Substation operation monitoring and early warning system based on 3D gaussian splashing and beidou RTK

Abstract

The application relates to the technical field of substation operation management, and particularly discloses a substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK, which comprises a coordinate and model construction module, a dangerous boundary representation generation module and a two-stage alarm triggering module, wherein the coordinate and model construction module is used for constructing a human tool envelope and generating a 3D Gaussian splashing scene model; the dangerous boundary representation generation module generates a dangerous boundary representation in the 3D Gaussian splashing scene model from a set of operation steps, so that the dangerous boundary can be synchronously updated with the changes of the operation steps and the effectiveness of safety measures, and the consistency of risk constraints and field states is improved; the two-stage alarm triggering module calculates the minimum clear distance of the dangerous boundary representation and the human tool envelope to form an alarm constraint set, and triggers a reminding stage alarm or a dangerous stage alarm according to the alarm constraint set, so that the hierarchical prompt for approaching risks is realized, and the explainability and executability of alarm triggering are enhanced.

Description

Technical Field

[0001] This invention relates to the field of substation operation management technology, specifically a substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK. Background Technology

[0002] 3D Gaussian splashing is a 3D representation method based on multi-view image reconstruction and real-time rendering. It uses a large number of 3D Gaussian volumes with attributes such as position, shape, and color to approximate the scene, thereby achieving near-realistic 3D scene reconstruction and browsing from any viewpoint. BeiDou RTK (Real-Time Kinematic) is a type of real-time dynamic carrier phase differential positioning technology that uses BeiDou as the positioning signal source. It performs real-time correction on the mobile end through reference station or network correction data, and can provide centimeter-level real-time position and trajectory data in engineering applications, providing a reliable absolute coordinate reference for personnel, vehicles, or robots.

[0003] In existing monitoring technologies, work permits or work orders are the main focus: before work begins, the work content, planned time, work scope, and areas requiring isolation to a safe state are clearly specified in the permit or order. The person in charge confirms each safety measure, including isolation, voltage testing, grounding, and tagging, and organizes personnel to sign off and conduct safety briefings. Simultaneously, a 3D model of the substation work scenario is established, representing spatial elements such as equipment, bays, passageways, and fences in three dimensions. Danger points, restricted areas, and risk boundaries are marked in the 3D model, mapping the permitted scope and safety measure status corresponding to the work order to a calculable permitted work area. During the work process, on-site supervisors and visual monitoring are relied upon. Frequency monitoring and access control are used for personnel access and area management. Combined with personnel location information, personnel identity, time, and real-time location are projected into a 3D model to continuously verify whether personnel are within the permitted range, whether they have crossed the restricted boundary, or are approaching the risk boundary. Combined with the completeness verification of safety measures, it judges whether the monitoring is compliant and triggers reminders or alarms. At the same time, visual algorithms are used to identify personnel's work status and dangerous actions, and ultra-wideband positioning electronic fences are used to provide real-time alarms for crossing the restricted area or dangerous distance. After the work is completed, on-site inventory, measure restoration verification, permission closure, and record archiving are performed to form a traceable closed loop of work process, alarm records, and ticketing process.

[0004] For example, Chinese invention patent application CN115879764A discloses a substation maintenance operation risk early warning and prevention system and method. The system compares the location information of maintenance personnel obtained by the positioning system with a regional grid-based risk level database through a backend system. When the maintenance personnel are in a high-risk area, an early warning is played through a playback device. The backend system also sends the corresponding operating procedures for the power equipment to the playback device based on the RFID tag information read by the RFID reader from the RFID tags on the power equipment around the maintenance personnel, prompting the maintenance personnel to operate according to the procedures. Furthermore, the backend system determines the equipment risk level of the corresponding power equipment based on the RFID tag information and, based on the distance information between the operator and the equipment collected by the radar rangefinder, plays an early warning through the playback device when the maintenance personnel approach power equipment with a high equipment risk level.

[0005] For example, Chinese invention patent CN119130148B discloses a method and device for early warning of substation operation risks, which relates to the field of power technology, and particularly to the field of substation technology. The method includes: acquiring data from N substations; determining a target 3D model of the substation based on the substation data, wherein the target 3D model changes with the substation data, and the target 3D model is used to represent the 3D scene when workers are performing construction operations at the substation; determining the operation risk information of the substation based on the target 3D model; and generating early warning information for the substation based on the operation risk information.

[0006] Based on the above technical solutions, the existing substation operation monitoring approach can be described as follows: using data such as positioning, video, multi-view vision, or point cloud to map personnel, equipment, and work areas into a unified three-dimensional scene, and then using preset danger boundaries or safe clearance thresholds to determine distance / collision on this three-dimensional model, or abstracting on-site risks into several risk factors and calculating risk levels, and then comparing them with preset thresholds to trigger early warnings; however, the actual substation operation monitoring process has strong uncertainty and randomness. The danger boundary is not a fixed geometric shell. It will continuously shrink or expand with the switching of equipment energization status and operating mode, the progress of work ticket steps and whether safety measures are truly in place, the changes in the activity envelope caused by personnel posture and tool extension, and the fluctuations in on-site environment and perception quality.

[0007] Therefore, the main technical drawbacks of existing technologies are that they oversimplify complex hazard boundaries into static thresholds, resulting in mostly statically set risk boundaries in 3D models that cannot be automatically adjusted according to changes in on-site conditions. Consequently, when calculating clearance, geometric approximations in equipment-level objects or models are often used to replace the actual hazard boundaries of live parts and the minimum clearance of personnel activity range, causing inconsistencies between the measurement objects used for calculation and the actual risk objects that need to be controlled. Finally, in order to output unified alarm results, the system often integrates multi-source evidence such as location, video, equipment status, and invoices into a single risk level or score. However, it is not perfect in terms of a stable triggering mechanism that continuously meets the conditions, resulting in negative impacts such as false alarms and false alarms coexisting, frequent alarm jumps, and difficulty in interpretation when on-site status changes, obstruction interference, or fluctuations in location / modeling quality. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK, which can effectively solve the problems mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: a substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK, comprising a coordinate and model construction module, used to generate the three-dimensional coordinates of the workers in the substation operation area using BeiDou RTK to construct the human tool envelope, and to collect multi-view image sequences of the substation operation area to generate a 3D Gaussian splashing scene model; a hazard boundary representation generation module, used by the edge computing unit to receive the substation operation work order and generate a set of operation steps, and to generate a hazard boundary representation of the substation operation in the 3D Gaussian splashing scene model based on the set of operation steps; and a two-level alarm triggering module, used to calculate the minimum clearance between the hazard boundary representation and the human tool envelope to form an alarm constraint set, and to trigger two levels of alarms based on the alarm constraint set, including a reminder alarm or a danger alarm, thus completing the substation operation monitoring and early warning.

[0010] Compared with the prior art, the embodiments of the present invention have at least the following advantages or beneficial effects: (1) This invention provides a substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK, comprising a coordinate and model construction module, a hazard boundary representation generation module, and a two-level alarm triggering module. The coordinate and model construction module obtains the three-dimensional coordinates of the workers and constructs the human tool envelope through Beidou RTK in the substation operation area. At the same time, it collects multi-view image sequences covering the operation area to generate and anchor a 3D Gaussian splashing scene model, thereby forming a unified spatial basis for net distance calculation under the real-world coordinates within the station and improving the integrity of the expression of personnel and tool occupancy. The hazard boundary representation generation module consists of an edge... The calculation unit receives the work order and generates a set of work steps. Based on the permitted area, prohibited objects, and the status of safety measures, it generates a hazard boundary representation in the 3D Gaussian splash scene model. This allows the hazard boundary to be updated synchronously with changes in the work steps and the effectiveness of safety measures, improving the consistency between risk constraints and the on-site status. The two-level alarm triggering module calculates the minimum clearance between the hazard boundary representation and the human tool envelope to form an alarm constraint set. Based on the alarm constraint set, it triggers a reminder-level alarm or a danger-level alarm, thereby achieving graded alerts and escalation control of approach risks and enhancing the interpretability and executability of alarm triggering.

[0011] (2) This invention constructs a dynamically updated human tool envelope by combining the three-dimensional coordinates of the operator with the reachable area of ​​the tools. This tool envelope can accurately reflect the actual space occupied by the operator and tools, taking into account factors such as changes in the operator's posture, the length range of the tools, and the range of hand operation, ensuring precise control of the safe distance during operation. In particular, through positioning accuracy and uncertainty indicators, the tool envelope can adaptively adjust its size and shape in a dynamic environment, thereby reducing unsafe factors caused by positioning errors. In addition, as the operation steps and safety measures change, the human tool envelope can be updated in real time to ensure that it always reflects the safe range of the operator's activities and is compared with the danger boundary of the work area, further enhancing safety.

[0012] (3) This invention employs dynamic hazard boundary representation, which can adjust the hazard boundary in real time according to different changes in the work steps. In traditional work monitoring systems, hazard boundaries are often static and cannot reflect the actual safety risks caused by changes in work permits or adjustments to safety measures during the work process. By dynamically generating hazard boundaries based on the set of work steps, the system can track changes in hazardous targets in real time, ensuring that workers are always within a controlled safety area. In addition, by combining a 3D Gaussian splash scene model and semantic labels of hazardous parts, each hazard source can be identified and isolated more accurately, greatly reducing safety hazards during the work process and improving the ability to respond to safety risks in complex environments.

[0013] (4) The two-level alarm triggering mechanism of the present invention effectively improves operational safety. By calculating the minimum clearance and dynamically triggering alert-level and danger-level alarms based on its changes, not only can the real-time distance between the operator and the dangerous target be accurately assessed, but potential hazards can also be responded to in the shortest possible time. The alert-level alarm issues a timely warning when the minimum clearance approaches the safety threshold, reminding the operator to pay attention to safety, while the danger-level alarm is triggered when the minimum clearance exceeds the set danger threshold, and emergency measures are taken immediately, such as adjusting the operator's position or suspending the operation. This mechanism can effectively prevent operators from approaching the danger source due to ignoring minor risks, and reduces false alarms and missed alarms through the step-by-step alarm method, improving the stability and reliability of the overall early warning system and ensuring that every link in the operation process is under high-level safety monitoring.

[0014] (5) Compared with the existing technology, the present invention achieves a more consistent and traceable closed loop in terms of on-site spatial reference and risk calculation caliber: On the one hand, based on Beidou RTK, a global coordinate reference is established in the substation operation area and the three-dimensional coordinates and positioning accuracy indicators of the operators are output in real time, so that the position of the personnel no longer depends only on the relative displacement in the video picture or the planar position of the electronic fence, thus maintaining stable spatial positioning and quality constraints under complex obstruction, lighting changes or limited viewing angle conditions; On the other hand, a 3D Gaussian splash scene model is constructed based on multi-view image sequences and anchored to the real-world coordinates of the substation, so that the spatial outer surface of dangerous targets, boundary point sets or distance fields and other dangerous boundary representations have a unified three-dimensional geometric expression, and can be dynamically updated with the work order operation steps and safety measure evidence status changes, ultimately improving the accuracy, stability and verifiability of substation operation monitoring and early warning. Attached Figure Description

[0015] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the system module connections of the present invention.

[0017] Figure 2 A schematic diagram of the human tool envelope.

[0018] Figure 3 This is a schematic diagram of the consistency fusion of hand pixels to 3D position candidates.

[0019] Figure 4 This is a schematic diagram of the minimum net distance and the nearest point pair. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0021] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0022] In actual substation maintenance, defect elimination, or work near energized compartments, the work is usually organized and implemented based on the work permit. Before starting work, the person in charge clarifies the work content, work scope, permitted areas and prohibited objects, and confirms each safety measure such as isolation, voltage testing, grounding, and tagging. After the work begins, the workers wear Beidou real-time dynamic positioning terminals to enter the work area within the station. Multi-view video acquisition devices deployed within the station continuously acquire images of the work process. The monitoring system constructs a 3D Gaussian splash scene model anchored to the real-world coordinates of the substation based on the multi-view images, mapping the permitted areas, prohibited objects, and safety measure evidence status in the work permit to work permitted boundaries and energized danger boundaries that change with each step.

[0023] During the operation, the monitoring system uses the area occupied by the personnel and the reachable area at the front end generated by the three-dimensional position of the hands and the length of the tools as the measurement objects. It continuously calculates the minimum clearance between the area and the danger boundary, triggers reminder-level or danger-level alarms based on the minimum clearance, and performs stable control on alarm escalation. When the status of the work order steps or safety measures evidence changes, causing the danger boundary to be updated, the system starts to freeze the window to update the alarm status.

[0024] It should be explained that the substation operation monitoring and early warning process in this embodiment of the invention uses an edge computing unit as the execution platform. The edge computing unit is deployed on the substation site side and is usually composed of an in-station industrial control computer, an edge server, or a gateway device with computing power. It establishes a data communication connection with the multi-view image acquisition unit and the Beidou RTK personnel positioning terminal. In this embodiment of the invention, the edge computing unit is used to receive image sequences uploaded by each camera and the three-dimensional coordinates and positioning accuracy indicators output by the personnel positioning terminal. On the site side, it completes core processing such as the construction and anchoring of the 3D Gaussian splash scene model, the generation of dangerous targets and dangerous boundaries, the estimation of the occupancy range of personnel and tools, the calculation of minimum clearance and nearest point pairs, and the alarm triggering and freezing window stability control. At the same time, it encapsulates alarm evidence packages and outputs them to the on-site alarm terminal or the background monitoring system, thereby achieving low latency response, on-site availability, and post-event traceability without relying on the cloud.

[0025] Example 1: See Figure 1As shown, this embodiment of the invention provides a technical solution: a substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK, including a coordinate and model construction module, a dangerous boundary representation generation module, and a two-level alarm triggering module.

[0026] The coordinate and model construction module is connected to the hazard boundary representation generation module, which in turn is connected to the two-level alarm triggering module.

[0027] The coordinate and model building module is used to generate the three-dimensional coordinates of the workers in the substation operation area using Beidou RTK to construct the human tool envelope, and to collect multi-view image sequences of the substation operation area to generate a 3D Gaussian splash scene model.

[0028] The three-dimensional coordinates of the workers are determined by deploying BeiDou real-time dynamic positioning base stations in the work area before the substation work begins. These base stations use BeiDou satellite observation data to model errors and generate correction information, continuously outputting the correction information as differential correction data. Simultaneously, the base stations use their coordinates as known control points to align with a preset global coordinate system, allowing edge computing units to calculate positions and output them to this global coordinate system, ensuring subsequent positioning results are expressed under the same coordinate reference. Workers wear BeiDou real-time dynamic positioning personnel terminals, which receive BeiDou satellite signals and synchronously receive differential correction data. Based on this data, corrections are made for errors affecting satellite clock bias, orbital error, ionospheric and tropospheric conditions, and carrier phase constraint calculations are performed to obtain the workers' three-dimensional coordinates under the preset global coordinate reference. During real-time calculations, the personnel terminal assesses the quality of the current calculation status, generating positioning accuracy indicators and outputting them synchronously with the three-dimensional coordinates.

[0029] It should be explained that the above-mentioned carrier phase constraint calculation to obtain the three-dimensional coordinate output is a typical existing technical process of differential RTK positioning. The above-mentioned positioning accuracy index is preferably fixed as a three-dimensional comprehensive positioning accuracy index to represent the degree of three-dimensional comprehensive dispersion of the current positioning result from the true position, which is obtained synchronously by the Beidou RTK personnel positioning terminal during the differential carrier phase calculation process.

[0030] In one specific embodiment, the positioning accuracy index is characterized by the three-dimensional dispersion radius obtained from the position sequence statistics within the most recent second: after calculating the mean of the horizontal coordinates and height within the most recent second, the horizontal distance and height distance from each sampling point to the mean are calculated and the root mean square is taken to obtain the horizontal dispersion radius and height dispersion amount, which are taken as the root mean square of the horizontal dispersion radius and height dispersion amount as the positioning accuracy index.

[0031] Construct a human tool envelope, specifically by using the three-dimensional coordinates of the operator as the center point of the human geometry, and generating a human occupancy geometry based on preset human occupancy parameters. The human occupancy geometry is a three-dimensional volume region that at least covers the movement outline of the operator's torso and limbs.

[0032] The specific construction is as follows: Figure 2 As shown, Figure 2 This diagram illustrates the human tool envelope, showing a unified geometric representation of the occupancy range of workers and their potentially used tools in a substation operation scenario. The diagram uses the worker's 3D positioning as the center to construct a worker occupancy geometry, covering the basic space occupied and sway margin of the human body during substation operations. Based on this geometry, and combining the tool type, length, grip point, or reachable area, a tool occupancy geometry or reachable area at the tool's tip is generated. This merges the human and tool into a unified human tool envelope within the same coordinate system. This envelope is used for subsequent spatial relationship queries and safe clearance calculations with hazard boundary representations, more closely reflecting the risk characteristics of actual operations where the human body is in a stable position but the tool tip may extend into the danger zone.

[0033] The edge computing unit further receives the coordinates of the tool's gripping end or the tool's reference point from the tool positioning terminal, and combines them with the tool's main axis direction, tool type, and corresponding length level obtained from multi-view image sequence recognition. When the positioning accuracy index is greater than the tool accuracy threshold, the gripping end coordinates are switched from a single point to the gripping point's reachable area. When the multi-view estimation deviation of the tool's main axis direction is greater than a preset direction deviation threshold, candidate angle regions for the tool's main axis direction are generated according to preset angle levels. Within the length range corresponding to the length level, the gripping point or the gripping point's reachable area is expanded along the candidate angle regions to generate the tool's front-end reachable area. The tool's front-end reachable area is a three-dimensional volume region covering the possible reachable positions of the tool's front end. Finally, the edge computing unit merges the geometry occupied by the human body and the tool's front-end reachable area into a single envelope output as the human-tool envelope, thereby achieving a unified safety clearance assessment of the space occupied by personnel and tools.

[0034] The aforementioned preset human body occupancy parameters may include the human body occupancy geometry type and the corresponding size parameters and margin configuration. The human body occupancy geometry type can be selected as a cylinder, elliptical cylinder, cuboid, or capsule. The size parameters include at least the horizontal occupancy scale and the vertical height range. At the same time, posture margins for covering the changes in the outline caused by the work action and personal item margins for covering the outlines of personal items such as safety helmets and tool bags can be configured, so that the human body occupancy geometry maintains a conservative and configurable consistent representation in different work areas and under different working conditions.

[0035] The geometry occupied by the human body and the reachable area of ​​the front end of the tools are both represented in the actual coordinates of the substation.

[0036] Constructing a human tool envelope also includes situations where only BeiDou RTK personnel positioning terminals exist, but tool positioning terminals do not: The edge computing unit performs human keypoint detection on multi-view images. Human keypoints include at least a portion of the following: head, torso center, shoulders, elbows, wrists, hips, knees, and ankles. Hand keypoints are represented by the left and right wrist points. The keypoint detection network outputs a probability heatmap for each type of keypoint. Each pixel in the hand keypoint heatmap has a response value representing the confidence score that the pixel is a hand keypoint. The edge computing unit takes the pixel with the largest response value as the hand pixel coordinates and quantifies the visibility confidence of this keypoint into a peak discrimination index, represented by the difference or ratio between the largest and second-largest response values. A larger difference or ratio indicates a more dominant peak and a more unique candidate position, while a smaller difference or ratio indicates multiple candidate ambiguities. Views with a visibility confidence score not less than a preset visibility threshold are determined as valid viewpoints.

[0037] Based on the imaging relationship corresponding to the effective viewpoint, the pixel coordinates of the hand are projected into the 3D Gaussian splash scene model. The imaging relationship is the pixel-to-spatial ray mapping relationship determined by the camera imaging parameters of the camera in the actual coordinates of the substation, so that the pixel coordinates of the hand can determine a spatial observation ray that starts from the optical center of the camera and passes through the pixel position. The edge computing unit searches for the spatial point that best matches the hand position in the 3D Gaussian splash scene model along the observation ray, and obtains the candidate three-dimensional position of the hand corresponding to the viewpoint.

[0038] Consistency determination is performed on each candidate 3D hand position. Consistency determination includes: the 3D Euclidean distance between candidate points is not greater than a preset consistency distance threshold; and this threshold is used as a neighborhood merging condition to determine the sets of candidate points that can be merged through neighborhood connectivity as the same candidate cluster, where a candidate cluster represents a set of candidate 3D points pointing to the same true hand position. Consistency is deemed successful when there is at least one candidate cluster with a number of members not less than a preset number threshold, and the spatial dispersion of the candidate cluster is not greater than a preset dispersion threshold; the spatial dispersion is obtained by calculating the root mean square distance from each member point within the candidate cluster to the cluster center.

[0039] Based on the consistency determination, edge computing units sum the results according to preset fusion weights to obtain the confidence score of the hand's 3D position. The confidence score of the hand's 3D position is a quantified value between zero and one, and is determined by at least the proportion of effective viewpoints, the consistency score, and the visibility confidence score of participating viewpoints. The proportion of effective viewpoints is the ratio of the number of effective viewpoints to the number of participating viewpoints. The consistency score is obtained by inverse normalization of spatial dispersion; the smaller the spatial dispersion, the larger the consistency score. Specifically, the proportion of effective viewpoints, the consistency score, and the visibility confidence score are multiplied by their respective fusion weights and then summed to obtain the confidence score of the hand's 3D position. A higher proportion, a higher consistency score, and a higher visibility confidence score result in a higher confidence score.

[0040] When the confidence level of the three-dimensional position of the hand is not less than the first confidence threshold and the number of effective viewpoints is not less than the preset number, the three-dimensional position of the hand that has passed the consistency judgment is taken as the gripping point of the tool. Otherwise, a spatial region with a preset radius is generated centered on the most recent effective gripping point to cover the gripping point uncertainty. Subsequently, the edge computing unit reads the corresponding length level according to the tool category recorded in the work order or work step set or the tool category obtained by multi-view image recognition, and expands the gripping point or the gripping point reachable region along the main axis of the tool within the length range corresponding to the length level to generate the reachable region of the tool front end.

[0041] Specifically, such as Figure 3 As shown, Figure 3 This diagram illustrates the consistency fusion of hand pixels to 3D position candidates. It shows the process of locating the hand from multiple viewpoint images and fusing them to obtain a stable 3D gripping point. Key hand points are detected on the image plane of each viewpoint to obtain hand pixel coordinates. These pixel observations are then projected onto a 3D scene model based on camera imaging relationships to generate candidate 3D hand positions for that viewpoint. Since occlusion, reflection, blurring, and false key point detection can cause single-viewpoint candidates to shift or become outliers, the diagram further demonstrates that after filtering effective viewpoints, candidates from each viewpoint are clustered and judged based on a consistency distance threshold. Candidates forming stable clusters are retained, while outliers are removed. Finally, the candidates that pass the judgment are fused to output the 3D hand position and calculate the confidence level. This is used to determine the gripping point or reachable area of ​​the tool, thus providing stable input for subsequent estimation of the reachable area of ​​the tool's front end and construction of the human tool envelope, reducing false alarms caused by single-viewpoint jitter.

[0042] Before the actual operation begins, multi-view image acquisition devices deployed around the substation operation area synchronously acquire image sequences covering the operation area. After receiving the image sequences from each viewpoint, the edge computing unit establishes the imaging relationship between the cameras at each viewpoint and the substation operation area, and performs three-dimensional reconstruction of the scene based on the multi-view image sequences. The stable structures and equipment outer surfaces in the scene are modeled using 3D Gaussian splash representation, resulting in a 3D Gaussian splash scene model that includes the position, scale, orientation, and appearance attributes of the Gaussian body.

[0043] Subsequently, the edge computing unit selects no fewer than three spatial control points with known coordinates within the scene. Specifically, it selects several fixed, immovable physical points that are easily and stably identifiable in multi-view images within the substation operation area as spatial control points, such as corner points of equipment foundations, centers of embedded parts, centers of bolt holes, obvious structural edges, or specially laid marker points. The three-dimensional coordinates of these control points are obtained by measurement methods and fixed into the configuration before the operation begins.

[0044] A correspondence is established between the control point coordinates in the 3D Gaussian splash scene model and the control point coordinates under the global coordinate reference. Specifically, the pixel position of the control point is marked in the multi-view image or the 3D position of the point is selected in the 3D reconstruction result, and then the 3D position is read as the control point coordinates in the scene model coordinate system. At the same time, the known coordinates of the physical control point under the global coordinate reference are read from the measurement configuration. The two are paired one-to-one according to the same control point identifier to obtain a set of control point coordinate correspondences. Based on the set of control point correspondences, the spatial transformation relationship from the scene model coordinate system to the global coordinate reference is solved. Specifically, the geometric center is calculated for the two sets of control points based on the set of control point coordinate correspondences, by analyzing the horizontal, vertical, and height coordinates of each control point. The average is obtained; then, a decentralization process is performed on each control point, that is, the geometric center of its point set is subtracted from the control point, so that the two sets of control points are aligned with their respective geometric centers as references; then, the optimal rotation is obtained by constructing the correlation matrix between the two sets of decentralized control points and performing matrix decomposition to obtain the rotation relationship that makes the orientation of the two sets of point clouds most consistent. Specifically, the correlation matrix is ​​obtained by performing a coordinate product table for each pair of decentralized control points, and then summing up the product tables of all point pairs to obtain a matrix of fixed size; finally, the translation amount is calculated based on the two sets of geometric centers, so that the geometric center on the scene model side coincides with the geometric center on the global reference side after rotation transformation and then translation, thus obtaining the anchoring transformation relationship that includes rotation and translation.

[0045] The control point positions after anchoring transformation are projected back to the images from each viewpoint and compared with the pixel labels of the control points in the images to obtain pixel-level deviations, which are then summarized as the reprojection error of the spatial control points. Anchoring is considered valid when the reprojection error is not greater than a preset error threshold; otherwise, it is considered invalid, triggering the reselection of control points and increasing the number of control points. This anchors the 3D Gaussian splash scene model to the substation's real-world coordinates, ensuring that the 3D position of any Gaussian body in the scene model has a corresponding substation real-world coordinate representation. This provides a unified coordinate basis for subsequent hazard boundary generation, minimum clearance calculation, and alarm location.

[0046] In one specific embodiment, the reprojection error of the spatial control points is calculated through the following process: Let Pk be the global 3D coordinates of the k-th spatial control point. Camera intrinsic and extrinsic parameters from a certain viewpoint are used to project Pk onto the pixel plane to obtain the projected pixel P'k. At the same viewpoint, the observed pixel of the control point is pk. The reprojection error of the control point at that viewpoint is defined as the Euclidean distance between the two points on the pixel plane, in pixels. The root mean square value of the reprojection error of a control point at multiple viewpoints is taken as the error of that control point; the average of the errors of all control points is then taken as the total reprojection error. If the total reprojection error is not greater than a preset error threshold, the geographic anchoring is considered valid; otherwise, the geographic anchoring is considered invalid.

[0047] The danger boundary representation generation module is used by the edge computing unit to receive the substation operation work order and generate a set of operation steps. Based on the set of operation steps, it generates the danger boundary representation of the substation operation in the 3D Gaussian splash scene model.

[0048] The set of work steps includes at least a set of permitted areas, a set of prohibited objects, and the status of safety measure evidence. The status of safety measure evidence includes at least one of the following: confirmed, unconfirmed, revoked, or expired.

[0049] The permitted area set is used to define the spatial range in which workers are permitted to enter, stay, and move under the current procedure. This set can be determined by the work area marked on the work permit, the area boundary enclosed by the fence or isolation zone inside the station, and the safe work area confirmed on site, and is represented by the area boundary under the actual coordinates of the substation.

[0050] The prohibited access object set is used to list dangerous objects that require maintaining a safe distance or are prohibited from being approached in the current step. These objects can be live parts, exposed conductors of primary equipment, busbar segments, terminal connection points, and other live dangerous targets, or dangerous spaces that have been designated as prohibited areas. A correspondence is established between these objects and the semantic identifiers of dangerous parts in the 3D Gaussian splash scene model.

[0051] The safety measure evidence status is used to characterize whether the safety measures related to the above-mentioned permitted stay areas and prohibited access objects, such as isolation, voltage testing, grounding, and tagging, have been completed and are in a valid state. Confirmed means that the corresponding safety measures have been completed and are within the validity period. Unconfirmed means that they have not been completed or lack traceable confirmation records. Revoked means that the implemented safety measures have been lifted or are no longer valid. Expired means that the confirmation record has exceeded the preset validity period or the site conditions have changed, making the original confirmation no longer applicable. This allows the system to dynamically adjust the work permission boundaries and danger boundaries as the steps change and trigger corresponding early warning controls.

[0052] Based on the set of prohibited objects and the status of safety measures evidence, the set of charged hazardous targets is extracted by semantic index in the 3D Gaussian splash scene model and a hazardous boundary representation is generated. The set of permitted stay areas is then mapped to the permitted work boundary.

[0053] Both the hazardous boundary and the permissible work boundary are represented using the actual coordinates of the substation.

[0054] The process of generating a hazardous boundary representation is as follows: Based on the set of prohibited access objects and the status of safety measure evidence, a set of energized hazardous targets is determined through preset constraints. These preset constraints include at least: determining each target object item in the set of work steps; the target object item can be either an equipment item or a confined space item; when a target object item belongs to the prohibited access object set, it is identified as an energized hazardous target; when the safety measure evidence status of a target object item is unconfirmed, revoked, or expired, it is identified as an energized hazardous target. When the safety measure evidence status of a target object item is confirmed, it is removed from the set of energized hazardous targets.

[0055] For each target in the set of charged hazardous targets, the semantic index library is called to retrieve the set of Gaussian bodies corresponding to the charged hazardous target in the 3D Gaussian splash scene model. The semantic index library records the correspondence between the Gaussian body index and the semantic identifier of the hazardous part. The retrieved target Gaussian body set is used to generate a hazardous boundary representation.

[0056] The hazardous boundary is represented by a semantic identifier for the corresponding hazardous location. Specifically, when constructing the hazardous boundary of the substation work area, the spatial boundary of each hazardous target is associated with its corresponding hazardous location semantic identifier. This ensures that each hazardous boundary not only represents its location and shape but also clearly indicates the specific hazardous location or equipment it corresponds to. Specifically, when constructing hazardous boundaries, the edge computing unit assigns a hazardous location semantic identifier to each hazardous boundary segment or geometry. This identifier indicates the category, degree of hazard, and work restrictions of the equipment or location represented by the hazardous boundary segment, such as "live equipment," "exposed conductor," or "disconnector." In this way, by combining hazardous boundaries with their semantic identifiers, the monitoring system can clearly identify the actual risk object behind each hazardous boundary when performing minimum clearance calculations, work area division, and alarm triggering, ensuring that workers can accurately understand the specific hazardous location they are facing.

[0057] In this embodiment, the semantic identifier of hazardous locations is an coded field used to uniquely identify equipment and locations within a substation. It consists of an equipment identifier field and a location identifier field. The equipment identifier field is a unique equipment number or bay number from the equipment ledger, and the location identifier field is a preset location enumeration value. The location enumeration value includes at least busbars, contacts, exposed conductors, disconnector gaps, fences, passageways, and ground structures. The semantic identifier of hazardous locations is stored in string or integer form and maintains uniqueness and stability within the same scenario model version.

[0058] To enable the 3D Gaussian splash scene model to have computable hazard target localization capabilities, this implementation establishes a semantic index library. The semantic index library records the mapping relationship between Gaussian volume indices and semantic identifiers: each Gaussian volume in the scene model is assigned a unique Gaussian volume index number and a semantic identifier. The semantic identifiers are generated through one of the following methods: (1) Offline annotation method: Select the equipment and parts in the three-dimensional view by three-dimensional area selection or three-dimensional brush annotation, and assign the same semantic identifier to the selected Gaussian volume index.

[0059] (2) Model-assisted method: Semantic segmentation is performed on multi-view images to obtain pixel-level categories, and then the pixel categories are back-projected onto a three-dimensional Gaussian volume set through camera pose. The Gaussian volume set with consistent multi-view views is assigned a corresponding semantic label. The above methods can be used alone or in combination.

[0060] The hazard boundary is represented in at least one of the following forms: a triangular mesh of the target's outer surface, used to characterize the outer surface of the hazard target and support the calculation of the nearest distance from point to surface; a set of target boundary points, used to characterize the outline of the hazard target with discrete boundary sampling points and support the calculation of the nearest distance from point to point set or from point to local fitted boundary; a set of target-occupied voxels, used to represent the space occupied by the hazard target with a set of voxel units and to use its boundary voxels as the hazard boundary for distance querying; and a target distance field boundary, used to provide the ability to query the distance to the hazard target for the location in the work space in the form of a distance field and to use the boundary of the distance isosurface as the hazard boundary for minimum clearance calculation, thereby achieving consistent clearance assessment and early warning triggering with replaceable boundary representation methods under different modeling accuracy and computing power conditions.

[0061] The set of permitted dwell areas corresponding to the current work step is mapped to the work permission boundary. This mapping includes at least: converting the set of permitted dwell areas from the area description in the work order or permit into a two-dimensional polygonal region or a three-dimensional volumetric region in the substation's real-world coordinates, thus obtaining the work permission boundary representation; when the set of permitted dwell areas includes height ranges or equipment level restrictions, the two-dimensional polygonal region is generated into a corresponding three-dimensional volumetric region according to the height range or equipment level restrictions. The equipment level restriction refers to dividing the work space into preset levels such as ground level, operation level, maintenance platform level, and upper equipment connection level based on the vertical layering of the equipment structure and the substation's work platform, and limiting the permitted dwell area to be valid only within the specified level. Therefore, when generating a three-dimensional volumetric region from the two-dimensional polygonal region, only the height range of the corresponding level is stretched or trimmed.

[0062] For each human body-occupied geometry and each dangerous boundary representation, perform the nearest distance calculation to obtain the minimum net distance between the pair of geometries and the corresponding nearest point pair; wherein, the minimum net distance is represented as the minimum Euclidean distance, and the nearest point pair includes at least the coordinates of the nearest point on the human body-occupied geometry and the coordinates of the nearest point on the dangerous boundary segment.

[0063] The minimum clearance is represented as follows: Figure 4 As shown, Figure 4 This diagram illustrates the minimum clearance and nearest point pair. It shows the calculation and interpretable output method of the minimum clearance between the human tool envelope and the hazard boundary representation. In the same three-dimensional coordinate system, the monitoring system performs a nearest distance query on the human tool envelope and the hazard boundary representation, obtains a pair of nearest points with the smallest distance, and calculates the spatial distance between the two points as the minimum clearance. The diagram visually represents the positional relationship of the nearest point pairs by connecting them, and can synchronously associate the semantic identifier of the hazard part corresponding to the nearest point pair. This is used to clarify which equipment boundary or which part of the human body or tool is approaching the most dangerous point, thus providing a traceable basis for triggering alert-level alarms and hazard-level alarms, and supporting the explanation of the alarm source and handling direction in alarm pop-ups or playback.

[0064] When the dangerous boundary segment is represented by a triangular mesh on the outer surface of the target or a set of target boundary points, a spatial index structure is used to traverse or search the boundary elements, calculate the minimum distance from the human-occupied geometry to the boundary elements, and output the nearest point pair. When the dangerous boundary segment is represented by a set of target-occupied voxels or a target distance field boundary, a distance field query or boundary voxel query is used to calculate the minimum distance from the human-occupied geometry to the boundary and output the nearest point pair.

[0065] Based on the time series of minimum clearances at continuous time points, the state of clearance change direction is obtained through robust statistics and difference determination within a sliding time window. The state of clearance change direction includes at least an approaching state, a moving away state, and a stable state. The state of clearance change direction is bound to the semantic identifier of the dangerous part corresponding to the nearest point pair. Specifically, when generating the state of clearance change direction, the minimum clearance time series is statistically analyzed and determined only on the premise that the semantic identifier of the dangerous part in the dangerous boundary segment into which the nearest point pair falls remains consistent. The approaching state, moving away state, or stable state obtained from the determination is output as the state of clearance change direction corresponding to the semantic identifier of the dangerous part, so as to avoid misjudgment caused by the switching of the nearest point pair between different dangerous targets.

[0066] The sliding time window refers to a continuous sample interval covering a preset duration, with the current time as the baseline, and is updated over time in preset steps. Robust statistics involve first removing outlier or low-confidence samples from the minimum net distance samples within the sliding time window based on positioning accuracy indicators, and then obtaining the representative net distance value for that time window using the median or truncated mean. Difference determination involves comparing the representative net distance value of the current sliding time window with that of the previous sliding time window. When the representative net distance value continuously decreases and the decrease is not less than a preset change threshold, it is determined to be in a close state; when the representative value continuously increases and the increase is not less than a preset change threshold, it is determined to be in a distant state; and when the change in the representative value is less than a preset change threshold, it is determined to be in a stable state.

[0067] The alarm constraint set includes at least the minimum clearance, clearance change direction status, positioning accuracy index, and uncertainty index of the reachable area of ​​the tool's front end. The uncertainty index is represented by the volume of the three-dimensional occupied space of the reachable area of ​​the tool's front end. The volume characterizes the space occupied by the set of possible positions of the tool's front end (tip, blade, or probe) in three-dimensional space: the edge computing end extracts and maps the hit voxels mapped to the three-dimensional voxel mesh as occupied, thus obtaining the three-dimensional occupied volume of the reachable area of ​​the tool's front end; finally, the volume value of the reachable area is obtained by multiplying the number of occupied voxels by the volume of a single voxel, which is used to quantify the magnitude of the uncertainty in the front end's position.

[0068] The two-level alarm triggering module is used to form an alarm constraint set by representing the danger boundary and calculating the minimum clearance between the human tool envelope. Based on the alarm constraint set, two-level alarms are triggered, including reminder alarms or danger alarms, to complete the substation operation monitoring and early warning.

[0069] The two-level alarm is triggered as follows: Based on the minimum clearance distance, when the monitoring system is in a non-alarm state, if the minimum clearance distance is less than the first entry clearance distance threshold and the duration is not less than the first entry duration threshold, then it enters an alert-level alarm state. When the system is in an alert-level alarm state, if the minimum clearance distance is greater than the first release clearance distance threshold and the duration is not less than the first release duration threshold, then the alert-level alarm is released and the system returns to a non-alarm state. The first entry clearance distance threshold and the first release clearance distance threshold satisfy a hysteresis relationship where the entry threshold is less than the release threshold.

[0070] When the monitoring system is in a non-hazardous alarm state, if the minimum clearance distance is less than the second entry clearance threshold and remains at least as long as the second entry duration threshold, a hazard alarm is triggered and the system enters a hazard alarm state. When the system is in a hazard alarm state, if the minimum clearance distance is greater than the second release clearance threshold and remains at least as long as the second release duration threshold, the hazard alarm is released and the system returns to a non-hazardous alarm state. The second entry clearance threshold and the second release clearance threshold satisfy a hysteresis relationship where the entry threshold is less than the release threshold, and the second entry clearance threshold is less than the first entry clearance threshold.

[0071] In addition to meeting the second entry clearance threshold and the second entry duration threshold, the triggering of a danger-level alarm also requires that the clearance change direction state is an approaching state and the duration is not less than the escalation release duration threshold. When the clearance change direction state is a distant state or a stable state, it is prohibited to directly jump from an alert-level alarm to a danger-level alarm.

[0072] Based on the positioning accuracy index, when the positioning accuracy index is not greater than the second accuracy threshold, the first entry threshold and the first release threshold are used to determine the reminder-level alarm.

[0073] When the positioning accuracy index is greater than the second accuracy threshold, the entry net distance threshold of the alert level alarm will be adjusted to the conservative entry net distance threshold, and the first entry duration threshold and the second entry duration threshold will be switched to the corresponding conservative duration threshold.

[0074] The conservative entry clearance threshold is specifically obtained by mapping the difference between the positioning accuracy index and the second accuracy threshold to obtain the entry clearance threshold correction amount, and then adding the entry clearance threshold correction amount to the first entry clearance threshold and the second entry clearance threshold respectively to obtain the conservative entry clearance threshold.

[0075] When the uncertainty index of the reachable area at the front end of the tool exceeds the preset uncertainty threshold, the danger level alarm will be downgraded to the alert level alarm until the uncertainty index is no greater than the preset uncertainty threshold and the continuous duration is no less than the uncertainty resolution duration threshold.

[0076] Generating a dangerous boundary representation also includes: When the status of safety measure evidence changes, the update of the hazard boundary representation is triggered and the freeze window is activated. The update of the hazard boundary representation includes at least redefining the set of live hazardous targets based on the set of work steps and generating the corresponding hazard boundary representation in the 3D Gaussian splash scene model.

[0077] In this embodiment of the invention, the freeze window is an alarm stabilization control period used to suppress the impact of instantaneous calculation mutations caused by boundary switching on alarm levels when the hazard boundary representation is updated. Specifically, when changes in the set of work steps or the status of safety measure evidence changes from confirmed to unconfirmed, revoked, or expired, resulting in adjustments to the permitted area or prohibited object, thereby triggering an update of the hazard boundary representation, the edge computing unit may obtain a mutation value of the minimum clearance in two adjacent calculations. This mutation is not necessarily caused by the actual approach of workers or tools to the hazardous target, but may be caused by changes in the geometric position, boundary segment division, or semantic binding of the hazard boundary. Therefore, the edge computing unit activates the freeze window after detecting an update of the hazard boundary representation. The freeze window lasts for one to two seconds, restricting rapid jumps in alarm levels within the freeze window. For example, it prohibits the direct escalation from a warning-level alarm to a hazard-level alarm, and prioritizes escalation only after the continuous decrease in the minimum clearance or the change in the clearance direction meets the release conditions. This reduces false escalations caused by hazard boundary updates and improves the stability and interpretability of alarm triggering.

[0078] Within the frozen window, the prohibition constraint restricts the alert level alarm from directly transitioning to the danger level alarm. If the minimum clearance calculated based on the output of the Beidou RTK personnel positioning terminal and the updated 3D Gaussian splash danger boundary representation meets the allowed triggering conditions, specifically, it is a continuous decrease with a decrease duration not less than the preset duration, then the prohibition constraint is lifted and the danger level alarm is allowed to trigger.

[0079] Specifically, when the displacement of the three-dimensional coordinates of the nearest point of the dangerous boundary corresponding to the minimum net distance between two adjacent calculations is greater than the preset jump threshold or the semantic identifier of the dangerous part changes, the time period is determined to be the boundary switching state and the prohibition constraint is maintained until the freeze window ends or the boundary switching state disappears and continues for a period of not less than the preset duration.

[0080] Example 2: With other conditions remaining unchanged in Real-Time Example 1, the minimum net distance calculation process can also be implemented through a distance field query: The edge computing unit discretizes the dangerous boundary into a three-dimensional spatial grid with a preset resolution under the real-world coordinates of the substation. It pre-calculates the shortest distance to the dangerous boundary for each spatial location within the grid and generates a target distance field. At the same time, it associates the coordinates of the nearest dangerous boundary point and the semantic identifier of the dangerous part with each grid cell in the distance field, which serves as index information for subsequent queries.

[0081] During operation monitoring, the edge computing unit acquires the geometry occupied by the human body and the reachable area of ​​the tool front end in each computing cycle. It selects a set of envelope surface sampling points on the outer surface of the human-tool envelope according to a preset sampling density, maps each sampling point to a distance field coordinate index, and reads the distance field value at the sampling point and its associated nearest dangerous boundary point and semantic identifier through interpolation. The edge computing unit selects the sampling point with the smallest distance field value from the query results of all sampling points as the nearest point on the envelope side, and takes the nearest dangerous boundary point corresponding to the sampling point as the nearest point on the dangerous side, thereby obtaining the minimum net distance and the nearest point pair, as well as the semantic identifier of the corresponding dangerous part.

[0082] When the dangerous boundary representation is updated, the edge computing unit performs a full recalculation of the distance field or only performs incremental updates on the affected local spatial regions. It can also increase the sampling density or refine the local grid resolution when the minimum clearance is close to the entry threshold to ensure the stability and accuracy of the minimum clearance and nearest point pair output.

[0083] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined by the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK, characterized in that, include: The coordinate and model building module is used to generate the three-dimensional coordinates of the workers in the substation operation area using Beidou RTK to construct the human tool envelope, and to collect multi-view image sequences of the substation operation area to generate a 3D Gaussian splash scene model. The hazardous boundary representation generation module is used by the edge computing unit to receive the substation operation work order and generate a set of operation steps, and generate the hazardous boundary representation of the substation operation in the 3D Gaussian splash scene model based on the set of operation steps. The two-level alarm triggering module is used to form an alarm constraint set by calculating the minimum clearance between the dangerous boundary representation and the human tool envelope. Based on the alarm constraint set, two-level alarms are triggered, including reminder alarms or danger alarms, to complete the substation operation monitoring and early warning.

2. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 1, characterized in that, Includes the following steps: The three-dimensional coordinates of the workers are specifically obtained by establishing a global coordinate reference based on the differential correction data output by the Beidou RTK base station in the substation operation area. The Beidou RTK personnel positioning terminal obtains the three-dimensional coordinates of the workers in real time based on the differential correction data and outputs the positioning accuracy index simultaneously. The 3D Gaussian splash scene model is generated by collecting multi-view image sequences covering the substation operation area, and the scene model is anchored to the substation real-world coordinates using a global coordinate reference.

3. The substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK as described in claim 2, characterized in that, Includes the following steps: The construction of the human tool envelope specifically involves using the three-dimensional coordinates of the operator as the center point of the human geometry to construct the geometry occupied by the human body, and outputting the reachable area of ​​the tool front end based on the tool positioning terminal and multi-view image sequence to form the human tool envelope. The anchoring specifically involves using a group of spatial control points to perform geographic anchoring on the scene model. The spatial control points are fixed structural points with known coordinates under a global coordinate reference. The reprojection error of the spatial control points is not greater than a preset error threshold as a valid anchoring criterion.

4. The substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK as described in claim 3, characterized in that: The construction of the human tool envelope also includes the presence of only BeiDou RTK personnel positioning terminals and no tool positioning terminals: The edge computing unit performs human key point detection on multi-view images, obtains the pixel coordinates of the worker's hand in the coordinate system of each view image, and outputs the visibility confidence score. Views with a visibility confidence score not less than a preset visibility threshold are determined as valid viewpoints. The effective viewpoint of the hand's pixel coordinates are projected onto the 3D Gaussian splash scene model to obtain candidate 3D hand positions. Consistency judgment is performed on each candidate 3D hand position and the confidence of the 3D hand position is calculated to determine the tool's grip point or the grip point's reachable area and generate the reachable area of ​​the tool's front end.

5. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 1, characterized in that, Includes the following steps: The set of work steps includes at least a set of permitted areas, a set of prohibited objects, and the status of safety measure evidence. The status of safety measure evidence includes at least one of the following: confirmed, unconfirmed, revoked, or expired. Based on the set of prohibited objects and the status of safety measures evidence, the set of charged dangerous targets is extracted by semantic index in the 3D Gaussian splash scene model and a dangerous boundary representation is generated. The set of allowed stay areas is then mapped to the work allowable boundary. Both the hazardous boundary representation and the work permit boundary are represented using real-world coordinates of the substation.

6. The substation operation monitoring and early warning system based on 3D Gaussian splashing and Beidou RTK as described in claim 5, characterized in that: The specific generation process for the generated hazardous boundary representation is as follows: Based on the set of prohibited objects and the status of safety measure evidence, a set of live hazardous targets is determined through preset constraints. The preset constraints include at least: judging each target object item in the set of work steps; when the target object item belongs to the set of prohibited objects, the target object item is determined as a live hazardous target; when the safety measure evidence status of the target object item is unconfirmed, revoked, or expired, the target object item is determined as a live hazardous target. For each target in the set of charged hazardous targets, the semantic index library is called to retrieve the set of Gaussian volumes corresponding to the charged hazardous target in the 3D Gaussian splash scene model, and the retrieved target Gaussian volume set is used to generate a hazardous boundary representation.

7. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 1, characterized in that, Includes the following steps: The minimum clearance is expressed as the minimum Euclidean distance between the danger boundary and the human tool envelope; Based on the time series formed by the minimum net distance at continuous time points, the direction of net distance change is obtained within the sliding time window through robust statistics and difference determination. The alarm constraint set includes at least the minimum clearance, the clearance change direction status, the positioning accuracy index, and the uncertainty index of the reachable area of ​​the tool front end. The states of the direction of change of net distance include approaching state, moving away state, and stable state, and the uncertainty index is the volume of the reachable area at the front end of the tool.

8. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 1, characterized in that, The triggering of the two-level alarm is specifically as follows: Based on the minimum clearance distance, when the minimum clearance distance is less than the first entry clearance distance threshold and is not less than the first entry duration threshold, an alert-level alarm is triggered; when the minimum clearance distance is greater than the first release clearance distance threshold and is not less than the first release duration threshold, the alert-level alarm is released. When the minimum clearance is less than the second entry clearance threshold and remains not less than the second entry duration threshold, a danger level alarm is triggered; when the minimum clearance is greater than the second release clearance threshold and remains not less than the second release duration threshold, the danger level alarm is released. In addition to satisfying the second entry clearance threshold and the second entry duration threshold, the triggering of the danger level alarm also requires that the clearance change direction state is in the approach state and lasts for a duration not less than the upgrade release duration threshold. When the clearance change direction state is in the distance state or stable state, it is prohibited to directly jump from the reminder level alarm to the danger level alarm.

9. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 8, characterized in that, Includes the following steps: Based on the positioning accuracy index, when the positioning accuracy index is not greater than the second accuracy threshold, the first entry threshold and the first release threshold are used to determine the reminder-level alarm. When the positioning accuracy index is greater than the second accuracy threshold, the entry net distance threshold of the alert level alarm will be adjusted to the conservative entry net distance threshold, and the first entry duration threshold and the second entry duration threshold will be switched to the corresponding conservative duration threshold. When the uncertainty index of the reachable area at the front end of the tool exceeds the preset uncertainty threshold, the danger level alarm will be downgraded to the alert level alarm until the uncertainty index is no greater than the preset uncertainty threshold and the continuous duration is no less than the uncertainty resolution duration threshold.

10. The substation operation monitoring and early warning system based on 3D Gaussian splashing and BeiDou RTK as described in claim 5, characterized in that, The generation of the dangerous boundary representation also includes: When the status of safety measure evidence changes, the update of the hazard boundary representation is triggered and the freeze window is activated. The update of the hazard boundary representation includes at least redetermining the set of live hazardous targets based on the set of work steps and generating the corresponding hazard boundary representation in the 3D Gaussian splash scene model. Within the frozen window, the prohibition constraint restricts the alert level alarm from directly transitioning to the danger level alarm. If the minimum clearance calculated based on the output of the Beidou RTK personnel positioning terminal and the updated 3D Gaussian splash danger boundary representation meets the allowed triggering conditions, then the prohibition constraint is lifted and the danger level alarm is allowed to trigger.

Images