Method and system for automatic classification of overhead transmission line equipment point cloud based on pty3

Abstract

The embodiment of the application provides a kind of overhead transmission line equipment point cloud automatic classification method and system based on PTV3, belong to three-dimensional point cloud data processing and artificial intelligence field.The method comprises: constructing the standardized point cloud sample library covering multiple voltage grades, multiple terrains, multiple tower types;Using the PTV3 deep learning model based on serialization attention and patch interaction mechanism to train the sample library, adopt the customization data enhancement strategy and composite loss function for the characteristics of electric power equipment in the training process, obtain the electric power special point cloud classification model;The trained model is integrated into the automation processing pipeline, realizes the whole process intelligent processing of point cloud data uploading, automatic classification, result checking and structured output;Based on high-precision classification result, dynamically combined to generate tree barrier analysis, working condition simulation, construction acceptance multiple business application scenarios.The classification precision is high, the generalization ability is strong, the processing efficiency is greatly improved, and the systematization and sustainable evolution are also realized.

Description

Technical Field

[0001] This invention relates to the fields of three-dimensional point cloud data processing and artificial intelligence technology, specifically to an automatic classification method and system for point clouds of overhead transmission line equipment based on PTV3. Background Technology

[0002] With the widespread adoption of drone-based LiDAR technology, power grid companies have acquired extensive and high-precision laser point cloud data of transmission lines for applications such as archive digitization, tree obstacle measurement, and flight path planning. However, the processing and application of massive point cloud data faces three major challenges: 1. Low data processing efficiency: Point cloud processing heavily relies on high-performance workstations and professional operators, and the processing flow is long (e.g., 2 days for a single line), which cannot meet the needs of large-scale and real-time applications.

[0003] 2. Insufficient classification accuracy and intelligence: Traditional point cloud processing software or general algorithms are not deeply optimized for power scenarios, resulting in low recognition rates for key small components such as conductors, insulators, and lead wires, and they cannot understand the spatial topological relationships of power equipment (such as conductor-tower connections).

[0004] 3. Rigid data management and application: Data is stored in a scattered manner (such as multiple external hard drives), making management difficult and retrieval time-consuming; application scenario development relies on manual labor, resulting in poor flexibility and insufficient value mining.

[0005] However, existing deep learning-based point cloud classification methods (such as PointNet++ and early Transformer models) have shortcomings when dealing with complex scenarios such as power transmission lines, which are long-distance, sparsely distributed, and have multiple scale targets (from macroscopic towers to microscopic insulators). These shortcomings include limited receptive field, weak ability to extract features from small targets, and sensitivity to class imbalance, resulting in inaccurate and discontinuous classification results. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic point cloud classification method and system for overhead transmission line equipment based on PTV3. It has high classification accuracy, strong generalization ability, greatly improved processing efficiency, and also achieves systematization and sustainable evolution.

[0007] To achieve the above objectives, embodiments of the present invention provide an automatic point cloud classification method for overhead transmission line equipment based on PTV3, the method comprising: Construct a standardized point cloud sample library covering multiple voltage levels, terrain types, and tower types; The PTV3 deep learning model based on sequential attention and patch interaction mechanism was used to train the sample database. During the training process, a customized data augmentation strategy and composite loss function were adopted for the characteristics of power equipment to obtain a power-specific point cloud classification model. The trained model is integrated into the automated processing pipeline to achieve intelligent processing of the entire process of point cloud data uploading, automatic classification, result verification and structured output; Based on high-precision classification results, various business application scenarios such as tree obstacle analysis, working condition simulation, and construction acceptance are dynamically combined and generated.

[0008] Preferably, constructing a standardized point cloud sample library covering multiple voltage levels, terrain types, and tower types includes: Collect raw point cloud data and perform data cleaning, including format unification conversion and filtering to remove noise. The cleaned data is structured and segmented according to the span of the transmission line tower. Each data segment contains one tower and its front and rear passages. The point cloud data after the data is cut is finely labeled according to at least 15 categories, including conductors, towers, ground wires, insulators, drain wires, vegetation, ground, buildings, roads, temporary buildings, overhead crossover lines, main grid under crossover lines, distribution grid under crossover lines, and noise. The voltage level, terrain, scene, tower type and collection time of each sample are recorded.

[0009] Preferably, the optimization of the PTV3 deep learning model includes: The input point cloud is serialized and encoded using a space-filling curve, mapping the disordered point cloud into a one-dimensional ordered sequence. The serialized point cloud is grouped into non-overlapping patches, and self-attention computation is performed within each patch. By employing patch interaction strategies such as extended displacement, shift patching, or shift sequence, information is exchanged between different patches, expanding the model's receptive field from local to global to capture continuous features of long-distance targets.

[0010] Preferably, the customized data augmentation strategy includes: During the model training phase, voxelization downsampling is used to control the amount of data, and random rotation, scaling, flipping, and coordinate jitter are used to enhance the model's generalization ability. For categories with small sample sizes, such as buildings and crossroads, manual cropping and copying / pasting are performed into the training set to alleviate the class imbalance problem. During the model inference stage, multi-angle ensemble inference is adopted to perform rotation predictions on the input point cloud from multiple angles and fuse the results to improve robustness to occlusion and complex overlapping scenes.

[0011] Preferably, the composite loss function includes: Cross-entropy loss is used as the basic classification loss to handle the class probability distribution; We introduce structured Lovasz loss as an auxiliary loss to directly optimize the IoU metric, thereby alleviating the class imbalance problem and improving the segmentation accuracy of small objects.

[0012] Preferably, the automated processing pipeline is built on the Workflow engine, which decomposes the point cloud classification task into multiple automated steps that are executed sequentially, including automatic pole positioning and cutting, automatic model classification, scene analysis and report generation, and supports light-weight manual interactive verification in specific steps, so as to realize the online and automated processing of the entire point cloud process.

[0013] Preferably, dynamically combining and generating business application scenarios includes: Based on the classified point cloud data labels, the Euclidean clustering algorithm is used to combine multiple features such as clearance distance and ground distance to generate at least one customized application in the following categories: high wind condition simulation, live-line working distance analysis, construction acceptance data verification, special working environment simulation, line hidden danger early warning and emergency repair plan generation.

[0014] On the other hand, the present invention provides an automatic point cloud classification system for overhead transmission line equipment based on PTV3 to implement the above method, the system comprising: The sample library management module is used to store and manage a standardized transmission line point cloud sample library, supporting data versioning, tagging management, and rapid retrieval; The PTV3 model training and optimization module is used to perform model training, parameter tuning, and version iteration. The online intelligent classification and processing module integrates a pre-trained PTV3 model and an automated processing pipeline. It is used to receive natural language task requests input by users, automatically parse tasks, retrieve data, call models, generate and output structured reports. The business application scenario generation module integrates a scenario-based application generation engine, which is used to convert classified structured point cloud data into analysis reports or decision-making suggestions that directly serve power transmission business scenarios.

[0015] Preferably, the system also includes an intensive data management module, which is used to centrally store the scattered point cloud data on an online server, breaking down physical data barriers and supporting the fusion and comparison of historical data and intelligent data query functions.

[0016] Preferably, the system also includes a model self-optimization and sample library iteration unit, which is used to collect false detection and false negative cases in practical applications, build a positive and negative sample library, and periodically trigger incremental training of the model to achieve continuous optimization and self-evolution of model performance.

[0017] Through the above technical solutions, a laser point cloud intelligent agent was constructed, integrating centralized data management, intelligent processing, scenario-based applications, and self-optimization iteration. This enabled the full-process, automated, and high-precision transformation of transmission line point clouds from raw data to business value. Simultaneously, it directly empowers business scenarios such as tree fault analysis, operational condition simulation, intelligent acceptance, and emergency repair, solving industry challenges related to the difficulty of managing massive point cloud data, slow processing, and limited application, and possesses sustainable self-evolution capabilities.

[0018] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating the automatic point cloud classification method for overhead transmission line equipment based on PTV3 provided by the present invention. Figure 2 This is a schematic diagram of the Point Transformer model; Figure 3 This is a schematic diagram of the point transformer layer in the Point Transformer model; Figure 4 This is a schematic diagram of the point transformer block in the Point Transformer model; Figure 5 This is a diagram illustrating the differences between PTv1 and PTv2; Figure 6 This is a diagram illustrating the patch grouping process. Detailed Implementation

[0020] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0021] It should be noted that the acquisition, transmission, storage, use, and processing of data in the technical solution of this application all comply with relevant laws and regulations. In the embodiments of this application, certain existing industry solutions such as software, components, and models may be mentioned. These should be considered exemplary, intended only to illustrate the feasibility of implementing the technical solution of this application, and do not imply that the applicant has already used or necessarily used such solutions.

[0022] See Figure 1 The first aspect of this invention provides an automatic point cloud classification method for overhead transmission line equipment based on PTV3, which specifically includes the following steps: S1. Construct a multi-level standardized point cloud sample library for transmission lines: Collect laser point cloud data of overhead transmission lines covering different voltage levels (110kV-1100kV), terrains (plains, mountains, hills), scenes (farmland, forests, towns, wetlands), seasonal variations, and tower types (straight towers, tension towers, angle towers, etc.). Standardize and preprocess the raw point cloud data, including: Format unification and cleaning: Extract key fields from the original LAS format file, convert it to NPY array format, and perform filtering and noise reduction to remove isolated noise points.

[0023] Structured segmentation: The transmission line towers are segmented, with each segment containing one tower and its preceding and following passages. The typical passage width is 100 meters, forming an independent unit that is easy to process.

[0024] Refined labeling: The point cloud after data segmentation is manually and precisely labeled, defining at least the following 15 categories: conductors, towers, ground wires, insulators, drain wires, vegetation, ground, buildings, roads, temporary structures, overhead crossovers, main grid crossovers, distribution grid crossovers, and noise. The labeling process follows strict geometric feature and topological relationship specifications (e.g., the hardware at the connection between the insulator and the conductor is classified as an insulator).

[0025] Meta-information recording: JSON files are used to record information such as voltage level, terrain, scene, region, collection time, channel characteristics, and tower type for each sample, ensuring data traceability and diversity.

[0026] S2. Training an automatic point cloud classification model based on the improved PTV3 model. The PTV3 model with U-Net as the skeleton and incorporating a sequential self-attention mechanism is used as the core classification network, and the following key optimizations are made for the power scenario: Serialization encoding and patch attention include: serializing unordered point clouds using Z-order or Hilbert space-filling curves to preserve spatial proximity; grouping non-overlapping patches and performing self-attention computation within each patch; and passing information between different patches through patch interaction strategies (such as expansion displacement, shifting patches, and shifting order) to extend the model's receptive field from local to global while maintaining computational efficiency, so as to effectively capture the continuous features of long-distance targets such as conductors and understand the topological structure between towers, conductors, and insulators.

[0027] Targeted data preprocessing and augmentation include: during the training phase, voxelized downsampling (grid size 0.8) is used to control the data volume while preserving the integrity of key components; random rotation, scaling, flipping, and coordinate jitter are used for data augmentation; for categories with small sample sizes, such as buildings, intersections, and drainage lines, manual cropping and copy-pasting are performed into the training set to alleviate class imbalance. During the inference phase, multi-angle ensemble inference is used to rotate the input point cloud at multiple angles (step size 0.5 degrees) and predict them separately, finally fusing the results to improve robustness to complex scenes such as intersection occlusion and vegetation cover.

[0028] The design of a composite loss function includes: Main loss function: Cross-entropy loss, which handles the probability distribution of the basic classification.

[0029] Auxiliary loss function: Structured Lovasz loss, which directly optimizes the IoU index, alleviates the class imbalance problem, and particularly enhances the model's segmentation accuracy for small targets (insulators, drain wires).

[0030] Thus, the model trained through the above steps can achieve a recognition rate of over 95% for ground features such as conductors and towers, and an overall mIoU of over 91%.

[0031] S3. Construct a point cloud classification intelligent processing system to integrate the trained PTV3 model into an automated, process-oriented system. This system includes: Centralized data management module: Centralizes point cloud data scattered on mobile hard drives to online servers, breaking down physical data barriers; establishes a fast search engine based on classification tags, so that finding a route that used to take 1 hour can now be done in 27 seconds by entering keywords; supports historical data fusion and comparison and intelligent data query functions.

[0032] Automated classification and processing pipeline: Based on the Workflow engine, the entire process, including file upload, point cloud classification, scene analysis, and report generation, is online. When a user inputs "query tree obstruction information for towers #1-#9 on the Lifu 5921 line," the system automatically performs point cloud analysis for that section and generates a tree obstruction report. Previously, processing 10 point cloud data points required two people and one day; now, one person can complete it within one hour. 14 out of 16 key steps have been automated.

[0033] Scenario-based application generation engine: Based on high-precision classification results, it uses algorithms such as Euclidean clustering to dynamically combine and generate customized applications that meet business needs, including but not limited to: tree and bamboo barrier clearance distance measurement, live-line working safety distance analysis, construction acceptance data verification, high wind condition simulation, special working environment simulation, line hidden danger early warning, and emergency repair plan generation.

[0034] S4. Model self-optimization and iterative updates of the sample library establish a closed-loop mechanism for sustainable optimization, including: Sample database update: Collect false positive cases from real-world applications as negative samples and false negative cases as positive samples, and construct a new training dataset with a positive-to-negative sample ratio of 2:8 or stratified sampling.

[0035] Incremental training: Incremental training is performed based on the new dataset to obtain a new version of the model.

[0036] Model testing and release: Compare the performance of different versions on the same test set, select the best one to release for production application, and realize the continuous evolution of the model.

[0037] A second aspect of the present invention provides an automatic point cloud classification system for overhead transmission line equipment based on PTV3 to implement the above-described method, the system comprising: Centralized data management module: Used to centrally store scattered point cloud data on an online server, breaking down physical data barriers and supporting the fusion and comparison of historical data and intelligent data query functions.

[0038] Sample Library Management Module: Used to store and manage the standardized transmission line point cloud sample library, supporting data versioning, tagging management, and rapid retrieval.

[0039] PTV3 Model Training and Optimization Module: Configured to perform the S2 step, supporting incremental training, parameter tuning, and version iteration of the model.

[0040] Online intelligent classification processing module: integrates the pre-trained PTV3 model and the automated processing pipeline, receives natural language task requests input by the user, automatically parses the task, retrieves data, calls the model, and generates and outputs a structured report.

[0041] Business application scenario generation module: Integrates a scenario-based application generation engine to convert the classified structured point cloud data into analysis reports or decision-making suggestions that directly serve power transmission business scenarios (such as tree obstruction reports, live-line working plans, construction acceptance reports, etc.).

[0042] Model self-optimization and sample library iteration unit: used to collect false positive and false negative cases in real-world applications, build positive and negative sample libraries, and periodically trigger incremental training of the model to achieve continuous optimization and self-evolution of model performance.

[0043] System Interface Unit: Provides standardized data and function interfaces for integration with external systems (such as Production Management System (PMS) and Geographic Information System (GIS).

[0044] The PTV3 model, primarily tailored to the needs of power transmission line scenarios, utilizes a Point Transformer-based point cloud classification algorithm to achieve fast and accurate automatic classification of power transmission line electrical clouds. This Point Transformer model is as follows: Figure 2 As shown, the overall structure is a typical U-Net, mainly composed of a point transformer layer, a transition down layer, and a transition up layer. The transition down layer performs downsampling, primarily through farthest point sampling and KNN. Specifically, it uses FPS to find several center points in the point cloud data, finds k neighbors for each center point, and aggregates the information of these k neighbors using maxpooling to achieve downsampling. The transition up layer performs upsampling, implemented in a manner similar to image processing. It first performs upsampling through trilinear interpolation, and then aggregates the features corresponding to the decoder and encoder layers through skip connections. The Point Transformer layer is the core of this model, mainly defining the point cloud self-attention calculation method, which is basically the same as the image self-attention calculation method, except that position encoding is added to α.

[0045] in, and Represents a set of points in the input point cloud. , , , represents the feature extraction module that uses a linear layer or an MLP layer, i.e., the QKV generation module in self-attention; Represents position code, This represents an MLP convolutional layer, which contains two linear layers and one ReLU activation layer; This represents normalization; the softmax function is used here.

[0046] Position encoding in point clouds differs from position encoding in NLP or images. Position encoding in NLP or images is typically constructed manually using operations such as sin and cos, while point cloud data already contains three-dimensional coordinates in its input. Therefore, position encoding can be represented as:

[0047] in, and Represents two points in a point cloud. This represents an MLP consisting of two linear layers and one ReLU activation layer, obtained through training.

[0048] Therefore, its point transformer layer can be represented as follows: Figure 3 Adding residual links constitutes the pointtransformer block, such as Figure 4 As shown.

[0049] In this way, Point Transformer continues the U-Net structure, introduces the idea of ​​self-attentive, constructs an operation similar to local attention, expands the receptive field with downsampling, and then gradually fuses features for upsampling to output classification results.

[0050] Based on PTv1, this invention proposes a new group weighted encoding layer by combining group vector attention and learnable weighted encoding. It enhances positional information through an additional position encoder and designs a new lightweight partitioned pooling method, PTv2. Figure 5 The differences between the two methods are illustrated, namely the main optimization methods mentioned above: group vector attention, Position Encoding Multiplexer, and Partition-based Pooling. Group vector attention divides the channels of the input vector v into g groups on average. The weight encoding layer outputs a group attention vector with g (1 <= g <= c) channels. Within the same attention group, the channels of v share the same scalar attention weights. The formula is expressed as:

[0051] Where γ is the relation function, ω is the group weight encoding, and the following is the group attention.

[0052] Position Encoding Multipler: In the transformer module, spatial information is obtained by adding pi–pj to the relation vector as a bias. Since attention in PTv1 is limited by generalization, PTv2 designs grouped attention to reduce overfitting and increase generalization ability. In this case, a multiplier bias is added to the relation vector to enhance position encoding, focusing on learning complex point cloud positional relationships. Figure 5 As shown on the left, the formula is expressed as:

[0053] in, This represents relational operations, specifically subtraction in this case. This represents a learnable weighted encoding used to weight v; and This represents the MLP position encoding function, which calculates the relative positional relationship between two points. This is the matrix dot product operation.

[0054] Partition-based pooling, a traditional pooling method, such as... Figure 5 The top right corner uses the farthest point for sampling and uses neighbor queries to aggregate adjacent points to achieve upsampling. In this pooling process, the point density and overlapping area of ​​each query set are uncontrollable, so the query sets of points are not spatially aligned.

[0055] Partition pooling divides the point cloud into non-overlapping partitions in space. For example... Figure 5 As shown in the lower right, a grid is used to divide the point cloud space, pooling is performed in each grid, and upsampling is performed using linear interpolation.

[0056] Serialized attention, evolved from window attention, is further defined in PTv3 as patch attention. This is a mechanism that groups points into non-overlapping patches and performs attention within each individual patch. The effectiveness of patch attention depends on two main design elements: patch grouping and patch interaction.

[0057] Patch grouping, based on the serialized point cloud, groups points into different patches according to certain rules. Since the point cloud has already been serialized and encoded using spatial filling curves (such as Z-order curves and Hilbert curves), the grouping process only needs to divide these points along the serialization order. PTv3 has various serialization modes, such as the standard Z-order, Hilbert curves, and their variants, trans-Z-order, trans-Hilbert, etc. Different patch grouping modes can be derived from these serialization modes. Combining serialization modes, patch grouping aims to expand the receptive field of the attention mechanism in 3D space as the patch size increases, while still preserving spatial neighborhood relationships within feasible limits.

[0058] Figure 6 The specific process of patchgrouping involves: a) reordering the point cloud according to the order derived from a specific serialization pattern; and b) filling the point cloud sequence by borrowing points from adjacent patches to ensure it is divisible by the specified patch size. The benefits and advantages of using patchgrouping are: Expanding the receptive field: By grouping patches appropriately, the attention mechanism can capture feature information of point clouds in a larger spatial range. Compared with traditional methods, the receptive field can be expanded from a small range to 1024 points, enabling the model to better understand long-distance dependencies and global structure in point cloud data.

[0059] Improved efficiency and scalability: While this grouping method may sacrifice some neighborhood search accuracy, it yields a significant improvement in efficiency and scalability. The attention mechanism can reweight the points within the patch, thereby compensating to some extent for the accuracy loss caused by grouping, making the model more efficient when processing large-scale point cloud data.

[0060] Preserving spatial neighborhood relationships: While expanding the receptive field, patch grouping can also preserve the spatial neighborhood relationships between points within a certain range. This allows the model to capture global information without ignoring local spatial features, which helps improve the model's ability to understand and analyze complex point cloud scenes.

[0061] It is evident that grouping operations have resulted in faster response times and higher metrics in point cloud semantic classification tasks.

[0062] In summary, this invention achieves centralized data management: it centralizes the storage of 214TB of scattered data, reduces retrieval efficiency from 1 hour to 27 seconds, supports historical data fusion and comparison, and completely solves the problem of data silos.

[0063] Meanwhile, it boasts high classification accuracy and strong generalization ability: The PTV3 model optimized for power scenarios significantly improves the classification accuracy (key component recognition rate >95%) for small components (insulators, drain wires) and long-distance targets (conductors) through sequential attention, patch interaction, and targeted data augmentation, as well as its generalization ability under different terrains and tower types.

[0064] Furthermore, the processing efficiency has been revolutionized: the fully online and automated pipeline design has reduced the processing time for 10 levels of point cloud from "2 person-days" to "1 person-hour", improving efficiency by more than 56%, and 14 out of 16 key processes have been fully automated.

[0065] Furthermore, its business empowerment is direct and widespread: the system not only outputs classification tags but also directly drives applications in over 15 business scenarios. In power grid construction, the output of survey reports has been shortened from several days to within one hour; in construction acceptance, the defect detection rate has increased by 60%, and the acceptance time has been reduced by more than 100%; in operation and maintenance, the number of line trips due to tree and bamboo obstructions has decreased by 30%; and in emergency repairs, the calculation efficiency of live-line working schemes has increased by 300%, assisting in the completion of 17 live-line operations.

[0066] Furthermore, it achieves systematization and sustainable evolution: it has built a complete closed-loop system from sample library construction, model training, automated processing to scenario application, and supports continuous optimization of the sample library and model through the "error collection" mechanism, possessing self-evolution capabilities.

[0067] In summary, this invention provides a high-precision, high-efficiency, highly intelligent, and self-evolving overall solution for automatic classification of point clouds of overhead transmission lines. It effectively solves the core pain points of point cloud data management, slow processing, and shallow application in the power industry, and provides key technical support for the digital and intelligent transformation of the power grid.

[0068] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0070] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0071] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0072] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0073] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0074] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0075] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0076] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. An automatic point cloud classification method for overhead transmission line equipment based on PTV3, characterized in that, The method includes: Construct a standardized point cloud sample library covering multiple voltage levels, terrain types, and tower types; The PTV3 deep learning model based on sequential attention and patch interaction mechanism was used to train the sample database. During the training process, a customized data augmentation strategy and composite loss function were adopted for the characteristics of power equipment to obtain a power-specific point cloud classification model. The trained model is integrated into the automated processing pipeline to achieve intelligent processing of point cloud data throughout the entire process, including uploading, automatic classification, result verification, and structured output. Based on high-precision classification results, various business application scenarios such as tree obstacle analysis, working condition simulation, and construction acceptance are dynamically combined and generated.

2. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The construction of a standardized point cloud sample library covering multiple voltage levels, terrain types, and tower types includes: Collect raw point cloud data and perform data cleaning, including format unification conversion and filtering to remove noise. The cleaned data is structured and segmented according to the span of the transmission line tower. Each data segment contains one tower and its front and rear passages. The point cloud data after the data is cut is finely labeled according to at least 15 categories, including conductors, towers, ground wires, insulators, drain wires, vegetation, ground, buildings, roads, temporary buildings, overhead crossover lines, main grid under crossover lines, distribution grid under crossover lines, and noise. The voltage level, terrain, scene, tower type and collection time of each sample are recorded.

3. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The optimization of the PTV3 deep learning model includes: The input point cloud is serialized and encoded using a space-filling curve, mapping the disordered point cloud into a one-dimensional ordered sequence. The serialized point cloud is grouped into non-overlapping patches, and self-attention computation is performed within each patch. By employing patch interaction strategies such as extended displacement, shift patching, or shift sequence, information is exchanged between different patches, expanding the model's receptive field from local to global to capture continuous features of long-distance targets.

4. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The customized data augmentation strategies include: During the model training phase, voxelization downsampling is used to control the amount of data, and random rotation, scaling, flipping, and coordinate jitter are used to enhance the model's generalization ability. For categories with small sample sizes, such as buildings and crossroads, manual cropping and copying / pasting are performed into the training set to alleviate the class imbalance problem. During the model inference stage, multi-angle ensemble inference is adopted to perform rotation predictions on the input point cloud from multiple angles and fuse the results to improve robustness to occlusion and complex overlapping scenes.

5. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The composite loss function includes: Cross-entropy loss is used as the basic classification loss to handle the class probability distribution; We introduce structured Lovasz loss as an auxiliary loss to directly optimize the IoU metric, thereby alleviating the class imbalance problem and improving the segmentation accuracy of small objects.

6. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The automated processing pipeline is built on the Workflow engine, which decomposes the point cloud classification task into multiple automated steps executed sequentially, including automatic pole positioning and archiving, automatic model classification, scene analysis and report generation, and supports light-weight manual interactive verification at specific steps, so as to realize the online and automated processing of the entire point cloud process.

7. The automatic point cloud classification method for overhead transmission line equipment based on PTV3 according to claim 1, characterized in that, The dynamically combined business application scenarios include: Based on the classified point cloud data labels, the Euclidean clustering algorithm is used to combine multiple features such as clearance distance and ground distance to generate at least one customized application in the following categories: high wind condition simulation, live-line working distance analysis, construction acceptance data verification, special working environment simulation, line hidden danger early warning and emergency repair plan generation.

8. An automatic point cloud classification system for overhead transmission line equipment based on PTV3, implementing the method of any one of claims 1-7, characterized in that, The system includes: The sample library management module is used to store and manage the standardized transmission line point cloud sample library, and supports data versioning, tagging management and fast retrieval; The PTV3 model training and optimization module is used to perform model training, parameter tuning, and version iteration. The online intelligent classification and processing module integrates the trained PTV3 model and the automated processing pipeline. It is used to receive natural language task requests input by users, automatically parse tasks, retrieve data, call models, generate and output structured reports. The business application scenario generation module integrates a scenario-based application generation engine, which is used to convert classified structured point cloud data into analysis reports or decision-making suggestions that directly serve power transmission business scenarios.

9. The automatic point cloud classification system for overhead transmission line equipment based on PTV3 according to claim 8, characterized in that, The system also includes an intensive data management module, which is used to centrally store scattered point cloud data on an online server, breaking down physical data barriers and supporting the fusion and comparison of historical data and intelligent data query functions.

10. The automatic point cloud classification system for overhead transmission line equipment based on PTV3 according to claim 8, characterized in that, The system also includes a model self-optimization and sample library iteration unit, which is used to collect false detection and false negative cases in practical applications, build a positive and negative sample library, and periodically trigger incremental training of the model to achieve continuous optimization and self-evolution of model performance.

Images