Substation three-dimensional inspection system and method, storage medium and computer device

Abstract

The application discloses a substation three-dimensional inspection system and method, a storage medium and a computer device. The method comprises the following steps: a three-dimensional modeling module constructs a three-dimensional real scene model of a substation; a data fusion module respectively fuses and processes multi-source operation monitoring data of each entity device in the substation to obtain fusion feature data of each entity device; a task planning module determines a target entity device to be inspected based on the fusion feature data of each entity device, generates an inspection path and an inspection range based on the geometric morphology, attribute label, first coordinate and second coordinate of the target entity device obtained by querying the three-dimensional real scene model of the substation; and a task execution module issues the inspection path and the inspection range to a physical inspection device, and receives time-space state data returned by the physical inspection device in real time, and synchronously updates a virtual inspection device corresponding to the physical inspection device in the three-dimensional real scene model of the substation, so as to monitor the execution state of the inspection task in real time.

Description

Technical Field

[0001] This application relates to the field of power equipment inspection technology, and in particular to a three-dimensional inspection system and method for substations, a storage medium, and a computer device. Background Technology

[0002] With the rapid development of my country's power industry and the continuous expansion of the power grid, substations, as key nodes in the power system, have numerous and complex equipment, placing extremely high demands on their safe and stable operation. Traditional manual inspection methods are not only inefficient, but also pose significant personal safety risks as maintenance personnel must be in close contact with high-voltage equipment. In recent years, intelligent inspection equipment such as drones and robots have begun to be applied in the power industry, improving the level of automation in inspections to some extent. However, how to achieve intelligent operation and maintenance management of the entire substation and all its equipment remains a challenging problem that the industry is currently striving to solve.

[0003] Existing intelligent inspection systems still have many limitations in practical applications. For example, inspection task planning is mostly based on two-dimensional maps or simple equipment lists, which cannot understand and utilize the complex three-dimensional spatial structure of substations. This makes inspection paths dependent on human experience, resulting in insufficiently precise planning, collision risks, and unstable data acquisition quality. Furthermore, substation equipment operating data comes from multiple sensors such as temperature, vibration, and images, but existing intelligent inspection systems mostly use single operating data points to assess the operating status of substation equipment, failing to fully explore the deep correlation characteristics of multi-source data, leading to insufficient accuracy in equipment operating status assessment. Summary of the Invention

[0004] In view of this, this application provides a substation three-dimensional inspection system and method, storage medium, and computer equipment. It utilizes a feedforward neural network to perform feature-level fusion of multi-source operational monitoring data of physical equipment, extracting deep correlation features of equipment operating status, which can effectively improve the accuracy of equipment status perception and fault prediction. Based on the fused feature data, the system dynamically determines inspection targets and, combined with the geometric shape, coordinate information, and attribute labels of physical equipment in the substation's three-dimensional real-scene model, generates inspection paths adapted to the substation layout and equipment operating status, ensuring the completeness and accuracy of inspection coverage. By synchronizing the spatiotemporal status of physical inspection equipment with virtual inspection equipment in the substation's three-dimensional real-scene model, the system intuitively displays the inspection progress and equipment status, and automatically generates fine-tuning instructions when path anomalies occur, forming a closed-loop control of "perception-decision-execution." This achieves precise execution and real-time correction of the inspection process, greatly improving the efficiency, safety, and intelligence level of inspection operations.

[0005] According to one aspect of this application, a three-dimensional inspection system for substations is provided, comprising:

[0006] The 3D modeling module is used to construct a 3D real-scene model of a substation with semantic information. Each entity equipment model in the 3D real-scene model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation.

[0007] The data fusion module is used to fuse the multi-source operation monitoring data of each physical device in the substation using a feedforward neural network to obtain the fused feature data of each physical device.

[0008] The task planning module is used to respond to task processing requests, determine the target physical equipment to be inspected based on the fused feature data of each physical equipment, query the geometric shape, attribute labels, first coordinates and second coordinates of the target physical equipment based on the three-dimensional real scene model of the substation, and generate the inspection path and inspection range of the target physical equipment according to the geometric shape, the attribute labels, the first coordinates and the second coordinates.

[0009] The task execution module is used to send the inspection path and inspection range to the physical inspection equipment, and receive the spatiotemporal status data returned by the physical inspection equipment in real time. It drives the virtual inspection equipment corresponding to the physical inspection equipment to be updated synchronously in the three-dimensional real scene model of the substation, so as to monitor the execution status of the inspection task in real time in the three-dimensional real scene model of the substation, and generate fine-tuning instructions when an abnormality is detected in the inspection path, and feed the fine-tuning instructions back to the physical inspection equipment.

[0010] According to another aspect of this application, a three-dimensional inspection method for substations is provided, comprising:

[0011] The 3D modeling module constructs a 3D real-scene model of the substation with semantic information. Each entity equipment model in the 3D real-scene model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation.

[0012] The data fusion module uses a feedforward neural network to fuse the multi-source operation monitoring data of each physical device in the substation to obtain the fused feature data of each physical device.

[0013] The task planning module responds to the task processing request, determines the target physical equipment to be inspected based on the fused feature data of each physical equipment, queries the geometric shape, attribute label, first coordinate and second coordinate of the target physical equipment based on the three-dimensional real scene model of the substation, and generates the inspection path and inspection range of the target physical equipment according to the geometric shape, attribute label, first coordinate and second coordinate.

[0014] The task execution module sends the inspection path and inspection range to the physical inspection equipment and receives the spatiotemporal status data returned by the physical inspection equipment in real time. It drives the virtual inspection equipment corresponding to the physical inspection equipment to be updated synchronously in the three-dimensional reality model of the substation, so as to monitor the execution status of the inspection task in real time in the three-dimensional reality model of the substation. When an abnormality is detected in the inspection path, a fine-tuning instruction is generated and fed back to the physical inspection equipment.

[0015] According to another aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the above-described substation three-dimensional inspection method.

[0016] According to another aspect of this application, a computer device is provided, including a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, wherein the processor executes the program to implement the above-described substation three-dimensional inspection method.

[0017] By employing the aforementioned technical solutions, this application provides a substation three-dimensional inspection system and method, storage medium, and computer equipment. It utilizes a feedforward neural network to perform feature-level fusion of multi-source operational monitoring data of physical equipment, extracting deep correlation features of equipment operating status. This effectively improves the accuracy of equipment status perception and fault prediction. Based on the fused feature data, it dynamically determines inspection targets and, combined with the geometric shape, coordinate information, and attribute labels of physical equipment in the substation's three-dimensional real-scene model, generates inspection paths adapted to the substation layout and equipment operating status, ensuring the completeness and accuracy of inspection coverage. By synchronizing the spatiotemporal status of physical inspection equipment with that of virtual inspection equipment in the substation's three-dimensional real-scene model, it intuitively displays the inspection progress and equipment status. Furthermore, it automatically generates fine-tuning instructions when path anomalies occur, forming a closed-loop control of "perception-decision-execution." This achieves precise execution and real-time correction of the inspection process, greatly improving the efficiency, safety, and intelligence level of inspection operations.

[0018] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0019] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0020] Figure 1This paper shows a schematic diagram of the structure of a substation three-dimensional inspection system provided in an embodiment of this application;

[0021] Figure 2 A flowchart illustrating a three-dimensional inspection method for a substation provided in an embodiment of this application is shown.

[0022] Figure 3 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0023] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0024] This embodiment provides a substation three-dimensional inspection system, such as Figure 1 As shown, the system includes:

[0025] The 3D modeling module is used to construct a 3D real-scene model of a substation with semantic information. Each entity equipment model in the 3D real-scene model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation.

[0026] The data fusion module is used to fuse the multi-source operation monitoring data of each physical device in the substation using a feedforward neural network to obtain the fused feature data of each physical device.

[0027] The task planning module is used to respond to task processing requests, determine the target physical equipment to be inspected based on the fused feature data of each physical equipment, query the geometric shape, attribute labels, first coordinates and second coordinates of the target physical equipment based on the three-dimensional real scene model of the substation, and generate the inspection path and inspection range of the target physical equipment according to the geometric shape, the attribute labels, the first coordinates and the second coordinates.

[0028] The task execution module is used to send the inspection path and inspection range to the physical inspection equipment, and receive the spatiotemporal status data returned by the physical inspection equipment in real time. It drives the virtual inspection equipment corresponding to the physical inspection equipment to be updated synchronously in the three-dimensional real scene model of the substation, so as to monitor the execution status of the inspection task in real time in the three-dimensional real scene model of the substation, and generate fine-tuning instructions when an abnormality is detected in the inspection path, and feed the fine-tuning instructions back to the physical inspection equipment.

[0029] The substation three-dimensional inspection system provided in this application realizes intelligent inspection through modular design. It mainly includes a three-dimensional modeling module, a data fusion module, a task planning module, and a task execution module. The modules work together to form a closed-loop management.

[0030] Specifically, the 3D modeling module first constructs a 3D reality model of the substation. This model not only includes the geometric shape of each physical device (such as the shape of transformers and the dimensions of switchgear), but also assigns attribute tags (such as equipment type, voltage level, and equipment identification) to each physical device model through semantic annotation technology. The 3D reality model of the substation adopts a dual-coordinate system positioning mechanism. Each physical device model is simultaneously labeled with a first coordinate in the world coordinate system (for global positioning) and a second coordinate in the local coordinate system within the substation (for precise positioning). This dual-coordinate design supports compatibility with external geographic information systems (GIS) and meets the precise navigation needs of densely populated equipment areas within the substation.

[0031] The data fusion module incorporates a feedforward neural network. This network takes multi-source operational monitoring data (such as infrared thermal imaging and partial discharge signals) from physical equipment within the substation as input to its input layer. Through nonlinear transformations in the hidden layers, it extracts correlation features from the multi-source monitoring data, ultimately enabling the output layer to generate fused feature data reflecting the health status of the physical equipment. For example, for transformer equipment, the feedforward neural network can simultaneously fuse oil chromatography analysis data, winding temperature data, and ultrasonic partial discharge data to generate fused feature data reflecting the transformer's health status. This data is more reliable than the judgment results from a single data source. In a specific embodiment, the data fusion module can deploy an independent feedforward neural network for each physical equipment.

[0032] Upon receiving a task processing request, the task planning module first filters the target physical equipment to be inspected based on fused feature data. For example, if the fused feature data of a certain physical equipment indicates that its health index is below a preset threshold, it is automatically included in the scope of target physical equipment to be inspected. Here, the task processing request can be manually triggered or automatically generated by the system at regular intervals; no limitation is made here. Subsequently, the task planning module queries the 3D reality model of the substation to obtain the geometric shape (such as equipment height, radius of danger zone), attribute information (such as equipment number, bay to which it belongs), and dual-coordinate data of the target physical equipment. Based on this information, the task planning module can generate the optimal inspection path and inspection range accordingly.

[0033] The task execution module can send inspection path and range data to physical inspection equipment (such as wheeled robots or drones). During the inspection task, the onboard sensors (such as visible light cameras and infrared thermal imagers) collect and transmit spatiotemporal status data in real time. Based on the transmitted spatiotemporal status data, the task execution module synchronously drives the corresponding virtual inspection equipment in the substation's 3D reality model. A coordinate mapping algorithm is used to convert the real-time position (world coordinates) of the physical inspection equipment into local coordinates in the substation's 3D reality model, achieving pose synchronization between the virtual and real worlds. When the physical inspection equipment deviates from the inspection path (e.g., due to obstacles) or temporary obstacles appear on the path, the task execution module can generate fine-tuning instructions based on reinforcement learning algorithms to adjust the physical inspection equipment's inspection speed or turning angle, ensuring task continuity.

[0034] By applying the technical solution of this embodiment, feature-level fusion of multi-source operation monitoring data of physical equipment is performed using a feedforward neural network to extract deep correlation features of equipment operation status, which can effectively improve the accuracy of equipment status perception and fault prediction. Based on the fused feature data, the inspection target is dynamically determined. Combined with the geometric shape, coordinate information and attribute labels of physical equipment in the three-dimensional real scene model of the substation, an inspection path adapted to the substation layout and equipment operation status is generated, ensuring the completeness and accuracy of inspection coverage. By synchronizing the spatiotemporal status of physical inspection equipment with virtual inspection equipment in the three-dimensional real scene model of the substation, the inspection progress and inspection equipment status are displayed intuitively. When the path is abnormal, fine-tuning instructions are automatically generated, forming a closed-loop control of "perception-decision-execution", realizing accurate execution and real-time correction of the inspection process, which greatly improves the efficiency, safety and intelligence level of inspection operations.

[0035] Optionally, in this embodiment, the 3D modeling module is used to: fuse point cloud data obtained by LiDAR scanning with texture data obtained by oblique photogrammetry to generate a 3D geometric model of the substation with a realistic appearance; perform semantic segmentation and entity reconstruction on the 3D geometric model of the substation to identify each independent entity equipment model; and generate attribute labels for each identified entity equipment model based on the equipment identifier, equipment model and parameter attributes in the digital engineering drawings of the substation, thus forming the 3D real-scene model of the substation with semantic information.

[0036] In this embodiment, the 3D modeling module can first generate a preliminary 3D geometric model of the substation with a realistic appearance. Specifically, point cloud data and texture data can be acquired first. Point cloud data acquired by LiDAR scanning contains spatial coordinate information of substation equipment and buildings, with an accuracy down to the millimeter level, but lacks color and material details; texture data acquired by oblique photogrammetry through multi-angle aerial photography contains rich RGB information and surface texture features, but its spatial resolution is limited by the shooting distance. The 3D modeling module can first perform denoising and registration processing on the point cloud data to eliminate overlap errors and motion distortions during the scanning process. Then, the texture image from the oblique photogrammetry is spatially aligned with the point cloud using the ICP (Iterative Closest Point) algorithm to ensure the geometric accuracy of the texture mapping. During the fusion process, a normal vector-based texture mapping method is used to select the texture fragment with the optimal viewing angle according to the surface normal direction of the point cloud, avoiding texture distortion caused by viewing angle differences, and finally generating a 3D geometric model of the substation that maintains both spatial accuracy and realistic appearance. For example, the heat sink structure of transformer equipment can be accurately restored through point cloud, while its surface paint color and rust marks are supplemented by texture data to form a complete visual representation.

[0037] Next, the 3D modeling module identifies the physical equipment models contained in the 3D geometric model of the substation through semantic segmentation and entity reconstruction. Specifically, an improved PointNet++ network can be used to perform point cloud-level semantic segmentation on the 3D geometric model of the substation. This network identifies the geometric features of different types of physical equipment (such as cylindrical equipment, rectangular switchgear, etc.) through a multi-scale feature extraction layer, and performs instance-level segmentation by combining spatial context information (such as the connection relationships between equipment), dividing the continuous point cloud into independent equipment entities. For complex structural equipment (such as GIS combined electrical appliances), a component decomposition algorithm based on graph neural networks can be used to identify sub-components such as circuit breakers and disconnectors, and reconstruct their topological relationships to achieve entity reconstruction.

[0038] Furthermore, attribute labels can be set for each physical equipment model. Here, the generation of attribute labels can be based on the digital engineering drawings of the substation. The 3D modeling module first performs OCR recognition and vectorization processing on the PDF or DWG format digital engineering drawings of the substation, extracting equipment identifiers (such as KKS codes), equipment models (such as SFP-72.5 / 2000), and parameter attributes (such as rated voltage and capacity). Subsequently, spatial association is established between the equipment identifiers and the physical equipment models. In a specific embodiment, the attribute labels corresponding to each physical equipment model can simultaneously contain static attributes (such as equipment type and serial number) and dynamic attributes (such as current operating status), forming a semantic 3D reality model that supports querying and reasoning. For example, by querying the "opening / closing status" attribute of a circuit breaker in the semantic 3D reality model, combined with its real-time operation monitoring data, it can be determined whether the equipment operating mechanism is functioning normally.

[0039] This application's embodiments achieve a unified approach to geometric modeling and appearance restoration through point cloud-texture fusion, while deep learning-driven semantic segmentation enhances the automation level of device recognition. Furthermore, attribute annotation based on digital engineering drawings enriches the semantic information of the physical device model. Compared to traditional modeling methods, this application's embodiments effectively improve modeling efficiency, model accuracy, and model information richness.

[0040] Optionally, in this embodiment, the data fusion module is configured to: acquire training samples, wherein the training samples include multi-source operation monitoring data samples and real fusion feature data corresponding to each entity device sample; set an input layer and an output layer of an initial feedforward neural network, wherein the input of the input layer is the multi-source operation monitoring data samples of the entity device samples, and the output of the output layer is the predicted fusion feature data of the entity device samples; construct a hidden layer of the initial feedforward neural network, wherein the hidden layer includes multiple neurons, and the output results of the multiple neurons constitute the activation matrix corresponding to the hidden layer; input the multi-source operation monitoring data samples corresponding to each entity device sample into the hidden layer through the input layer of the initial feedforward neural network, perform fusion processing on the multi-source operation monitoring data samples based on the activation matrix of the hidden layer, and output the predicted fusion feature data corresponding to each entity device sample through the output layer; and train the initial feedforward neural network according to the predicted fusion feature data and real fusion feature data corresponding to each entity device sample through an error processing function to obtain the final feedforward neural network.

[0041] The excitation matrix is ​​as follows:

[0042] ;

[0043] Indicates the first i The input feature vector is composed of multi-source operation monitoring data samples of each physical device sample. Indicates the first i The input feature vector of each physical device sample is in the th... j The activation function outputs the results of each neuron. m This indicates the number of entity device samples included in the training samples. n This indicates the number of neurons in the hidden layer;

[0044] The error processing function is as follows:

[0045] ;

[0046] E represents the overall loss value of the feedforward neural network. Indicates the firsti Real-world fusion feature data of individual device samples Indicates the hidden layer number 1 j The output weights of each neuron This indicates that the current feedforward neural network is working on the first... i Predictive fusion feature data for individual physical device samples.

[0047] In this embodiment, the data fusion module can first train a feedforward neural network before deployment. First, the data fusion module can acquire training samples. Acquiring training samples is the foundation for building the feedforward neural network. Specifically, multi-source operational monitoring data samples of physical equipment can be collected from the substation's historical database, including but not limited to temperature sensor data, current transformer data, vibration acceleration data, partial discharge signals, and environmental temperature and humidity data. This data needs to cover the entire lifecycle of the physical equipment samples, including normal operation, anomaly warnings, and fault occurrences, to ensure the representativeness of the samples. Simultaneously, real fused feature data corresponding to each physical equipment sample is generated through expert annotation or fault diagnosis algorithms. This data comprehensively reflects the core indicators of the physical equipment sample's health status.

[0048] The initial feedforward neural network architecture follows a three-layer structure of "input-hidden-output". The hidden layer, as the core of feature extraction, adopts a fully connected structure, and the number of neurons can be determined through experimental tuning. Data fusion processing can achieve nonlinear transformation of multi-source operational monitoring data samples for each physical device sample through the activation matrix of the hidden layer. When the multi-source operational monitoring data samples of the physical device samples are input into the initial feedforward neural network, the input layer converts the multi-source operational monitoring data samples into vector form and passes them to the hidden layer. Each neuron in the hidden layer performs a weighted summation of the input vector (weights are training parameters) and generates activation values ​​through an activation function. The activation values ​​of all neurons constitute the activation matrix. This matrix extracts the correlation features between data through nonlinear transformation, such as simultaneously capturing the coupling relationship between temperature increase and vibration intensification. The output layer linearly combines the activation matrices to generate predictive fusion feature data. In this process, the initial feedforward neural network maps the original low-dimensional monitoring data to a high-dimensional feature space through deep feature extraction in the hidden layer, realizing the transformation from data to knowledge. For example, for circuit breaker equipment, a feedforward neural network can fuse data such as opening and closing coil current, contact temperature, and operating mechanism vibration to output predictive fused feature data that reflects the degree of contact wear.

[0049] The network training employs a backpropagation algorithm and an error optimization mechanism. Predicted fused feature data and true fused feature data are input into an error processing function (such as the mean squared error function), and the loss value between them is calculated. Using a chain rule, the loss value propagates back along the network, adjusting the connection weights between neurons layer by layer to make the prediction result approximate the true value. Finally, when the number of iterations reaches a preset threshold, or the loss value is less than a preset loss threshold, the final feedforward neural network is obtained.

[0050] Optionally, in this embodiment, the task planning module is configured to: calculate a three-dimensional inspection safety zone surrounding the target physical device based on its geometric shape; determine historical inspection data and common defect types corresponding to the attribute tags from a preset knowledge base based on the attribute tags, and determine the device parts to be inspected and the observation angles based on the historical inspection data and the common defect types; generate the inspection range of the target physical device based on the device parts to be inspected and the observation angles; and use an improved BI-RRT based on the first coordinates, the second coordinates, the three-dimensional inspection safety zone, and the inspection range. The path planning algorithm, combining a dynamic target bias strategy and a bias sampling strategy, generates an inspection path from the current position of the physical inspection equipment to the three-dimensional inspection safety area, covering the inspection range.

[0051] In this embodiment, the task planning module can determine the inspection range and inspection path based on the following process. Specifically, the task planning module first calculates the three-dimensional inspection safety area for each target entity device. The calculation of the three-dimensional inspection safety area is constrained by the geometry of the target entity device. For each target entity device, a three-dimensional bounding box centered on the target entity device can be constructed based on its geometry (such as length, width, height, key component positions, and device boundary contours) and the minimum safety distance between devices specified in the substation safety regulations (such as maintaining a distance of at least 0.5 meters between live equipment and the inspection robot). For example, for a cylindrical transformer, a cylindrical safety area extending along its axis can be generated, with a radius equal to the device radius plus the safety distance and a height covering the entire height of the device. Simultaneously, collisions between the safety area and the surrounding environment (such as walls and other equipment) can be detected. If spatial conflicts exist, the safety distance parameters are iteratively adjusted to generate a minimum safety area that satisfies the collision constraints, providing a safety boundary for subsequent path planning. The positioning of the three-dimensional inspection safety area is precisely determined based on the first and second coordinates of the entity device model.

[0052] Furthermore, the task planning module can retrieve historical inspection data (such as inspection cycle, key inspection areas, recommended observation angles, etc.) and common defect types (such as transformer bushing cracks, circuit breaker contact erosion, etc.) corresponding to each target entity device from a pre-set knowledge base through semantic matching based on the target entity device's attribute tags (such as device type, model, operating parameters, etc.). Based on this information, a rule-based reasoning engine is used to determine the equipment parts to be inspected. For example, for GIS switchgear, high-risk areas such as disconnector contacts and density relays are marked as key inspection targets based on historical inspection data. At the same time, combining equipment structural characteristics (such as component observability) and common defect types (such as surface defects requiring frontal observation, internal defects requiring side viewing), the optimal observation angle is selected from a pre-set observation angle library (such as 0°, 45°, 90°, etc.). Finally, the inspection area to be inspected and the observation angle are combined to generate an inspection range covering all critical parts of the equipment. For example, for circuit breakers, task instructions such as "contact area - frontal 45° observation" and "operating mechanism - side 90° observation" are generated.

[0053] The planning of inspection routes can adopt an improved BI-RRT. The algorithm integrates dynamic target bias and biased sampling strategies. First, it acquires the current position coordinates of physical inspection equipment (such as robots or drones) and the coordinates of the entrance point of the 3D inspection safety area (determined based on the first and second coordinates), converting the inspection range into a set of target points in 3D space. This leads to the improved BI-RRT algorithm. The algorithm, based on the traditional Bidirectional Rapidly Expanding Random Tree (BI-RRT), introduces a dynamic target bias strategy. This means that during path expansion, the path grows directly towards the target point (a key observation point within the inspection range) with a certain probability (e.g., 30%), accelerating convergence. Simultaneously, a biased sampling strategy is employed, adjusting the sampling probability in free space based on obstacle areas to prioritize exploring sparse obstacle regions and reduce ineffective searches. After path generation, a collision detection algorithm verifies the path's conformity to the safe zone. If a collision risk exists, local replanning adjusts the path nodes, ultimately generating an optimal path from the current position into the safe zone, covering the entire inspection range without collisions.

[0054] This application embodiment ensures the safety of physical inspection equipment during the inspection process through three-dimensional inspection safety area calculation. The knowledge-driven inspection range generation achieves accurate coverage of key parts and defect types. At the same time, the improved path planning algorithm can significantly improve path search efficiency and adaptability.

[0055] Optionally, in this embodiment, the task planning module is further configured to: set the current position of the physical inspection equipment as the starting point of path planning in the three-dimensional reality model of the substation, and map the inspection range into multiple ordered path target points to be covered in three-dimensional space; and adopt an improved BI-RRT. The path planning algorithm sequentially performs bidirectional path search on the starting point and the first path target point, and then on every two adjacent path target points. For each search, two random expansion trees are constructed, pointing from the starting point to the ending point of the path segment, and a path is explored between them through bidirectional search. After the bidirectional search between the last set of adjacent path target points is completed, the random expansion trees are connected according to the ordered path target points to generate the inspection path. During each bidirectional path search, the following two strategies are simultaneously executed to guide the growth direction of the random expansion trees: a dynamic target bias strategy, which dynamically calculates the target bias based on the number of search iterations and obstacle distribution density, and controls the probability of sampling points growing towards the path segment ending point during the path search based on the target bias; and a biased sampling strategy, which calculates the distance sampling probability and obstacle area sampling probability corresponding to each sampling area, and controls the sampling weight corresponding to each sampling area during the path search based on the distance sampling probability and the obstacle area sampling probability.

[0056] In this embodiment, the task planning module can generate inspection paths based on the following process.

[0057] First, the current coordinates of the physical inspection equipment (such as robots and drones) are used as the starting point for path planning, while the inspection area is decomposed into multiple ordered path target points. These path target points are generated through the following steps: Based on the inspection task requirements (such as the inspection of key parts of the equipment, the angle of defect observation, etc.), the inspection locations to be covered (such as the top of the transformer bushing, the side of the circuit breaker contact, etc.) are marked in the 3D reality model of the substation; then, combined with the equipment geometry and safety distance constraints (such as the robot needing to maintain a distance of more than 0.5 meters from live equipment), the marked points are spatially adjusted to ensure that they are located in reachable free space; finally, the path target points are sorted according to the inspection process logic (such as from the top to the bottom of the equipment, from the front to the back) to generate an ordered sequence of path target points.

[0058] Improved BI-RRT The algorithm achieves efficient path exploration through bidirectional search and random tree expansion. The task planning module uses the path start point as the root node and sequentially performs bidirectional path search between the start point and the first path target point, and then between every two adjacent path target points. For each path segment (e.g., from the start point to path target point A, or from path target point A to path target point B), the algorithm constructs two randomly expanded trees: one starting from the start point of the path segment (e.g., the start point or path target point A), and the other starting from the end point of the path segment (e.g., path target point A or path target point B). The two trees guide each other during the search process through a bidirectional expansion mechanism: in each iteration, the start point tree generates a random sampling point towards the end point, and the end point tree generates a random sampling point towards the start point. Then, a nearest neighbor search is used to find a connection point between the two trees. If a feasible connection path (no collisions and satisfying distance constraints) is found, the two trees are merged to form a partial path from the start point to the end point.

[0059] Once the bidirectional search between all adjacent path target points is complete, the task planning module connects the random expansion trees corresponding to each path segment according to the sequence of path target points. Finally, all path segments are pieced together into a complete inspection path that starts from the starting point and sequentially covers all path target points.

[0060] During path search, intelligent guidance using dynamic target bias and biased sampling strategies can improve path search efficiency. In each bidirectional search, the task planning module executes both strategies simultaneously: the dynamic target bias strategy dynamically adjusts the target bias based on the number of search iterations and obstacle density. For example, in the early stages of the search (few iterations) or in areas with sparse obstacles, the bias is increased (e.g., increasing the probability of a sampling point pointing towards the destination from 30% to 60%) to accelerate path convergence; in the later stages of the search or in areas with dense obstacles, the bias is decreased to enhance local exploration capabilities. The biased sampling strategy calculates the distance sampling probability and obstacle area sampling probability (the smaller the obstacle area, the higher the sampling probability) for each sampling region, comprehensively determining the sampling weight to guide the expansion of the random tree. For example, when the robot needs to navigate around a dense group of devices, the biased sampling strategy can reduce the sampling weight in densely populated areas, making the random tree more inclined to pass through gaps between devices, reducing the risk of collision.

[0061] The embodiments of this application ensure complete coverage of the inspection task through ordered path target point mapping, significantly improve path generation efficiency through an improved bidirectional search algorithm, and enhance the algorithm's adaptability in complex environments through dynamic target bias strategy and bias sampling strategy.

[0062] Optionally, in this embodiment of the application, the dynamic target bias strategy dynamically calculates the target bias amount in the following manner:

[0063] ;

[0064] in, This represents the target bias. This represents the initial target bias coefficient. Indicates the density of obstacle distribution. Indicates the current search iteration number. Indicates the maximum number of iterations. This represents the iteration limit value.

[0065] In this embodiment, This is the initial target bias coefficient, which controls the probability of selecting the target point during each sampling. This represents the density of obstacles near the target point, which can be obtained using the density function density() during calculation. This is the maximum number of iterations, which can be adjusted based on the complexity of the map environment. This is the iteration limit. When the current number of search iterations exceeds the iteration limit, sampling will only consider the obstacle distribution density. The higher the obstacle distribution density, The larger the value, the corresponding The smaller the value, the smaller the corresponding target bias, which guides the sampling points to explore the random space and get rid of the influence of the target bias.

[0066] The key to the dynamic target bias strategy is that the target bias amount adjusts the sampling probability of the target point in a timely manner according to changes in obstacle distribution density and the number of search iterations. In the early stages of path search, the target bias amount is larger, increasing the probability of sampling towards the target point, thereby guiding the search closer to the target point and accelerating the search process. As the number of search iterations increases, the target bias amount gradually decreases, so that the algorithm no longer relies excessively on target point sampling, but instead performs more random sampling, enhancing exploratory nature and avoiding premature convergence. Obstacle distribution density affects the magnitude of the target bias amount; in areas with higher obstacle distribution density, the target bias amount decreases, meaning the search will focus more on gaps in the environment rather than the target point.

[0067] Optionally, in this embodiment, the biased sampling strategy assigns sampling weights to the sampling region in the following manner: dividing the search region into multiple sampling regions; for each sampling region, calculating the Euclidean distance between the center point of the sampling region and the virtual gravity line; based on the Euclidean distance, calculating the distance sampling probability of the center point of the sampling region using a normal distribution function, wherein the virtual gravity line is a straight line connecting the start and end points of a path segment; calculating the obstacle area occupancy rate within the sampling region; inputting the obstacle area occupancy rate into a negative exponential function to calculate the obstacle area sampling probability of the sampling region; and weighting and fusing the preset base probability, the distance sampling probability, and the obstacle area sampling probability to obtain the comprehensive sampling probability of the sampling region, using the comprehensive sampling probability as the corresponding sampling weight.

[0068] In this embodiment, the division of the sampling region is a fundamental operation of the biased sampling strategy. First, the 3D space to be searched (such as a robot's mobile area or a drone flight corridor) can be divided into multiple regular or irregular sampling regions. The division method can be selected based on the complexity of the environment: in structurally simple areas (such as open corridors), a uniform grid is used to ensure that each region has a similar area; in structurally complex areas (such as densely populated equipment areas), a Voronoi diagram based on obstacle boundaries is used to ensure that the region boundaries fit the obstacle contours, reducing invalid sampling. After division, the coordinates of the geometric center point of each sampling region are recorded to provide a spatial reference for subsequent probability calculations.

[0069] The distance sampling probability is calculated using a virtual gravity line and a normal distribution function. First, a virtual gravity line connecting the start and end points of a path segment is constructed; this straight line represents the ideal direction of the path search. For the center point of each sampling region, its Euclidean distance (i.e., perpendicular distance) to the virtual gravity line is calculated; the shorter the distance, the closer the region is to the ideal path direction. Then, a normal distribution function (such as a Gaussian function) is used to map the Euclidean distance to the distance sampling probability of that center point.

[0070] The obstacle area sampling probability is calculated using a negative exponential function to quantify the obstacle's influence. First, the obstacle area occupancy rate (the ratio of the obstacle's projected area to the total area of ​​the region) is calculated within each sampling region. This value ranges from [0,1], with a larger value indicating a denser obstacle density. Then, the obstacle area occupancy rate is input into the negative exponential function, mapping it to a sampling probability. This strategy penalizes densely obstacle-prone areas, guiding the random tree to preferentially expand in free space.

[0071] Finally, the preset base probability (e.g., 0.2, to ensure that all areas have a chance to be sampled), distance sampling probability, and obstacle area sampling probability are input into the weighted fusion formula: Comprehensive probability = base probability × α + distance sampling probability × β + obstacle area sampling probability × γ, where α, β, and γ are weight coefficients, to obtain the comprehensive sampling probability. This value is used as the sampling weight of the sampling area. The higher the weight, the greater the probability of generating sampling points in the area.

[0072] In a specific embodiment, the normal distribution function can be:

[0073] ;

[0074] In the biased sampling strategy, the sampling weight of the sampling region is obtained by weighting the distance sampling probability, the obstacle area sampling probability, and the base probability. The comprehensive sampling probability expression can be:

[0075] ;

[0076] To preset the base probability, This is an adjustable weighting factor for the distance sampling probability. The distance sampling probability representing the center point of the sampling region is expressed as follows:

[0077] ;

[0078] This represents the Euclidean distance from the center point of the sampling region to the virtual gravity line. The larger the Euclidean distance, the smaller the sampling probability.

[0079] The bias in calculating the distance sampling probability is dynamically adjusted based on the number of iterations. The initial bias is relatively large, and it gradually decreases as the number of iterations increases. The specific formula is:

[0080] ;

[0081] To control the coefficient of the sampling probability difference in different regions based on the number of iterations, the larger a is, the larger σ is, and the smaller the probability difference will be. This setting is to avoid invalid sampling of areas with dense obstacles during path search, thus avoiding wasting time and resources.

[0082] The probability of sampling the obstacle area in the sampling region is expressed as follows:

[0083]

[0084] This indicates the area percentage of obstacles in each sampling region.

[0085] Furthermore, as Figure 1 The specific implementation of the system, as described in this application embodiment, provides a three-dimensional inspection method for substations, such as... Figure 2 As shown, the method includes:

[0086] Step 101: The 3D modeling module constructs a 3D real-world model of the substation with semantic information. Each entity equipment model in the 3D real-world model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation.

[0087] Step 102: The data fusion module uses a feedforward neural network to fuse the multi-source operation monitoring data of each physical device in the substation to obtain the fused feature data of each physical device.

[0088] Step 103: In response to the task processing request, the task planning module determines the target physical equipment to be inspected based on the fused feature data of each physical equipment. Based on the three-dimensional real scene model of the substation, it queries the geometric shape, attribute label, first coordinate and second coordinate of the target physical equipment. Based on the geometric shape, attribute label, first coordinate and second coordinate, it generates the inspection path and inspection range of the target physical equipment.

[0089] Step 104: The task execution module sends the inspection path and inspection range to the physical inspection equipment and receives the spatiotemporal status data returned by the physical inspection equipment in real time. In the three-dimensional real-scene model of the substation, it drives the virtual inspection equipment corresponding to the physical inspection equipment to update synchronously, so as to monitor the execution status of the inspection task in real time in the three-dimensional real-scene model of the substation. When an abnormality is detected in the inspection path, a fine-tuning instruction is generated and fed back to the physical inspection equipment.

[0090] Optionally, the "construction of a 3D real-scene model of the substation with semantic information" includes:

[0091] Point cloud data obtained through LiDAR scanning is fused with texture data obtained through oblique photogrammetry to generate a three-dimensional geometric model of the substation with a realistic appearance.

[0092] The three-dimensional geometric model of the substation is semantically segmented and reconstructed to identify the individual entity equipment models contained therein.

[0093] Based on the equipment identification, equipment model and parameter attributes in the digital engineering drawings of the substation, attribute labels are generated for each identified physical equipment model to form the three-dimensional real-scene model of the substation with semantic information.

[0094] Optionally, before step 102, the method further includes:

[0095] The data fusion module acquires training samples, which include multi-source operation monitoring data samples and real fusion feature data corresponding to each entity device sample;

[0096] An initial feedforward neural network is defined with an input layer and an output layer, wherein the input of the input layer is a multi-source operation monitoring data sample of the physical device sample, and the output of the output layer is the predicted fusion feature data of the physical device sample;

[0097] Construct the hidden layer of the initial feedforward neural network, wherein the hidden layer includes multiple neurons, and the output results of the multiple neurons constitute the activation matrix corresponding to the hidden layer;

[0098] The multi-source operation monitoring data samples corresponding to each physical device sample are input into the hidden layer through the input layer of the initial feedforward neural network. The multi-source operation monitoring data samples are fused based on the activation matrix of the hidden layer, and the predicted fusion feature data corresponding to each physical device sample are output through the output layer.

[0099] Based on the predicted fusion feature data and the actual fusion feature data corresponding to each physical device sample, the initial feedforward neural network is trained through an error processing function to obtain the final feedforward neural network.

[0100] The excitation matrix is ​​as follows:

[0101] ;

[0102] Indicates the first i The input feature vector is composed of multi-source operation monitoring data samples of each physical device sample. Indicates the first i The input feature vector of each physical device sample is in the th... j The activation function outputs the results of each neuron. m This indicates the number of entity device samples included in the training samples. n This indicates the number of neurons in the hidden layer;

[0103] The error processing function is as follows:

[0104] ;

[0105] E represents the overall loss value of the feedforward neural network. Indicates the first iReal-world fusion feature data of individual device samples Indicates the hidden layer number 1 j The output weights of each neuron This indicates that the current feedforward neural network is working on the first... i Predictive fusion feature data for individual physical device samples.

[0106] Optionally, the step of "generating the inspection path and inspection range of the target entity device based on the geometric shape, the attribute label, the first coordinate, and the second coordinate" includes:

[0107] Based on the geometry of the target physical device, calculate the three-dimensional inspection safety zone surrounding the target physical device;

[0108] Based on the attribute tags, historical inspection data and common defect types corresponding to the attribute tags are determined from a preset knowledge base. Based on the historical inspection data and the common defect types, the equipment parts to be inspected and the observation angle are determined. Based on the equipment parts to be inspected and the observation angle, the inspection range of the target physical equipment is generated.

[0109] Based on the first coordinates, the second coordinates, the three-dimensional inspection safety area, and the inspection range, an improved BI-RRT is adopted. The path planning algorithm, combining a dynamic target bias strategy and a bias sampling strategy, generates an inspection path from the current position of the physical inspection equipment to the three-dimensional inspection safety area, covering the inspection range.

[0110] Optionally, the phrase "adopting an improved BI-RRT" The path planning algorithm, combining a dynamic target bias strategy and a bias sampling strategy, generates an inspection path from the current position of the physical inspection equipment to the three-dimensional inspection safety area, covering the inspection range, including:

[0111] In the three-dimensional real-scene model of the substation, the current position of the physical inspection equipment is set as the starting point of the path planning, and the inspection range is mapped in three-dimensional space as multiple ordered path target points to be covered;

[0112] Adopting an improved BI-RRT The path planning algorithm sequentially performs bidirectional path search on the starting point and the first path target point, as well as on every two adjacent path target points thereafter. For each search, two randomly expanded trees are constructed, pointing from the starting point of the path segment to the ending point of the path segment, and the path is explored between the two trees through bidirectional search.

[0113] After the bidirectional search between the last group of adjacent path target points is completed, the random expansion trees are connected in the order of the ordered path target points to generate the inspection path.

[0114] During each bidirectional path search, the following two strategies are executed simultaneously to guide the growth direction of the randomly expanding tree:

[0115] Based on the dynamic target bias strategy, the target bias is dynamically calculated according to the number of search iterations and the obstacle distribution density, and the probability of the sampling point growing toward the end of the path segment is controlled based on the target bias during the path search process.

[0116] Based on the biased sampling strategy, the distance sampling probability and obstacle area sampling probability corresponding to each sampling region are calculated respectively. Based on the distance sampling probability and the obstacle area sampling probability, the sampling weight corresponding to each sampling region is controlled during the path search process.

[0117] Optionally, the dynamic target bias strategy dynamically calculates the target bias in the following manner:

[0118] ;

[0119] in, This represents the target bias. This represents the initial target bias coefficient. Indicates the density of obstacle distribution. Indicates the current search iteration number. Indicates the maximum number of iterations. This represents the iteration limit value.

[0120] Optionally, the biased sampling strategy assigns sampling weights to the sampling region in the following manner:

[0121] The area to be searched is divided into multiple sampling areas;

[0122] For each sampling region, the Euclidean distance between the center point of the sampling region and the virtual gravity line is calculated. Based on the Euclidean distance, the distance sampling probability of the center point of the sampling region is calculated using a normal distribution function. The virtual gravity line is a straight line connecting the start and end points of a path segment.

[0123] Calculate the obstacle area occupancy rate within the sampling area, input the obstacle area occupancy rate into a negative exponential function, and calculate the obstacle area sampling probability of the sampling area.

[0124] The sampling probability of the sampling area is obtained by weighted fusion of the preset base probability, the distance sampling probability, and the obstacle area sampling probability, and the comprehensive sampling probability is used as the corresponding sampling weight.

[0125] It should be noted that other corresponding descriptions of the functional units involved in the substation three-dimensional inspection method provided in this application embodiment can be found in the following references. Figure 1 The corresponding descriptions in the system will not be repeated here.

[0126] This application also provides a computer device, which may specifically be a personal computer, a server, a network device, etc. Figure 3 As shown, the computer device includes a bus, a processor, memory, and a communication interface, and may also include an input / output interface and a display device. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores location information. The network interface allows communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the various method embodiments.

[0127] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0128] In one embodiment, a computer-readable storage medium is provided, which may be non-volatile or volatile, having stored thereon a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0129] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0130] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0131] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0132] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0133] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A three-dimensional inspection system for substations, characterized in that, include: The 3D modeling module is used to construct a 3D real-scene model of a substation with semantic information. Each entity equipment model in the 3D real-scene model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation. The data fusion module is used to fuse the multi-source operation monitoring data of each physical device in the substation using a feedforward neural network to obtain the fused feature data of each physical device. The task planning module is used to respond to task processing requests, determine the target physical equipment to be inspected based on the fused feature data of each physical equipment, query the geometric shape, attribute labels, first coordinates and second coordinates of the target physical equipment based on the three-dimensional real scene model of the substation, and generate the inspection path and inspection range of the target physical equipment according to the geometric shape, the attribute labels, the first coordinates and the second coordinates. The task execution module is used to send the inspection path and the inspection range to the physical inspection equipment, and receive the spatiotemporal status data returned by the physical inspection equipment in real time. It drives the virtual inspection equipment corresponding to the physical inspection equipment to be updated synchronously in the three-dimensional real scene model of the substation, so as to monitor the execution status of the inspection task in real time in the three-dimensional real scene model of the substation, and generate fine-tuning instructions when an abnormality of the inspection path is detected, and feed the fine-tuning instructions back to the physical inspection equipment. The 3D modeling module is used for: Point cloud data obtained through LiDAR scanning is fused with texture data obtained through oblique photogrammetry to generate a three-dimensional geometric model of the substation with a realistic appearance. The three-dimensional geometric model of the substation is semantically segmented and reconstructed to identify the individual entity equipment models contained therein. Based on the equipment identification, equipment model and parameter attributes in the digital engineering drawings of the substation, attribute labels are generated for each identified entity equipment model to form the three-dimensional real scene model of the substation with semantic information. The task planning module is used for: Based on the geometry of the target physical device, calculate the three-dimensional inspection safety zone surrounding the target physical device; Based on the attribute tags, historical inspection data and common defect types corresponding to the attribute tags are determined from a preset knowledge base. Based on the historical inspection data and the common defect types, the equipment parts to be inspected and the observation angle are determined. Based on the equipment parts to be inspected and the observation angle, the inspection range of the target physical equipment is generated. Based on the first coordinates, the second coordinates, the three-dimensional inspection safety area, and the inspection range, an improved BI-RRT is adopted. The path planning algorithm, combining a dynamic target bias strategy and a bias sampling strategy, generates an inspection path from the current position of the physical inspection equipment to the three-dimensional inspection safety area, covering the inspection range.

2. The system according to claim 1, characterized in that, The data fusion module is used for: Obtain training samples, wherein the training samples include multi-source operation monitoring data samples and real fused feature data corresponding to each entity device sample; An initial feedforward neural network is defined with an input layer and an output layer, wherein the input of the input layer is a multi-source operation monitoring data sample of the physical device sample, and the output of the output layer is the predicted fusion feature data of the physical device sample; Construct the hidden layer of the initial feedforward neural network, wherein the hidden layer includes multiple neurons, and the output results of the multiple neurons constitute the activation matrix corresponding to the hidden layer; The multi-source operation monitoring data samples corresponding to each physical device sample are input into the hidden layer through the input layer of the initial feedforward neural network. The multi-source operation monitoring data samples are fused based on the activation matrix of the hidden layer, and the predicted fusion feature data corresponding to each physical device sample are output through the output layer. Based on the predicted fusion feature data and the actual fusion feature data corresponding to each physical device sample, the initial feedforward neural network is trained through an error processing function to obtain the final feedforward neural network. The excitation matrix is ​​as follows: ； Indicates the first i The input feature vector is composed of multi-source operation monitoring data samples of each physical device sample. Indicates the first i The input feature vector of each physical device sample is in the th... j The activation function outputs the results of each neuron. m This indicates the number of entity device samples included in the training samples. n This indicates the number of neurons in the hidden layer; The error processing function is as follows: ； E represents the overall loss value of the feedforward neural network. Indicates the first i Real-world fusion feature data of individual device samples Indicates the hidden layer number 1 j The output weights of each neuron This indicates that the current feedforward neural network is working on the first... i Predictive fusion feature data for individual physical device samples.

3. The system according to claim 1, characterized in that, The task planning module is also used for: In the three-dimensional real-scene model of the substation, the current position of the physical inspection equipment is set as the starting point of the path planning, and the inspection range is mapped in three-dimensional space as multiple ordered path target points to be covered; Adopting an improved BI-RRT The path planning algorithm sequentially performs bidirectional path search on the starting point and the first path target point, as well as on every two adjacent path target points thereafter. For each search, two randomly expanded trees are constructed, pointing from the starting point of the path segment to the ending point of the path segment, and the path is explored between the two trees through bidirectional search. After the bidirectional search between the last group of adjacent path target points is completed, the random expansion trees are connected in the order of the ordered path target points to generate the inspection path. During each bidirectional path search, the following two strategies are executed simultaneously to guide the growth direction of the randomly expanding tree: Based on the dynamic target bias strategy, the target bias is dynamically calculated according to the number of search iterations and the obstacle distribution density, and the probability of the sampling point growing toward the end of the path segment is controlled based on the target bias during the path search process. Based on the biased sampling strategy, the distance sampling probability and obstacle area sampling probability corresponding to each sampling region are calculated respectively. Based on the distance sampling probability and the obstacle area sampling probability, the sampling weight corresponding to each sampling region is controlled during the path search process.

4. The system according to claim 3, characterized in that, The dynamic target bias strategy dynamically calculates the target bias in the following way: ； in, This represents the target bias. This represents the initial target bias coefficient. Indicates the density of obstacle distribution. Indicates the current search iteration number. Indicates the maximum number of iterations. This represents the iteration limit value.

5. The system according to claim 3, characterized in that, The biased sampling strategy assigns sampling weights to the sampling region in the following manner: The area to be searched is divided into multiple sampling areas; For each sampling region, the Euclidean distance between the center point of the sampling region and the virtual gravity line is calculated. Based on the Euclidean distance, the distance sampling probability of the center point of the sampling region is calculated using a normal distribution function. The virtual gravity line is a straight line connecting the start and end points of a path segment. Calculate the obstacle area occupancy rate within the sampling area, input the obstacle area occupancy rate into a negative exponential function, and calculate the obstacle area sampling probability of the sampling area. The sampling probability of the sampling area is obtained by weighted fusion of the preset base probability, the distance sampling probability, and the obstacle area sampling probability, and the comprehensive sampling probability is used as the corresponding sampling weight.

6. A three-dimensional inspection method for substations, characterized in that, include: The 3D modeling module constructs a 3D real-scene model of the substation with semantic information. Each entity equipment model in the 3D real-scene model of the substation has a corresponding geometric shape and attribute label, as well as a first coordinate in the world coordinate system and a second coordinate in the local coordinate system within the substation. The data fusion module uses a feedforward neural network to fuse the multi-source operation monitoring data of each physical device in the substation to obtain the fused feature data of each physical device. The task planning module responds to the task processing request, determines the target physical equipment to be inspected based on the fused feature data of each physical equipment, queries the geometric shape, attribute label, first coordinate and second coordinate of the target physical equipment based on the three-dimensional real scene model of the substation, and generates the inspection path and inspection range of the target physical equipment according to the geometric shape, attribute label, first coordinate and second coordinate. The task execution module sends the inspection path and inspection range to the physical inspection equipment and receives the spatiotemporal status data returned by the physical inspection equipment in real time. It drives the virtual inspection equipment corresponding to the physical inspection equipment to be updated synchronously in the three-dimensional real scene model of the substation, so as to monitor the execution status of the inspection task in real time in the three-dimensional real scene model of the substation. When an abnormality is detected in the inspection path, a fine-tuning instruction is generated and fed back to the physical inspection equipment. The construction of a 3D reality model of a substation with semantic information includes: Point cloud data obtained through LiDAR scanning is fused with texture data obtained through oblique photogrammetry to generate a three-dimensional geometric model of the substation with a realistic appearance. The three-dimensional geometric model of the substation is semantically segmented and reconstructed to identify the individual entity equipment models contained therein. Based on the equipment identification, equipment model and parameter attributes in the digital engineering drawings of the substation, attribute labels are generated for each identified entity equipment model to form the three-dimensional real scene model of the substation with semantic information. The step of generating the inspection path and inspection range of the target entity device based on the geometric shape, the attribute label, the first coordinate, and the second coordinate includes: Based on the geometry of the target physical device, calculate the three-dimensional inspection safety zone surrounding the target physical device; Based on the attribute tags, historical inspection data and common defect types corresponding to the attribute tags are determined from a preset knowledge base. Based on the historical inspection data and the common defect types, the equipment parts to be inspected and the observation angle are determined. Based on the equipment parts to be inspected and the observation angle, the inspection range of the target physical equipment is generated. Based on the first coordinates, the second coordinates, the three-dimensional inspection safety area, and the inspection range, an improved BI-RRT is adopted. The path planning algorithm, combining a dynamic target bias strategy and a bias sampling strategy, generates an inspection path from the current position of the physical inspection equipment to the three-dimensional inspection safety area, covering the inspection range.

7. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 6.

8. A computer device, comprising a storage medium, a processor, and a computer program stored on the storage medium and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of claim 6.

Images