anomalies and continuously developing defects. By extracting 3D point cloud data and combining it with boundary depth information to construct the defect geometry, and using curvature mapping and laser triangulation techniques to analyze the spatial distribution, deformation characteristics, depth changes, and expansion rate of defects, the robot can accurately acquire 3D features and quantified physical parameters of defects, enhancing the objective characterization of defect severity. By integrating defect evolution data, maintenance priority ranking, and future evolution probability calculation results, a closed-loop processing system is achieved, from defect detection, evolution analysis, parameter quantification to risk warning and maintenance decision-making. This improves the intelligence level of the inspection robot and the timeliness and targeting of industrial equipment operation and maintenance. By combining multimodal perception, 3D reconstruction, depth measurement and probabilistic assessment, a technology chain for intelligent detection and risk assessment of equipment surface defects has been formed in complex industrial inspection scenarios, providing reliable technical support for predictive maintenance and safe operation of industrial equipment. Attached Figure Description

[0010] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a flowchart of the intelligent detection method for inspection robots based on multimodal sensor fusion, as described in this application. Figure 2 This is a schematic diagram of the curvature distribution of the defect area in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the influence range of curvature distribution and evolution in an embodiment of this application; Figure 4 This is a schematic diagram of the curvature cross-section along the main direction of an embodiment of this application; Figure 5 This is a schematic diagram of the structure of the intelligent inspection system for inspection robots based on multimodal sensor fusion, as described in this application. Detailed Implementation

[0012] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0013] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of the intelligent detection method for inspection robots based on multimodal sensor fusion provided by the present invention. The flowchart specifically includes the following steps: S1. Acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image.

[0014] In one specific embodiment, the process of performing step S1 may specifically include the following steps: The inspection robot uses a multispectral sensor to scan the target surface from multiple angles, acquiring a sequence of original multimodal images covering multiple preset spectral bands. The original multimodal image sequence is subjected to hierarchical analysis, and the multimodal images are separated according to spectral channels to obtain independent image sets under different spectra; For each image in an independent image set, adaptive filtering is performed on the image boundary region to obtain a preliminarily denoised image group; If residual noise exists in the image group, cross-comparison of the multispectral channels of the image group is performed to identify the significant noise region; Local smoothing optimization is performed on areas with significant noise to generate a clear surface image after denoising.

[0015] Specifically, the inspection robot is equipped with a multispectral sensor to scan the target surface from multiple angles. As it moves along the inspection path, it continuously acquires images of the same area in multiple preset spectral bands, covering both visible and near-infrared wavelengths. Images in each band are correlated according to acquisition time, observation angle, and channel index, forming an original multimodal image sequence. To avoid image incomparability caused by robot movement, the acquisition time, robot pose information, and distance measurement information from the sensor to the target surface are recorded synchronously for each frame. The displacement between adjacent frames is controlled to not exceed the spatial sampling interval corresponding to a preset pixel.

[0016] Layered analysis includes radiometric response correction, dark response compensation, outlier replacement, channel separation, and channel registration. Radiometric response correction converts the original responses at different wavelengths to a unified grayscale response benchmark. Dark response compensation eliminates fixed sensor bias, and outlier replacement performs neighborhood median replacement for bad pixels and isolated outliers. After basic correction, the composite image or time-series acquired image is separated into independent image sets under different spectra according to the spectral channel index. If the multispectral sensor uses an array-type filter structure, channel reconstruction is performed based on the filter arrangement; if a time-division exposure structure is used, the images are reconstructed according to the acquisition order and channel number. Considering that multi-angle acquisition and robot posture changes may cause positional deviations between images of different channels, geometric registration is performed on the images of each channel. During registration, translational deviation is first calculated based on the phase correlation method, and then rotation and scale errors are corrected based on the feature point matching method to obtain an independent image set under a unified spatial reference frame.

[0017] Boundary regions are local areas in an image where grayscale changes are significant and may correspond to defect contours. For each image, gradient magnitude and grayscale variance are calculated for local windows. When the mean gradient magnitude of a local window exceeds a preset boundary threshold and the grayscale variance exceeds a preset texture threshold, the corresponding pixel is marked as a boundary candidate pixel, and a boundary region mask is obtained through connected component filtering. Edge-preserving adaptive filtering is applied to pixels within the boundary region mask, while conventional smoothing filtering is applied to non-boundary regions. For example, edge-preserving adaptive filtering can use bilateral filtering or guided filtering. When using bilateral filtering, the unit of the spatial domain parameter is pixels (px), the grayscale domain parameter is consistent with the pixel grayscale unit, and the grayscale domain parameter value within the boundary region is smaller than that in the non-boundary region to reduce edge diffusion caused by cross-boundary averaging. When using guided filtering, the original image is used as the guide map; a smaller window is selected within the boundary region, and a larger window is selected within the non-boundary region, with the regularization term adjusted based on the local variance. Preferably, adaptive median filtering can be performed on local burst noise points first. The noise judgment criteria are that the gray-level variance of the target pixel neighborhood exceeds a preset multiple of the statistical variance of the whole image or the local area. The filtering window starts from a preset starting window size and gradually expands according to a preset step size until the median stability condition is met or the preset upper limit size is reached.

[0018] Residual noise is determined based on the consistency of response across spectral channels at the same spatial location. Local windows are established at corresponding positions of each channel in the image set within a unified spatial reference frame. The mean grayscale value, grayscale dispersion, edge response intensity, and principal direction distribution of each channel within the local window are calculated. When the same local region exhibits continuous edges or textures in most adjacent channels, and the principal direction distribution remains consistent across channels, the local region is identified as a real surface structure region. When the same local region exhibits isolated bright spots, isolated dark spots, discontinuous peaks, or abrupt edge responses only in a single or a few channels, and adjacent channels do not show continuous edges, continuous textures, or consistent principal direction structural responses at the corresponding positions, the local region is identified as a candidate region for residual noise. The candidate regions for residual noise are statistically analyzed. If the number of candidate regions exceeds a preset threshold, or the ratio of the total area of ​​candidate regions to the corresponding detection area exceeds a preset area threshold, or the density of abnormal pixels within the candidate regions exceeds a preset density threshold, residual noise is determined to exist in the image set.

[0019] When identifying significant noise regions, pixel responses at the same spatial location in each channel image are compared point-by-point. This is combined with the edge response map, principal direction distribution, and local mean vector within the corresponding local window to filter out abnormal response locations. For example, the pixel value of the current channel can be compared with the local mean of the corresponding location in adjacent channels. If the deviation exceeds a preset multiple of the local statistical dispersion of adjacent channels, and the connected area of ​​the pixel's region is lower than a preset area threshold, or the region of the pixel does not meet the condition of continuous extension in space, the pixel is marked as a significant noise candidate point. Eight-neighbor connected component merging is performed on the significant noise candidate points, and isolated connected components with areas lower than a preset area threshold are removed to obtain a significant noise region mask.

[0020] Local smoothing optimization is performed only within significant noise regions, with the processing method selected based on the noise's manifestation. If the significant noise region exhibits isolated impulse noise, median replacement or sorting filtering is used for correction. If the significant noise region exhibits speckle noise or small-scale random fluctuations, anisotropic diffusion is used to suppress noise and preserve boundaries. If the significant noise region exhibits abnormally bright areas caused by local strong reflections, restricted grayscale backfilling is performed by combining the response values ​​of adjacent effective bands at corresponding positions and the background values ​​of the current channel's neighborhood to maintain spectral continuity between channels. Preferably, bilateral filtering can also be used for significant noise regions, with the filtering window, spatial domain parameters, and grayscale domain parameters adaptively determined based on the area of ​​the significant noise region and the local grayscale distribution. After local smoothing optimization, the consistency of the boundary principal direction, connectivity length, and edge width before and after correction is checked. When the corrected boundary offset, width change, or connectivity change exceeds the corresponding preset limit, the local smoothing intensity is reduced and reprocessed until the boundary preservation condition is met. Through the above processing, a clear surface image after denoising is generated.

[0021] S2. Based on the analysis of the boundary gray-scale gradient and contrast intensity between pixels in the clear surface image, optical flow tracing technology is used to lock the dynamically changing area and determine the boundary information of potential defects.

[0022] In one specific embodiment, the process of performing step S2 may specifically include the following steps: The clear surface image is divided into blocks, and the local change characteristics of the gray-level gradient of each image block are calculated to obtain gradient distribution data. Based on gradient distribution data, analyze the spatial differences in boundary contrast intensity, identify areas where the difference between the contrast intensity and the surrounding background area exceeds a preset deviation threshold, and determine the dynamic change range. For dynamic variations, optical flow tracing technology is used to analyze the motion trajectory in the image sequence to obtain preliminary boundary continuity information of potential defects; Multi-frame comparison of preliminary boundary continuity information is performed to calculate the fluctuation of the boundary width range and identify irregular boundary regions; The pixel data of irregular boundary regions are locally magnified to obtain the boundary information of potential defects.

[0023] Specifically, the clear surface image is divided into blocks according to a preset size, with overlapping areas retained between adjacent image blocks to prevent feature truncation caused by defect boundaries falling at block boundaries. The block size can be determined jointly based on the minimum detectable defect width, image resolution, and the size of the subsequent optical flow calculation window. For each image block, the horizontal and vertical grayscale changes are calculated separately to obtain the pixel-level gradient magnitude and gradient direction distributions. A sliding window is set within the image block, and the gradient mean, gradient variance, and gradient direction consistency within the window are statistically analyzed to form the local grayscale gradient change features of that image block. The local grayscale gradient change features of each image block are mapped to the entire image according to their spatial location to obtain gradient distribution data. For example, the Sobel operator is used to calculate the grayscale gradient, and the sliding window size is set according to the preset width range of the defect boundary.

[0024] Within each image block, pixel regions with gradient magnitudes higher than a preset gradient threshold and spatially continuous distribution are designated as candidate boundary regions. Background reference bands are extracted on both sides along the normal direction of the candidate boundary regions, and the gray-level mean difference, local contrast, and gradient energy difference between the candidate boundary regions and the surrounding background regions are calculated. When the above differences exceed the corresponding preset deviation threshold, and the difference remains spatially continuous between adjacent image blocks, the corresponding region is marked as a region with significant appearance changes. The preset deviation threshold is determined using a partitioned adaptive method, and the threshold is corrected within each image block based on the background gray-level mean, background fluctuation degree, and local reflection uniformity. Morphological connectivity processing is performed on regions with significant appearance changes, merging adjacent candidate regions with a distance less than a preset distance threshold and removing isolated regions with an area less than a preset area threshold to obtain a dynamic change range. This range represents the region in the same scanned continuous images where the boundary response, texture response, or local contrast relationship changes.

[0025] The inspection target is the surface of stationary equipment. Optical flow tracing reflects the apparent displacement, edge drift, and texture response changes of defect boundaries in consecutive scanned images, used for boundary temporal consistency verification and identification of unstable boundary regions. Before performing optical flow calculation, the image sequence is temporally registered based on the robot pose information, ranging information, and channel registration results recorded in S1, ensuring that the same surface position falls into the same spatial reference coordinate system in consecutive frames. Corner points, edge points, or gradient-significant points are extracted as tracking points within the dynamic range, and the displacement vector, displacement direction, and displacement continuity of each tracking point are calculated between adjacent frames. For example, the Lucas-Kanade sparse optical flow method is used to perform optical flow calculation. The size of the optical flow window and the number of pyramid layers are determined based on the defect boundary scale and the inter-frame displacement range. When the displacement direction of the tracking points within the window is consistent, the displacement amplitude changes smoothly, and they persist in consecutive frames, the corresponding tracking points are merged into the same boundary trajectory segment. When a tracking point appears only in a single frame or the displacement direction jumps beyond a preset angle threshold, it is judged as an unstable point and removed. After the above processing, preliminary boundary continuity information of potential defects is obtained, including the boundary trajectory location, boundary extension direction, boundary connectivity length, and the displacement change relationship of the boundary over time.

[0026] For each boundary trajectory segment, multiple sampling sections are established along its normal direction. The boundary width, grayscale transition distance, and edge energy distribution on the corresponding sections are measured in consecutive frame images. The boundary widths in consecutive frames are arranged into a width sequence in chronological order, and the mean width, width variation, width dispersion, and width increment between adjacent frames are calculated. When the same boundary trajectory segment exhibits continuous width expansion, local contraction followed by expansion, asymmetrical transition on both sides, or an edge energy distribution deviation exceeding a preset threshold in consecutive frames, the corresponding location is identified as an irregular boundary region. Preferably, the width sequence is temporally smoothed, and the boundary irregularity determination simultaneously satisfies both the width variation condition and the trajectory continuity condition. For example, when the standard deviation of the boundary width of the corresponding section in consecutive frames exceeds a preset fluctuation threshold, and the corresponding trajectory segment persists for at least three consecutive frames, it can be identified as an irregular boundary region. Here, the irregular boundary region reflects the apparent instability of the boundary in consecutive scanned images.

[0027] A local sub-image is cropped centered on an irregular boundary region and resampled according to a preset magnification factor. After resampling, the gradient peak position, gray-level transition center position, and edge connectivity are recalculated in the magnified local image. For example, bicubic or bilinear interpolation is used to complete the local magnification and resampling. Boundary points are extracted by combining Canny edge detection or local gradient peak search. Eight-neighborhood connectivity analysis is performed on the boundary points to form continuous boundary lines. The edge response is calculated on the multi-scale images before and after magnification, and the position with stable response is selected as the sub-pixel boundary point. For the magnified boundary region, the boundary start and end positions, boundary width range, boundary curvature change, boundary break position, and boundary connectivity index are extracted to form the boundary information of potential defects. The boundary information is mainly used to characterize the spatial morphology and local anomaly features of the defect boundary during the current inspection process.

[0028] S3. If the boundary of a potential defect is irregular and the dynamic changes of the boundary exceed a preset change threshold, then extract the three-dimensional point cloud data of the corresponding area, combine the boundary depth information to construct the geometric shape of the defect, and obtain the spatial distribution information of the defect.

[0029] In one specific embodiment, the process of performing step S3 may specifically include the following steps: If the boundary of a potential defect is detected to be irregular and the dynamic change of the boundary exceeds a preset change threshold, then a three-dimensional scan is performed on the area corresponding to the potential defect to obtain the corresponding three-dimensional point cloud data. Multi-view registration and fusion of 3D point cloud data are performed, and 3D structure analysis is carried out in combination with boundary depth information to construct the preliminary geometric morphology of the defect and determine the basic shape information of the defect. Based on the preliminary geometric shape, the spatial distribution characteristics of the defects are analyzed, and point cloud segmentation technology is used to divide the defect distribution area to obtain the defect location range; For the location range of the defect, multi-view depth information is integrated for multi-angle registration and optimization to reconstruct the three-dimensional geometric features of the defect and obtain the spatial distribution information of the defect.

[0030] Specifically, if the boundary curvature change of a potential defect exceeds a preset curvature threshold, the boundary width dispersion exceeds a preset dispersion threshold, and the temporal displacement change of the boundary exceeds a preset change threshold, then the boundary of the potential defect is determined to be irregular and the dynamic change of the boundary exceeds the preset change threshold. For example, the irregular shape can be determined using fractal dimension. The fractal dimension of the boundary line is calculated using box counting. When the fractal dimension exceeds a preset fractal threshold, the boundary is determined to be irregular. The dynamic change of the boundary is taken as the average displacement of the boundary region obtained by optical flow tracing in S2. When the average displacement exceeds a preset displacement threshold, the dynamic change of the boundary is determined to exceed the preset change threshold.

[0031] When both of the above conditions are met, the inspection robot uses the boundary of the potential defect as a reference and expands outward by a preset distance to determine the scanning range. It then uses a structured light 3D scanner to perform a 3D scan and acquire the corresponding 3D point cloud data. The 3D point cloud data includes the 3D spatial coordinates of each sampling point within the scanning area, the laser reflection intensity value, the acquisition viewpoint information, and the acquisition timestamp. For example, the 3D scan is performed using a line laser scanning module or a structured light scanning module, and the scanning frame rate is set according to the inspection robot's moving speed.

[0032] A single scan can only acquire point clouds from one viewpoint, therefore, it is necessary to scan the defect area from multiple angles. Multi-view registration and fusion of the 3D point cloud data involves initial registration of point cloud data from different acquisition viewpoints, extracting scale-invariant feature points for matching, and obtaining the initial transformation matrix between point clouds from different viewpoints. Based on this, fine registration is performed. During fine registration, candidate matching regions are filtered by combining the boundary position information, boundary trajectory information, or boundary constraint regions output by S2, prioritizing the structural features corresponding to the defect boundary and its neighborhood. Simultaneously, the depth gradient information of the boundary neighborhood acquired by the current 3D scan is used to prioritize points with significant depth changes within the boundary neighborhood as key matching points, thereby improving the registration accuracy of the defect boundary region. Finally, fused point cloud data in a unified spatial coordinate system is obtained. For example, fine registration can be implemented using an iterative nearest-point algorithm. The iteration termination condition can be set as the mean square error between two adjacent iterations being lower than a preset convergence threshold, the maximum number of iterations not exceeding a preset upper limit, and the Euclidean distance threshold for corresponding point pairs set to 1.5 times the average spacing of the point clouds. By employing the above methods, a stable spatial correspondence can be established between point clouds from multiple perspectives, reducing the interference of non-defective or weakly featured regions on the registration results. When performing stereoscopic structural analysis on the fused point cloud data, a local reference plane on the device surface can be fitted first, and then the elevation difference, surface normal vector, and local curvature value of each sampling point relative to the reference plane can be calculated. Based on this, defect depression regions, protrusion regions, and reference plane regions can be distinguished, thereby constructing the preliminary geometric shape of the defect and determining its basic shape information. The basic shape information includes the defect's contour range, initial depth range, initial volume value, and morphological type.

[0033] For the initial geometric shape of the defect, based on the preceding fused point cloud data and the corresponding 3D contour range of the initial geometric shape, non-defect point clouds within a preset neighborhood outside the contour range are selected as a reference point set. Local surface fitting is then performed on the reference point set to obtain a local reference surface for the equipment surface. For each sampling point in the fused point cloud, the elevation difference of the sampling point relative to the local reference surface is calculated. Based on the neighborhood point set of the sampling point, the local normal vector and local curvature are calculated. The normal vector deviation, curvature difference, and distance relative to the boundary of the 3D contour range between the sampling point and its neighboring sampling points are also calculated to form the spatial distribution characteristic parameters of the defect. The neighborhood point set is determined through fixed radius search or nearest neighbor search. Sampling points located within the 3D contour range, with an absolute elevation difference higher than a preset elevation difference threshold and satisfying the curvature condition, are selected as seed points and added to the defect region. Using the growth boundary point in the current defect region as the center, a search is conducted within a preset neighborhood for sampling points adjacent to the growth boundary point to be judged, which are then selected as candidate points. For each candidate point, if the deviation of its normal vector from the corresponding growth boundary point is lower than a preset angle threshold, the curvature difference is lower than a preset curvature difference threshold, the adjacent distance is lower than a preset distance threshold, and the elevation difference is lower than a preset elevation difference threshold, then the candidate point is merged into the defect region, and the newly merged candidate point is updated as a new growth boundary point. If the above conditions are not met, it is divided into a background reference region or an adjacent interference region based on its elevation difference relative to the local reference surface and its spatial relationship with the three-dimensional contour range. The candidate point search and judgment process is repeated until no new candidate point that meets the growth criteria is found. The point cloud subset corresponding to the defect region is extracted, and the defect location range is determined based on the spatial coordinate range of the point cloud subset and the three-dimensional contour range.

[0034] For the defect region point cloud subset corresponding to the defect location range, multi-view depth information is fused for multi-angle fine registration and optimization. Taking the defect region point cloud subset and its corresponding depth images from various views as the processing objects, supplementary view depth data that were not included in the fusion during the initial registration stage due to occlusion, insufficient view coverage, or low matching confidence are introduced. The boundary 3D mapping results established during the previous point cloud fusion process are used as boundary constraints. Combined with the calibration parameters of each view, the relative pose of each view is optimized using the bundle adjustment method. The reprojection error is used as the optimization target to complete the secondary fine registration, obtaining the registered defect region point cloud data. Historical depth data of the same area collected in different inspection cycles are introduced to perform temporal consistency verification. The local reference surface of the equipment surface is used as a reference to align the spatial coordinate system of the depth data at different times. Sampling points that have significant inconsistencies between the current multi-view data and deviate from the historical depth data by more than a preset threshold are judged as abnormal data points and removed. Preprocessing is performed on the corrected point cloud data. Outliers are removed based on the rule that the average distance between a sampling point and its neighborhood is greater than a preset multiple of the neighborhood distance standard deviation. For data holes within the defect contour area with a scale lower than a preset hole size threshold, neighborhood-weighted interpolation is used to complete the data, constrained by the defect boundary. The interpolation neighborhood range is set according to the average spacing of the point cloud, resulting in optimized defect point cloud data. Based on the optimized defect point cloud data and using a local reference surface as a calculation reference, the three-dimensional geometric features of the defect are reconstructed. The maximum depth, average depth, actual volume, surface area, contour size, and spatial distribution range of the defect are extracted and integrated to form the spatial distribution information of the defect.

[0035] S4. Based on the spatial distribution information of defects, the curvature mapping technique is used to evaluate the deformation characteristics of the defect area, and the geometric deformation and depth changes of the defect area are analyzed to obtain the deformation characteristic evaluation results.

[0036] In one specific embodiment, the process of performing step S4 may specifically include the following steps: Based on the spatial distribution information of defects, the initial geometric data of the surface region corresponding to the defects are obtained, and the curvature mapping technology is used to perform spatial quantitative analysis of the deformation amplitude to obtain the deformation quantification data. Based on the deformation data, analyze the geometric deformation and depth variation characteristics of the defect area to determine the range of influence of defect evolution; Based on the range of evolutionary influence, the depth deterioration pattern is analyzed by combining time-series depth change data, and compared with the preset standard defect evolution rate change trend to obtain the evolution trend judgment result. By integrating deformation data, the scope of evolutionary impact, and the results of evolutionary trend determination, deformation characteristic assessment results are generated.

[0037] Specifically, based on the spatial distribution information of defects, a subset of point clouds in the defect region, the three-dimensional contour range of the defect, local reference surface parameters, elevation difference of each sampling point relative to the local reference surface, local normal vector, local curvature, maximum depth, average depth, surface area, and volume data are extracted. These data are used as the initial geometric data of the surface region corresponding to the defect. To ensure the comparability of deformation analysis results under different inspection cycles, the subset of point clouds in the defect region is resampled and triangulated to reconstruct a mesh, ensuring that the spacing between adjacent sampling points meets the preset spacing conditions. For locations with local voids within the contour range, constrained point filling is performed based on the spatial position and normal vector of adjacent valid points to obtain a continuous surface model. Using a local reference surface, the principal curvature, mean curvature, and Gaussian curvature are calculated for each mesh node or mesh element in the continuous surface model. These curvature values ​​are then mapped to their corresponding spatial locations to form a curvature distribution map of the defect region. Combining this with the elevation difference of each location relative to the local reference surface, the area proportions of curvature peak regions, curvature abrupt change regions, curvature continuously changing regions, and curvature level regions are statistically analyzed. The deformation amplitude, edge transition zone width, and central concavity or convexity amplitude of the defect region in the normal direction are calculated to obtain deformable data. This deformable data includes at least the curvature peak distribution, curvature gradient distribution, deformation amplitude distribution, local concavity / convexity depth distribution, and the corresponding spatial coordinate range.

[0038] When analyzing the geometric deformation and depth variation characteristics of the defect region, multiple cross-sections are established along the main extension direction and normal direction of the defect. On each cross-section, the degree of contour curvature, cross-section width variation, cross-section depth variation, and edge slope variation are extracted. Simultaneously, by combining the curvature difference, elevation difference variation, and their spatial continuous distribution between adjacent grid nodes or grid cells, the continuous expansion characteristics of defect deformation and depth variation are analyzed. When the curvature difference and elevation difference variation between adjacent grid cells continuously expand in space, and this continuous expansion exceeds a preset neighborhood range outside the initial contour boundary of the defect, the corresponding expansion area is determined as the evolution influence area of ​​the defect. Then, based on the attenuation boundaries of the curvature change gradient and depth change gradient within the evolution influence area, the evolution influence range of the defect is determined. For example, a region growth method can be used to expand outward from the curvature peak region (a region where the absolute value of curvature is higher than a preset curvature peak threshold and is locally continuous) or the depth abrupt change region (a region where the change in elevation difference between adjacent grid nodes or grid cells is higher than a preset depth change threshold and extends continuously). Adjacent grid cells that meet the following criteria are incorporated into the influence region: curvature difference value lower than a preset curvature difference threshold, normal vector angle lower than a preset angle threshold, adjacent distance lower than a preset distance threshold, and elevation difference changes extending in the same direction. Expansion stops when the curvature change amplitude of newly added grid cells falls below a preset lower limit, or when the elevation difference with the local reference surface returns to within the background fluctuation range. The evolutionary influence range is output in the form of a spatial envelope boundary, area range, and depth influence zone.

[0039] Historical depth data for the same area across multiple inspection cycles is retrieved, and the depth data at each moment is aligned with coordinates based on a local reference surface to form a temporal depth sequence corresponding to each sampling location within the evolutionary influence range. The depth change, depth change rate per unit time, and depth change acceleration are calculated according to spatial location, and each location is classified based on the changing direction, amplitude, cumulative change, and depth change rate trend over multiple consecutive inspection cycles. The depth change acceleration characterizes the rate of change of depth per unit time; preferably, it is calculated based on the depth change rate of at least three consecutive inspection cycles. When the time intervals between adjacent inspection cycles are inconsistent, the actual time interval is used for calculation. For the same location, if it changes along the direction of defect deepening in multiple consecutive inspection cycles, and the rate of change of depth per unit time is continuously higher than a preset rate of change threshold, then the location is determined to be a continuously deepening location; if its depth change acceleration is continuously positive and higher than a preset acceleration threshold, then it is further determined to have an accelerating deterioration trend; if its rate of change of depth per unit time is within a preset fluctuation range, and the cumulative depth change does not exceed a preset cumulative threshold, then it is determined to be a fluctuating stable location; if its depth change acceleration is within a preset fluctuation range, then it can be used as an auxiliary criterion for fluctuating stable state; if its depth measurement value relative to the local reference surface shows a decreasing trend in multiple consecutive inspection cycles, and continuously reaches the preset cycle condition, then it is determined to be an apparent mitigation location; the apparent mitigation location indicates that the depth measurement value at the corresponding location is reduced, the change slows down, or the local morphology response weakens relative to the previous inspection results, and does not necessarily indicate that the defect has undergone physical repair; if its depth change acceleration is continuously negative and the influence range no longer expands, then it can be used as an auxiliary criterion for apparent mitigation state. Spatial connectivity merging is performed on adjacent locations of the same type to obtain regions that are continuously deepening, regions that are stable in fluctuation, and regions that appear to be alleviated.

[0040] The study statistically analyzes the area proportion and spatial distribution of various regions within the evolutionary influence range, and examines whether defects spread from the center to the edge or extend along a specific direction based on the spatial migration direction of the continuously deepening zone. Simultaneously, it averages the unit-time depth change rate of the continuously deepening zone to obtain the actual depth deterioration rate. Pre-defined standard defect evolution rate trends can be established in advance according to defect type; preferably, they can be constructed offline based on historical known defect samples, industry standards, or experimental calibration data. Specifically, standard depth growth curve intervals, standard diffusion modes, and standard rate fluctuation ranges can be established for corrosion pits, cracks, and erosion depressions, respectively. The unit-time depth change rate of each zone within the evolutionary influence range is statistically analyzed under continuous inspection cycles and arranged chronologically to form a temporal depth change rate sequence for the current defect. This temporal depth change rate sequence is then matched with the corresponding type of standard trend template to determine the defect evolution trend category. If the current sequence falls into the accelerated deterioration zone and the depth change acceleration is continuously positive or the average value is higher than the preset acceleration threshold, it is determined to be accelerated evolution; if the current sequence changes within the preset fluctuation range and the depth change acceleration is within the preset fluctuation range, it is determined to be stable evolution; if the depth change rate of the current sequence continues to decrease, the depth change acceleration is continuously negative and the influence range no longer expands, it is determined to be moderate evolution; if the current sequence exceeds the preset threshold corresponding to the rapid failure zone, it is determined to be critical deterioration.

[0041] The peak curvature distribution, curvature gradient distribution, deformation amplitude distribution, local depth gradient distribution, evolutionary influence range boundary, influence area, influence zone width, actual depth deterioration rate, and trend matching results are correlated and encoded to form the defect deformation feature assessment result. The deformation feature assessment result includes at least the defect deformation type, deformation amplitude level, geometric deformation concentration location, depth deterioration state, actual depth deterioration rate, evolutionary influence range, and evolutionary trend category.

[0042] Figures 2 to 4 The process of determining the curvature mapping and evolution influence range of a defect region is illustrated by way of example. Figure 2 The curvature distribution map of the defect area uses pseudo-color to represent the curvature values, and the white dashed line represents the profile position line AA′. Figure 3 An evolutionary influence range boundary is superimposed on the curvature distribution. The boundary corresponds to a closed contour area where the curvature value reaches a preset evolutionary influence threshold, and the influence area is highlighted. Figure 4 This represents the curvature profile curve along the profile location line AA′. The gray dashed line indicates the evolutionary influence threshold, and the gray filled area indicates the influence range where the curvature exceeds the threshold. Combined with... Figures 2 to 4 It is evident that curvature mapping technology can quantify the deformation amplitude and spatial distribution of defect areas and provide a basis for determining the range of evolutionary influence.

[0043] S5. Based on the deformation characteristic assessment results, for defect areas where the evolution influence range exceeds the preset range, laser triangulation technology is used to obtain depth and height data. Combined with the evolution time span of the defect, the defect expansion rate is calculated, and the quantitative physical parameters of the defect are determined.

[0044] In one specific embodiment, the process of performing step S5 may specifically include the following steps: Based on the deformation characteristic assessment results, the defect area whose evolution influence range exceeds the preset range is identified as the target area; Laser triangulation technology is used to perform high-precision scanning of the target area to obtain the depth and height data at the current moment; Retrieve historical depth and height data for the target area and calculate the difference in depth change between the current time and historical times; The defect propagation rate is calculated based on the depth variation difference and the evolution time span of the defect; By combining depth and height data with defect propagation rate, the quantitative physical parameters of the defect are determined.

[0045] Specifically, the preset range can be composed of at least one of the following: an influence area threshold, an influence band width threshold, a boundary expansion distance threshold, or a depth influence ratio threshold. When the area, expansion width, or depth influence band of the evolving influence range exceeds the corresponding preset threshold, the corresponding defect area is determined as the target area. If the same defect has multiple separate influence sub-regions, the multiple influence sub-regions are merged according to spatial connectivity and boundary adjacency to form the target area boundary. To ensure that the subsequent measurement range covers the true expansion direction of the defect, a preset safety margin is extended outward based on the target area boundary to generate a laser triangulation measurement scanning window, and the scanning window is mapped to the current pose coordinate system of the inspection robot.

[0046] The inspection robot is controlled to adjust the laser emitter and imaging receiver to a preset angle, ensuring that the linear laser stripes cover the target area. After the laser stripes are projected onto the target surface, a stripe offset image is formed on the receiver's imaging plane. Based on the laser emission baseline length, receiver imaging parameters, projection angle, and stripe center position, the depth values ​​of each sampling point within the target area relative to the reference datum are calculated, forming the current depth and height data. To reduce the impact of metal reflections, surface deposits, and local occlusions on stripe center extraction, the stripe image is first normalized for brightness, enhanced for stripes, and sub-pixel center extracted. Then, the depth results of multiple cut-off lines are stitched together to form a depth map of the target area according to the scanning path. Simultaneously, based on the inspection robot's pose information and the target area boundary coordinates, the current depth map is registered to the unified spatial reference system used by S3 and S4. The current depth and height data includes at least the spatial coordinates, depth values, and acquisition timestamps of each sampling position within the target area.

[0047] Historical depth and height data corresponding to the spatial location of the target area are retrieved from the historical inspection database. Baseline historical depth and height data are selected based on temporal proximity, data integrity, stable area coverage, and registration reliability. Using local reference surfaces and stable areas within the target area that have not undergone significant deformation as references, coordinate unification and boundary alignment are performed between the current depth and height data and the baseline historical depth and height data. For non-overlapping sampling locations caused by differences in inspection angles, equipment attitude fluctuations, or local occlusion, corresponding sampling point pairs are established using neighborhood interpolation or nearest neighbor matching. The depth difference is calculated point-by-point for each registered corresponding sampling point, forming a depth change difference distribution map of the current time relative to the baseline historical time within the target area. To avoid isolated outliers interfering with subsequent rate calculations, outlier removal is performed on the depth change difference distribution map; outlier removal can be based on the degree of deviation of neighborhood depth differences and spatial continuity conditions. Subsequently, the maximum depth change, average depth change, and width of the continuous deepening zone within the target area are statistically analyzed.

[0048] The time interval between the current acquisition time and the baseline historical time is used as the evolution time span. The depth change difference of corresponding sampling points within the target area is divided by the evolution time span to obtain the unit time depth change rate at each sampling location. Then, the area within the target area where the depth change difference is greater than a preset effective change threshold and meets the spatial connectivity condition is defined as the extended analysis area. The unit time depth change rate at each sampling location within the extended analysis area is averaged to obtain the depth expansion rate of the defect. In addition to depth expansion, the boundary expansion rate of the defect in the planar direction can also be calculated based on the outward expansion distance between the target area boundary at the current time and the baseline historical time. Preferably, the target area boundary at the current time and the target area boundary at the baseline historical time are registered to the same spatial reference frame. The normal outward expansion distance between multiple corresponding boundary sampling point pairs is calculated, and the normal outward expansion distance is statistically analyzed to obtain the average outward expansion distance. The average outward expansion distance is divided by the time interval between the current acquisition time and the baseline historical acquisition time to obtain the boundary expansion rate of the defect in the planar direction. The defect expansion rate is obtained based on the combination relationship between the depth expansion rate and the planar boundary expansion rate. Preferably, when the defect is mainly characterized by deepening depressions, the depth propagation rate is used as the main rate; when the defect is mainly characterized by boundary propagation, the planar boundary propagation rate is used as the main rate; when both change significantly at the same time, the depth propagation rate and the planar boundary propagation rate are output separately and combined according to a preset weight. The preset weight can be set according to the defect type, such as using the depth propagation rate preferentially for corrosion pits and the boundary propagation rate preferentially for cracks and grooves.

[0049] Based on the current depth and height data, extract the maximum depth, average depth, depth gradient distribution, cross-sectional width, target area, and volume change of the defect. The target area can be represented by the projected area of ​​the target area boundary on a local reference surface, and the volume change can be obtained by accumulating the volume micro-elements between each sampling point within the target area and the local reference surface. Combine the defect propagation rate to extract the maximum depth increase per unit time, average depth increase per unit time, boundary expansion per unit time, and the area ratio of the expansion analysis zone. These parameters are combined into a set of quantized physical parameters, including at least the current maximum depth, current average depth, maximum depth change, average depth change, depth propagation rate, boundary expansion rate, target area, area ratio of the expansion analysis zone, and volume loss change. If there are sub-regions with different evolution intensities within the target area, the corresponding local quantized physical parameters can also be calculated separately for each sub-region.

[0050] In a preferred embodiment, determining the quantitative physical parameters of the defect includes: Time-series analysis of defect propagation rate is performed to determine the dynamic trend of defect propagation rate. Based on dynamic changes, the target area is divided into layers, the degree of evolutionary impact of each area is analyzed, and risk levels are classified. For high-risk areas where the degree of evolutionary impact exceeds a preset threshold, a high-density supplementary scan is performed to extract detailed depth and height data of the high-risk areas, update the high-precision expansion rate of the defects, and combine the detailed depth and height data with the high-precision expansion rate to determine the final quantitative physical parameters of the defects.

[0051] Specifically, when performing time-series analysis on defect propagation rates, the depth propagation rate and planar boundary propagation rate of the same target area under multiple consecutive inspection cycles are retrieved, and the propagation rates corresponding to each inspection cycle are arranged in chronological order to form a time-series sequence of propagation rates. The rate difference, rate change amplitude, and rate change direction between adjacent inspection cycles are calculated, and trend fitting is performed on the time-series sequence of propagation rates to determine the dynamic trend of defect propagation rates. For example, linear regression is performed on the time-series sequence of propagation rates to fit the trend slope of the propagation rate changing with time. When the trend slope is greater than a preset rising threshold and the rate increment of adjacent inspection cycles is continuously positive, it is determined to be an accelerating propagation trend. When the trend slope is within a preset fluctuation range and the propagation range does not continue to increase, it is determined to be a stable propagation trend. When the trend slope is less than a preset falling threshold and the rate increment of adjacent inspection cycles is continuously negative, it is determined to be a slowing propagation trend. To reduce the impact of fluctuations in a single measurement, the expansion rate time series is smoothed using a moving average or weighted average, with the smoothing window set according to the number of inspection cycles. Simultaneously, the acceleration of rate change is calculated based on the rate difference between adjacent cycles to help determine if there are abrupt changes in the expansion rate. If both depth expansion rate and planar boundary expansion rate exist in the target area, separate time series for depth expansion rate and planar boundary expansion rate are generated, and a joint judgment is made based on preset weights set according to the defect type. Through time series analysis, the instantaneous expansion rate obtained from a single measurement is transformed into a dynamic trend reflecting the rhythm of defect evolution, solving the problem that it is difficult to determine whether a defect has entered an accelerated deterioration stage based solely on the rate at a single moment.

[0052] The target area is divided into multiple analysis sub-regions, which can be defined based on the depth gradient distribution in the current depth and height data, the distribution of previous depth change differences, the boundary of the extended analysis area, and local contour changes. For each analysis sub-region, the current maximum depth, average depth, depth change, depth expansion rate, boundary expansion rate, area proportion, projected area, and volume loss change are extracted, and the evolutionary influence degree of each analysis sub-region is calculated in conjunction with dynamic change trends. The evolutionary influence degree is obtained by comprehensively considering the morphological parameters and rate parameters of each analysis sub-region. For example, the current maximum depth, depth change, depth expansion rate, boundary expansion rate, area proportion, and volume loss change are normalized, and then the comprehensive influence value (i.e., the evolutionary influence degree) is calculated according to preset weights. Preferably, for depth-dominant defects, the weights of depth expansion rate, depth change, and maximum depth are increased; for boundary-spreading defects, the weights of boundary expansion rate, area proportion, and projected area are increased. The comprehensive impact value of each analysis sub-region is compared with the preset risk level range. Sub-regions with a comprehensive impact value in the first level range are classified as low-risk areas, those in the second level range as medium-risk areas, and those in the third level range with an accelerating expansion trend are classified as high-risk areas. If adjacent analysis sub-regions have the same risk level and are spatially continuous, they are merged to form a continuous risk stratification area.

[0053] After risk level classification, the spatial boundaries, risk levels, and dynamic trends of high-risk areas are read, and the inspection robot is controlled to move the laser triangulation device to the corresponding position in the high-risk area. Then, the single scan step distance is reduced, the scan line density is increased, and the spacing between adjacent measurement lines is decreased, so that the number of sampling points per unit area is higher than the sampling density in the previous target area scanning stage, obtaining finer depth sampling results for the high-risk area. During supplementary scanning, the local reference surface is still used as a reference, and the coordinate unification and boundary registration are performed between the depth map obtained from the supplementary scan and the depth map of the previous target area. For local noise, reflection anomalies, and missing measurement voids that appear in the supplementary scan, correction or interpolation is performed based on the continuity of neighborhood depth and boundary constraints to obtain detailed depth and height data of the high-risk area. The detailed depth and height data includes at least the spatial coordinates, depth values, local depth gradients, local boundary positions, and acquisition timestamps of each subdivided sampling position in the high-risk area. High-density supplementary scanning can improve the extraction accuracy of local depth changes and solve the problem of difficulty in accurately quantifying micro-cracks, grooves, sharp pit edges, and narrow deepening zones under conventional scanning density.

[0054] The detailed depth and height data of the current high-risk area are registered with the historical depth and height data of the corresponding area at the reference historical time, and the high-precision depth change rate per unit time of each subdivided sampling position within the high-risk area is calculated. Areas within the high-risk area where the depth change difference is greater than a preset high-precision effective change threshold and meets the spatial connectivity condition are defined as high-precision extended analysis areas. The high-precision depth change rate per unit time of each subdivided sampling position within the high-precision extended analysis area is averaged to obtain the high-precision depth extension rate of the high-risk area. Simultaneously, the outward expansion distance between the current high-risk area boundary and the high-risk area boundary at the reference historical time is extracted, and the high-precision boundary extension rate of the high-risk area is calculated. Preferably, the current high-risk area boundary and the high-risk area boundary at the reference historical time are registered to the same spatial reference frame, the normal outward expansion distance between multiple corresponding boundary sampling point pairs is calculated, and the normal outward expansion distance is averaged to obtain the average boundary outward expansion distance. The average boundary outward expansion distance is divided by the time interval between the current acquisition time and the reference historical acquisition time to obtain the high-precision boundary extension rate of the high-risk area. Based on the defect type in the high-risk region, the high-precision depth expansion rate and the high-precision boundary expansion rate are fused according to a preset combination to obtain the high-precision expansion rate. Preferably, if the high-risk region belongs to the depth-dominant defect type, the high-precision depth change rate per unit time is used as the main rate; if the high-risk region belongs to the boundary-dominant defect type, the high-precision boundary expansion rate is used as the main rate; if both change significantly at the same time, they are combined according to a preset weight. The high-precision expansion rate is used to replace the coarse-grained expansion rate at the previous target region level, or is superimposed as a local correction amount into the original quantized physical parameters to improve the regional representativeness of the final parameters.

[0055] Based on detailed depth and height data of high-risk areas, the maximum depth, average depth, local depth gradient distribution, cross-sectional width, projected area, area proportion, and volume loss change of high-risk areas are recalculated. Then, combined with high-precision expansion rates, the maximum depth increase per unit time, average depth increase per unit time, boundary expansion per unit time, and rapid deterioration zone width are extracted. The updated local parameters of the high-risk areas are integrated with the original quantitative physical parameters of the rest of the target area to form the final quantitative physical parameters of the defect. These parameters include at least the current maximum depth, current average depth, maximum depth change, average depth change, volume loss change, depth expansion rate, boundary expansion rate, high-risk area proportion, and high-precision expansion rate of the high-risk areas for the entire target area. If the target area contains multiple high-risk areas, the local parameters of each high-risk area are calculated separately and summarized according to area proportion or risk level to form the final set of quantitative physical parameters for the entire defect.

[0056] S6. If the quantified physical parameters exceed the preset safety threshold, an alarm signal is generated, and the defect evolution data is integrated to obtain the maintenance priority ranking.

[0057] In one specific embodiment, the process of performing step S6 may specifically include the following steps: If the quantized physical parameters exceed the preset safety threshold, an alarm signal will be triggered; Extract the historical evolution history of defects, simultaneously acquire environmental impact factor data from the inspection site, and integrate them to generate a comprehensive impact analysis basis; Based on the comprehensive impact analysis, the evolutionary history was compared over time, and combined with changes in environmental factors, the driving factors of defect evolution were identified, and high-risk monitoring areas were determined. Based on the high-risk monitoring areas, adjust and determine the maintenance priority order, and simultaneously generate the corresponding resource allocation priority order.

[0058] Specifically, each quantitative physical parameter is compared with its corresponding safety threshold. When any parameter exceeds its corresponding safety threshold, or when multiple parameters, although not individually exceeding their individual thresholds, have a combined risk value exceeding a preset combined threshold, the current defect is determined to be in an alarm state, and a corresponding alarm signal is generated. For example, depth thresholds, propagation rate thresholds, area thresholds, and volume loss thresholds can be set according to defect type. For corrosion pit defects, the weighting of the depth propagation rate and volume loss change is increased; for crack and groove defects, the weighting of the boundary propagation rate and the proportion of high-risk area is increased. The alarm signal includes at least the defect identifier, corresponding equipment location, trigger time, trigger parameter category, alarm level, and threshold exceedance magnitude, and is sent to the local control terminal of the inspection robot and the operation and maintenance management platform.

[0059] The historical inspection records corresponding to the current defect are retrieved from the inspection database. The evolution history records include at least the spatial distribution information of the defect in each historical inspection cycle, deformation characteristic assessment results, maximum depth, average depth, depth change, depth expansion rate, boundary expansion rate, area ratio of high-risk areas, volume loss change, historical alarm records, and historical maintenance records. Simultaneously, environmental impact factor data for the inspection site is obtained from the environmental monitoring unit or on-site monitoring system mounted on the inspection robot. This includes at least one of the following: ambient temperature, ambient humidity, equipment surface temperature, corrosive medium concentration, vibration intensity, load fluctuation state, and operation duration. Vibration intensity refers to the strength of mechanical vibration in the target area or related structural parts during equipment operation, and can be characterized by vibration amplitude, effective value of vibration velocity, or effective value of vibration acceleration. Load fluctuation state refers to the change in equipment workload within a preset time window, and can be characterized by changes in torque, pressure, flow rate, current, rotational speed, axial force, or radial force. The evolution history records and environmental impact factor data are aligned along a unified time axis, and the equipment operating condition information corresponding to the spatial location of the defect is written into the analysis dataset to form a comprehensive impact analysis basis. To avoid temporal misalignment caused by inconsistent sampling frequencies between different collection periods, evolutionary history records and environmental influencing factor data can be resampled or aggregated within a preset time window.

[0060] When comparing the evolution history of the same defect over multiple consecutive inspection cycles, the time series sequence of defect evolution parameters is formed by extracting the depth change, depth expansion rate, boundary expansion rate, volume loss change, and high-risk area proportion for each inspection cycle. Simultaneously, the time series sequence of environmental factors is formed by extracting the ambient temperature, ambient humidity, equipment surface temperature, corrosive medium concentration, vibration intensity, load fluctuation state, and operation duration for each inspection cycle. The time series sequences of defect evolution parameters and environmental factors are compared to determine the direction, magnitude, and duration of change for each environmental factor between adjacent inspection cycles. When an environmental factor continuously increases or fluctuates more strongly over multiple consecutive inspection cycles, and the corresponding defect evolution parameters increase synchronously within the same cycle or increase within a preset lag period, this environmental factor is identified as a candidate driving factor. A time-series correlation analysis is performed on candidate driving factors and their corresponding defect evolution parameters to calculate the correlation between them within a preset time window. If the correlation is higher than a preset correlation threshold, and the change in the candidate driving factor continues to affect at least one preset analysis period, then the candidate driving factor is identified as a driving factor for defect evolution. For example, the correlation between the change sequence of corrosive medium concentration and the change sequence of volume loss, and the correlation between the change sequence of vibration intensity and the change sequence of boundary expansion rate can be calculated separately. When the concentration of corrosive medium continuously increases and the change in volume loss increases synchronously, the concentration of corrosive medium can be identified as a driving factor; when the vibration intensity continuously increases and the boundary expansion rate increases synchronously, the vibration intensity can be identified as a driving factor. For identified driving factors, the contribution of the driving factor to defect evolution can also be calculated based on the fitting relationship between its change and the change in the corresponding defect evolution parameter. Preferably, for identified driving factors, their change sequences over multiple consecutive inspection cycles are extracted, along with their corresponding defect evolution parameter change sequences. A fitting analysis is performed on the driving factor change sequences and their corresponding defect evolution parameter change sequences to obtain the fitting slope, and the correlation and the proportion of the driving factor's duration of action are calculated. The sign of the fitting slope indicates the direction of the driving factor's effect on defect evolution: a positive slope indicates that the corresponding defect evolution parameter increases when the driving factor strengthens, classifying it as a deterioration-promoting driving factor; a negative slope indicates that the corresponding defect evolution parameter decreases when the driving factor strengthens, classifying it as an inhibition-type driving factor. To quantify the strength of the driving factor, the absolute value of the fitting slope, the correlation, and the proportion of the duration of action are normalized, and then weighted and summed according to preset weights to obtain the contribution of the driving factor.Preferably, for deterioration-promoting driving factors with a positive fitting slope, their contribution is used as one of the bases for identifying high-risk monitoring areas, adjusting maintenance priorities, or ranking risks; for inhibition-promoting driving factors with a negative fitting slope, their direction and intensity are recorded separately to assist in the analysis of defect evolution conditions, but they are not directly included in the accumulation as positive risk contribution in the determination of high-risk monitoring areas.

[0061] Based on the source location, range of action, and spatial distribution information of the driving factors, the corresponding area of ​​action of the driving factors is determined. Then, considering the continuous distribution of high-risk areas and their correlation with adjacent structural parts of the defect, the portion of the area where the current defect is located that is continuously affected by the driving factors is identified as a high-risk monitoring area. If the range of action of the driving factors extends to structural connection points, force transmission points, or media propagation paths adjacent to the current defect, these areas are also identified as high-risk monitoring areas. A high-risk monitoring area can be the area where the current defect is located, or it can be an extended monitoring area adjacent to the current defect and affected by the same driving factors.

[0062] For each defect requiring maintenance, its alarm level, quantitative physical parameters, evolution trend category, driving factor type, driving factor contribution, high-risk monitoring area range, and the importance level of the equipment location are extracted. A maintenance priority value is then calculated according to a unified evaluation rule. Multiple defects are ranked based on their maintenance priority values ​​to obtain a maintenance priority ranking. For example, alarm level, depth expansion rate, boundary expansion rate, high-risk monitoring area ratio, driving factor contribution, and equipment location importance level (e.g., determined according to the preset location importance classification in the equipment management ledger) can be normalized and summed according to preset weights to obtain the maintenance priority value. Preferably, when a defect is located in a critical load-bearing area, its high-risk monitoring area is large, and its expansion rate continues to increase, its maintenance priority value is increased; when a defect has a large depth but its evolution trend has slowed down and the high-risk monitoring area has not continued to expand, its maintenance priority value is decreased. A resource allocation priority order is simultaneously generated based on the maintenance priority ranking. The resource allocation priority order includes at least the order of supplementary inspections, the order of dispatching maintenance personnel, the order of spare parts retrieval, and the order of recommended shutdown maintenance.

[0063] S7. Based on quantitative physical parameters and maintenance priority ranking, the evolution rate and depth deterioration data of defects are fused using probabilistic update technology to calculate the future evolution probability of defects and obtain the final risk assessment result.

[0064] In one specific embodiment, the process of performing step S7 may specifically include the following steps: Based on maintenance priority ranking, defects with a priority higher than a preset priority threshold are selected as target evaluation defects, and evolution rate change data and evolution depth deterioration data of the corresponding defects are obtained according to the quantified physical parameters. Using a probabilistic update technique, data on changes in evolution rate and data on deterioration in evolution depth are fused together to calculate the future evolution probability of defects; If the future evolution probability shows a continuous upward trend, then the future evolution probability is weighted and adjusted to obtain the adjusted evolution probability. Based on the modified evolution probability, a defect risk distribution map is generated by comparing and matching through a preset risk mapping table, and then the final risk assessment result is obtained.

[0065] Specifically, defects with a maintenance priority higher than a preset priority threshold are selected as target evaluation defects. Based on the defect identifier and spatial location corresponding to the target evaluation defect, historical records of quantitative physical parameters from multiple consecutive inspection cycles are retrieved. Based on these historical records, the evolution rate change data and evolution depth deterioration data of the corresponding defects are calculated. The evolution rate change data includes at least a depth expansion rate change sequence, a boundary expansion rate change sequence, and a high-precision expansion rate change sequence. The evolution depth deterioration data includes at least a maximum depth change sequence, an average depth change sequence, a depth increment sequence, and a volume loss change sequence. The above data are time-aligned according to a unified inspection time, and missing inspection cycles, duplicate records, and abnormal jump values ​​are removed to form an evaluation input dataset for probability updates. The preset priority threshold can be set based on maintenance resource capacity, equipment importance level, or the number of defects that can be handled per unit time.

[0066] Based on historical inspection records of similar defects, a priori probabilities of future defect evolution events are established (differentiating similar defects in historical samples into categories such as continued deterioration, rapid expansion, and regional deterioration according to their evolution results within a preset future assessment period, and statistically analyzing the frequency of each category in the total sample as the prior probability). Conditional probability distributions of evolution rate changes and evolution depth deterioration data under different evolution states are also statistically analyzed. These conditional probability distributions can be empirical distributions obtained from historical sample statistics, parameterized probability distributions, or frequency distributions after discrete segmentation. The depth expansion rate, boundary expansion rate, high-precision expansion rate, maximum depth change, average depth change, and volume loss change of the target assessment defect in the current inspection period and several previous inspection periods are used as observation inputs. A Bayesian update method is used to update the prior probabilities periodically, obtaining the posterior probabilities of the target assessment defect experiencing continued deterioration, rapid expansion, and regional deterioration within the preset future assessment period. Evolution rate change data characterizes changes in the defect expansion rhythm, and evolution depth deterioration data characterizes changes in the degree of geometric damage to the defect; both serve as the observation basis for future evolution probabilities. Preferably, the changes in depth expansion rate and boundary expansion rate are used together to update the posterior probability of rapid expansion events; the changes in maximum depth, average depth, and volume loss are used together to update the posterior probability of depth deterioration events; and the changes in the area of ​​locally deteriorated regions, the expansion of high-risk regions, or the changes in the range of evolutionary influence are used to update the posterior probability of regional deterioration events. After obtaining the posterior probabilities of various evolutionary events, the probabilities of continued deterioration events, rapid expansion events, and regional deterioration events are weighted and summed according to preset weights to obtain the future evolution probability of the target assessment defect within a preset future assessment period; or, the cumulative probability of the target assessment defect reaching preset failure conditions within the future assessment period is used as the future evolution probability. The aforementioned preset weights can be pre-set based on the defect type, historical statistical results, or experimental calibration results.

[0067] The future evolution probability sequence of the target assessment defect is extracted over multiple consecutive assessment periods, and the probability increment between adjacent assessment periods is calculated. If the future evolution probability continuously increases over multiple consecutive assessment periods, and the probability increments between adjacent assessment periods are all positive, or the average value of the probability increments is higher than a preset increment threshold, the future evolution probability is determined to show a continuous upward trend. Combined with auxiliary quantities that characterize the increasing degree of the defect's continuous deterioration trend, the future evolution probability is weighted and corrected to obtain a corrected evolution probability. The auxiliary quantities preferably include the change rate of depth expansion rate, the change rate of boundary expansion rate, the expansion rate of the evolutionary influence range, the proportion of high-risk monitoring area, and the change in driving intensity directly related to defect deterioration within the most recent inspection period. The greater the correction weight, the more the change rate of depth expansion rate is continuously positive and the increase is increasing; the more the change rate of boundary expansion rate is continuously positive and the increase is increasing; the more the evolutionary influence range expands; the more the proportion of high-risk monitoring area increases; or the more the change in driving intensity directly related to defect deterioration increases. The corrected evolution probability represents the future evolution possibility after further considering the recent continuous deterioration trend on top of the basic future evolution probability. For example, the change rates of depth expansion rate, boundary expansion rate, evolution influence range expansion rate, high-risk monitoring area area ratio, and driving intensity can be normalized first, and then summed according to preset weights to obtain a trend correction factor. The trend correction factor is then used to incrementally correct the future evolution probability to obtain the corrected evolution probability. The preset weights can be pre-set according to the defect type, historical statistical results, or experimental calibration results.

[0068] A pre-defined risk mapping table stores the mapping relationships between modified evolution probability intervals and risk levels, risk identifiers, and corresponding treatment categories. The modified evolution probability of each target assessment defect is matched against the pre-defined risk mapping table to obtain the corresponding risk level. Then, the risk level is associated with the spatial location of the defect, the range of high-risk monitoring areas, equipment location information, and maintenance priority ranking to generate a defect risk distribution map. The defect risk distribution map includes at least the spatial location, modified evolution probability, risk level, and high-risk monitoring area distribution of each target assessment defect within the equipment target area, representing the risk distribution status of different defects on the equipment surface. For multiple target assessment defects on the same equipment, the risk levels and corresponding spatial locations of multiple defects can be overlaid to form a risk distribution result for the entire machine or entire pipeline section. The risk assessment result is output based on the defect risk distribution map, including at least the modified evolution probability of the target assessment defect, its corresponding risk level, risk distribution location, and corresponding maintenance priority.

[0069] In a preferred embodiment, the process of performing step S7 may specifically include the following steps: Based on maintenance priority ranking, defects with a priority higher than a preset priority threshold are selected as target evaluation defects, and evolution rate change data and evolution depth deterioration data of the corresponding defects are obtained according to the quantified physical parameters. Using a probabilistic update technique, data on changes in evolution rate and data on deterioration in evolution depth are fused together to calculate the future evolution probability of defects; If the probability of future evolution shows a continuous upward trend, a trend warning signal is generated to determine the evolutionary stability level of the defect. Based on the evolutionary stability level, a defect risk distribution map is generated by comparing and matching the results through a preset risk mapping table, and the final risk assessment result is obtained.

[0070] Specifically, the screening of target assessment defects, the acquisition of evolution rate change data and evolution depth deterioration data, the calculation of future evolution probability, and the trend of future evolution probability can adopt the same processing method as in the aforementioned implementation method, and will not be repeated here.

[0071] The trend warning signal includes at least the defect identifier, the corresponding equipment location, the trend trigger time, the number of consecutive rising cycles, the probability change, the current future evolution probability, and the corresponding maintenance priority. Based on the fluctuation amplitude, duration of consecutive rises, and rate of change of the probability change sequence, the evolution stability level of the defect is determined, which characterizes the stability of the defect's future evolution state. A stable evolution level is defined as follows: when the future evolution probability maintains low fluctuations within a preset statistical window, and the probability change between adjacent evaluation cycles is zero or alternates; a fluctuating evolution level is defined as follows: when the future evolution probability fluctuates continuously within a preset interval, and the probability change direction is inconsistent between adjacent evaluation cycles or the average probability change is close to a preset trend threshold; and an unstable evolution level is defined as follows: when the future evolution probability rises continuously, and both the number of consecutive rising cycles and the average probability change are higher than the corresponding thresholds. For example, the standard deviation, average probability change, and number of consecutive upward cycles of the future evolution probability within a preset statistical window can be calculated first, and then the classification can be carried out accordingly. When the standard deviation is in the low fluctuation range and the average probability change does not exceed the preset trend threshold, it is determined to be a stable evolution level. When the standard deviation is in the fluctuation range or the average probability change is close to the preset trend threshold, it is determined to be a fluctuating evolution level. When the average probability change exceeds the preset trend threshold and the number of consecutive upward cycles reaches the preset number of cycles, it is determined to be an unstable evolution level.

[0072] A pre-defined risk mapping table establishes a correspondence between evolutionary stability level, future evolution probability range, and maintenance priority range, and risk level, risk identifier, and disposal category. The evolutionary stability level of the target assessment defect is used as the primary mapping condition, while future evolution probability and maintenance priority are used as secondary mapping conditions. These are input into the pre-defined risk mapping table to match the corresponding risk level and disposal category. The evolutionary stability level reflects the stability of the defect's future trend; the future evolution probability reflects the likelihood that the defect will continue to deteriorate or reach failure conditions within a pre-defined future assessment period; and the maintenance priority reflects the order of disposal for defects in scenarios with multiple defects coexisting. All three factors participate in risk mapping but do not substitute for each other. For example, a stable evolutionary level can correspond to a low-risk or low-to-medium-risk range, a fluctuating evolutionary level can correspond to a medium-risk range, and an unstable evolutionary level can correspond to a medium-to-high-risk or high-risk range. When the evolutionary stability level is the same but the future evolution probability or maintenance priority is higher, it can be mapped to a higher risk level or a more urgent disposal category. The risk level is correlated with the spatial location of the target assessment defect, the scope of the high-risk monitoring area, equipment location information, and defect identifier to generate a defect risk distribution map. This map includes at least the spatial location, evolution stability level, future evolution probability, risk level, and handling category of each target assessment defect, and is marked on the equipment structure diagram, pipeline layout diagram, or surface unfolded diagram of the target area. For cases where multiple target assessment defects exist in the same equipment or area, the risk identifiers corresponding to multiple defects can be superimposed according to their spatial location to form a risk distribution result for the entire machine or target area. Then, based on the defect risk distribution map, the risk assessment result is output. The result includes at least the evolution stability level, future evolution probability, risk level, risk distribution location, corresponding maintenance priority, and handling category of the target assessment defect, and is used for subsequent inspection frequency adjustments, on-site warnings, maintenance personnel dispatch, and shutdown maintenance arrangements.

[0073] The above describes the intelligent detection method for inspection robots based on multimodal sensor fusion in the embodiments of this application. The following describes the intelligent detection system for inspection robots based on multimodal sensor fusion in the embodiments of this application. Please refer to [link / reference]. Figure 5 The present application provides a schematic diagram of the structure of an intelligent inspection robot system based on multimodal sensor fusion. The system includes: The image acquisition module 10 is used to acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image.

[0074] The boundary analysis module 20 is used to analyze the boundary gray-scale gradient and contrast intensity between pixels based on a clear surface image, and uses optical flow tracing technology to lock the dynamically changing area and determine the boundary information of potential defects.

[0075] The point cloud construction module 30 is used to extract the three-dimensional point cloud data of the corresponding area when the boundary of a potential defect is irregular and the dynamic change of the boundary exceeds a preset change threshold, and to construct the geometric shape of the defect by combining the boundary depth information to obtain the spatial distribution information of the defect.

[0076] The deformation assessment module 40 is used to assess the deformation characteristics of the defect region based on the spatial distribution information of the defect, using curvature mapping technology, and to analyze the geometric deformation and depth changes of the defect region to obtain the deformation characteristic assessment results.

[0077] The parameter quantification module 50 is used to obtain depth and height data by using laser triangulation technology for defect areas whose evolution influence range exceeds the preset range, based on the deformation characteristic evaluation results. It then calculates the defect expansion rate by combining the defect evolution time span and determines the quantitative physical parameters of the defect.

[0078] The early warning and sorting module 60 is used to generate an alarm signal when the quantified physical parameters exceed the preset safety threshold, integrate defect evolution data, and obtain a maintenance priority sort.

[0079] The probability prediction module 70 is used to calculate the future evolution probability of defects by combining the evolution rate and depth deterioration data of defects according to the quantified physical parameters and maintenance priority sorting, and to obtain the final risk assessment result.

[0080] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A multimodal sensor fusion-based intelligent detection method for inspection robots, characterized in that, The method includes: S1. Acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image; S2. Based on the clear surface image, analyze the boundary grayscale gradient and contrast intensity between pixels, and use optical flow tracing technology to lock the dynamically changing area and determine the boundary information of potential defects; S3. If the boundary of a potential defect is irregular and the dynamic change of the boundary exceeds the preset change threshold, then extract the three-dimensional point cloud data of the corresponding area, combine the boundary depth information to construct the geometric shape of the defect, and obtain the spatial distribution information of the defect. S4. Based on the spatial distribution information of the defects, the deformation characteristics of the defect region are evaluated using curvature mapping technology, and the geometric deformation and depth changes of the defect region are analyzed to obtain the deformation characteristic evaluation results. S5. Based on the deformation characteristic evaluation results, for defect areas whose evolution influence range exceeds the preset range, laser triangulation technology is used to obtain depth and height data, and the defect expansion rate is calculated by combining the defect evolution time span to determine the quantitative physical parameters of the defect. S6. If the quantified physical parameters exceed the preset safety threshold, an alarm signal is generated, defect evolution data is integrated, and maintenance priority ranking is obtained. S7. Based on the quantified physical parameters and the maintenance priority ranking, the evolution rate and depth deterioration data of the defects are fused using probabilistic update technology to calculate the future evolution probability of the defects and obtain the final risk assessment result.

2. The method according to claim 1, characterized in that, S1 includes: The inspection robot uses a multispectral sensor to scan the target surface from multiple angles, acquiring a sequence of original multimodal images covering multiple preset spectral bands. The original multimodal image sequence is subjected to hierarchical analysis, and the multimodal images are separated according to spectral channels to obtain independent image sets under different spectra; For each image in the independent image set, adaptive filtering is performed on the image boundary region to obtain a preliminarily denoised image group; If residual noise exists in the image group, cross-comparison of the multispectral channels of the image group is performed to determine the significant noise region; Local smoothing optimization is performed on the significant noise region to generate the denoised, clear surface image.

3. The method according to claim 1, characterized in that, S2 include: The clear surface image is divided into blocks, and the local change characteristics of gray-level gradient of each image block are calculated to obtain gradient distribution data. Based on the gradient distribution data, the spatial differences in boundary contrast intensity are analyzed, and areas where the difference between the contrast intensity and the surrounding background area exceeds a preset deviation threshold are identified to determine the dynamic change range. For the aforementioned dynamic range of change, optical flow tracing technology is used to analyze the motion trajectory in the image sequence to obtain preliminary boundary continuity information of potential defects; The preliminary boundary continuity information is compared across multiple frames to calculate the fluctuation of the boundary width range and identify irregular boundary regions. The pixel data of the irregular boundary region is locally magnified to obtain the boundary information of potential defects.

4. The method according to claim 1, characterized in that, S3 include: If the boundary of a potential defect is detected to be irregular and the dynamic change of the boundary exceeds a preset change threshold, then a three-dimensional scan is performed on the area corresponding to the potential defect to obtain the corresponding three-dimensional point cloud data. The three-dimensional point cloud data is registered and fused from multiple perspectives, and the three-dimensional structure is analyzed in combination with the boundary depth information to construct the preliminary geometric shape of the defect and determine the basic shape information of the defect. Based on the preliminary geometric shape, the spatial distribution characteristics of the defects are analyzed, and point cloud segmentation technology is used to divide the defect distribution area to obtain the defect location range; For the location range of the defect, multi-view depth information is fused to perform multi-angle registration and optimization, and the three-dimensional geometric features of the defect are reconstructed to obtain the spatial distribution information of the defect.

5. The method according to claim 1, characterized in that, S4 includes: Based on the spatial distribution information of the defects, the initial geometric data of the surface region corresponding to the defects are obtained, and the deformation amplitude is spatially quantified by curvature mapping technology to obtain deformation quantification data. Based on the deformation data, analyze the geometric deformation and depth variation characteristics of the defect area to determine the range of influence of defect evolution; Based on the aforementioned evolutionary influence range, the depth deterioration pattern is analyzed by combining time-series depth change data, and compared with the preset standard defect evolution rate change trend to obtain the evolution trend determination result. By integrating the deformation data, the range of evolutionary influence, and the evolutionary trend determination results, a deformation feature evaluation result is generated.

6. The method according to claim 1, characterized in that, S5 include: Based on the deformation characteristic evaluation results, the defect area whose evolution influence range exceeds the preset range is identified as the target area; The target area is scanned with high precision using laser triangulation technology to obtain the depth and height data at the current moment; Retrieve historical depth and height data of the target area and calculate the difference in depth change between the current time and the historical time. Based on the depth variation difference and the evolution time span of the defect, the defect propagation rate is calculated; By combining the depth and height data with the defect propagation rate, the quantitative physical parameters of the defect are determined.

7. The method according to claim 6, characterized in that, The quantitative physical parameters for determining the defect include: A time-series analysis was performed on the defect propagation rate to determine the dynamic trend of the defect propagation rate. Based on the dynamic change trend, the target area is divided into layers, the degree of evolutionary impact of each area is analyzed, and the risk level is classified. For high-risk areas where the degree of evolutionary impact exceeds a preset threshold, a high-density supplementary scan is performed to extract detailed depth and height data of the high-risk areas, update the high-precision expansion rate of the defects, and combine the detailed depth and height data with the high-precision expansion rate to determine the final quantitative physical parameters of the defects.

8. The method according to claim 1, characterized in that, S6 include: If the quantized physical parameters exceed a preset safety threshold, an alarm signal is triggered; Extract the historical evolution history of defects, simultaneously acquire environmental impact factor data from the inspection site, and integrate them to generate a comprehensive impact analysis basis; Based on the comprehensive impact analysis, the evolutionary history was compared over time, and combined with changes in environmental factors, the driving factors of defect evolution were identified, and high-risk monitoring areas were determined. Based on the high-risk monitoring areas, the maintenance priority order is adjusted and determined, and a corresponding resource allocation priority order is generated synchronously.

9. The method according to claim 1, characterized in that, S7 includes: Based on the maintenance priority ranking, defects with a priority higher than the preset priority threshold are selected as target evaluation defects, and the evolution rate change data and evolution depth deterioration data of the corresponding defects are obtained according to the quantified physical parameters. The evolution rate change data and the evolution depth deterioration data are fused using a probabilistic update technique to calculate the future evolution probability of the defect; If the future evolution probability shows a continuous upward trend, the future evolution probability is weighted and corrected to obtain the corrected evolution probability; Based on the modified evolution probability, a defect risk distribution map is generated by comparing and matching the results through a preset risk mapping table, thereby obtaining the final risk assessment result.

10. A multimodal sensor fusion-based intelligent detection system for inspection robots, used to implement the method as described in any one of claims 1 to 9, characterized in that, The system includes: The image acquisition module is used to acquire the original multimodal image sequence of the target surface through a multispectral sensor and perform adaptive filtering to obtain a clear surface image; The boundary analysis module is used to analyze the boundary grayscale gradient and contrast intensity between pixels based on the clear surface image, and to use optical flow tracing technology to lock the dynamically changing area and determine the boundary information of potential defects. The point cloud construction module is used to extract the 3D point cloud data of the corresponding area when the boundary of a potential defect is irregular and the dynamic change of the boundary exceeds a preset change threshold. It then combines the boundary depth information to construct the geometric shape of the defect and obtain the spatial distribution information of the defect. The deformation assessment module is used to assess the deformation characteristics of the defect region based on the defect spatial distribution information using curvature mapping technology, and to analyze the geometric deformation and depth changes of the defect region to obtain the deformation characteristic assessment results. The parameter quantification module is used to obtain depth and height data by using laser triangulation technology for defect areas whose evolution influence range exceeds the preset range, based on the deformation feature evaluation results, and to calculate the defect expansion rate by combining the defect evolution time span, thereby determining the quantitative physical parameters of the defect. The early warning and sorting module is used to generate an alarm signal when the quantified physical parameters exceed a preset safety threshold, integrate defect evolution data, and obtain a maintenance priority sorting. The probability prediction module is used to calculate the future evolution probability of the defect based on the quantified physical parameters and the maintenance priority sorting, and to obtain the final risk assessment result by fusing the evolution rate and depth deterioration data of the defect using probability update technology.

Images