Intelligent decision-making system for power supply operation of generator car

Abstract

The embodiment of the invention discloses an intelligent decision-making system for power supply operation of a generator car, and relates to the technical field of electric power operation safety, and the system comprises an edge node and a cloud management platform. Wherein the edge node is used for collecting a multi-view image sequence of a power generation vehicle power supply operation site; the cloud management platform comprises an on-site identification sub-platform which is used for constructing three-dimensional point cloud data by using the multi-view image sequence and carrying out semantic segmentation on the three-dimensional point cloud data so as to identify on-site parameters related to power supply operation of the power generation vehicle; and the operation decision sub-platform is used for generating a power generation vehicle power supply operation analysis decision based on a pre-constructed multi-dimensional operation condition specification index system and the field parameters. According to the embodiment of the invention, on-site parameters are accurately obtained through an intelligent means, a unified evaluation standard is established, the risk detection and early warning capability is improved, the power supply operation risk caused by human factors is effectively avoided, and the safety, reliability and high efficiency of power supply operation are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of power operation safety, and in particular to an intelligent decision-making system for power supply operation of a power generation vehicle. Background Art

[0002] As the power system continues to evolve, generator trucks are playing an increasingly critical role in key scenarios such as temporary power supply, emergency repairs, and distributed energy access. Power supply operations require the coordination of numerous tools and equipment, as well as complex operations such as cable connection, load matching, and grid-connection and off-grid switching. This imposes stringent safety, reliability, and efficiency requirements. Any deviations in parameter configuration, incorrect operation sequences, or inadequate safety precautions can lead to equipment damage, power outages, and even serious casualties.

[0003] However, the current power generation process for generator trucks still largely relies on manual field surveys and empirical judgment, which presents significant drawbacks. Manual surveys are limited by factors such as topography and visual bias, making it difficult to accurately grasp on-site power load data and spatial layout. Experience-based operational plans lack standardized, quantitative assessment mechanisms, leading to significant differences in risk identification and control capabilities among different operational teams. Manual verification of complex operational steps is not only inefficient and labor-intensive, but also prone to omissions and misjudgments due to fatigue and negligence, making it impossible to promptly detect and warn of potential risks. Summary of the Invention

[0004] In view of this, an embodiment of the present invention provides an intelligent decision-making system for power supply operations of power generation vehicles, so as to effectively avoid the risks of power supply operations caused by human factors and ensure the safety, reliability and efficiency of power supply operations.

[0005] An embodiment of the present invention provides an intelligent decision-making system for power supply operations of a power generation vehicle, including an edge node and a cloud management platform; wherein:

[0006] The edge node is used to collect multi-view image sequences of the power supply operation site of the power generation vehicle;

[0007] The cloud management platform includes:

[0008] The on-site recognition sub-platform is used to: construct three-dimensional point cloud data using the multi-view image sequence; perform semantic segmentation on the three-dimensional point cloud data to identify on-site parameters related to the power supply operation of the power generation vehicle;

[0009] The operation decision sub-platform is used to generate power supply operation analysis decisions of the power generation vehicle based on the pre-built multi-dimensional operation condition specification indicator system and the above-mentioned on-site parameters.

[0010] Furthermore, the field recognition sub-platform adopts a heterogeneous architecture consisting of a central processing unit (CPU), a graphics processing unit (GPU), and a field programmable gate array (FPGA); wherein:

[0011] The CPU schedules the GPU to execute a point cloud construction task: constructing three-dimensional point cloud data using the multi-view image sequence;

[0012] After completing the point cloud construction task, the GPU notifies the FPGA to perform a semantic segmentation task: performing semantic segmentation on the three-dimensional point cloud data to identify on-site parameters related to the power supply operation of the power generation vehicle;

[0013] After the FPGA completes the semantic segmentation task, it notifies the CPU that the task is completed.

[0014] Furthermore, the GPU executes a point cloud construction task, including: the GPU reads the multi-view image sequence from a shared memory, constructs three-dimensional point cloud data using the multi-view image sequence, and writes the three-dimensional point cloud data into the shared memory after adding a self-description data header;

[0015] The FPGA performs semantic segmentation tasks, including:

[0016] The hardware parses the self-describing data header and dynamically configures semantic segmentation pipeline parameters based on the information carried therein; the semantic segmentation pipeline is started based on the configuration result: the three-dimensional point cloud data is semantically segmented to identify on-site parameters related to the power supply operation of the power generation vehicle; and the identification results are stored in the shared memory.

[0017] Furthermore, the FPGA hardware parses the self-describing data header and dynamically configures semantic segmentation pipeline parameters based on the information carried therein, including:

[0018] The hardware parses the self-describing data header to extract the point cloud data format and point cloud scale attribute information therein;

[0019] Acquire the FPGA load and power supply operation type, and select a semantic segmentation model to be used by the semantic segmentation pipeline from a plurality of semantic segmentation models pre-built by the CPU according to the acquisition result;

[0020] Configure the parameters of the selected semantic segmentation model based on the extracted point cloud data format and point cloud scale attribute information.

[0021] Furthermore, the FPGA selects a semantic segmentation model to be used by the semantic segmentation pipeline from a plurality of semantic segmentation models pre-built by the CPU according to the acquisition result, including:

[0022] Determining a load state of the FPGA;

[0023] If the FPGA load does not reach the set upper limit, selecting a refined semantic segmentation model that matches the power supply operation type from the multiple semantic segmentation models pre-built by the CPU;

[0024] If the FPGA load reaches a set upper limit, a general lightweight semantic segmentation model is selected from multiple semantic segmentation models pre-built by the CPU;

[0025] The selected semantic segmentation model is loaded as the semantic segmentation model to be used in the semantic segmentation pipeline.

[0026] Furthermore, the edge node includes a drone equipped with a camera;

[0027] The edge node is used to collect a multi-view image sequence of the power supply operation site of the power generation vehicle, including:

[0028] Shoot the power supply operation site of the generator truck from multiple angles;

[0029] When photographing the tower structure at the power supply operation site of the power generation vehicle, the viewing angle range of adjacent photographed images is controlled to be within the range of 10° to 20°.

[0030] Furthermore, the edge node is specifically used to: shoot the power supply operation site of the power generation vehicle at multiple time periods and multiple angles;

[0031] The cloud management platform also includes an image fusion sub-platform for performing spatiotemporal registration and feature fusion on multi-view image sequences taken at multiple time periods to generate a unified multi-view image sequence.

[0032] Furthermore, the image fusion sub-platform is also used to:

[0033] An image preprocessing operation including denoising and contrast enhancement is performed on the unified multi-view image sequence.

[0034] Furthermore, the job decision sub-platform includes:

[0035] The data access layer is used to access the following data: the multi-dimensional operating condition specification indicator system and on-site parameters related to the power supply operation of the generator truck, and convert the accessed data into a standardized format;

[0036] The decision-making core layer is used to combine the on-site parameters related to the power supply operation of the generator truck, analyze each indicator in the multi-dimensional operating condition specification indicator system one by one, evaluate whether the power supply operation site of the generator truck meets the operation requirements, generate an operation plan, and output the evaluation results and generation results;

[0037] The dynamic optimization layer is used to collect decision-making effect feedback data and optimize the decision-making core layer according to the collected results.

[0038] Furthermore, the site identification sub-platform in the cloud management platform is further used to: construct a three-dimensional real scene model of the power supply operation site of the power generation vehicle based on the three-dimensional point cloud data;

[0039] The dynamic optimization layer is specifically used to: display the three-dimensional real scene model and the output results of the decision core layer; receive decision effect feedback data input by the user for the displayed content, and optimize the decision core layer according to the collected results.

[0040] The technical solution provided by the embodiments of the present invention effectively avoids the parameter collection distortion common in traditional manual surveys by automatically acquiring three-dimensional point cloud data and analyzing key site parameters from it. Furthermore, it integrates a pre-set multi-dimensional operational condition standard indicator system for intelligent analysis and decision-making, transforming the previous decision-making model that relied on empirical judgment and establishing a standardized, quantitative evaluation mechanism. This solution not only enhances the scientific nature and consistency of operational decisions, but also significantly improves verification efficiency and enhances the dynamic identification and early warning capabilities of operational risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 A schematic diagram of the architecture of an intelligent decision-making system for power supply operations of a power generation vehicle provided by an embodiment of the present invention;

[0043] Figure 2 A schematic diagram of a heterogeneous architecture of a field identification sub-platform in a cloud management platform provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of a process for identifying power facility parameters by a field identification sub-platform in a cloud management platform provided by an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of the layered architecture of a job decision sub-platform in a cloud management platform provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0047] It should be understood that the embodiments described are only a portion of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0048] The embodiment of the present invention provides an intelligent decision-making system for power supply operation of power generation vehicles, which can be applied to complex power operation scenarios such as temporary power supply, emergency repair, and distributed energy access. Figure 1 The system includes an edge node 100 and a cloud management platform 101.

[0049] The edge node 100 is used to collect multi-view image sequences of the power supply operation site of the power generation vehicle;

[0050] Cloud Management Platform 101, including:

[0051] The on-site identification sub-platform 1011 is used to: construct 3D point cloud data using the multi-view image sequences collected by the edge node 100; perform semantic segmentation on the 3D point cloud data to identify on-site parameters related to the power supply operation of the power generation vehicle;

[0052] The operation decision sub-platform 1012 is used to generate power generation vehicle power supply operation analysis decisions based on the pre-built multi-dimensional operation condition specification indicator system and the field parameters.

[0053] The following is a detailed introduction to the various components in the system and their functional implementation.

[0054] 1. Edge Node 100

[0055] The acquisition of multi-perspective image sequences is a basic link in constructing the three-dimensional space of the power supply operation site of the power generation vehicle. Specifically, the edge node 100 may include a drone equipped with a camera, and the power supply operation site of the power generation vehicle is photographed from multiple perspectives by the drone. The shooting range needs to fully cover the power facilities (including the distribution of poles, wires, insulators, etc.), topography (geographical environment characteristics and surface structure) and channel environment (distribution of surrounding obstacles and direction of adjacent lines, etc.) at the power supply operation site of the power generation vehicle to ensure the accuracy and completeness of the three-dimensional reconstructed scene data. Of course, the edge node 100 may also include multi-angle sensing equipment deployed at the operation site, such as wide-angle cameras, laser scanners, etc., through which multi-perspective image sequences of the power supply operation site of the power generation vehicle are collected in real time. This embodiment does not make specific limitations on this. Exemplarily, the acquisition of multi-perspective image sequences is performed in accordance with the following requirements:

[0056] ①Shooting angle control

[0057] Adopting multi-angle shooting modes such as orthographic, oblique, and surround to enhance the detail expression and reconstruction accuracy of the 3D spatial model;

[0058] In particular, when photographing the tower structure at the power generation site, the viewing angle of adjacent images should be controlled within the range of 10° to 20°.

[0059] ②Shooting environment requirements

[0060] Shooting should be avoided under complex lighting or weather conditions such as strong light, shadows, rain and fog to ensure the clarity and color reproduction of the captured images; when complex lighting conditions cannot be avoided, a multi-time batch shooting strategy can be adopted, and the images collected at different times can be integrated through image fusion algorithms to improve image quality and reconstruction effects.

[0061] In the above requirements, the viewing angle range of adjacent images of the tower structure is controlled within the range of 10° to 20°. The reasons are:

[0062] An angle of 10° to 20° allows adjacent images to maintain a large overlapping area, ensuring that a sufficient number of points with the same name are captured during feature extraction, thus avoiding feature point matching failures caused by a large viewing angle span.

[0063] This angle range meets the baseline-distance ratio requirement of triangulation (baseline length / shooting distance ≈ tan(10°-20°)), and can control the reconstruction error of 3D points to the centimeter level (when the angle is >20°, the depth calculation error increases exponentially with the angle).

[0064] Compared with smaller angles (such as <10°), the 20° upper limit can reduce the number of shots and avoid redundant data; compared with larger angles (such as >20°), the 10° lower limit can reduce the computational complexity of the hole completion algorithm.

[0065] Furthermore, in actual power generation truck operations, the operation process is dynamic, and the real-time operating status of the equipment and surrounding environmental conditions are not static. For example, as the operation progresses, the load on the truck may be constantly adjusted, the connection status of the cables may also change, and environmental factors such as on-site lighting and weather may also affect the operation and safety of the equipment. Images captured at a single time interval only reflect static information at a moment in time and fail to capture the dynamic details of the operation process. This can lead to missing key parameters or misjudgment, which in turn affects the accuracy and safety of subsequent power supply operation decisions. To this end, the edge node 100 can be specifically used to capture the power generation truck power supply operation site from multiple perspectives at multiple time intervals. Accordingly, the cloud management platform 101 also includes an image fusion sub-platform 1010, which is responsible for performing spatiotemporal registration and feature fusion on the multi-perspective image sequences captured at multiple time intervals to generate a unified multi-perspective image sequence. This improves the integrity and temporal continuity of environmental perception at the power generation truck power supply operation site, overcomes the information gaps and limitations of single-perspective or single-period image acquisition, and provides richer and more complete data support for subsequent precise analysis and decision-making.

[0066] Among them, the image fusion sub-platform 1010 is used to perform spatiotemporal registration and feature fusion on multi-view image sequences taken at multiple time periods to generate a unified multi-view image sequence, which can be achieved through the following operations 1-3.

[0067] 1. Data Collection and Preprocessing

[0068] Data collection: Acquire multi-period, multi-view image sequences from the edge node 100. These images can come from different devices and different time points.

[0069] Preprocessing: Preprocess the original image sequence, including denoising, color correction, image cropping and scale adjustment, to ensure the consistency and accuracy of subsequent processing.

[0070] 2. Spatiotemporal Registration

[0071] Time synchronization: Add an accurate timestamp to each frame of image and arrange the images in sequence according to the timestamp to ensure timing consistency.

[0072] Spatial Registration:

[0073] Feature point extraction: Use algorithms such as SIFT (Scale-Invariant Feature Transform), SURF (Speeded-Up Robust Features), or ORB (Oriented FAST and Rotated BRIEF) to extract key feature points from images at each viewpoint.

[0074] Feature matching: Find the correspondence between images from different perspectives by matching feature descriptors.

[0075] Geometric transformation: Use the RANSAC (Random Sample Consensus) algorithm to estimate the geometric transformation matrix (such as affine transformation and homography transformation) and align all perspective images to the same coordinate system.

[0076] Global optimization: Apply global optimization methods (such as BundleAdjustment) to further refine the camera pose and 3D point positions to improve registration accuracy.

[0077] 3. Feature Fusion

[0078] Feature extraction: Extract multi-level features from the registered image, including low-level features (such as color, texture) and high-level features (such as depth information).

[0079] Feature fusion strategy:

[0080] Weighted average method: Perform weighted averaging on image features of the same area at different time periods but with the same viewing angle to reduce the impact of noise.

[0081] Maximum response method: select the maximum response value of each pixel position as the final feature value to highlight significant features.

[0082] Deep learning fusion: Use convolutional neural networks (CNN) or Transformer architectures to automatically learn how to best fuse feature information from images of the same viewpoint at multiple time intervals.

[0083] The image of each perspective will go through the above feature extraction and fusion steps to obtain a unified multi-perspective image sequence after fusion.

[0084] More preferably, the image fusion sub-platform 1010 is further configured to perform image preprocessing operations on the unified multi-view image sequence, including denoising and contrast enhancement operations. This helps further improve image clarity and feature visibility, thereby enhancing the accuracy and stability of subsequent 3D reconstruction and semantic segmentation.

[0085] 2. Cloud Management Platform 101

[0086] 1. On-site identification sub-platform 1011

[0087] After the edge node 100 has collected the multi-view image sequence of the power generation vehicle power supply operation site, it sends it to the cloud management platform 101. The field recognition sub-platform 1011 on the cloud management platform 101 can use these image sequences to construct the three-dimensional point cloud data of the power generation vehicle power supply operation site. Subsequently, the field recognition sub-platform 1011 automatically identifies and classifies different objects (such as poles, wires, insulators, ground, obstacles, etc.) in the three-dimensional point cloud data based on semantic segmentation technology. On this basis, it extracts the field parameters related to the power generation vehicle power supply operation, including the structural parameters of the power facilities (such as pole height, cross arm spacing, insulator size) and the environmental spatial distribution parameters (such as terrain slope, channel width and obstacle distribution, etc.), providing data support and decision-making basis for the power generation vehicle power supply operation.

[0088] For details, see Figure 2 The field recognition sub-platform 1011 utilizes a heterogeneous architecture consisting of a central processing unit (CPU), a graphics processing unit (GPU), and a field-programmable gate array (FPGA). The CPU and GPU can be connected via the PCIe (Peripheral Component Interconnect Express) bus, the CPU and FPGA can be connected via the PCIe bus or a high-speed serial interface (such as Avalon / AXI), and the GPU and FPGA can be connected via the PCIe bus or a dedicated high-speed interconnect. All three can access the same shared memory, and the collaborative working process is as follows.

[0089] (1) The CPU schedules the GPU to perform the point cloud construction task: constructing three-dimensional point cloud data using multi-view image sequences.

[0090] For example, multiple GPUs and FPGAs are deployed on the scene recognition sub-platform 1011. Based on the operating status of these GPUs and FPGAs, the CPU can select an idle GPU and FPGA, control these idle GPUs and FPGAs to prepare for collaborative operation, and store the multi-view image sequences captured by the edge node 100 in shared memory. The CPU then sends a point cloud construction task execution notification to the selected GPU.

[0091] After receiving the execution notification, the GPU reads the multi-view image sequence from shared memory, constructs 3D point cloud data using the multi-view image sequence, and writes the 3D point cloud data to the shared memory after adding a self-describing data header for subsequent semantic segmentation by the FPGA. The self-describing data header can include information about the 3D point cloud data format and point cloud scale attributes (such as the number of points, point density, and point cloud data capacity), and optionally includes a checksum for integrity verification of the 3D point cloud data.

[0092] Those skilled in the art will appreciate that the GPU can employ any existing 3D point cloud construction technology to generate high-precision 3D point cloud data based on a multi-view image sequence. As an example, the construction process includes the following operations ① and ②.

[0093] ①Process the multi-view image sequences and construct sparse point cloud data.

[0094] First, image preprocessing is performed: denoising and contrast enhancement are performed on the acquired multi-view image sequences to eliminate uneven illumination and sensor noise interference, thereby improving image quality.

[0095] Next, feature extraction and matching are performed: feature descriptors such as SIFT, SURF, or ORB are used to detect feature points (such as corner points and edge features) in the preprocessed image, and matching relationships between feature points across the image views are established.

[0096] Then, the camera parameters are estimated: using the feature point matching results, the intrinsic parameters (such as focal length, principal point coordinates) and extrinsic parameters (such as position and attitude matrix) of the camera that captured the image are estimated through optimization algorithms such as bundle adjustment.

[0097] Finally, sparse point cloud data is generated through triangulation: based on the matched feature points and estimated camera parameters, the position of the point in three-dimensional space is calculated to form sparse point cloud data.

[0098] ② Based on the set depth map fusion algorithm, dense point cloud data is generated based on the sparse point cloud data.

[0099] Specifically, the implementation scheme of the above operation ② is as follows:

[0100] Depth map initialization: Initialize the depth map of each viewpoint based on the camera intrinsic parameters and sparse point cloud data;

[0101] Patch initialization: select a local area (Patch) in the depth map for initialization;

[0102] Depth map propagation optimization: gradually improve the details of the depth map through parallax propagation and optimization algorithms;

[0103] Dense point cloud data generation: Based on the depth map and camera parameters corresponding to each viewpoint, dense 3D point cloud data is generated. The dense 3D point cloud data consists of multiple points, each of which includes the 3D coordinates of the point and its visual attribute characteristics (such as RGB values).

[0104] (2) After the GPU completes the point cloud construction task, it notifies the FPGA to perform the semantic segmentation task: semantic segmentation of the three-dimensional point cloud data to identify the on-site parameters related to the power supply operation of the power generation vehicle.

[0105] In this embodiment of the present invention, after completing the point cloud construction task, the GPU directly sends a completion notification to the FPGA (or, if multiple are present, a CPU selects one), triggering the FPGA to execute the semantic segmentation task. Instead of sending the notification to the CPU, which then instructs the FPGA to perform semantic segmentation, the GPU directly sends the notification to the FPGA. This has the advantages of reducing the intermediate steps in task scheduling, lowering system latency, and improving the real-time and efficiency of task execution. It also reduces the scheduling burden on the CPU and enables efficient collaboration and pipelined processing across multiple hardware platforms.

[0106] For example, after completing the point cloud construction task, the GPU queries the address mapping table of the selected FPGA preconfigured by the CPU to obtain the FPGA's MSI (Message Signaled Interrupt) address. An MSI-X interrupt is sent to the FPGA corresponding to the MSI address via the PCIe root complex. The interrupt message carries a pointer value used to point to the three-dimensional point cloud data storage address. The FPGA detects the MSI-X interrupt and reads the three-dimensional point cloud data packet from shared memory based on the pointer value. Specifically, when the FPGA fails to perform a CRC (cyclic redundancy check) on the three-dimensional point cloud data packet, it sends a negative acknowledgement (NACK) signal to the GPU via the reverse interrupt channel and updates the NACK signal counter. After receiving the NACK signal, the GPU retransmits the pointer value used to point to the three-dimensional point cloud data storage address. When the error counter value exceeds a preset threshold, the FPGA triggers a hardware interrupt signal and reports the communication anomaly to the CPU, which then takes over the data communication interaction process between the GPU and FPGA to achieve seamless switching between fault response and task scheduling.

[0107] The FPGA performs semantic segmentation tasks, including: hardware parsing of the self-describing data header of the three-dimensional point cloud data stored in the shared memory, and dynamically configuring the semantic segmentation pipeline parameters based on the information carried therein; starting the semantic segmentation pipeline based on the configuration results: semantically segmenting the three-dimensional point cloud data stored in the shared memory; identifying on-site parameters related to the power supply operation of the power generation vehicle based on the semantic segmentation results; and storing the identification results in the shared memory. Exemplarily, a circular queue is stored in the shared memory for storing the three-dimensional point cloud data and its self-describing data header. The position of the self-describing data header can be directly attached to the beginning of the three-dimensional point cloud data. After the GPU generates the three-dimensional point cloud data, it inserts the self-describing data header into the data body header to form a complete data packet, writes the complete data packet into the circular queue, updates the production pointer, and notifies the FPGA that the point cloud construction task is completed.

[0108] The FPGA dynamically configures the semantic segmentation pipeline parameters based on the parsed self-describing data header information, including but not limited to the following Table 1:

[0109] Table 1

[0110]

[0111]

[0112] In an embodiment of the present invention, in a semantic segmentation pipeline, semantic segmentation can be performed on three-dimensional point cloud data based on a semantic segmentation model. The model can be a deep learning model, such as a convolutional neural network architecture model. When the scale of the three-dimensional point cloud (such as the data capacity) is large, it means that the data contains richer details and complex spatial information. In this case, the convolution layer of the semantic segmentation model uses a larger convolution kernel (such as 5x5) to capture the global structural features of the three-dimensional point cloud within a larger receptive field, which helps to extract macroscopic spatial layout information. When the scale of the three-dimensional point cloud (such as the data capacity) is small, the convolution layer of the semantic segmentation model uses a smaller convolution kernel (such as 3x3) to effectively extract local detail features while reducing the amount of calculation, avoiding information redundancy and waste of computing resources caused by excessively large convolution kernels. Therefore, a threshold can be set according to the capacity of the three-dimensional point cloud data. When the three-dimensional data capacity exceeds the threshold, a larger convolution kernel is selected; otherwise, a smaller convolution kernel is used. The fully connected layer is responsible for integrating the features extracted by the previous convolution layer and mapping them to the final output. The large scale (e.g., data capacity) of 3D point clouds generally contains higher-dimensional features, requiring more fully connected layer channels to process and express these features to avoid information loss. For example, based on experience or extensive research and experimentation, a linear or nonlinear mapping model can be established between the scale of a 3D point cloud and the number of fully connected layer channels. Based on the specific scale of the 3D point cloud, the appropriate number of channels can be calculated. This allows the model to fully utilize the data features while maintaining reasonable computational efficiency and preventing overfitting or underfitting.

[0113] In short, all parameters related to the attributes of 3D point cloud data in the semantic segmentation pipeline can be dynamically configured according to the 3D point cloud scale attribute information contained in the self-describing data header, thereby achieving a precise match between the pipeline parameters and the input data characteristics, ensuring the efficiency and accuracy of the semantic segmentation task.

[0114] As a preferred embodiment, FPGA hardware parses the self-describing data header and dynamically configures the semantic segmentation pipeline parameters based on the information carried therein, including:

[0115] The hardware parses the self-describing data header and extracts the point cloud data format and point cloud scale attribute information;

[0116] Obtaining the FPGA load and power supply operation type, and selecting a semantic segmentation model to be used in the semantic segmentation pipeline from multiple semantic segmentation models pre-built by the CPU based on the obtained results;

[0117] Configure the parameters of the selected semantic segmentation model based on the extracted point cloud data format and point cloud scale attribute information.

[0118] Among them, the parameters for configuring the selected semantic segmentation model include but are not limited to model structure parameters, which are specifically related to the selected semantic segmentation model type. For example, when the selected semantic segmentation model type is a CNN (Convolutional Neural Network) architecture model, the configured parameters may include the convolution kernel size and network depth. The point cloud scale attribute information can be the total size of the three-dimensional point cloud data, or it can be information such as the number of points and point density that characterizes the point cloud volume. For example, when the point density of the three-dimensional point cloud data exceeds the set density threshold, a set large-size convolution kernel (7×7×7) is used to enhance the local feature perception capability, and at the same time, the network depth is expanded to 12 layers to fit the complex device topology and avoid missing high-detail components such as insulators; when the point density does not exceed the set density threshold, a set small-size convolution kernel (3×3×3) is enabled to improve the sensitivity of small wire recognition, and the network depth is compressed to 8 layers to adapt to FPGA resource constraints, ensuring relatively low segmentation delay.

[0119] For another example, when the selected semantic segmentation model type is the PointNet++ architecture model, the configured parameters may include the MLP (Multilayer Perceptron) layer dimension and sampling rate. For example, when the number of points exceeds the set threshold and the point density reaches the set distribution density condition, a deep MLP architecture (such as the dimension sequence [64, 128, 256, 512]) is used to enhance the feature extraction capability of complex equipment, while reducing the sampling rate to 0.2 to retain the fine point cloud of key components (such as circuit breaker operating mechanisms) to avoid oversampling and loss of details. When the number of points does not exceed the set threshold or the point density does not reach the set distribution density condition, a lightweight MLP architecture (dimensional sequence [32, 64, 128]) is enabled to adapt to FPGA resource constraints, and the sampling rate is increased to 0.5 to cover the continuity of long-distance conductors, ensuring the complete identification of key components such as insulator strings.

[0120] In the preferred embodiment described above, the CPU pre-builds different refined semantic segmentation models for different power supply operation types (such as emergency repair operations, maintenance and repair operations, disaster relief power supply, and power supply for major events), and establishes a mapping relationship between power supply operation types and refined semantic segmentation models. For example, for emergency repair and disaster relief power supply, a refined semantic segmentation model with strong interference resistance and high real-time performance, such as ENet4 or a modified Mask R-CNN7, is required. Such models can maintain good segmentation results under complex lighting and weather conditions. For maintenance and repair operations, a high-precision semantic segmentation model such as HRNet1 or U-Net3+1 can be built. Such models can accurately segment equipment details and ensure operational safety. For power supply maintenance operations, multiple models can be integrated, such as combining the scene understanding capabilities of DeepLab and the multi-scale processing advantages of PSPNet to build a hybrid model. Refined semantic segmentation models are high-precision models specifically designed for specific types of generator truck power supply operations. These models are trained and optimized for a specific type of operation scenario, such as tower access, cable connection, or transformer bypass operations, and have stronger recognition capabilities and higher classification accuracy. It enables more detailed segmentation of power equipment and environmental elements within a 3D point cloud, such as distinguishing detailed structures like insulators, crossarms, conductors, ground, and obstacles. Refined models typically employ complex network structures and rely on large amounts of high-quality annotated data for training, resulting in relatively high resource consumption. They are suitable for scenarios with low FPGA load and high recognition accuracy requirements.

[0121] The CPU also pre-builds a general-purpose lightweight semantic segmentation model. This model is a general-purpose model applicable to various power supply operation types. Its goal is to reduce computational complexity and improve inference speed and real-time performance while maintaining basic recognition capabilities. This model typically uses a simplified network architecture, with training data covering a variety of operation scenarios, and possesses strong generalization capabilities. While its recognition granularity is not as fine as that of a refined model, it can provide stable and efficient semantic segmentation results even under resource-constrained conditions, making it suitable for use when the FPGA is under heavy load.

[0122] The FPGA determines its load status. If the FPGA load does not reach a set upper limit, it selects a refined semantic segmentation model that matches the power supply operation type from multiple pre-built semantic segmentation models on the CPU. If the FPGA load reaches the set upper limit, it selects a general lightweight semantic segmentation model from multiple pre-built semantic segmentation models on the CPU. The selected semantic segmentation model is then loaded and used as the semantic segmentation model for the semantic segmentation pipeline. This mechanism ensures recognition accuracy and detailed representation when resources are sufficient, accurately identifying complex details such as equipment configuration and cable layout at the power generation truck operation site, effectively improving the accuracy of on-site parameter recognition and providing a reliable basis for subsequent operation decisions. It also switches to a low-resource consumption model when the FPGA is highly loaded, significantly reducing computational complexity and resource consumption. While ensuring stable system operation, it avoids processing speed degradation or even system crashes caused by excessive model computation, ensuring rapid response and continuous progress of semantic segmentation tasks. Furthermore, this mechanism implements intelligent and adaptive model scheduling, improving the flexibility, efficiency, and reliability of the overall system, effectively supporting efficient perception and intelligent decision-making at the power generation truck operation site.

[0123] Of course, the entire semantic segmentation model selection process can also be completed by the CPU: that is, the FPGA reports its own load status to the CPU. The CPU comprehensively considers the current power supply operation type and FPGA load status, decides on the optimal model, and then sends it to the FPGA for execution, thereby realizing an organic combination of centralized intelligent scheduling and distributed efficient processing.

[0124] Furthermore, those skilled in the art will appreciate that other methods can be used to select the semantic segmentation model to be used in the semantic segmentation pipeline. For example, a joint decision matrix for load and power supply operation type can be established to classify the FPGA load status into three levels: low, medium, and high. The power supply operation type can be further subdivided into categories such as routine maintenance, emergency power supply, and complex terrain operations. Through pre-set rule combinations, the corresponding semantic segmentation model can be quickly indexed. Taking the combination of medium load and complex terrain operations as an example, a medium-sized model that balances accuracy and computational efficiency can be directly selected. A reinforcement learning dynamic optimization strategy can also be adopted to allow the FPGA to continuously accumulate operational performance data for various semantic segmentation models under different load and power supply operation types during task execution. A reward mechanism can be used to guide semantic segmentation model selection, gradually training the optimal model selection strategy for specific scenarios, achieving more intelligent and adaptive model selection, and further improving the system's flexibility and performance.

[0125] After the FPGA completes semantic segmentation of the 3D point cloud data, it then uses the segmentation results to identify on-site parameters related to the generator truck's power supply operation, completing the semantic segmentation task. These on-site parameters include power facility parameters (such as the physical dimensions of power equipment like towers, crossarms, and insulators) and channel environment parameters (such as terrain slope, ground type and load-bearing capacity, and the location and distribution of surrounding obstacles). The following details the identification process for these two types of parameters.

[0126] ① Based on the power facility point cloud data identified in the semantic segmentation results, the geometric fitting and projection analysis methods are used to calculate the power facility parameters related to the power supply operation of the generator truck. For details, see Figure 3 .

[0127] a. Calculate tower height

[0128] Based on the tower point cloud data identified in the semantic segmentation results: fitting the ground plane; calculating the distance from the tower point cloud to the ground plane to determine the tower height.

[0129] First, a ground plane is fitted from the tower point cloud data (i.e., the point cloud data of the tower in the three-dimensional point cloud data of the power generation vehicle power supply operation site) as a reference for calculating the tower height. Then, based on the tower point cloud data, a minimum bounding box of the tower point cloud is constructed, and the vertical distance from each tower point to the fitted ground plane is calculated. Finally, the largest distance from all calculated vertical distances is selected as the tower height.

[0130] b. Calculate crossarm length

[0131] Based on the crossarm point cloud data identified in the semantic segmentation results: fitting the crossarm plane; projecting the crossarm point cloud onto the crossarm plane; fitting the minimum circumscribed circle on the crossarm point cloud projection to determine the crossarm length.

[0132] First, a plane representing the crossarm's position, known as the crossarm plane, is fitted from the crossarm point cloud data (i.e., the crossarm point cloud data from the 3D point cloud data of the generator truck's power supply operation site). The crossarm point cloud is then projected onto the fitted crossarm plane to eliminate any tilt or rotation errors. Next, a minimum circumscribed circle is fitted to the projected crossarm point cloud. The diameter of this circumscribed circle is the length of the crossarm, as crossarms are generally assumed to have a circular or approximately circular cross-section.

[0133] Optionally, the cross-arm spacing may be further calculated: based on the voxelized cross-arm point cloud data, the center position of the cross-arm is identified, and the distance between adjacent cross-arms is calculated.

[0134] c. Calculate the insulator disk diameter

[0135] Based on the insulator point cloud data identified in the semantic segmentation results: fit the axial direction of the insulator; calculate the maximum spacing of the insulator point cloud in the axial direction, and determine the insulator height based on the maximum spacing; project the insulator point cloud onto a plane perpendicular to the axial direction, and fit the minimum circumscribed circle on the insulator point cloud projection to determine the insulator disk diameter.

[0136] Fitting the insulator's axial direction includes constructing a minimum bounding box for the insulator point cloud based on the insulator point cloud data (i.e., the insulator point cloud data in the three-dimensional point cloud data of the power generation vehicle power supply operation site), and determining the axial direction of the insulator point cloud data using principal component analysis. The maximum spacing of the insulator point cloud in the axial direction is the insulator height. The diameter of the minimum circumscribed circle fitted on the insulator point cloud projection is the insulator disc diameter.

[0137] ② Based on the point cloud data of environmental objects identified in the semantic segmentation results, determine the channel environmental parameters related to the power supply operation of the power generation vehicle.

[0138] a. Determine the width and slope of the working channel pavement based on the point cloud data of the working channel pavement identified in the semantic segmentation results.

[0139] Among them, the working channel pavement point cloud data identified in the semantic segmentation results can be plane fitted to extract the plane equation of the pavement; the working channel pavement point cloud data can be projected onto the fitting plane and sliced ​​along the direction of travel to generate a two-dimensional contour line; the pavement width is calculated based on the horizontal distance between the boundary points of the two-dimensional contour line, and the slope is calculated based on the elevation change.

[0140] b. Based on the obstacle point cloud data identified in the semantic segmentation results, determine the obstacle removability judgment result and spatial distribution parameters.

[0141] Non-removable obstacles include houses, light poles, etc., which must be avoided during operations. Removable obstacles include trees, vehicles, etc., which can be removed to optimize the operation path and vehicle parking location when necessary.

[0142] During specific operations, the geometric features and spatial position information of the obstacles identified in the semantic segmentation results can be extracted, and combined with the preset removability judgment rules (such as size, material, fixing method, etc.) to determine whether the obstacles are removable; at the same time, the spatial distribution parameters of the obstacles, including position coordinates, density distribution and relative distance from the operation path, are counted to provide an environmental analysis basis for subsequent operation analysis and decision-making.

[0143] It should be noted that in this embodiment of the present invention, the two key steps of semantic segmentation—semantic segmentation and field parameter identification—can be independently executed by different FPGAs. The CPU dynamically schedules and allocates these tasks based on the FPGA's load, computing resources, and task priority, thereby optimizing hardware resource utilization and maximizing task processing efficiency.

[0144] (3) After the FPGA completes the semantic segmentation task, it notifies the CPU that the semantic segmentation task is completed. At this point, the field identification sub-platform 1011 processes the input three-dimensional point cloud data and finally outputs the identified field parameters related to the power supply operation of the power generation vehicle for access by the operation decision sub-platform 1012.

[0145] In summary, the field recognition sub-platform 1011 adopts a heterogeneous architecture of CPU+GPU+FPGA, which has the following advantages:

[0146] First, this architecture fully leverages the strengths of each computing unit: the CPU is responsible for overall task scheduling and resource management, serving as the control center of the field recognition sub-platform 1011; the GPU excels at parallel processing of large-scale regular data and is suitable for compute-intensive tasks such as point cloud data preprocessing; and the FPGA, with its low latency, high energy efficiency, and reconfigurability, is particularly suitable for deploying customized semantic segmentation pipelines to meet the needs of scenarios with high real-time requirements.

[0147] Secondly, inter-module communication is achieved through high-speed interconnects such as PCIe and Avalon / AXI. This not only ensures data transmission bandwidth and efficiency, but also supports collaboration between multiple hardware platforms, improving the overall system response speed and stability. In particular, the direct data exchange between the GPU and FPGA via high-speed interconnect reduces reliance on CPU intervention and further reduces system latency.

[0148] In addition, the architecture has good flexibility and scalability, which facilitates dynamic adjustment of model deployment strategies according to actual operation scenarios, supports switching of multiple algorithms and function upgrades, and provides an efficient and reliable technical foundation for intelligent on-site identification and decision-making of power supply operations of power generation vehicles in complex environments.

[0149] 2. Operation decision sub-platform 1012

[0150] The exemplary job decision sub-platform 1012 is specifically used for:

[0151] Pre-build a multi-dimensional operating condition specification indicator system that integrates quantitative indicators from multiple dimensions such as environment, equipment, and personnel (such as load matching thresholds and safety distance thresholds). These indicators can be set according to relevant safety standards and operating specifications;

[0152] Based on the actual situation, the on-site parameters related to the power supply operation of the generator truck are identified, and the operating condition analysis indicators of each dimension are checked item by item to determine whether the current operating environment meets the conditions for the use of the bucket truck and the generator truck;

[0153] With the support of the indicator engine, the operation analysis and decision-making results are generated, including: operation condition compliance judgment (whether safe operation conditions are met), recommended operation plans (such as vehicle parking location, operation path planning), and shielding plans.

[0154] A specific implementation is provided below.

[0155] (1) The operation decision sub-platform 1012 evaluates the applicability of the boom truck and generates an operation plan based on the pre-built boom truck usage condition specification indicators and the field parameters related to the power supply operation of the power generation truck.

[0156] Specifically, based on the on-site parameters related to the generator truck's power supply operation and the on-site 3D point cloud data, the actual operating conditions at the generator truck's power supply operation site can be determined from two aspects: on-site equipment and environmental conditions. This allows for a suitability assessment of the boom truck. The specific assessment content can be as shown in the example given in Table 2. If all assessment items are qualified, the boom truck is qualified for use; if one or more items are not met, the boom truck is not qualified for use.

[0157] Table 2

[0158]

[0159] Based on the boom truck suitability assessment, boom truck usage condition specification indicators can be set to be implemented by at least the following set specification category constraints: motion range verification rules, parking area screening rules, and safe operation rules. Table 3 below shows some examples of boom truck usage condition specification indicators provided in this embodiment.

[0160] Table 3

[0161]

[0162] Exemplarily, the operation decision sub-platform 1012 is used to generate an operation plan for the boom truck, which specifically includes the following operations ①-④.

[0163] ①Determine the specifications of the bucket truck

[0164] From the pre-configured bucket trucks of various specifications, select the bucket truck model that meets the requirements of the bucket truck usage conditions and specifications.

[0165] ② Determine the reach of the bucket truck's working bucket

[0166] Combined with the selected boom truck model and on-site parameters related to the power supply operation of the generator truck, the reachable range of the working bucket that meets the requirements of the boom truck's operating conditions and specifications is determined from the three-dimensional point cloud data of the power supply operation site of the generator truck.

[0167] As a specific implementation method, the key parameters of the selected boom truck model, such as boom length, rotation angle, etc., are collected, and a three-dimensional kinematic model is constructed based on this to define the theoretical reachable space; at the same time, the three-dimensional point cloud coordinates of the power supply operation site of the power generation vehicle (referred to as the on-site point cloud coordinates) and the boom truck model are unified in the coordinate system, and the three-dimensional point cloud coordinates of the obstacles in the on-site parameters are obtained (referred to as the obstacle point cloud coordinates). Then, based on the kinematic positive solution, the coordinate range of all possible postures of the working bucket is calculated, and then a Boolean operation is performed with the obstacle point cloud coordinates to exclude the occluded area, and the boundary points of the remaining space are extracted to generate the reachable range outline. Afterwards, the boom truck model, on-site point cloud coordinates and reachable range are visualized in three dimensions with the help of professional software, and verified and corrected in combination with the actual operation scene, uncovered obstacles are excluded, and the reachable range of the boom truck working bucket is output to meet the requirements of the boom truck use condition specification indicators.

[0168] ③Determine the optimal working path of the bucket truck

[0169] A three-dimensional model of the bucket truck's motion envelope, including the working bucket, multi-section arm, and rotating base, is constructed. Interference detection is performed based on this model and point cloud data of environmental obstacles to identify risk areas where the working bucket and arm may collide with surrounding obstacles during movement. Based on the risk area identification results and combined with the constraints required by the bucket truck's usage specification indicators, a path planning algorithm (such as RRT or artificial potential field method) is used to search for the optimal working bucket operation trajectory that meets safety and accessibility, ensuring that the bucket truck dynamically avoids obstacles during movement, and ultimately determining the optimal operating path of the bucket truck.

[0170] ④Determine the parking range of the bucket truck

[0171] According to the on-site parameters related to the power supply operation of the generator truck, the parking range of the boom truck is determined from the three-dimensional point cloud data of the power supply operation site of the generator truck, so that it meets the requirements of the standard indicators of the use conditions of the boom truck.

[0172] (2) The operation decision sub-platform 1012 evaluates the applicability of the power generation vehicle and generates an operation plan based on the pre-built power generation vehicle usage condition specification indicators and the on-site parameters related to the power supply operation of the power generation vehicle.

[0173] The power generation capacity of a generator truck is a key factor in ensuring reliable power supply operations. However, its actual use depends not only on its own performance indicators but also on a combination of factors such as environmental conditions, access point selection, and parking location. Specifically, based on the on-site parameters related to the generator truck's power supply operations and the on-site 3D point cloud data, the actual operating conditions at the generator truck's power supply site can be determined from two perspectives: on-site equipment and environmental conditions. This allows for a suitability assessment of the generator truck. Specific assessment details can be shown in the example given in Table 4. If all assessment items are qualified, the generator truck is qualified for use; if one or more items are not met, the generator truck is not qualified for use.

[0174] Table 4

[0175]

[0176] Based on the applicability assessment of the power generation vehicle, the power generation vehicle usage condition specification indicators can be set to be implemented by at least the following set specification category constraints: electrical connection rules, space layout rules, and cable laying rules. Table 5 below is an example of a power generation vehicle usage condition specification indicator provided in this embodiment.

[0177] Table 5

[0178]

[0179]

[0180] Exemplarily, the operation decision sub-platform 1012 is used to generate an operation plan for a power generation vehicle, specifically including: determining the access point and parking location of the power generation vehicle. Specifically, based on pre-established power generation vehicle usage condition specification indicators, combined with the identified parameters such as the topography, obstacle distribution, and location of the conductor to be operated at the power generation vehicle power supply operation site, the area that meets the spatial layout constraints (such as minimum operating space 8m×3m, minimum turning radius ≥6m, slope ≤5°, load-bearing capacity ≥150kPa) is first evaluated to preliminarily determine the parking location of the power generation vehicle; then, according to electrical access rules such as voltage, phase, and harmonics, suitable access points are matched, and the cable path is planned in combination with cable laying rules and constraints; finally, verification and optimization are carried out based on the safety distance, operating range, and environmental constraints to ensure that the access point and parking location of the power generation vehicle are optimally configured in terms of electrical matching, operational safety, and operational efficiency.

[0181] (3) The operation decision sub-platform 1012 generates an insulation shielding plan based on pre-built insulation shielding specification indicators and combined with on-site parameters related to the power supply operation of the power generation vehicle.

[0182] During implementation, the insulation shielding scope, materials, and operating standards that meet the insulation shielding specification requirements can be determined based on the 3D point cloud data of the generator truck power supply operation site and relevant on-site parameters of the generator truck power supply operation to ensure electrical safety during the operation. Table 6 below provides some examples of insulation shielding specification indicators.

[0183] Table 6

[0184]

[0185] Preferably, the operation decision sub-platform 1012 is used to generate an insulation shielding plan, specifically including:

[0186] Detect charged objects within the operating range based on on-site parameters related to the power supply operation of the generator truck;

[0187] Determine insulation shielding scope, sequence, and tools based on pre-built insulation shielding specifications and test results.

[0188] Typically, the operation decision sub-platform 1012 is used to determine the insulation shielding range, specifically including:

[0189] Determine whether there is a potential safety risk based on the distance between the working position of the power supply operator on the generator truck and the charged object;

[0190] Calculate the range of motion of the power supply operator on the power generation vehicle: A = B + T + M + N, where B represents the arm span of the power supply operator on the power generation vehicle, T represents the length of the operating tool of the power supply operator on the power generation vehicle, M represents the amplitude of the operating movement of the power supply operator on the power generation vehicle, and N represents the movable distance of the power supply operator on the working bucket of the boom truck;

[0191] Determine the insulation shielding range based on the risk assessment results and activity range calculation results.

[0192] Among them, if the distance between the working position of the power supply operator of the power generation vehicle and the charged object is greater than the set distance threshold, it is judged that there is no potential safety risk. At this time, the internal area of ​​the sphere with the waist of the power supply operator of the power generation vehicle as the center and R as the radius is used as the insulation shielding range, where R=A+D, and D represents the safety distance specified by the insulation shielding specification indicators. If the distance between the working position of the power supply operator of the power generation vehicle and the charged object is less than or equal to the set distance threshold, it is judged that there is a potential safety risk. At this time, the internal area of ​​the sphere with the waist of the power supply operator of the power generation vehicle as the center and (R+extended distance) as the radius is used as the insulation shielding range. The extended distance can be set by those skilled in the art based on experience, for example, 0.3m.

[0193] Based on the above solution, as a preferred embodiment, the operation decision sub-platform 1012 is used to generate power generation vehicle power supply operation analysis and decision based on the pre-built multi-dimensional operation condition specification indicator system and the field parameters, including:

[0194] The data access layer is used to access the following data: a multi-dimensional operating condition specification indicator system and on-site parameters related to the power supply operation of the generator truck, and convert the accessed data into a standardized format;

[0195] The decision-making core layer is used to combine the on-site parameters related to the power supply operation of the generator truck, analyze each indicator in the multi-dimensional operating condition specification indicator system one by one, evaluate whether the power supply operation site of the generator truck meets the operation requirements, generate an operation plan, and output the evaluation results and generation results;

[0196] The dynamic optimization layer is used to collect decision-making effect feedback data and optimize the decision-making core layer based on the collected results.

[0197] Among them, the dynamic optimization layer is specifically used to: collect decision-making effect feedback data from multiple sources, such as decision-making effect evaluation results and decision modification opinions; classify, count and model the collected feedback data, and identify under what conditions the decision-making strategy of the current decision-making core layer has deviations or misjudgments; according to the feedback data analysis results, the decision-making strategy of the decision-making core layer is parameter tuned or logically corrected, for example, adjusting each indicator in the multi-dimensional indicator system and adding new constraints; introducing machine learning algorithms (such as reinforcement learning) to automatically optimize decision-making strategies.

[0198] Optionally, the site identification sub-platform within the cloud management platform is further configured to construct a 3D real-world model of the power generation truck's power supply operation site based on the 3D point cloud data. Accordingly, the dynamic optimization layer is specifically configured to display the 3D real-world model and the output of the decision-making core layer; receive user-inputted decision-making feedback on the displayed content; and optimize the decision-making core layer based on the collected feedback.

[0199] The on-site identification sub-platform can employ an implicit surface representation method to obtain a 3D real-world model of the power generation truck's power supply operation site based on 3D point cloud data. In specific implementation, an implicit surface can be constructed by solving the Poisson equation based on the 3D point cloud data and its normal vector field. The on-site identification sub-platform then uses the MarchingCubes algorithm to extract a triangular mesh from the implicit surface and simplifies and optimizes the extracted triangular mesh. For example, edge collapse and Laplacian smoothing algorithms are applied to reduce mesh complexity while maintaining geometric features, eliminate triangular facet distortion, and improve model computational efficiency and visualization quality. Finally, the texture of the original multi-view image is mapped onto the triangular mesh, resulting in a 3D real-world model with rich details and a sense of realism.

[0200] In summary, the technical solution provided by the embodiments of the present invention automatically identifies key on-site parameters by collecting and analyzing three-dimensional point cloud data from the generator truck power supply operation site. This, combined with a pre-built multi-dimensional operating condition specification indicator system, enables intelligent analysis and assessment of the operating environment, thereby generating scientific and reliable operational decision-making plans. This solution effectively overcomes the data errors and empirical judgment biases associated with traditional reliance on manual surveys, improves the standardization and automation of operational assessments, and provides dynamic risk warning capabilities, significantly enhancing the safety, standardization, and efficiency and accuracy of decision-making responses during generator truck power supply operations.

[0201] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0202] In embodiments of the present invention, the term "and / or" describes the association relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally indicates that the associated objects are in an "or" relationship.

[0203] Each embodiment in this specification is described in a related manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments.

[0204] In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0205] For the convenience of description, the above device is described as being divided into various units / modules based on their functions. Of course, when implementing the present invention, the functions of each unit / module can be implemented in the same or multiple software and / or hardware.

[0206] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing related hardware through a computer program. The program can be stored in a computer-readable storage medium, and when executed, the program can include the processes in the above-described method embodiments. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0207] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. An intelligent decision-making system for power supply operation of a power generation vehicle, characterized in that: The system includes edge nodes and a cloud management platform; wherein: The edge node is used to collect multi-view image sequences of the power supply operation site of the power generation vehicle; The cloud management platform includes: The on-site recognition sub-platform is used to: construct three-dimensional point cloud data using the multi-view image sequence; perform semantic segmentation on the three-dimensional point cloud data to identify on-site parameters related to the power supply operation of the power generation vehicle; The operation decision sub-platform is used to generate power supply operation analysis decisions of the power generation vehicle based on the pre-built multi-dimensional operation condition specification indicator system and the above-mentioned on-site parameters.

2. The system according to claim 1, wherein: The on-site identification sub-platform adopts a heterogeneous architecture consisting of a central processing unit (CPU), a graphics processing unit (GPU), and a field programmable gate array (FPGA); wherein: The CPU schedules the GPU to execute a point cloud construction task: constructing three-dimensional point cloud data using the multi-view image sequence; After completing the point cloud construction task, the GPU notifies the FPGA to perform a semantic segmentation task: performing semantic segmentation on the three-dimensional point cloud data to identify on-site parameters related to the power supply operation of the power generation vehicle; After the FPGA completes the semantic segmentation task, it notifies the CPU that the task is completed.

3. The system according to claim 2, characterized in that The GPU executes a point cloud construction task, including: the GPU reads the multi-view image sequence from a shared memory, constructs three-dimensional point cloud data using the multi-view image sequence, and writes the three-dimensional point cloud data into the shared memory after adding a self-description data header; The FPGA performs semantic segmentation tasks, including: The hardware parses the self-describing data header and dynamically configures semantic segmentation pipeline parameters based on the information carried therein; the semantic segmentation pipeline is started based on the configuration result: the three-dimensional point cloud data is semantically segmented to identify on-site parameters related to the power supply operation of the power generation vehicle; and the identification results are stored in the shared memory.

4. The system according to claim 3, characterized in that The FPGA hardware parses the self-describing data header and dynamically configures semantic segmentation pipeline parameters based on the information carried therein, including: The hardware parses the self-describing data header to extract the point cloud data format and point cloud scale attribute information therein; Acquire the FPGA load and power supply operation type, and select a semantic segmentation model to be used by the semantic segmentation pipeline from a plurality of semantic segmentation models pre-built by the CPU according to the acquisition result; Configure the parameters of the selected semantic segmentation model based on the extracted point cloud data format and point cloud scale attribute information.

5. The system according to claim 4, characterized in that The FPGA selects a semantic segmentation model to be used by the semantic segmentation pipeline from the multiple semantic segmentation models pre-built by the CPU according to the acquisition result, including: Determining a load state of the FPGA; If the FPGA load does not reach the set upper limit, selecting a refined semantic segmentation model that matches the power supply operation type from the multiple semantic segmentation models pre-built by the CPU; If the FPGA load reaches a set upper limit, a general lightweight semantic segmentation model is selected from multiple semantic segmentation models pre-built by the CPU; The selected semantic segmentation model is loaded as the semantic segmentation model to be used in the semantic segmentation pipeline.

6. The system according to claim 1, wherein: The edge node includes a drone equipped with a camera; The edge node is used to collect a multi-view image sequence of the power supply operation site of the power generation vehicle, including: Shoot the power supply operation site of the generator truck from multiple angles; When photographing the tower structure at the power supply operation site of the power generation vehicle, the viewing angle range of adjacent photographed images is controlled to be within the range of 10° to 20°.

7. The system according to claim 6, characterized in that The edge node is specifically used to: shoot the power supply operation site of the power generation vehicle at multiple time periods and multiple angles; The cloud management platform also includes an image fusion sub-platform for performing spatiotemporal registration and feature fusion on multi-view image sequences taken at multiple time periods to generate a unified multi-view image sequence.

8. The system according to claim 7, characterized in that The image fusion sub-platform is also used for: An image preprocessing operation including denoising and contrast enhancement is performed on the unified multi-view image sequence.

9. The system according to any one of claims 1 to 8, characterized in that The job decision sub-platform includes: The data access layer is used to access the following data: the multi-dimensional operating condition specification indicator system and on-site parameters related to the power supply operation of the generator truck, and convert the accessed data into a standardized format; The decision-making core layer is used to combine the on-site parameters related to the power supply operation of the generator truck, analyze each indicator in the multi-dimensional operating condition specification indicator system one by one, evaluate whether the power supply operation site of the generator truck meets the operation requirements, generate an operation plan, and output the evaluation results and generation results; The dynamic optimization layer is used to collect decision-making effect feedback data and optimize the decision-making core layer according to the collected results.

10. The system according to claim 9, characterized in that The site identification sub-platform in the cloud management platform is further used to: construct a three-dimensional real-scene model of the power supply operation site of the power generation vehicle based on the three-dimensional point cloud data; The dynamic optimization layer is specifically used to: display the three-dimensional real scene model and the output results of the decision core layer; receive decision effect feedback data input by the user for the displayed content, and optimize the decision core layer according to the collected results.

Images