Intelligent risk identification method for low-voltage uninterruptible operation site based on multi-modal image identification

Abstract

The invention discloses an intelligent risk identification method for a low-voltage uninterruptible operation site based on multi-modal image identification, and the method comprises the following steps: constructing three-dimensional point cloud data based on a depth map, carrying out the normal direction consistency analysis of each frame of point cloud, extracting the change value of a normal included angle between continuous patches, and carrying out the calculation of the change value of the normal included angle; a spatial gradient tensor is generated to characterize local geometry fluctuations. According to the method, a distortion region is identified by constructing a structural integrity scoring matrix and fusing point cloud density, a distortion feature vector is generated and mapped into a point cloud distortion index, and a perception-feedback closed-loop mechanism is established. When the distortion index exceeds the threshold value, the system automatically adjusts the infrared pulse parameter, enhances the perception capability of a low reflection area, significantly improves the three-dimensional reconstruction stability and risk identification accuracy in complex environments such as night and shadow, and has a high engineering practical value.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power operation safety monitoring and intelligent identification, and particularly relates to a low-voltage live-line operation site intelligent risk identification method based on multi-modal image recognition. BACKGROUND

[0002] The low-voltage live-line operation site intelligent risk identification based on multi-modal image recognition refers to, in a low-voltage live-line operation environment, using multiple image sensing modes such as RGB visible light, infrared thermal imaging and depth map, performing stereoscopic and multi-angle image acquisition and fusion processing on the operation site through a terminal device such as AR smart glasses, using a locally deployed deep learning model to perform real-time identification and judgment on the wearing condition of safety protection equipment of the operation personnel (such as safety helmets, insulating gloves, etc.) and high-risk elements in the operation environment (such as exposed wires, high-temperature equipment, obstacles, etc.), and once the system identifies an abnormal state, issuing a warning message to the operation personnel in the form of graphic superposition and voice broadcast, and marking a virtual danger boundary in the AR field of view, so as to improve the safety perception ability and risk response efficiency in the operation process. This method breaks the limitations of traditional single image monitoring and manual inspection, and realizes the digitization, intelligentization and visualization of operation site risk identification.

[0003] The prior art has the following disadvantages: under low-light interference conditions, especially in night operation, shadow shielding or insufficient indoor lighting, the active projection of structured light or infrared signals of the depth camera for acquiring three-dimensional space information of the operation site may not be accurately captured by the receiver due to insufficient environmental light reflection or increased light noise, resulting in failure of distance data calculation in the corresponding area of the depth map. In this case, the point cloud data generated by the system is prone to point cloud distortion in local areas, including collapse (i.e. the surface geometry is incorrectly identified as a low-lying area), distortion (i.e. the spatial structure is distorted), or void (i.e. the effective depth information is missing, forming a data blank area). Such distortion will seriously interfere with the real reconstruction of the spatial relationship between the operation personnel and the site equipment, and may lead to misidentification or missed identification of the personnel and live bodies, safety boundaries or high-risk areas, thereby causing risk prompt failure or safety judgment deviation, and potential hazards to the safety of the personnel in live-line operation. SUMMARY

[0004] In order to overcome the above-mentioned shortcomings of the prior art, the application provides a low-voltage live-line operation site intelligent risk identification method based on multi-modal image recognition, which constructs a structure integrity score matrix and fuses point cloud density to identify abnormal areas, generates an abnormal feature vector and maps it to a point cloud abnormality index, and establishes a perception-feedback closed loop mechanism.

[0005] The technical scheme adopted by the present application to solve its technical problems is: a low-voltage uninterrupted operation site intelligent risk identification method based on multi-modal image recognition, comprising the following steps: S1, constructing three-dimensional point cloud data based on a depth map, performing normal direction consistency analysis on each frame of point cloud, extracting the normal angle change value between continuous patches, generating a spatial gradient tensor to represent local geometric morphological fluctuations; S2, constructing a local consistency scoring model based on the spatial gradient tensor, calculating the deviation between the normal direction of each patch and the average normal direction of its periphery, and normalizing it to generate a structure integrity scoring matrix; S3, fusing the scoring matrix and the point cloud density distribution calculated from the three-dimensional point cloud data to identify the overlapping part of the normal mutation area and the low-density hollow area as the preliminary judgment result of the point cloud distortion area; S4, extracting the spatial distribution range, shape complexity and scale ratio parameters of the point cloud distortion area, constructing a distortion feature vector, and performing weighted fusion according to a preset weight function; S5, inputting the weighted and fused distortion feature vector into an exponential mapping function to generate a single-valued point cloud distortion index, which is used to evaluate the stability of the point cloud structure; S6, comparing the point cloud distortion index with a reference threshold value, when it is greater than the threshold value, adjusting the infrared pulse width of the depth camera according to the distortion area feedback to match the infrared energy and the target surface reflection ability, and enhance the detectability and anti-interference ability of the echo signal.

[0006] Preferably, step S1 comprises the following steps: Obtain the depth map data collected by the depth camera, and convert the pixel-level depth information into three-dimensional point coordinates through the depth and camera internal parameter model to construct point cloud data; Each frame of point cloud performs triangular patch reconstruction, calculates the normal vector of each patch, and extracts the normal angle change value between adjacent patches; Construct a spatial gradient tensor based on the normal change value, which is used to represent local geometric fluctuations; Z-score normalization and local variance sliding window smoothing processing are performed on the spatial gradient tensor to output the normalized tensor result.

[0007] Preferably, step S2 comprises the following steps: Construct a local neighborhood set for each triangular patch, and extract the normal vectors of each patch and the neighborhood patches; Calculate the angle deviation value between the normal direction of the target patch and the average normal direction of the neighborhood, and normalize the deviation value to 0 to 1; Form a structure integrity scoring matrix, and perform spatial Gaussian smoothing processing on the scoring matrix to filter out local extreme points and retain continuous trends.

[0008] Preferably, step S3 comprises the following steps: The original point cloud data is voxelized, the point density in each voxel unit is calculated, and Gaussian weighted smoothing is performed to generate a density atlas; The normalized structural integrity score matrix is spatially mapped and bidirectionally interpolated with the density atlas; The potential distortion region is identified based on a fusion function of the high value of the structural score and the low value of the density value, and false positives are excluded through neighborhood consistency verification; The distortion region labeling result is output for subsequent processing and visual display.

[0009] Preferably, in step S4, the spatial distribution range parameter, shape complexity parameter and scale ratio parameter of the distortion region are extracted, including the bounding box volume, curvature variation coefficient, concave-convex area ratio and principal axis proportion coefficient.

[0010] Preferably, step S4 specifically comprises the following steps: The spatial distribution range parameter, shape complexity parameter and scale ratio parameter of the distortion region are extracted; The parameters are combined to construct a distortion feature vector, and a corresponding weight function is set according to the sensitivity of each parameter in distortion discrimination; The distortion feature vector is linearly weighted and synthesized to generate a fusion risk score as the basis input for point cloud distortion index calculation.

[0011] Preferably, step S5 comprises the following steps: The weighted and fused distortion feature vector is normalized to construct an index mapping input structure and input into a nonlinear index mapping function; The mapping operation is performed to output a point cloud distortion index normalized to the range of 0 to 1, which is used to represent the degree of structural abnormality; The point cloud distortion index is used as an input index for subsequent safety warning and infrared parameter adaptive adjustment.

[0012] Preferably, the point cloud distortion index is compared with a preset reference threshold value, and when the point cloud distortion index is greater than the reference threshold value, the infrared pulse width of the depth camera is dynamically adjusted based on the feedback information of the distortion region as follows: According to the generated point cloud distortion index and the preset reference threshold value, a normalized energy response factor reflecting the current distortion intensity is constructed, which is used to measure the change trend of the distortion region's demand for infrared energy, and the calculation expression is as follows: , In the formula, is the point cloud distortion index, is the point cloud distortion index reference threshold value, is the distortion energy response factor; After obtaining the distortion energy response factor Then, the optical physical characteristics of the target surface in the construction operation environment are combined with the actual echo quality, and an infrared pulse dynamic adjustment factor is further constructed to quantize the adjustment requirement of the infrared pulse width, and the calculation expression is as follows: ， In the formula, is the reflectivity dispersion of the target surface, is the infrared echo signal residual, is the reflectivity dispersion adjustment weight, is the infrared echo signal residual weight, is the infrared pulse dynamic adjustment factor; After the infrared pulse dynamic adjustment factor The pulse width of the infrared projection unit of the depth camera is dynamically updated to realize the perception compensation and signal enhancement of the abnormal area, and the update formula of the pulse width is as follows: ， In the formula, is the current cycle infrared pulse width, is the adjustment sensitivity coefficient, is the next cycle infrared pulse width, is the hyperbolic tangent function.

[0013] Compared with the prior art, the beneficial effects of the present application are: The present application starts from multi-dimensional geometric features, constructs a structure integrity score matrix, and fuses point cloud density to identify distortion areas, forms a distortion feature vector with spatial perception capability, and then generates a quantifiable point cloud distortion index through exponential mapping, and establishes a closed-loop mechanism from perception analysis to feedback adjustment. When the system detects that the distortion index is higher than the reference threshold, the infrared pulse parameters of the depth camera can be adjusted adaptively based on the area feedback to enhance the signal response capability to low reflection or interference areas, so as to continuously provide stable and reliable three-dimensional information in complex operation environments. This scheme significantly improves the robustness and real-time response capability of the risk identification system in night operation, shadow shielding and scenes with large material reflection difference, effectively reduces the misidentification and missed identification rate, ensures the spatial safety perception and operation decision accuracy of the operation personnel, and has significant engineering practical value and popularization prospect. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description.

[0015] Figure 1A flowchart of a low-voltage live-line work site intelligent risk identification method based on multi-modal image recognition of an embodiment of the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0017] In the description of the present application, it should be noted that the terms “center”, “upper”, “lower”, “left”, “right”, “vertical”, “horizontal”, “inner”, “outer” and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0018] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms “mounting”, “connection”, “connecting” should be understood in a broad sense, for example, can be fixed connection, can also be detachable connection, or integrally connected; can be mechanical connection, can also be electrical connection; can be directly connected, can also be indirectly connected through an intermediate medium, can be the communication inside two elements. For a person of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In addition, the technical features involved in the different embodiments of the present application described below can be combined with each other as long as they do not conflict with each other.

[0019] As Figure 1 shown, the low-voltage live-line work site intelligent risk identification method based on multi-modal image recognition of an embodiment of the present application comprises the following steps: S1, constructing three-dimensional point cloud data based on a depth map, performing normal direction consistency analysis on each frame of point cloud data, extracting the normal angle change value between continuous patches, and constructing a spatial gradient tensor for representing the fluctuation degree of local geometric morphology; Aiming at the point cloud distortion problem caused by low light interference in the multi-modal image acquisition environment of low-voltage non-stop power operation site, a spatial geometric structure analysis method based on depth map is proposed to construct three-dimensional point cloud data from the depth map and perform normal direction consistency analysis on the point cloud data. Further, the normal angle change value between continuous patches is extracted to construct a spatial gradient tensor to quantitatively describe the fluctuation degree of local geometric morphology, thereby providing accurate geometric index support for point cloud distortion identification and subsequent adaptive processing. The technical implementation process includes the following steps: Firstly, the depth map data collected by the depth camera is obtained. The pixel-level depth information in the depth map is used to complete the coordinate conversion from two-dimensional image to three-dimensional space point through the depth-camera internal parameter model, and then the three-dimensional point cloud structure of the operation site is reconstructed. In the reconstruction process, in order to ensure the geometric continuity and surface smoothness of the point cloud data, the grid method is preferably adopted for triangular patch reconstruction of the original point cloud, so that the topological relationship between points and points is clear, and the normal calculation basis can be established between patches. After completing the point cloud grid construction, the normal vector of each triangular patch in each frame of point cloud data is calculated. The normal calculation adopts the standard cross product method, and the three vertex vectors of each patch are substituted into the vector cross product formula to obtain the normal vector direction of the patch in the three-dimensional space.

[0020] Secondly, after the preliminary normal vector extraction is completed, the direction consistency of the normal vector of each patch and its directly adjacent patch is analyzed. Specifically, the geometric direction continuity is represented by calculating the cosine value of the normal vector between any pair of adjacent patches. In this embodiment, in order to improve the calculation efficiency and suppress noise interference, five nearest patches are selected as the local neighborhood range, and the average cosine error is used as the normal consistency index of the patch, so that the patches in the edge mutation area or light interference area in the spatial geometric structure can be effectively distinguished. The normal angle change value as a key geometric change feature can reflect whether there is a spatial mutation or structural anomaly in the point cloud region, and is especially suitable for detecting geometric distortion caused by low light or poor infrared reflection.

[0021] Thirdly, based on the above-obtained normal angle change value of the facet, a spatial gradient tensor is further constructed to represent the fluctuation degree of the local geometric morphology from the perspective of the tensor. The tensor is based on the local curvature change and the normal change of each point in the three-dimensional space, and a multi-dimensional tensor structure is established, wherein each dimension respectively represents the normal difference rate between the facets, the normal distribution covariance of the adjacent topological structure, and the gradient direction projection rate of the normal change along the spatial coordinate axis. Through the spatial gradient tensor, the surface undulation and the abruptness of the entire point cloud structure can be modeled in a high dimension on a macro level, and then the local distortion can be marked with high sensitivity. In particular, when the regional void or reconstruction failure caused by infrared signal attenuation occurs in the depth map, the corresponding gradient abnormal value in the tensor will be significantly increased, thereby forming an obvious distortion mark in the feature space.

[0022] Finally, in order to ensure that the extracted spatial gradient tensor can truly and accurately reflect the local distortion degree of each frame of point cloud in the three-dimensional structure, the tensor result needs to be standardized and noise suppressed. In the embodiment, the Z-score normalization method is used to standardize the fluctuation data in each tensor dimension, so that the gradient values from different regions and different scales are comparable. At the same time, a local variance sliding window mechanism is introduced to perform a local smoothing filtering operation in the gradient tensor, effectively eliminating extreme gradient noise values caused by acquisition errors or single-point drift. After completing the standardization process, the spatial gradient tensor can be used as the input basis in the subsequent distortion region identification, distortion feature vector extraction, and adaptive control module, to realize dynamic monitoring and structural stability measurement of the three-dimensional reconstruction quality of the work site.

[0023] In summary, the method of constructing a point cloud based on a depth map, performing normal direction consistency analysis, and constructing a spatial gradient tensor described in the present step not only has strong geometric sensitivity and spatial analysis capability, but also can realize higher-precision distortion detection by combining the structural abnormal features in the low-light scene, thereby providing a stable and reliable geometric data basis and discrimination basis for the intelligent risk identification system of the entire low-voltage uninterrupted power operation, and has significant technical creativity and application value.

[0024] S2, a local consistency scoring model is established based on the spatial gradient tensor, the deviation between the normal direction of each facet and the average normal direction of the surrounding facets is calculated, and the deviation is normalized to generate a structural integrity scoring matrix; To realize the quantitative and structured evaluation of the geometric integrity of each local region in the three-dimensional point cloud structure, a construction method of a local consistency scoring model based on a spatial gradient tensor is provided. The method can fully utilize the spatial gradient tensor information calculated in the previous step, generate a set of structural integrity evaluation matrix at the patch level through normal direction consistency analysis, and thus provide more refined basic data support for the identification of distorted regions. The specific implementation process includes the following four steps: First, for each triangular patch identified by the spatial gradient tensor, a local neighborhood set is constructed in the three-dimensional point cloud. The neighborhood set is preferably composed of directly adjacent patches that share at least one edge with the current patch, with a typical number of 5 to 8, to ensure that the analysis range is local enough to reflect microscopic geometric features and has certain statistical robustness. After the neighborhood is established, the normal vectors of each target patch and its adjacent patches are uniformly extracted, and all normal vectors are standardized in the unit sphere to eliminate the scale effect and ensure the rationality and consistency of the direction comparison.

[0025] Second, the angle deviation value between the normal direction of the target patch and the average normal direction of its adjacent patches is calculated. To avoid misjudgment caused by the extreme sensitivity of the simple angle difference, the present invention uses the cosine of the angle calculation method, that is, the cosine value of the angle between the normal directions is calculated by vector dot product, and is further converted into the deviation angle as the basic index of the patch consistency. In the specific calculation process, first, the weighted average of all normal vectors of the neighborhood patches is calculated, and the weight is determined according to the relative spatial distance from the target patch, and the weight is higher for closer distance, to highlight the contribution of local features; then the angle between the normal vector of the target patch and the weighted average vector is calculated, and the obtained deviation angle is the initial consistency score value of the patch. The smaller the score value, the closer the normal direction to the overall trend of the neighborhood, i.e., the more stable the local geometric structure.

[0026] Third, the initial consistency score values of all patches are normalized to form a standardized structural integrity score matrix. In the present invention, the linear normalization method is used to map the deviation angle to the closed interval of 0 to 1, where 0 represents complete consistency and 1 represents complete inconsistency; this normalization process helps to compatible and joint modeling between the scoring matrix and other module feature indicators. At the same time, to enhance the resolution ability of the scoring results to geometric details, a local adaptive normalization parameter is introduced, which dynamically adjusts the normalization range according to the spatial gradient tensor amplitude of different regions, so that more structural change information is retained in high curvature or high gradient regions, and the differentiation effect is enhanced in flat regions, thereby effectively improving the sensitivity and discrimination of the scoring matrix in structural anomaly detection.

[0027] Finally, the normalized score results are filled into the structural integrity score matrix to form a set of local structure quality evaluation system at the surface sheet level. The score matrix maintains consistency with the point cloud surface sheet topology in space and can be directly visualized as a spatial heat map for visual analysis, or can be used as an input data source for subsequent distortion area extraction and feature vector construction. In order to enhance the stability and noise resistance of the score matrix, the present application further performs spatial Gaussian smoothing processing on the score matrix to filter out local isolated extreme points and retain the trend changes in the continuous region. The score matrix plays a key role in the system as an intermediate expression, which not only has a significant response to the distortion caused by the depth map error calculation under low light conditions, but also can accurately quantify the structural discontinuity caused by occlusion, reflection, curved surface reconstruction and other complex scenes, and is the key basis for realizing the subsequent point cloud distortion index calculation and infrared projection adaptive adjustment control.

[0028] In summary, the present application realizes accurate description and quantitative expression of the local surface sheet geometric stability of three-dimensional point cloud by constructing a local consistency scoring model based on spatial gradient tensor, has strong directionality, high adaptability and strong computability, improves the detectability and positioning accuracy of the point cloud distortion area, and provides effective support for depth information correction in low light environment, and has obvious innovation and engineering application value.

[0029] S3, the structural integrity score matrix and the density distribution of the original point cloud are fused and analyzed to identify the overlapping part of the normal mutation area and the low-density hollow area as the preliminary judgment result of the point cloud distortion area; In order to realize efficient detection of geometric structure abnormal area in three-dimensional point cloud data, a method of fusing and analyzing the structural integrity score matrix and the original point cloud density distribution is proposed, which accurately identifies the key area with distortion risk in the point cloud by considering the normal mutation feature and the sparse feature of the spatial point distribution. The method overcomes the limitations of traditional single feature judgment, improves the accuracy and robustness of distortion detection based on multi-dimensional data fusion. The technical implementation process includes the following four steps: First, the density distribution of the original point cloud data is calculated to obtain the point density value corresponding to each triangular surface sheet or grid element. In order to ensure the locality and resolution of the density distribution, the present application adopts a fixed radius voxel division method to divide the three-dimensional space into multiple regular cubic elements, counts the number of points contained in each voxel element, and takes the number as the local density index of the region. At the same time, considering the problem of error extreme value in the sparse area of point cloud, a density smoothing strategy is introduced to perform Gaussian weighted average processing on the density values of adjacent voxels to eliminate noise caused by accidental defects or device jitter, thereby constructing a more stable density distribution atlas.

[0030] Secondly, the normalized structure integrity score matrix is spatially mapped and matched with the point cloud density map to establish a corresponding relationship. Specifically, on the basis of the spatial coordinates of the point cloud triangular facets or sampling points, each score unit in the structure score matrix is aligned with the corresponding density voxel unit to ensure consistency in the spatial dimension. To avoid registration errors caused by different resolutions, the present application adopts a bidirectional spatial interpolation strategy, that is, while aligning the structure score to the density map, the reverse projection of the density value in the structure score grid is performed to achieve bidirectional compensation and accurate alignment, ensuring the accuracy of subsequent fusion analysis. This mapping process establishes the basis for joint analysis of the local structure quality and spatial distribution characteristics of the point cloud.

[0031] Third, on the basis of the above spatial alignment, the structure score and the density value are jointly fused and analyzed to determine the distortion region. The specific steps are as follows: first, the joint index of the structure integrity score and the density value of each facet is calculated, and the present application designs a multi-dimensional fusion function, taking the high value of the structure score (indicating severe normal mutation) and the low value of the density value (indicating that the number of points in this region is sparse) as risk amplification factors. When both conditions are met, the facet is marked as a potential distortion region. This joint index can be realized by logical product, weighted average or index combination, and can be flexibly selected according to the actual application scene. The present application preferably uses the exponential superposition method to give stronger recognition weight to high-risk areas, thereby improving the overall detection sensitivity. In the fusion determination process, a neighborhood consistency verification mechanism is also introduced, which requires that a plurality of adjacent facets simultaneously satisfy the distortion determination condition to prevent single-point misjudgment caused by accidental errors.

[0032] Finally, the identified distortion region result is spatially labeled and output to form a preliminary point cloud distortion region discrimination result and serve as the input basis for subsequent distortion feature vector extraction and dynamic adjustment mechanism. To improve the spatial visualization effect, the present application uses the heat map superposition method to map the fusion analysis result to the point cloud model and intuitively display the high-risk distortion region on the AR terminal or visualization console used by the operator. At the same time, the discrimination result can also be used to dynamically update the point cloud reconstruction strategy, such as automatically adjusting the scanning angle, increasing the image sampling frequency, etc. in the distortion region to realize real-time correction and self-optimization. In addition, to further improve the stability of the discrimination result, the present application also supports time series filtering analysis of multiple frame point cloud distortion region results to exclude temporary fluctuation data interference and only retain the continuous distortion trend as the final output, improving the system reliability.

[0033] In summary, by fusing the structural integrity score and the point cloud density distribution characteristics, and introducing spatial registration, multi-dimensional index fusion and neighborhood verification mechanism, the application realizes high-precision preliminary discrimination of point cloud distortion area. This method has the technical advantages of strong structure sensitivity, accurate density analysis and wide adaptability, can significantly improve the stability of three-dimensional data and the intelligent identification ability of operation safety risk under low light conditions, and embodies high technical creativity and engineering application value.

[0034] S4, extracting spatial distribution range, shape complexity and scale ratio parameters for the distortion area, constructing a distortion feature vector, and setting a weight function according to the importance of each parameter to perform weighted fusion processing on the distortion feature vector; In order to further improve the accuracy and classification ability of point cloud distortion recognition, a high-dimensional geometric feature modeling method for distortion area is proposed. This method performs fine analysis on the identified distortion area, extracts its key geometric properties such as spatial distribution range, shape complexity and scale ratio, constructs a discriminative distortion feature vector, and introduces a multi-dimensional weight function to assign different weights to each attribute, finally forms a weighted fusion index for quantifying the distortion degree, providing core input basis for subsequent distortion index calculation and system adaptive control. This method embodies the complete modeling process from geometric information abstraction to structural index fusion, and its implementation steps include the following four aspects: First, for each identified point cloud local segment as a distortion area, the spatial distribution range extraction operation is performed. In this embodiment, the minimum bounding box (Minimum Bounding Box) algorithm is used to fit the three-dimensional space envelope of all points in the distortion area, calculate the length, width and height parameters of the region in X, Y and Z three axial directions, and obtain the spatial occupied volume and maximum circumscribed sphere radius according to the above parameters, so as to fully characterize the spatial expansion ability of the region. This index can be used to judge whether the distortion area has a serious spatial diffusion tendency, and is of great significance for distinguishing local disturbance and structural collapse. In addition, the center coordinate position of the distortion area is also recorded in this step, which is used for subsequent spatial clustering and distribution pattern recognition.

[0035] Secondly, the shape complexity of each distortion region is calculated. As an important indicator reflecting the degree of geometric distortion of the region, the curvature variation coefficient, surface roughness index and concave-convex area ratio are used as the main parameters in this embodiment. The curvature variation coefficient is obtained by statistical analysis of the normal curvature of each face in the region, and is used to measure the fluctuation strength of the surface. The surface roughness index is calculated based on the average residual of the point to the local fitting plane, and reflects the overall flatness. The concave-convex area ratio refers to the ratio of the area of the concave region to the total area, and is used to reveal whether there are abnormal shapes such as collapse or hole in the region. The above parameters together constitute the shape descriptor of the distortion region, and are an important part of the subsequent feature vector.

[0036] Thirdly, the scale ratio parameters of each distortion region are extracted. Scale ratio is used to describe the difference in extension of the distortion region in each spatial dimension, mainly including the aspect ratio (the ratio of the length of X and Y axes), the height ratio (the ratio of the height of Z axis to the horizontal dimension), and the flatness index (the principal axis proportion coefficient obtained by principal component analysis). Through these parameters, different types of distortion modes such as stretching deformation, lateral expansion or vertical collapse can be effectively identified. For example, when a region shows an extreme longitudinal elongation trend, it may indicate that the region is caused by image stretching artifacts due to insufficient scanning angle; while flat regions are more likely to be caused by insufficient projection energy. As the geometric contour features of distortion morphology, scale ratio parameters play an important role in distinguishing the model.

[0037] Finally, the spatial distribution range, shape complexity and scale ratio parameters extracted above are combined into a unified distortion feature vector, and a weight function is set according to the sensitivity of each parameter in actual distortion discrimination for weighted fusion processing. Specifically, according to experimental data statistics and field feedback experience, different weight values are assigned to each type of parameter in this embodiment, for example, higher weights are given to surface roughness and concave-convex area ratio, and medium weights are given to enclosed volume and flatness. Finally, a weight coefficient vector is formed. In the fusion process, linear weighted synthesis method is used, and the fusion risk score of the distortion region is obtained by multiplying each feature parameter and its corresponding weight and then summing them. This score will be used as the basis for constructing the point cloud distortion index in the subsequent module, reflecting the possible distortion degree of the current region in the geometric structure restoration. At the same time, the feature vector can also be used for clustering analysis, abnormal pattern classification and time trend tracking of the distortion region, providing data support and model basis for building a more intelligent and adaptive risk perception system.

[0038] In summary, the present application realizes multi-angle and multi-level quantitative modeling of the point cloud distortion area by constructing a distortion feature vector that fuses spatial distribution, geometric complexity, and scale structure, and performing weighted processing according to a multi-dimensional weight function. This method takes into account global morphology and local details, has strong interpretability, high classification ability, and wide adaptability, and has important application value and obvious technical innovation in real-time risk identification systems in low-voltage uninterrupted operation environments.

[0039] S5, input the weighted and fused distortion feature vector into an exponential mapping function to calculate a single-valued point cloud distortion index representing the abnormality degree of the point cloud data structure restoration, which is used for quantitative evaluation of the point cloud structure stability; To realize quantitative evaluation of the three-dimensional point cloud structure integrity and stability, a point cloud distortion index construction method based on distortion feature vector exponential mapping is proposed. This method performs unified reduction and mapping on the weighted and fused multi-dimensional distortion feature vector to output a single-valued distortion index, which is used to measure the abnormality degree of the point cloud structure in the restoration process, and provides intuitive and controllable quantitative basis for subsequent intelligent adjustment and risk response modules. The technical process includes the following four key steps: First, the weighted and fused distortion feature vector generated in the previous stage is received and preprocessed. The feature vector is composed of multiple dimensional features such as spatial distribution range, geometric shape complexity, and scale ratio, and after weighted fusion, a multi-dimensional vector with comprehensive discriminability is formed. In this embodiment, to ensure the input normalization and adaptability of the mapping function, the dimensional values of the input feature vector are standardized to ensure that they are normalized to the same numerical interval range and that the comparison and processing between different dimensions are consistent. In addition, to reduce the interference of individual extreme values on the overall judgment result, an outlier suppression mechanism is also introduced, which uses median replacement or local smoothing to correct abnormal parameters that deviate greatly from the median, thereby enhancing the stability and generalization ability of the distortion index.

[0040] Secondly, the input structure of the exponential mapping model is constructed, and the standardized distortion feature vector is input into the mapping function processing module. In this embodiment, the exponential mapping function is a nonlinear enhancement mechanism that can amplify the small differences in the feature vector into a significant exponential change, thereby improving the response sensitivity of the model to the boundary critical state. When constructing the input structure, the dimensions of the feature vector are arranged in order, and the loss or rearrangement between dimensions is ensured to maintain the physical meaning of each parameter in the final distortion index calculation. In addition, in order to improve the generality of the model in multiple scenarios, the present application predefines multiple exponential mapping parameter templates, which are dynamically selected according to different conditions of the working environment, so that the output distortion index has adjustability and adaptability. For example, in a low-light environment at night, the model tends to give higher response weight to low-density and curvature fluctuation features; while in a scene with serious device occlusion, the response ability to spatial distribution offset is enhanced.

[0041] Thirdly, the mapping operation of the distortion feature vector is performed, and a single-valued point cloud distortion index is output. In actual operation, the mapping function inputs the entire vector as a whole, integrates the numerical characteristics of each dimension, and outputs a real number representing the abnormality degree of the region structure after nonlinear transformation, which is called "point cloud distortion index". The index value reflects the possible deformation risk, geometric distortion or data incompleteness of the current region in the three-dimensional structure restoration process, and the larger the value, the more serious the distortion. In order to enhance the interpretability of the results and the compatibility of subsequent processing, the output point cloud distortion index value is normalized to the closed interval of 0 to 1, where 0 represents no distortion and 1 represents a high distortion risk. In order to improve the response ability of the system, the calculation frequency of the distortion index is consistent with the point cloud refresh frequency, ensuring that the updated distortion evaluation result can be obtained in each frame of point cloud data processing period, thereby supporting high-frequency real-time analysis of the scene.

[0042] Finally, the post-processing and multi-target correlation of the point cloud distortion index are completed, which are used to drive the subsequent decision and risk response of the system. In this embodiment, the distortion index is introduced into the operation safety warning module, and is jointly analyzed with historical distortion trend data, spatial position of the operator and dynamic factors of the operation environment to form a risk response determination mechanism. This mechanism compares the real-time distortion index with the preset risk threshold to realize intelligent judgment of the stability of the current point cloud structure. At the same time, the distortion index also serves as an input trigger for the depth camera working parameter adjustment module, participating in the adaptive adjustment process such as infrared pulse width adjustment and image fusion priority setting, thereby improving the perception ability and reconstruction accuracy of the acquisition system in abnormal areas. In addition, the time series of the distortion index can also be used for long-term stability trend analysis, providing an important reference for equipment maintenance, system update and fault prediction.

[0043] In summary, by inputting the weighted fused distortion feature vector into the exponential mapping function, a single-valued point cloud distortion index is generated, and the application realizes the accurate quantitative expression of structural stability anomalies. This method not only has high sensitivity, high response rate and good robustness, but also can be efficiently integrated with other system modules to form a complete point cloud structure quality evaluation and intelligent control closed-loop system, with obvious technical innovation and practical application value.

[0044] S6, compare the point cloud distortion index with the preset reference threshold, when the point cloud distortion index is greater than the reference threshold, dynamically adjust the infrared pulse width of the depth camera based on the feedback information of the distortion region, to match the relationship between the infrared emission energy and the reflection ability of the target surface on site, enhance the detectability and anti-interference ability of the echo signal, and improve the structure restoration accuracy and spatial reconstruction stability of the point cloud data; The role of this step is to realize the closed-loop control and adaptive compensation mechanism in the point cloud data acquisition process, to improve the three-dimensional structure restoration quality and spatial reconstruction stability in complex working environments such as low light, large reflection rate difference or serious interference. The point cloud distortion index, as a quantitative index of structural anomalies output by the previous analysis module, reflects the degree of possible geometric distortion, void or collapse in the current collected data. When the index is higher than the preset reference threshold, it means that the point cloud data currently acquired by the system has a perceptible structural instability, and if the default acquisition parameters are maintained, it is easy to cause distortion in subsequent risk identification and judgment. Therefore, by dynamically monitoring the distortion index, this step triggers the automatic adjustment of the infrared pulse width of the depth camera, so that the infrared projection intensity and the target surface reflection characteristics are more matched, especially when facing targets with strong light absorption, complex materials or position offset, the effective reception rate of the infrared echo signal can be significantly enhanced. This adaptive control not only effectively reduces the signal loss rate and structural distortion probability in point cloud acquisition, but also improves the continuity and consistency between multiple frames of point cloud, providing a solid foundation for high-precision spatial modeling and real-time risk assessment, and is one of the key links for the system to realize intelligent perception and dynamic optimization capability.

[0045] The specific steps of dynamically adjusting the infrared pulse width of the depth camera based on the feedback information of the distortion region when the point cloud distortion index is greater than the reference threshold are as follows: According to the generated point cloud distortion index and the preset reference threshold, a normalized energy response factor reflecting the current distortion intensity is constructed, which is used to measure the change trend of the distortion region's demand for infrared energy, and the calculation expression is as follows: , In the formula, is a point cloud distortion index, representing a single-valued distortion index obtained after feature analysis, feature fusion, and mapping function processing of the current frame point cloud data, with a value range of , is a point cloud distortion index reference threshold, representing a structural distortion tolerance reference threshold set by the system according to preset experience, historical statistics, or task scene, is a distortion energy response factor, which maps the structural distortion index to the continuous control signal intensity of the infrared projection control module, limited by , and the common output range is , which is used for subsequent combination with the infrared energy gain model; Through the above formula, the distortion energy response factor smoothly maps the growth of in logarithmic form, ensuring that the response range has both sensitivity and does not risk uncontrollable energy response in high distortion situations.

[0046] After obtaining the distortion energy response factor , the optical and physical properties of the target surface in the work environment and the actual echo quality are combined to further construct an infrared pulse dynamic adjustment factor, which is used to quantify the adjustment requirement of the infrared pulse width, and the calculation expression is as follows: , wherein is the reflectivity dispersion of the target surface, representing the difference in infrared reflectivity of different surface points in the distortion area, and is a quantitative indicator of material non-uniformity or incident angle change, which is the standard deviation of the reflectivity of multiple points in the area, and the unit is a dimensionless ratio, is the infrared echo signal residual error, representing the error between the infrared receiving signal and the theoretical modeling signal, used to measure the decrease in infrared perception quality caused by shielding, dust, shading, and external light interference, with a value range of [0, 1], and the closer to 1 indicates that the echo is severely attenuated or lost, is the reflectivity dispersion adjustment weight, controlling the influence of the reflectivity dispersion of the target surface on infrared compensation, with a value range of: , adjusted according to actual engineering tests, is the infrared echo signal residual error weight, controlling the influence of the infrared echo residual value on pulse width adjustment, used to adjust the compensation response degree of the system to signal distortion or echo attenuation, and the greater the value, the more sensitive the system is to interference phenomena at the infrared receiving end, with a value range of: , is an infrared pulse dynamic adjustment factor, representing the required infrared emission intensity enhancement of the current operation frame, the larger the value, the greater the energy input required to resist the situation of weak reflection, heavy distortion or strong signal interference, used as a control factor of the infrared projection module, determines the real-time adjustment ratio of the infrared pulse width, is the key bridge between structural distortion and perception compensation; The above steps fuse the geometric distortion intensity and the reflection physical characteristics to model and dynamically generate personalized adjustment weights , so that the infrared emission behavior has regional adaptability and target specificity, providing real-time decision basis for the control link.

[0047] After the infrared pulse dynamic adjustment factor , the pulse width of the infrared projection unit of the depth camera is dynamically updated to realize the perception compensation and signal enhancement of the abnormal area, and the update formula of the pulse width is as follows: , In the formula, is the current cycle infrared pulse width, representing the emission time length of the infrared light pulse used by the depth camera in the current frame (or cycle), usually in nanoseconds (ns) or microseconds (μs), determining the energy of the unit infrared pulse, the larger the width, the stronger the infrared energy, suitable for low reflectivity, long distance or strong interference scenes, is the adjustment sensitivity coefficient, a positive number less than 1, used to control the adjustment amplitude range of the pulse width, usually taking the value range of [0.01, 0.2], used to prevent the infrared pulse width from fluctuating greatly due to sudden changes in distortion index, providing flexibility and safety buffer for adjustment, is the next cycle infrared pulse width, the infrared pulse emission width set for the next frame or the next acquisition cycle, is the final execution parameter of the adjustment control logic; directly determines the energy output of the infrared light source, and then affects the echo signal quality and depth map acquisition accuracy, is the hyperbolic tangent function.

[0048] In the infrared pulse width dynamic adjustment mechanism proposed in the present application, the hyperbolic tangent function plays a core role in realizing the nonlinear mapping of the gain adjustment factor, thereby enhancing the flexibility and stability of the adjustment process. Since the function has the mathematical properties of smoothness, continuity, symmetry and limited upper and lower boundaries, its output is always in the interval (-1, 1), which makes it effectively suppress the violent response of the system caused by abnormal mutations when processing the infrared pulse dynamic adjustment factor . In the case of light distortion, The near-linear response makes the adjustment behavior gentle and fine, and when the distortion degree is high, the function output tends to a saturation value, avoiding the infrared pulse width from increasing indefinitely, and ensuring that the system can still maintain adjustment stability and hardware safety in a strong interference environment. Therefore, In the system, the key role of suppressing burst gain, smoothing the adjustment curve, and strengthening the controllability of adjustment is the mathematical buffer and behavior stabilizer that connects the point cloud structure analysis and the hardware execution feedback.

[0049] Through the three-stage closed-loop mechanism, the system can adjust the infrared emission behavior in real time according to the structural abnormality degree and optical feedback, improve the stability, integrity, and reliability of point cloud reconstruction in complex environments, and thus enhance the intelligent perception level and risk prevention and control capability in the overall operation scene.

[0050] Through the above-mentioned intelligent risk identification method for low-voltage uninterrupted operation site based on multi-modal image recognition, the structural restoration accuracy and spatial reconstruction stability of point cloud data can be effectively improved under low-light interference conditions. Starting from multi-dimensional geometric features, the method constructs a structural integrity score matrix, and fuses point cloud density for distortion region identification, forming a distortion feature vector with spatial perception capability, and then generating a quantifiable point cloud distortion index through exponential mapping, establishing a closed-loop mechanism from perception analysis to feedback adjustment. When the system detects that the distortion index is higher than the reference threshold, it can adaptively adjust the infrared pulse parameters of the depth camera based on regional feedback, enhancing the signal response capability to low-reflective or interference areas, thereby continuously providing stable and reliable three-dimensional information in complex operation environments. This scheme significantly improves the robustness and real-time response capability of the risk identification system in night operation, shadow shielding, and scenes with large material reflection differences, effectively reduces the false positive and false negative rates, ensures the spatial safety perception and operation decision accuracy of the operation personnel, and has significant engineering practical value and promotion prospects.

[0051] The above formulas are dimensionless numerical calculations, and the formulas are obtained by software simulation of a large amount of data to obtain a formula closest to the actual situation. The preset parameters in the formula are set by a person skilled in the art according to the actual situation.

[0052] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions within the technical scope disclosed by the present application can be easily thought of by those skilled in the art, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for intelligent risk identification in low-voltage uninterrupted power supply operations based on multimodal image recognition, characterized in that: Includes the following steps: S1: Construct 3D point cloud data based on depth map, perform normal direction consistency analysis on each frame of point cloud, extract the change value of normal angle between continuous patches, and generate spatial gradient tensor to characterize local geometric morphology fluctuation. S2, based on the spatial gradient tensor, constructs a local consistency scoring model, calculates the deviation between the normal direction of each patch and the average normal direction of its surroundings, and normalizes it to generate a structural integrity scoring matrix. S3, the point cloud density distribution calculated by fusing the scoring matrix and 3D point cloud data, identifies the overlapping part of the normal abrupt region and the low-density void region, as the preliminary discrimination result of the point cloud distortion region; S4. Extract the spatial distribution range, shape complexity and scale ratio parameters of the point cloud distortion region, construct the distortion feature vector, and perform weighted fusion according to the preset weight function; S5, input the weighted and fused distortion feature vector into the exponential mapping function to generate a single-valued point cloud distortion index, which is used to evaluate the stability of the point cloud structure. S6 compares the point cloud distortion index with a reference threshold. When it exceeds the threshold, the infrared pulse width of the depth camera is dynamically adjusted based on the feedback from the distortion area to match the infrared energy with the reflectivity of the target surface, thereby enhancing the detectability and anti-interference capability of the echo signal.

2. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, Step S1 includes the following steps: Acquire depth map data captured by a depth camera, and convert pixel-level depth information into three-dimensional point coordinates using a depth and camera intrinsic parameter model to construct point cloud data; Each frame of point cloud performs triangular patch reconstruction, calculates the normal vector of each patch, and extracts the change value of the normal angle between adjacent patches; A spatial gradient tensor is constructed based on the normal variation value to characterize local geometric fluctuations; The spatial gradient tensor is normalized by Z-score and smoothed by local variance sliding window to output the standardized tensor result.

3. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, Step S2 includes the following steps: Construct a local neighborhood set for each triangular facet and extract the normal vectors between each facet and its neighboring facets; Calculate the angular deviation between the target patch normal direction and the neighborhood average normal direction, and normalize the deviation value to between 0 and 1. A structural integrity scoring matrix is ​​generated, and spatial Gaussian smoothing is performed on the scoring matrix to filter out local extrema and retain continuous trends.

4. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, Step S3 includes the following steps: The original point cloud data is divided into voxels, the point density in each voxel is calculated and Gaussian weighted smoothing is performed to generate a density distribution map. The normalized structural integrity score matrix and density map are spatially mapped and bidirectionally interpolated and registered. Potential distortion regions are identified based on a fusion function of high structural score and low density value, and false judgments are eliminated through neighborhood consistency verification. The output of the distorted region annotation results is used for subsequent processing and visualization.

5. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, In step S4, the spatial distribution range parameters, shape complexity parameters, and scale ratio parameters of the distorted region are extracted, including the bounding box volume, curvature variation coefficient, concave-convex area ratio, and principal axis scale coefficient, respectively.

6. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, Step S4 specifically includes the following steps: Extract the spatial distribution range parameters, shape complexity parameters, and scale ratio parameters of the distorted region; The parameters are combined to construct the distortion feature vector, and the corresponding weight function is set according to the sensitivity of each parameter in distortion discrimination. A linear weighted synthesis is performed on the distortion feature vectors to generate a fusion risk score, which serves as the basic input for calculating the point cloud distortion index.

7. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, Step S5 includes the following steps: The distorted feature vector after weighted fusion is standardized, and an exponential mapping input structure is constructed and input into a nonlinear exponential mapping function; Perform mapping operations and output a point cloud distortion index normalized to the range of 0 to 1, which is used to characterize the degree of structural anomaly; The point cloud distortion index is used as an input indicator for subsequent safety early warning and adaptive adjustment of infrared parameters.

8. The intelligent risk identification method for low-voltage uninterrupted power supply operations based on multimodal image recognition according to claim 1, characterized in that, In step S6, the point cloud distortion index is compared with a preset reference threshold. When the point cloud distortion index is greater than the reference threshold, the specific steps for dynamically adjusting the infrared pulse width of the depth camera based on the feedback information of the distorted region are as follows: Based on the generated point cloud distortion index and a preset reference threshold, a normalized energy response factor reflecting the current distortion intensity is constructed to measure the changing trend of the infrared energy demand of the distorted region. The calculation expression is as follows: ， In the formula, It is the point cloud distortion index. It is the reference threshold for point cloud distortion index. It is the distortion energy response factor; In obtaining the distortion energy response factor Subsequently, combining the optical and physical properties of the target surface in the working environment with the actual echo quality, an infrared pulse dynamic adjustment factor is further constructed to quantify the adjustment requirements of the infrared pulse width. The calculation expression is as follows: ， In the formula, It is the dispersion of the reflectivity of the target surface. It is the residual of the infrared echo signal. It is the reflectivity dispersion adjustment weight. It is the weight of the infrared echo signal residual. It is the infrared pulse dynamic adjustment factor; Infrared pulse dynamic adjustment factor Then, the pulse width of the infrared projection unit of the depth camera is dynamically updated to achieve perception compensation and signal enhancement for abnormal areas. The formula for updating the pulse width is as follows: ， In the formula, It is the width of the infrared pulse in the current cycle. It is to adjust the sensitivity coefficient. It is the width of the infrared pulse in the next cycle. It is the hyperbolic tangent function.

Images