A substation maintenance safety early warning method based on a real scene three-dimensional model

Abstract

The application relates to a substation maintenance safety early warning method based on a real scene three-dimensional model. The method comprises the following steps: obtaining three-dimensional point cloud and image data of a substation scene in a manner that a UAV is combined with tilt photography and laser scanning, and fusing the two to establish a three-dimensional real scene model; presetting a plurality of dangerous points and operation areas in the three-dimensional real scene model; establishing a three-dimensional bounding box or capsule for each mobile object and dangerous area on site, and calculating the shortest axial distance of each mobile object and dangerous area on X, Y and Z three axial directions; when the distance on any axial direction is smaller than a preset threshold value, it is determined that there is a potential collision risk. The method is fast in response, high in precision, and can effectively improve the substation operation safety guarantee level.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology for substation operating environments, specifically to a safety early warning and collision detection method. Background Technology

[0002] Substation environments contain high-voltage live equipment and large hoisting machinery, posing extremely high safety requirements for on-site personnel and mobile equipment. Currently, most substations rely on traditional surveillance cameras or simple pre-set warning zones for safety monitoring, methods that struggle to accurately assess dynamic risks in complex environments. Existing technologies attempt to visualize the site using 3D models, but these typically suffer from insufficient modeling accuracy, difficulties in data fusion, and poor real-time performance. Furthermore, existing collision detection algorithms often rely on full-field scanning or Cartesian distance calculations, resulting in high computational costs, low efficiency, and an inability to adequately address the diverse risks associated with abnormal personnel postures or proximity to hazardous areas. Therefore, a high-precision, high-efficiency safety early warning and collision detection method is urgently needed to adapt to the characteristics of substation operating environments and achieve accurate monitoring and timely alarms for personnel and equipment in multiple scenarios. Summary of the Invention

[0003] The present invention is made in view of the above-mentioned situation of the prior art, in order to solve or alleviate one or more problems existing in the prior art, and at least provide an advantageous alternative.

[0004] To achieve the above objectives, this invention provides a substation maintenance safety early warning method based on a real-scene 3D model, comprising the following steps: acquiring 3D point cloud and image data of the substation scene through a combination of UAV oblique photography and laser scanning, and fusing the two to establish a 3D real-scene model; pre-setting several danger points and work areas in the 3D real-scene model; establishing 3D bounding boxes or capsules for each moving object and danger area on site, calculating the shortest axial distance of each moving object and danger area in the X, Y, and Z axes, and determining a potential collision risk when the distance in any axis is less than a preset threshold.

[0005] The embodiments of this invention combine multi-source data acquisition, spatial modeling, and intelligent early warning to establish a risk identification and alarm system with a clear structure, fast response, and user-friendly interface, achieving the following significant technical effects: By combining drones and laser scanning for modeling, the overall structure and key details are taken into account, significantly improving the accuracy and completeness of scene modeling. By introducing a spatial indexing and labeling mechanism, the computational load of collision detection is effectively reduced, enabling high-frequency, low-latency risk assessment. The system is integrated into a 3D visualization platform, is compatible with multiple terminals, and allows operators and managers to intuitively perceive operational risks, supporting real-time intervention and post-event analysis.

[0006] In summary, this invention provides a safety early warning and collision detection technology solution that can be widely applied in high-risk work environments such as substation maintenance, and has good engineering adaptability and promotion value. Attached Figure Description

[0007] Figure 1 This is a flowchart illustrating a safety early warning and collision detection method for a substation operating environment according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the construction of a 3D reality model; Figure 3 This is a schematic diagram of dynamic data mapping for personnel / equipment; Figure 4 This is a schematic diagram of the danger point triggering and early warning mechanism; Figure 5 This is a schematic diagram of 3D bounding box construction and collision detection; and Figure 6 It is a visual response diagram of the interface after the alarm is triggered. Detailed Implementation

[0008] To more clearly illustrate the technical content of the present invention, the embodiments of the present invention will be described in detail below with reference to the accompanying drawings. These embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the scope of protection of the present invention.

[0009] Figure 1 A flowchart illustrating a substation maintenance safety early warning method based on a real-world 3D model, according to an embodiment of the present invention, is shown. Figure 1 As shown, the present invention discloses a safety early warning and collision detection method for substation operating environment, which includes the following steps.

[0010] S101: High-precision 3D reality model construction According to one implementation method, multi-source spatial data of the substation operation site is first acquired by combining UAV oblique photography with ground laser scanning. In practice, this can be done as follows: the UAV takes multi-angle photos of the entire substation area along a preset flight path, collecting high-resolution oblique image data for constructing a macroscopic structural model; simultaneously, a ground laser scanner is deployed around key equipment to collect high-precision point cloud data of transformers, circuit breakers, and disconnecting switch facilities, compensating for the lack of detail in aerial imagery.

[0011] The system uses feature extraction and matching algorithms to register and fuse sparse point clouds generated from UAV imagery with dense point clouds obtained from laser scanning. During the fusion process, the following method is employed to ensure spatial uniformity and geometric consistency between data from different sources:

[0012] in, and These are the matching points in the source point cloud and the target point cloud. It is a rotation matrix. It is a translation vector. Through continuous iterative optimization... and This minimizes the matching error between point clouds, thereby ensuring spatial uniformity and geometric consistency of data from different sources.

[0013] After data fusion is completed, point cloud smoothing, hole filling, and boundary smoothing are performed on the fused point cloud.

[0014] According to one implementation method, the following algorithm is used to smooth point clouds:

[0015] in, The filtered point cloud The mean, The standard deviation is denoted as .

[0016] According to one implementation, hole filling is performed using the following formula, and interpolation techniques are used to handle the missing areas:

[0017] in, It is a fitting function. These are known data points. This represents the Euclidean distance, and what is minimized is the sum of squared errors across all points. By performing the above processing on all data points, the optimal fitting curve or surface can be obtained, thereby filling in the missing regions.

[0018] According to one implementation, boundary smoothing is performed using the following formula:

[0019] in, These are points on the fitted curve. It is a control point. These are Bessel basis functions, defined as follows:

[0020] in, These are the parameters of the curve, which typically vary within the interval [0,1]. It is a combination number. Indicates the index of the control point. This is the total number of control points.

[0021] To smooth the boundaries, control points can be optimized. The position of [a point] minimizes the curvature of the curve. The formula for calculating curvature is:

[0022] in, and These are the first and second derivatives of the Bezier curve, respectively. Using the method of this invention, minimizing curvature makes the boundary smoother, reduces irregularities, and ultimately yields a smooth model boundary.

[0023] Using the methods described above, a 3D reality model with realistic textures, complete structure, and geographic coordinates can be generated. This model not only has centimeter-level accuracy but also preserves the geometric contours and layout relationships of on-site equipment, providing high-quality foundational support for subsequent personnel positioning, collision detection, and visual early warning.

[0024] Figure 2 This is a schematic diagram of the construction of a 3D reality model. Figure 2 The image shows an exemplary interface for processing a three-dimensional model of a substation using the method of the present invention. Figure 2 As shown, the interface comprises two main working areas: the left side displays the texture coordinate unfolded view, and the right side displays the 3D modeling view. The left image is an unfolded view of the texture mapping on the model surface, used to show the projection relationship of each part of the 3D model onto the 2D plane. Each image area corresponds to a part of the model surface, and the color and graphic information represent the data arrangement of the material texture. This view helps developers check for issues such as texture misalignment, overlap, or stretching, thereby ensuring the visual effect and structural consistency of the final 3D model. The right side is the 3D modeling view, displaying a station area model generated based on the fusion of oblique photogrammetry and laser scanning. The grid lines in the image represent the coordinate reference system. The overall model has been initially constructed, and the main outlines include the plant boundary, road structure, and equipment foundation platform. The currently selected area in the center of the screen is presented as a circular operation indicator area, indicating that the user is performing local geometric editing.

[0025] According to one implementation, the model can be further processed by removing irrelevant or redundant structures (such as temporary construction facilities and obstructions); correcting geometric errors in the model, such as holes, overlapping surfaces, or reversed normals; and adding key details to structurally complex equipment (such as circuit breakers and transformers). Specifically, these tasks can be accomplished through the following technical steps: 1. Redundant structure removal: Using computer vision and image processing algorithms, such as deep learning and semantic segmentation, the system automatically identifies and removes irrelevant structures such as temporary construction sites and obstructions. Through machine learning models, it analyzes redundant parts in the model and uses clustering algorithms to detect redundant structures that do not conform to the building or equipment specifications, then automatically removes them.

[0026] 2. Hole Repair and Geometric Error Correction: Hole-filling algorithms, such as least-squares interpolation, are employed to automatically identify and repair voids in the model by calculating geometric errors on the model surface. Overlapping surfaces are automatically detected, and algorithms such as geometric matching and deep learning models are used to identify and correct overlapping areas, ensuring structural continuity.

[0027] Normal reversal can be repaired using automated normal detection and reversal algorithms. A common repair method is to calculate the normal direction of each point on the surface and reverse the normals based on the adjacency relationship of the surrounding geometry.

[0028] 3. Detailed Supplements and Improved Precision: For complex equipment (such as circuit breakers and transformers), computer vision technology and 3D reconstruction algorithms can be used to automatically detect missing details on the surface of the equipment. Then, by extracting the corresponding geometric data and material information from the database or library, the details in the model can be automatically supplemented, thereby improving the model's accuracy. 4. Automation tools and algorithms: Algorithm-based intelligent optimization tools, such as graphics-based automatic texture mapping and geometric repair tools, can automatically optimize the structure of 3D models and ensure seamless integration of textures and geometry. For the repair of complex structures, geometric analysis and optimization algorithms, such as bounding box-based optimization techniques and point cloud analysis, can be used to automatically calculate and repair structural accuracy. This prepares the model for subsequent animation rigging or export to a 3D visualization platform. The entire modeling process aims to recreate the real scene, ensuring the model possesses good geometric accuracy, structural continuity, and visual expressiveness, laying the foundation for subsequent early warning algorithm integration and interactive display.

[0029] Figure 3This section showcases the model management interface within the system's safety detection module, primarily used for loading, binding, and maintaining information on dynamic equipment or personnel models. A model type selection bar is located at the top of the interface, allowing users to choose the target model type, such as personnel, vehicle, or equipment, via a drop-down menu for easy categorization and management. The "Stop Adding Models" button on the right terminates the current model entry operation, suitable for interactive control in batch addition scenarios. The lower area displays a list of model information; the system assigns unique identifiers (IDs) to loaded models and displays their corresponding type, longitude, latitude, and other key attributes. Each row of data includes "Delete" and "Select" buttons, facilitating targeted operations on specific models. The example in the image shows a vehicle model whose ID and spatial location information have been successfully recorded, placing it in a manageable state. This interface provides the foundational data entry point and visualization support for the system's subsequent establishment of model spatial relationships, dynamic location binding, and risk assessment. The spatial location and status of each dynamic object are continuously updated within the platform and serve as input for subsequent collision detection and warning algorithms.

[0030] S102: Hazardous Area Designation and Safety Early Warning Mechanism To achieve proactive prevention and control of high-risk scenarios, the system pre-defines several dangerous areas and key work areas in the 3D model and constructs multi-dimensional early warning trigger rules.

[0031] Hazardous areas include areas with high-voltage live equipment, restricted access areas, and hot work sites. During the model building phase, system maintenance personnel accurately model these areas and set buffer boundaries based on the site layout and safety regulations.

[0032] To achieve prevention and control in high-risk scenarios, the system accurately models hazardous areas based on the site layout and safety regulations. The specific methods are as follows: 1. Model based on site layout and safety specifications Site Layout: Based on the equipment location and geometry, a three-dimensional coordinate system is used to determine the position and relative relationship between the equipment and hazardous areas. Safety Regulations: Based on the equipment's working principle and hazards, safety boundaries are established for areas such as high-voltage equipment and hot work sites in accordance with safety regulations. For example, the safety distance for high-voltage equipment areas may be 3 meters, and the safety distance for hot work sites may be 10 meters.

[0033] 2. Buffer boundary settings Buffer Calculation: The buffer size is based on the device's safety distance and risk factor. For device areas, a circular buffer can be used, calculated using the following formula:

[0034] It is a safety factor (for example, a setting of 1.5 means that a safety distance of 1.5 times the size of the device needs to be maintained around it). This refers to the actual dimensions of the equipment or the standard safety distance. Dynamic adjustment: For different hazardous areas (such as restricted access areas and hot work sites), the size of the buffer zone should be adjusted according to the type of work to ensure safe isolation. For different types of hazardous areas (such as restricted access areas and hot work sites), the buffer boundary should be dynamically adjusted according to the actual situation: For high-voltage live equipment areas, the buffer zone should be a circular area with a radius calculated based on the safe distance of the equipment. For hot work sites, the buffer zone can be a polygonal area, with a larger buffer zone set considering different work locations, wind directions, and types of fire sources.

[0035] 3. Implementation Method Modeling tools: Use Blender or Cesium for 3D modeling to define the spatial relationship between hazardous areas and equipment. Data import: Optimize the model using a GIS platform or dedicated software, display hazardous areas in real time, and set buffer boundaries.

[0036] The system continuously calculates the spatial relationship between personnel and equipment and the danger zone based on their real-time location and status information. Once personnel enter the danger zone buffer zone, or their posture parameters become abnormal (such as significant tilting or falling), or the equipment's operating trajectory deviates from the set range, the system immediately triggers the corresponding warning.

[0037] The early warning mechanism is divided into three triggering methods: 1. Distance trigger: For example, the horizontal distance between the operator and the high-voltage equipment is less than 0.5 meters; 2. Posture trigger: If a person's pitch angle exceeds ±45°, it is considered a possible fall. 3. Path deviation trigger: If the hoisting path deviates from the preset line by more than 0.3 meters.

[0038] like Figure 4 The figure illustrates the relationship between a target model and its bounding box in a 3D scene. The central model is located inside a cube-shaped frame, which is the axis-aligned bounding box automatically generated by the system, used to indicate the actual boundary of the model within the scene. The scene background is a remote sensing base map of a substation area, clearly showing the model's location and the complete distribution of surrounding features. The bounding box is drawn with solid lines, and its four boundaries are easily identifiable in the view, allowing for subsequent spatial calculations for collision detection and risk analysis. This figure reflects the system's ability to spatially locate dynamic models in a 3D coordinate system, demonstrating its fundamental capabilities for target modeling, bounding box calculation, and spatial association in a real-world environment.

[0039] S103: 3D Bounding Box Construction and Distance Calculation A three-dimensional bounding box or capsule is created for each moving object and hazardous area on site. The shortest axial distance between each moving object and the hazardous area in the X, Y, and Z axes is calculated. When the distance in any axis is less than a preset threshold, it is judged as a potential collision risk.

[0040] like Figure 5 The image illustrates the system's ability to construct capsules and bounding boxes for 3D models within a real-world scene. In the diagram, the central device model is completely enclosed within a transparent capsule, which consists of a cylinder and two circular end faces. The cylinder's length matches the target device's dimensions, and the diameter of the circular end faces matches the target device's width. This design allows for more precise adaptation to elongated, curved, or circular device shapes. A separate vehicle, using an AABB bounding box, is shown on the right side of the image. The boundaries of both the capsule and the AABB bounding box are represented by solid lines, providing clear boundaries and high visual recognizability, facilitating rapid identification of device boundaries in complex scenes. This image clearly demonstrates the system's ability to construct accurate spatial boundaries for multiple device models, providing fundamental support for subsequent tasks such as collision detection and spatial analysis.

[0041] To achieve more efficient collision detection under different equipment and scenario conditions, the embodiments of this invention can combine AABB and capsule collision methods, dynamically selecting based on the target's shape and motion state: For regularly shaped equipment (such as upright electrical boxes, standard equipment frames): the AABB collision detection method is continued for regular rectangular or cubic shapes. The AABB method is computationally simple, has low performance consumption, and is well-applicable to many standard devices. For complex-shaped or elongated equipment (such as cables, pipes, large equipment supports): the capsule collision method can be used for detection. The capsule collision method can better adapt to the shape of the equipment, reduce unnecessary space waste caused by AABB, and improve detection accuracy. In each frame of data, the system first extracts the corresponding point set based on the object classification result, and then traverses the point set to calculate its boundary coordinates in the X, Y, and Z directions, i.e., the minimum value (Xmin, Ymin, Zmin) and the maximum value (Xmax, Ymax, Zmax). Based on these six values, the system constructs a hexahedron aligned with the coordinate axes, i.e., the object's AABB bounding box. To improve processing efficiency, this process is executed under a GPU parallel architecture, employing a block-level reduction method to quickly extract boundary values ​​for large-scale point sets. This implementation supports simultaneous processing of multiple objects, with an average latency of less than 20 milliseconds per frame, meeting the real-time requirements of high-frequency, multi-object simultaneous tracking. The system calculates the minimum axial distance between any two bounding boxes in the X, Y, and Z axes in real time, serving as the basis for collision risk determination; and the collision threshold can be defined by setting the `collisionThreshold` parameter to ensure the accuracy and adaptability of the calculation. For the capsule collision method, the minimum and maximum cylindrical boundaries of the target device, as well as its central axis, are used to determine whether the device has collided with other objects. The capsule method determines whether a collision has occurred by calculating the minimum distance between the target object and the detected object; if the minimum distance is less than the set threshold, a collision is considered to have occurred. When combining AABB with the capsule collision method, spatial indexing structures (such as octrees, BSP trees, etc.) are used to accelerate collision detection, ensuring that the system maintains high computational efficiency even under high concurrency.

[0042] According to one implementation, the system divides the substation workspace into multiple logical partitions and sets label attributes (such as "operators", "equipment", "vehicles") for objects. Before detection, the system automatically filters out object combinations with non-overlapping spatial intervals or inconsistent label logic, retaining only object pairs with a high probability of collision for calculation.

[0043] When the bounding box distance of a certain object combination in any direction is detected to be lower than a set threshold, the system determines it to be a potential collision object and enters the early warning response stage.

[0044] S104: Early Warning Response and 3D Interface Display like Figure 6 As shown, once the system identifies a potential collision risk, it immediately activates the early warning response mechanism and displays the relevant information in real time on the 3D visualization interface. Objects deemed high-risk (such as people and vehicles) will be highlighted in color in the model.

[0045] To enhance the on-site response of early warnings, the system can simultaneously trigger an audible and visual alarm module, using a buzzer to guide workers to immediately evacuate. The platform supports alarm volume adjustment, language selection, and multi-channel output to adapt to the needs of different working environments.

[0046] In addition, all alarm events are automatically recorded in the database, including alarm timestamp, related object ID, risk type, and duration fields. Users can use the history playback function to query the alarm records of any object and its movement trajectory in the 3D scene by time period, assisting in accident tracing and safety assessment.

[0047] Figure 6 The image demonstrates the system's visual response after detecting a spatial collision between models. In the image, the two 3D models in the center of the scene represent different object categories, their spatial boundaries marked with solid-line bounding boxes. The two models are currently in contact, and the system displays a gray-on-white message at the bottom of the screen: "Model 'car' and model 'transformer.glb' have collided!" clearly informing the colliding entities. The entire image is presented from a bird's-eye view, facilitating observation of the spatial relationships between the models and the collision area, showcasing the system's real-time monitoring and visual feedback capabilities for dynamic risk events.

[0048] The numbering of the method steps in this invention is for ease of description only and is not a specification or explanation of their execution order. Those skilled in the art should understand that the order of these steps can be adjusted or some steps can be executed in parallel according to the actual situation.

[0049] The above detailed description of the invention is merely intended to provide those skilled in the art with further information for carrying out preferred aspects of the invention, and does not limit the scope of the invention. Only the claims are used to define the scope of protection of the invention. Therefore, the combination of features and steps in the foregoing detailed description is not necessary for carrying out the invention in the broadest possible sense, and is alternatively taught only for representative embodiments of the invention as described in particularly detailed terms. Furthermore, various different features taught in the specification can be combined in various ways to obtain additional useful embodiments of the invention; however, these ways are not specifically enumerated.

Claims

1. A substation maintenance safety early warning method based on a real-scene 3D model, characterized in that, Includes the following steps: The three-dimensional point cloud and image data of the substation scene were obtained by combining drone oblique photography and laser scanning, and the two were then integrated to build a three-dimensional reality model. Several danger points and work areas are pre-defined in the three-dimensional reality model; A three-dimensional bounding box or capsule is created for each moving object and hazardous area on site. The shortest axial distance between each moving object and hazardous area in the X, Y, and Z axes is calculated. When the distance in any axis is less than a preset threshold, it is judged as a potential collision risk. Provide real-time early warnings and issue alerts to the target or area.

2. The method according to claim 1, characterized in that, The process of establishing the three-dimensional real-scene model includes: acquiring panoramic images of the substation using oblique photography by a drone and generating preliminary point clouds through aerial triangulation; simultaneously using a laser scanner to perform intensive scanning of key equipment; and then registering and fusing the point clouds to form a complete three-dimensional scene model.

3. The method according to claim 1, characterized in that, The fusion of 3D point cloud and image data includes: Register and fuse sparse point clouds generated from UAV imagery with dense point clouds obtained from laser scanning. Perform point cloud smoothing on the fused point cloud; Hole filling is performed on the smoothed point cloud to obtain a fitted curve or surface; and The fitted curve or surface is then smoothed at the boundary.

4. The method according to claim 3, characterized in that, The sparse point cloud generated from UAV imagery and the dense point cloud obtained from laser scanning are registered and fused using the following formula: in, and These are the matching points in the source point cloud and the target point cloud. It is a rotation matrix. It is a translation vector, which is optimized through continuous iteration. and This minimizes the matching error between point clouds.

5. The method according to claim 3, characterized in that, Point cloud smoothing is performed using the following formula: in, The filtered point cloud The mean, The standard deviation is denoted as .

6. The method according to claim 3, characterized in that, Use the following formula to fill the void: in, It is a fitting function. These are known data points. The Euclidean distance is minimized by the sum of squared errors across all points.

7. The method according to claim 3, characterized in that, The fitted curve or surface is smoothed at the boundary using the following formula: in, It is a point on the fitted curve or plane. It is a control point. These are Bessel basis functions. in, These are the parameters of the curve, varying within the interval [0,1]. It is a combination number. Indicates the index of the control point. This is the total number of control points; The control points are optimized using the following formula. The position of makes the curvature of the curve Minimum: in, and These are the first and second derivatives of the Bezier curve, respectively.

8. The method according to claim 1, characterized in that, When assessing potential collision risks, spatial partitioning and object labeling are used for screening. The spatial partitioning is based on a predefined grid or logical region, grouping objects located in the same region together and performing collision detection only on objects within the group; the object label is used to identify the object type or security level, and to filter out objects that do not need to be detected.

9. The method according to claim 1, characterized in that, For regular rectangular or cubic moving objects or hazardous areas, the AABB collision detection method is used. For complex or elongated devices, the capsule collision method is used for detection. The capsule consists of a cylindrical section and two hemispheres. The length of the cylinder corresponds to the main shaft length of the target device, and the radius of the hemisphere matches the maximum cross-sectional radius of the device.

10. The method according to claim 1, characterized in that, It also includes outputting the results of the above methods to an integrated 3D visualization platform, which supports real-time monitoring and historical playback. Alarm thresholds, display colors, and prompt sound parameters are user-defined and applicable to various terminal devices such as PCs, tablets, and mobile devices.

Images