Power equipment point cloud interpretable segmentation method and system based on rule guide hierarchical clustering
By using a rule-guided hierarchical clustering method, combined with geometric clustering, semantic segmentation, and rule base optimization, the problem of high-precision equipment segmentation in complex power scenarios was solved, achieving transparent and interpretable high-precision segmentation and domain knowledge fusion.
Patent Information
- Application Number
- CN202511863519.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-11
- Publication Date
- 2026-02-27
AI Technical Summary
Existing technologies struggle to achieve high-precision device segmentation in complex power scenarios, while ensuring a transparent and explainable decision-making process and effectively embedding domain knowledge.
A rule-guided hierarchical clustering method is adopted, which performs geometric clustering through an adaptive similarity metric function, generates a supervoxel set by combining an improved seed region growing algorithm, performs initial semantic segmentation using an improved distance-aware Point Transformer network, and constructs a structured power equipment rule base for iterative correction and optimization.
High-precision segmentation was achieved in complex occlusion and adhesion scenarios. The decision-making process is transparent and interpretable. It incorporates domain knowledge and improves the credibility of the results in high-risk power applications.
Smart Images
Figure CN121582585A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of computer vision and digital power, and particularly relates to a power equipment point cloud explainable segmentation method and system based on rule-guided hierarchical clustering. BACKGROUND
[0002] Power plants (such as substations and distribution station buildings) are the core hubs of the power grid, and the intelligent inspection and digital management of their equipment are the key to building a new power system. Three-dimensional laser scanning technology can quickly obtain high-precision point cloud data of power scenes, providing a data basis for equipment contour reconstruction, state monitoring, and digital twin construction. However, the equipment in the plant is dense and complex in structure, and the point cloud data generally has serious occlusion and adhesion phenomena, making automatic and high-precision equipment segmentation a long-term challenge.
[0003] Current point cloud segmentation methods mainly include traditional geometric methods and deep learning methods. Traditional methods (such as Euclidean clustering and region growing) heavily rely on manually set thresholds and lack generalization ability in the face of diverse equipment forms and complex spatial relationships, easily leading to serious under-segmentation or over-segmentation. Deep learning methods (such as PointNet++ and PointTransformer) can learn semantic features through data-driven methods, but their "black box" nature makes the decision-making process opaque, and the segmentation results may violate the inherent physical structure and engineering specifications of the equipment (such as equipment size and installation spacing), making it difficult to effectively integrate rich domain prior knowledge, which is unacceptable in the high-risk and high-reliability power industry applications.
[0004] Therefore, the existing technology lacks a method that can achieve high-precision segmentation in complex power scenes while ensuring a transparent and explainable decision-making process and effectively embedding domain knowledge. SUMMARY
[0005] The purpose of the present application is to provide a power equipment point cloud explainable segmentation method and system based on rule-guided hierarchical clustering that can improve point cloud segmentation accuracy.
[0006] Technical solution: The power equipment point cloud explainable segmentation method based on rule-guided hierarchical clustering comprises:
[0007] S1: Input the original three-dimensional point cloud data of the power plant, calculate the local geometric features of each point, and based on spatial proximity, surface smoothness, and geometric consistency, construct an adaptive similarity measurement function, combine an improved seed region growing algorithm to cluster the point cloud, and generate an over-segmented voxel set;
[0008] S2: input the super voxel set into an improved distance-aware Point Transformer network, output the initial probability distribution of each super voxel belonging to each type of power equipment, and complete the semantic initial segmentation; the improved distance-aware Point Transformer network generates enhanced features for each super voxel by introducing a position encoding function of spatial distance information, and enhances the perception ability of the self-attention mechanism to the relative position between power equipment;
[0009] S3: construct a structured power equipment rule base, activate the corresponding rule subset based on the semantic initial segmentation result using the lazy loading method; calculate the conflict score of the initial segmentation result and the rule subset, and iteratively correct and optimize the segmentation boundary or class label according to the rule subset, record the decision log, and output the final explainable segmentation result.
[0010] Further, the adaptive similarity measurement function in step S1 is:
[0011]
[0012] wherein, and are points in the point cloud, and are the local curvatures of the points, and are the normal vectors of the points, is a distance scale parameter adaptively calculated according to the density of the point neighborhood, is a weight coefficient and satisfies .
[0013] Further, the enhanced feature in step S2 is calculated as follows:
[0014]
[0015] wherein, , are query, key and value vectors respectively, is the Euclidean distance between the super voxel and the centroid, is a learnable position encoding function, is a core weight calculation function used to convert all original scores into a probability distribution with a sum of 1, the neighbor with a high score has a large corresponding weight, and the proportion is high in the final aggregation; is a scaling factor used to stabilize the gradient when the dot product result is too large, to prevent the function from entering the saturation zone; represents the matrix transposition operation, is the super voxel The neighborhood set.
[0016] Furthermore, the structured power equipment rule base mentioned in step S3 includes size rules, topology rules, and functional rules extracted based on national standards, industry regulations, and expert experience; the lazy loading method means that only after a specific type of equipment is initially identified, a subset of rules related to that type of equipment is loaded and calculated.
[0017] The method for calculating the conflict score between the initial segmentation result and the rule subset is as follows:
[0018]
[0019] in, For segmentation objects Conflict score, For the subset of rules that are activated, It is a rule Confidence weights It is the object of measurement Violation of rules A function of degree.
[0020] Based on the same inventive concept, this invention also provides an interpretable segmentation system for power equipment point clouds based on rule-guided hierarchical clustering, comprising:
[0021] The geometric clustering module is used to input the original 3D point cloud data of power plants, calculate the local geometric features of each point, and construct an adaptive similarity metric function based on spatial proximity, surface smoothness and geometric consistency. Combined with an improved seed region growing algorithm, the point cloud is clustered to generate an oversegmented supervoxel set.
[0022] The semantic segmentation module is used to input the set of supervoxels into the improved distance-aware Point Transformer network and output the initial probability distribution of each supervoxel belonging to various types of power equipment, thus completing the initial semantic segmentation. The improved distance-aware Point Transformer network generates enhanced features for each supervoxel by introducing a position encoding function with spatial distance information, thereby enhancing the self-attention mechanism's ability to perceive the relative positions between power equipment.
[0023] The optimized output module is used to construct a structured power equipment rule base. Based on the initial semantic segmentation result, the corresponding rule subset is activated using a lazy loading method. The conflict score between the initial segmentation result and the rule subset is calculated, and the segmentation boundary or category label is iteratively corrected and optimized according to the rule subset. At the same time, the decision log is recorded, and the final interpretable segmentation result is output.
[0024] Furthermore, the adaptive similarity metric function of the geometric clustering module is:
[0025]
[0026] in, and For points in a point cloud, and Let be the local curvature of a point. and Let be the normal vector of the point. The distance scale parameter is adaptively calculated based on the point's neighborhood density. The weighting coefficients are satisfied. .
[0027] Furthermore, the enhanced features described in the semantic segmentation module The calculation is as follows:
[0028]
[0029] in, , These are query, key, and value vectors, respectively. It is a supervoxel and Euclidean distance between the centroids It is a learnable positional encoding function. It is the core weight calculation function, which is used to convert all the raw scores into a probability distribution that sums to 1. The neighbors with higher scores have higher weights and thus a higher proportion in the final aggregation. This is a scaling factor used to stabilize the gradient when the dot product result is too large, preventing... The function enters the saturation region; This represents the matrix transpose operation. It is a supervoxel The neighborhood set.
[0030] Furthermore, the structured power equipment rule base of the optimized output module includes size rules, topology rules, and functional rules extracted based on national standards, industry regulations, and expert experience; the lazy loading method means that only after a specific type of equipment is initially identified, a subset of rules related to that type of equipment is loaded and calculated.
[0031] The method for calculating the conflict score between the initial segmentation result and the rule subset is as follows:
[0032]
[0033] in, For segmentation objects Conflict score, is a subset of rules activated, is a confidence weight of the rule is a function measuring the degree of violation of the rule by the object.
[0034] Based on the same inventive concept, the present application also provides a computing device, comprising one or more processors, one or more memories, and one or more programs stored in the memories and configured for execution by the processors, which, when loaded into the processors, implement the steps of the rule-guided hierarchical clustering-based power equipment point cloud interpretable segmentation method according to any one of the above.
[0035] Based on the same inventive concept, the present application also provides a storage medium storing a computer program, the computer program comprising program instructions which, when executed by a processor, cause the processor to perform the steps of the rule-guided hierarchical clustering-based power equipment point cloud interpretable segmentation method according to any one of the above.
[0036] Advantages: Compared with the prior art, the present application processes point clouds layer by layer through the three-level cooperation of "geometry-semantic-rule", suppresses noise through geometric clustering, understands high-level features through semantic network, and performs final calibration through a rule base, thereby achieving leading segmentation accuracy in complex scenes with serious occlusion and adhesion; by explicitly formalizing domain knowledge as rules and dynamically checking, the "black box" output of deep learning is converted into a transparent and auditable decision-making process; the system can explicitly inform the user why a certain correction is made and on which rule, greatly improving the credibility of the results in high-risk power applications; the structured rule base constructed by the present application systematically encapsulates prior knowledge in the power field and realizes efficient calculation through a lazy loading mechanism, so that expert experience can be seamlessly and low-costly integrated into the automated processing flow; the method has been verified on real power plant data and has the ability to be deployed on the edge side, and can directly support practical businesses such as intelligent inspection and digital twin construction, and has broad engineering application prospects. BRIEF DESCRIPTION OF DRAWINGS
[0037] Figure 1 is a flowchart of the method of an embodiment of the present application;
[0038] Figure 2 is a structural diagram of the improved distance-aware Point Transformer network of an embodiment of the present application. DETAILED DESCRIPTION
[0039] In order for those skilled in the art to better understand the technical scheme of the present application, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.
[0040] The rule-based hierarchical clustering power equipment point cloud explainable segmentation method of the present embodiment, as shown in Embodiment 1, Figure 1 includes:
[0041] S1: input the original three-dimensional point cloud data of the power station, calculate the local geometric features of each point, and construct an adaptive similarity measurement function based on spatial proximity, surface smoothness and geometric consistency, combine the improved seed region growing algorithm to cluster the point cloud, and generate an over-segmented super voxel set;
[0042] S2: input the super voxel set into the improved distance-aware Point Transformer network, output the initial probability distribution of each super voxel belonging to each type of power equipment, and complete the semantic primary segmentation; the improved distance-aware Point Transformer network generates enhanced features for each super voxel by introducing a position encoding function of spatial distance information, and enhances the self-attention mechanism to perceive the relative position between power equipment;
[0043] S3: construct a structured power equipment rule base, activate the corresponding rule subset based on the semantic primary segmentation result using the lazy loading method; calculate the conflict score of the primary segmentation result and the rule subset, and iteratively correct and optimize the segmentation boundary or class label according to the rule subset, record the decision log, and output the final explainable segmentation result.
[0044] Specifically, in step S1, the original three-dimensional point cloud data of the power station is input , and the local geometric features of each point , including the normal vector and the curvature , are calculated. An adaptive similarity measurement function is constructed, which integrates spatial distance, curvature consistency and normal vector consistency. Then an improved seed region growing algorithm is used for clustering, which determines the attribution of points according to the comparison between the average similarity of points and class clusters and the dynamic threshold , and finally outputs a set of well-structured over-segmented super voxels . This step provides a less noisy and more structured input for subsequent semantic segmentation.
[0045] The adaptive similarity measurement function is:
[0046] ;
[0047] and For points in a point cloud, and Let be the local curvature of a point. and Let be the normal vector of the point. The distance scale parameter is adaptively calculated based on the point's neighborhood density. The weighting coefficients are satisfied. .
[0048] The clustering algorithm in this step determines whether the first point should be added to the currently growing second point cluster (supervoxel). Each decision calls the function built in the second step to calculate the similarity score between the first point and representative points within the second point cluster. Only when the score is higher than a set threshold is the first point included. Through countless decisions, the originally scattered points are finally aggregated into meaningful 'supervoxel' blocks.
[0049] In step S2, the supervoxel set output in step S1 is... As input, the data is fed into an improved PointTransformer encoder-decoder network for semantic segmentation. The improvement of this invention lies in the introduction of a distance-aware self-attention mechanism. Specifically, when calculating attention weights, in addition to the dot product of the query vector and the key vector, an Euclidean distance between supervoxels is added. Learnable positional encoding This allows the network to not only focus on semantic similarity when aggregating features, but also to explicitly perceive spatial structural information, thus more effectively distinguishing spatially close but semantically different devices (such as stacked insulator strings and connecting fittings). The network ultimately outputs the initial probability distribution of each supervoxel belonging to various types of power equipment (such as transformers, circuit breakers, disconnectors, etc.).
[0050] Improved distance-aware Point Transformer network such as Figure 2 As shown, the enhanced feature f̂ k The calculation is as follows:
[0051]
[0052] in, , , These are query, key, and value vectors, respectively. It is a supervoxel and Euclidean distance between the centroids It is a learnable positional encoding function. is the core weight computation function, which converts all raw scores into a probability distribution (weights) with sum 1. Neighbors with higher scores will have larger weights, and thus more influence in the final aggregation. is a scaling factor to stabilize the gradient when the dot product result is too large, preventing the function from entering the saturation region. represents the matrix transpose operation, is the neighborhood set of a super-voxel .
[0053] Within the neural network of this step, is a core computation step that bridges the previous and the following steps. The purpose is to generate an enhanced feature descriptor for each point cloud block (super-voxel) that not only contains its own semantics but also encodes its spatial relationship with surrounding components. This enhanced feature is the direct basis for the subsequent classification decision of the network. Through this formula, the improved distance perception mechanism is realized, allowing the network to utilize both shape information and spatial position information for decision-making, thus achieving more accurate segmentation than ordinary networks.
[0054] Step S3 is the core innovation of the present application, aiming to convert domain knowledge into computable and executable optimization power. First, a structured power equipment rule base R is constructed, which contains multiple rules (e.g., 287 rules) extracted from national standards (GB / T), industry regulations, and expert experience, mainly divided into:
[0055] Size rules R_size: define the reasonable size range of equipment (e.g., "The height of current transformers is usually between 1.5m and 3.0m").
[0056] Topology rules R_topo: define the connection and spatial relationship between devices (e.g., "The moving contact and static contact of disconnectors should maintain a minimum distance X in the open state").
[0057] Function rules R_func: based on the constraints of device functions (e.g., "The number of insulator discs should be odd").
[0058] The optimization process is an iterative "check-correction" cycle:
[0059] Rule triggering: based on the initial segmentation result of step S2, an inert loading mechanism is adopted, only activating the rule subset R_active ⊆ R related to the current recognized device type, greatly reducing the computational overhead.
[0060] Conflict detection and quantification: for each segmented object , calculate its conflict score with all related rules:
[0061] wherein, is the activated subset of rules, is a rule with a confidence weight, is a function measuring the degree of rule violation by the object. The function is a quantitative process based on geometric calculation that connects abstract rules with concrete point cloud data. Taking the 'isolated open closed state rule' as an example, the calculation of the function is a complete, automated sub-process: 1. Input: two point cloud objects after semantic initial segmentation (labeled as'moving contact' and'static contact'). 2. Processing: The system will call the spatial indexing algorithm (KD-Tree) to efficiently calculate the three-dimensional Euclidean distance between all pairs of points in the two point cloud clusters, and find the minimum value . 3. Judgment: Compare with the minimum safety distance specified in the safety regulations. 4. Output: If , then , indicating compliance with the rule; if , then , output a specific positive number whose value directly reflects the degree of 'danger' or 'violation'.
[0062] Explainable correction: For objects with high conflict scores, the system does not simply reclassify, but gives a correction suggestion based on the specific rule triggered. For example, if an object identified as an 'insulator' does not meet the rule in the number of discs, the algorithm will locally adjust the segmentation boundary based on the contour information, and generate a decision log such as 'the segmentation boundary of the object has been optimized according to the rule '. This process ensures that the final result not only has high accuracy, but is also physically reasonable, in line with engineering specifications, and the entire decision chain is traceable and auditable.
[0063] Embodiment 2: This embodiment is implemented on a workstation equipped with an NVIDIA RTX 4090 GPU, and the software is based on the PyTorch deep learning framework. The rule base is stored in a structured format using JSON. The rule-guided hierarchical clustering-based power equipment point cloud explainable segmentation method of this embodiment is as follows:
[0064] (1) Data preparation and preprocessing: Obtain power station point cloud data (LAS format) from ground-based or unmanned aerial vehicle LiDAR, and perform preprocessing such as denoising and downsampling (optional).
[0065] (2) Perform geometric perception spatial hierarchical clustering (S100):
[0066] ① Compute the normal vector and curvature for each point in the point cloud P.
[0067] ② Set the weight parameters in the adaptive similarity measure function .
[0068] ③ Execute the improved region growing algorithm, with a dynamic threshold T_cluster set to an initial value based on scene complexity and fine-tuned through experiments. Output the set of super voxels {V_k}.
[0069] (3) Perform semantic initial segmentation (S200):
[0070] ① Load the pre-trained distance-aware Point Transformer model. This model has been trained on the PowerGrid3D-2025 dataset.
[0071] ② Input the super voxels and their average features into the network, perform forward propagation, and obtain the initial semantic label for each super voxel.
[0072] (4) Perform rule-guided optimization (S300):
[0073] ① The system reads the structured rule base (JSON file).
[0074] ② According to the device types identified by the initial segmentation (such as "circuit breaker" and "insulator"), load the corresponding rule subset .
[0075] ③ Traverse each segmented object , calculate its conflict score with each rule in . Set a conflict threshold , when , trigger the correction process.
[0076] ④ Correction process: for example, for objects that violate size rules, the system will re-evaluate their boundaries or adjust their class labels based on the geometric characteristics of the point cloud cluster (such as the principal axis size obtained by PCA analysis). All correction actions and the rule IDs they are based on are recorded in the decision log.
[0077] (5) Result output:
[0078] ① The output module visualizes the final optimized segmented point cloud (different devices are displayed in different colors).
[0079] ② At the same time, a text format decision log can be provided to the user, detailing the rules triggered during optimization and the operations performed.
[0080] Embodiment 3, based on the same inventive concept, also provides a power equipment point cloud explainable segmentation system based on rule-guided hierarchical clustering, comprising:
[0081] a geometric clustering module for inputting raw three-dimensional point cloud data of a power station, calculating local geometric features of each point, and constructing an adaptive similarity measurement function based on spatial proximity, surface smoothness and geometric consistency, combining an improved seed region growing algorithm to cluster the point cloud and generate an over-segmented super voxel set;
[0082] a semantic segmentation module for inputting the super voxel set into an improved distance-aware Point Transformer network to output an initial probability distribution of each super voxel belonging to each type of power equipment, and completing semantic initial segmentation; the improved distance-aware Point Transformer network generates enhanced features for each super voxel by introducing a position encoding function of spatial distance information, and enhances the self-attention mechanism's ability to perceive the relative position between power equipment;
[0083] an optimization output module for constructing a structured power equipment rule base, activating a corresponding rule subset based on the semantic initial segmentation result using a lazy loading method; calculating the conflict score of the initial segmentation result and the rule subset, and iteratively correcting and optimizing the segmentation boundary or class label according to the rule subset, while recording the decision log and outputting the final explainable segmentation result.
[0084] Further, the adaptive similarity measurement function of the geometric clustering module is:
[0085]
[0086] wherein, and are points in the point cloud, and are the local curvatures of the points, and are the normal vectors of the points, is a distance scale parameter adaptively calculated according to the density of the point neighborhood, is a weight coefficient and satisfies .
[0087] Further, the enhanced features in the semantic segmentation module are calculated as follows:
[0088]
[0089] wherein, , are query, key and value vectors, respectively, is a super voxel with the Euclidean distance between the centroids, is a learnable positional encoding function, is a core weight computation function that converts all raw scores into a probability distribution with sum 1, neighbors with higher scores have larger weights and thus higher importance in the final aggregation; is a scaling factor to stabilize the gradient when the dot product result is too large, preventing the function from entering the saturation region; denotes the matrix transpose operation, is a super voxel 's neighborhood set.
[0090] Further, the structured power equipment rule base of the optimization output module comprises size rules, topology rules and function rules refined based on national standards, industry regulations and expert experience; the lazy loading manner refers to: only when a specific type of equipment is preliminarily identified, a rule subset related to the type of equipment is loaded and calculated;
[0091] The method for calculating the conflict score of the initial segmentation result and the rule subset is:
[0092]
[0093] wherein, is the conflict score of the segmentation object , is the activated rule subset, is the confidence weight of the rule , is a function measuring the degree of violation of the rule by the object .
[0094] Embodiment 4: Based on the same inventive concept, the embodiment further provides a computing device, comprising: one or more processors, one or more memories, and one or more programs stored in the memories and configured to be executed by the processors, the programs, when loaded into the processors, implement the steps of the rule-guided hierarchical clustering-based power equipment point cloud interpretable segmentation method according to any one of the above embodiments.
[0095] Embodiment 5: Based on the same inventive concept, the embodiment further provides a storage medium, the storage medium stores a computer program, the computer program comprises program instructions, the program instructions, when executed by a processor, cause the processor to perform the steps of the rule-guided hierarchical clustering-based power equipment point cloud interpretable segmentation method according to any one of the above embodiments.
Claims
1. An interpretable segmentation method for power equipment point clouds based on rule-guided hierarchical clustering, characterized in that, include: S1: Input the original 3D point cloud data of the power plant, calculate the local geometric features of each point, and construct an adaptive similarity metric function based on spatial proximity, surface smoothness and geometric consistency. Combine the improved seed region growth algorithm to cluster the point cloud and generate an oversegmented supervoxel set. S2: Input the set of supervoxels into the improved distance-aware Point Transformer network, output the initial probability distribution of each supervoxel belonging to various types of power equipment, and complete the initial semantic segmentation; the improved distance-aware Point Transformer network generates enhanced features for each supervoxel by introducing a position encoding function with spatial distance information, thereby enhancing the self-attention mechanism's ability to perceive the relative positions between power equipment. S3: Construct a structured rule base for power equipment. Based on the initial semantic segmentation results, activate the corresponding rule subsets using a lazy loading method. Calculate the conflict score between the initial segmentation results and the rule subsets. Iteratively correct and optimize the segmentation boundaries or category labels based on the rule subsets. Simultaneously, record the decision log and output the final interpretable segmentation results.
2. The interpretable segmentation method for power equipment point clouds based on rule-guided hierarchical clustering according to claim 1, characterized in that, The adaptive similarity metric function mentioned in step S1 is: ； in, and For points in a point cloud, and Let be the local curvature of a point. and Let be the normal vector of the point. The distance scale parameter is adaptively calculated based on the point's neighborhood density. The weighting coefficients are satisfied. .
3. The interpretable segmentation method for power equipment point clouds based on rule-guided hierarchical clustering according to claim 1, characterized in that, The enhanced features described in step S2 The calculation is as follows: ； in, , These are query, key, and value vectors, respectively. It is a supervoxel and Euclidean distance between the centroids It is a learnable positional encoding function. It is the core weight calculation function, which is used to convert all the raw scores into a probability distribution that sums to 1. The neighbors with higher scores have higher weights and thus a higher proportion in the final aggregation. This is a scaling factor used to stabilize the gradient when the dot product result is too large, preventing... The function enters the saturation region; This represents the matrix transpose operation. It is a supervoxel The neighborhood set.
4. The interpretable segmentation method for power equipment point clouds based on rule-guided hierarchical clustering according to claim 1, characterized in that, The structured power equipment rule base mentioned in step S3 includes size rules, topology rules, and functional rules extracted based on national standards, industry regulations, and expert experience; the lazy loading method means that only after a specific type of equipment is initially identified will a subset of rules related to that type of equipment be loaded and calculated. The method for calculating the conflict score between the initial segmentation result and the rule subset is as follows: ； in, For segmentation objects Conflict score, For the subset of rules that are activated, It is a rule Confidence weights It is the object of measurement Violation of rules A function of degree.
5. An interpretable segmentation system for power equipment point clouds based on rule-guided hierarchical clustering, characterized in that, include: The geometric clustering module is used to input the original 3D point cloud data of power plants, calculate the local geometric features of each point, and construct an adaptive similarity metric function based on spatial proximity, surface smoothness and geometric consistency. Combined with an improved seed region growing algorithm, the point cloud is clustered to generate an oversegmented supervoxel set. The semantic segmentation module is used to input the set of supervoxels into the improved distance-aware Point Transformer network and output the initial probability distribution of each supervoxel belonging to various types of power equipment, thus completing the initial semantic segmentation. The improved distance-aware Point Transformer network generates enhanced features for each supervoxel by introducing a position encoding function with spatial distance information, thereby enhancing the self-attention mechanism's ability to perceive the relative positions between power equipment. The optimized output module is used to construct a structured power equipment rule base. Based on the initial semantic segmentation result, the corresponding rule subset is activated using a lazy loading method. The conflict score between the initial segmentation result and the rule subset is calculated, and the segmentation boundary or category label is iteratively corrected and optimized according to the rule subset. At the same time, the decision log is recorded, and the final interpretable segmentation result is output.
6. The interpretable segmentation system for power equipment point clouds based on rule-guided hierarchical clustering according to claim 5, characterized in that, The adaptive similarity metric function described in the geometric clustering module is: ； in, and For points in a point cloud, and Let be the local curvature of a point. and Let be the normal vector of the point. The distance scale parameter is adaptively calculated based on the point's neighborhood density. The weighting coefficients are satisfied. .
7. The interpretable segmentation system for power equipment point clouds based on rule-guided hierarchical clustering according to claim 5, characterized in that, Enhanced features described in the semantic segmentation module The calculation is as follows: ； in, , These are query, key, and value vectors, respectively. It is a supervoxel and Euclidean distance between the centroids It is a learnable positional encoding function. It is the core weight calculation function, which is used to convert all the raw scores into a probability distribution that sums to 1. The neighbors with higher scores have higher weights and thus a higher proportion in the final aggregation. This is a scaling factor used to stabilize the gradient when the dot product result is too large, preventing... The function enters the saturation region; This represents the matrix transpose operation. It is a supervoxel The neighborhood set.
8. The interpretable segmentation system for power equipment point clouds based on rule-guided hierarchical clustering according to claim 5, characterized in that, The optimized output module includes a structured power equipment rule base containing size rules, topology rules, and functional rules derived from national standards, industry regulations, and expert experience. The lazy loading method means that a subset of rules related to a specific type of equipment is loaded and calculated only after a specific type of equipment has been initially identified. The method for calculating the conflict score between the initial segmentation result and the rule subset is as follows: ； in, For segmentation objects Conflict score, For the subset of rules that are activated, It is a rule Confidence weights It is the object of measurement Violation of rules A function of degree.
9. A computing device, characterized in that, include: One or more processors, one or more memories, and one or more programs, said programs being stored in the memory and configured to be executed by the processors, said programs being loaded onto the processors to implement the steps of the rule-guided hierarchical clustering-based interpretable segmentation method for power equipment point clouds according to any one of claims 1 to 4.
10. A storage medium, characterized in that, The storage medium stores a computer program, which includes program instructions that, when executed by a processor, cause the processor to perform the steps of the rule-guided hierarchical clustering-based interpretable segmentation method for power equipment point clouds according to any one of claims 1 to 4.