Anomaly detection method and system for pumped storage power station construction based on UAV image analysis

Abstract

The present invention discloses a method and system for detecting anomalies in the construction of pumped storage power stations based on drone image analysis, which relates to the field of engineering construction monitoring technology, including constructing a hybrid knowledge model; planning the acquisition path of the construction area, acquiring multi-view images, processing multimodal sensor data, and generating a construction site feature tensor; inputting the construction site feature tensor into a densely connected convolutional network for processing, combining a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities to establish a four-dimensional construction dynamic model; using a two-layer contrastive learning network to calculate the feature distance between the hybrid knowledge model and the four-dimensional construction dynamic model, calibrating the abnormal area and its uncertainty; performing causal reasoning and tracing analysis on the detected abnormal area, and screening the optimal compensation solution through a case-enhanced Bayesian analyzer. The present invention realizes the effective integration of CAD drawings and expert knowledge through a deep cross-modal knowledge distillation network, providing a more comprehensive and standardized knowledge foundation for anomaly detection.

Description

Technical Field

[0001] The present invention relates to the technical field of engineering construction monitoring, and in particular to a method for detecting abnormalities in the construction of a pumped-storage power station based on unmanned aerial vehicle (UAV) image analysis. Background Art

[0002] Pumped-storage power stations are a crucial component of clean energy systems, and their construction quality is directly linked to their safe operation. Currently, commonly used construction anomaly detection systems primarily employ rule-based expert systems or simple machine learning models. These approaches have significant limitations when processing high-dimensional, multimodal construction scenario data. Existing technologies struggle to accurately identify anomalies and generate effective compensation plans, particularly when faced with the uncertainties of construction sites, the complexity of the construction process, and the diversity of safety risks. Furthermore, due to the lack of intelligent decision-making support mechanisms, anomaly handling often relies on manual judgment, which can easily lead to processing delays and secondary risks. Summary of the Invention

[0003] In view of the problems of the existing pumped storage power station construction anomaly detection system, such as insufficient knowledge fusion, limited data collection, lack of dynamic analysis and weak decision support, the present invention is proposed.

[0004] Therefore, the problem to be solved by the present invention is how to construct a pumped storage power station construction anomaly detection method that integrates expert knowledge, multi-source data acquisition, dynamic modeling analysis and intelligent decision support to achieve intelligent supervision and anomaly handling of the construction process.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] In the first aspect, an embodiment of the present invention provides a method for detecting anomalies in the construction of a pumped-storage power station based on drone image analysis, which includes using a deep cross-modal knowledge distillation network to map the CAD drawings of the pumped-storage power station and the expert knowledge base into graph-structured training samples, performing semantic segmentation and annotation steps, and outputting a hybrid knowledge model with specification constraints; using a terrain-adaptive multi-drone collaborative strategy to plan the construction area acquisition path, acquiring multi-perspective images based on a multimodal sensor array, processing the multimodal sensor data through a hierarchical spatiotemporal attention network, and generating a construction site feature tensor; inputting the construction site feature tensor into a densely connected convolutional network for processing, combining a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities to establish a four-dimensional construction dynamic model; using a two-layer contrastive learning network to calculate the feature distance between the hybrid knowledge model and the four-dimensional construction dynamic model, and calibrating the abnormal area and its uncertainty through a safety level adaptive anomaly detector; performing causal reasoning and traceability analysis on the detected abnormal area, generating multiple compensation schemes in combination with multi-agent reinforcement learning, and screening the optimal compensation scheme through a case-enhanced Bayesian analyzer.

[0007] As a preferred solution of the pumped storage power station construction anomaly detection method based on drone image analysis described in the present invention, the hybrid knowledge model is constructed as follows: the pumped storage power station CAD drawings are standardized through an image preprocessing network, and the geometric features, spatial relationships and engineering parameters in the drawings are extracted to form a drawing feature vector; the text knowledge in the expert knowledge base is parsed using a natural language processing model to extract construction specifications, anomaly judgment standards and processing methods, and generate a knowledge feature vector; a deep cross-modal knowledge distillation network is constructed, the drawing feature vector is used as the teacher network input, and the knowledge feature vector is used as the student network input, and the feature representations of the two modalities are aligned through a cross-attention mechanism; the aligned features are mapped to a graph structure space based on a graph neural network, where the graph nodes represent the key components of the construction area, and the graph edges represent the spatial and functional relationships between the components, forming an initial graph structure; the initial graph structure is fine-grainedly annotated using a semantic segmentation network, and a correspondence is established between the construction specification requirements and the physical components to generate an annotated graph structure; the annotated graph structure is input into a specification constraint encoder to extract the temporal dependency and quality control standards of the construction process, and output a specification-constrained hybrid knowledge model.

[0008] As a preferred solution of the pumped storage power station construction anomaly detection method based on drone image analysis described in the present invention, generating a construction site feature tensor includes the following steps: using a laser radar to scan the terrain of the construction area to generate a three-dimensional point cloud map, dividing the construction area into multiple sub-areas through a point cloud segmentation algorithm, and calculating the terrain complexity index of each sub-area; constructing a multi-UAV task allocation model based on the terrain complexity index, using an adaptive particle swarm algorithm to assign UAV formations and flight altitudes to each sub-area, and generating an initial path planning scheme; based on the initial path planning scheme, combining the obstacle distribution and weather conditions in the construction area, generating a collaborative flight path that takes into account safety distances through a trajectory optimization algorithm; equipped with a multimodal sensor array of visible light cameras, infrared cameras, and multispectral cameras, collecting a construction site image sequence along the collaborative flight path and recording the corresponding position and posture information; performing multi-perspective image registration on the construction site image sequence, extracting image features using a deep learning network, and constructing a four-dimensional feature map containing time series information; inputting the four-dimensional feature map into a hierarchical spatiotemporal attention network, and generating a construction site feature tensor through multi-level feature extraction and time series modeling.

[0009] As a preferred solution of the pumped storage power station construction anomaly detection method based on drone image analysis described in the present invention, the invention further comprises the following steps: inputting the construction site feature tensor into a densely connected convolutional network, generating a hierarchical feature map through a self-calibrated multi-scale feature extraction module; constructing a heterogeneous graph structure of the construction scene, treating construction equipment, building components and construction personnel as different types of graph nodes, establishing edge connections based on spatial position relationships and construction process dependencies, and generating a scene relationship graph; performing feature fusion on the hierarchical feature representation and the scene relationship graph, calculating the dynamic correlation weights between nodes through a hierarchical adaptive graph attention mechanism, and generating a temporal scene graph; using a causal reasoning module to analyze the causal relationship between nodes in the temporal scene graph, constructing a causal dependency network of the construction scene, and identifying key influencing factors and propagation paths; inputting the causal dependency network into a gated recurrent unit, performing temporal modeling on key indicators, and generating a state evolution sequence; based on the state evolution sequence, combining information on spatial dimension, time dimension, construction element dimension and quality dimension, establishing a four-dimensional construction dynamic model that describes the entire construction process.

[0010] As a preferred solution of the pumped storage power station construction anomaly detection method based on drone image analysis described in the present invention, the method includes the following steps: constructing a feature-adapted two-layer contrastive learning network, dimensional alignment and feature standardization of the hybrid knowledge model and the four-dimensional construction dynamic model through a preprocessing module, and generating a normalized feature vector using a hierarchical feature extraction and residual preservation mechanism; calculating a semantic similarity matrix for the normalized feature vector, using a multi-granularity cross-attention mechanism to highlight key feature differences, and generating a feature deviation map; constructing a hierarchical safety assessment system based on engineering specifications and historical cases, mapping the feature deviation map to different safety levels, and generating a safety assessment vector; using an adaptive threshold mechanism to dynamically adjust the sensitivity of the anomaly detector according to the safety assessment vector, and determining a multi-scale detection threshold; using the multi-scale detection threshold to perform hierarchical screening on the feature deviation map, locate potential abnormal areas, and output a set of abnormal areas; combining the Monte Carlo sampling method to evaluate the detection confidence of each area in the abnormal area set, and generating an anomaly calibration result including the abnormal location, type, and uncertainty.

[0011] As a preferred solution of the pumped storage power station construction anomaly detection method based on drone image analysis described in the present invention, wherein: screening the optimal compensation solution through a case-enhanced Bayesian analyzer includes the following steps: using a dynamic graph causal attention network to process the anomaly calibration results, establishing a spatiotemporal causal graph of abnormal events, extracting the correlation measurement values between abnormal features and construction elements, and outputting an anomaly propagation matrix; initializing the parameters of the multi-agent interaction model through a hierarchical knowledge distillation architecture, constructing a teacher network containing construction expert rules, setting hierarchical reward indicators, and forming an agent training environment; deploying an adversarial reinforcement learning model to train professional agents, and The domain divides the roles of intelligent agents, executes the competitive game training process, and generates differentiated compensation strategy combinations; builds a graph neural network case retrieval model, encodes historical compensation plans into a time-series graph structure, calculates the case graph matching metric, and extracts empirical feature maps; configures a two-stream Bayesian neural network to process cost and duration indicators and risk reliability indicators respectively, applies attention weight fusion, and outputs a plan evaluation vector; performs combinatorial optimization calculations under uncertainty constraints, jointly considers evaluation scores and resource constraints, and generates a set of compensation plans with emergency plans; sorts the compensation plan set based on the plan evaluation vector, and selects the plan with the highest comprehensive score as the optimal compensation plan.

[0012] As an optimal solution for the pumped storage power station construction anomaly detection method based on drone image analysis of the present invention, the method comprises the following steps: constructing a deep cross-modal knowledge distillation network, comprising the following steps: constructing a multi-head self-attention encoder, performing feature enhancement on the drawing feature vector and the knowledge feature vector respectively, and generating drawing attention features and knowledge attention features; designing a dual-stream cross-attention module, taking the drawing attention features as the query vector of the teacher network, and the knowledge attention features as the key-value pairs, and calculating the cross-modal attention score matrix; establishing a feature distillation loss function, comprising a knowledge distillation loss term and a feature consistency loss term, and adopting an adaptive temperature adjustment strategy to adjust the knowledge distillation process; constructing a feature fusion module, fusing the features output by the teacher network and the student network through a residual connection method, adopting a multi-layer feature pyramid network to extract feature representations of different scales, and adaptively selecting the optimal feature combination through the attention mechanism; normalizing the fused features, eliminating the scale differences of the features of different modalities through a batch normalization layer, and outputting the aligned feature representations.

[0013] In a second aspect, an embodiment of the present invention provides a pumped storage power station construction anomaly detection system based on drone image analysis, which includes a knowledge modeling module for mapping the pumped storage power station CAD drawings and the expert knowledge base into graph-structured training samples using a deep cross-modal knowledge distillation network, performing semantic segmentation and annotation steps, and outputting a hybrid knowledge model with specification constraints; a data acquisition module for planning the construction area acquisition path using a terrain-adaptive multi-drone collaborative strategy, acquiring multi-perspective images based on a multimodal sensor array, processing the multimodal sensor data through a hierarchical spatiotemporal attention network, and generating a construction site feature tensor; a dynamic modeling module for inputting the construction site feature tensor into a densely connected convolutional network for processing, combining a heterogeneous graph attention module with causal reasoning capabilities and a gated recurrent unit to establish a four-dimensional construction dynamic model; an anomaly detection module for calculating the feature distance between the hybrid knowledge model and the four-dimensional construction dynamic model using a two-layer contrastive learning network, and calibrating the abnormal area and its uncertainty using a safety level-adaptive anomaly detector; a solution generation module for performing causal reasoning and traceability analysis on the detected abnormal area, generating multiple compensation solutions in combination with multi-agent reinforcement learning, and screening the optimal compensation solution through a case-enhanced Bayesian analyzer.

[0014] The beneficial effects of the present invention are as follows: the present invention realizes the effective integration of CAD drawings and expert knowledge through a deep cross-modal knowledge distillation network, providing a more comprehensive and standardized knowledge basis for anomaly detection; based on terrain-adaptive multi-UAV collaborative strategy and multimodal sensor data fusion technology, it solves the problems of difficulty and low efficiency in information collection at construction sites in complex terrain environments; a four-dimensional construction dynamic model is constructed by using a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities, which realizes the precise positioning and traceability analysis of construction anomalies; by combining multi-agent reinforcement learning with case-enhanced Bayesian analysis, a more feasible and adaptable compensation solution is provided, which significantly improves the intelligence level of construction quality management. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. 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.

[0016] Figure 1 This is a framework flow chart of the anomaly detection method for pumped storage power station construction based on UAV image analysis.

[0017] Figure 2 A flow chart is constructed for the hybrid knowledge model of anomaly detection method for pumped storage power station construction based on UAV image analysis.

[0018] Figure 3 Flowchart of anomaly detector calibration for the pumped storage power station construction anomaly detection method based on UAV image analysis. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0022] Example 1, reference Figures 1 to 3 , which is the first embodiment of the present invention, provides a method for detecting abnormalities in the construction of a pumped storage power station based on drone image analysis. The framework flow chart is as follows Figure 1 Shown, including,

[0023] S1: Utilize a deep cross-modal knowledge distillation network to map the pumped storage power station CAD drawings and the expert knowledge base into graph-structured training samples, perform semantic segmentation and annotation steps, and output a hybrid knowledge model with specification constraints.

[0024] Specifically, the hybrid knowledge model construction flow chart is as follows: Figure 2 As shown, the following steps are included:

[0025] S1.1: Standardize the pumped storage power station CAD drawings through an image preprocessing network, extract the geometric features, spatial relationships, and engineering parameters in the drawings, and form drawing feature vectors.

[0026] Among them, the image preprocessing network adopts a multi-scale convolutional neural network structure, including four key modules: image denoising, geometric correction, edge enhancement and feature extraction. The preprocessing parameters are adaptively optimized through the validation set.

[0027] S1.2: Use the natural language processing model to parse the text knowledge in the expert knowledge base, extract construction specifications, anomaly judgment standards and processing methods, and generate knowledge feature vectors.

[0028] Among them, the natural language processing model adopts a BERT-based bidirectional encoder architecture and combines domain-adaptive pre-training to improve the ability to understand professional terminology.

[0029] S1.3: Construct a deep cross-modal knowledge distillation network, take the drawing feature vector as the teacher network input and the knowledge feature vector as the student network input, and align the feature representations of the two modalities through the cross-attention mechanism.

[0030] Specifically, the construction process of the deep cross-modal knowledge distillation network is as follows: a multi-head self-attention encoder is constructed to enhance the drawing feature vector and knowledge feature vector respectively to generate drawing attention features and knowledge attention features; a two-stream cross-attention module is designed to use the drawing attention features as the query vector of the teacher network and the knowledge attention features as key-value pairs to calculate the cross-modal attention score matrix; a feature distillation loss function is established, which includes a knowledge distillation loss term and a feature consistency loss term. The knowledge distillation loss guides the learning of the student network through the teacher network, and the feature consistency loss constrains the semantic alignment of the two modal features; an adaptive temperature adjustment strategy is adopted to regulate the knowledge distillation process. The temperature parameter is dynamically adjusted according to the KL divergence of the teacher network and the student network. When the divergence is greater than the preset threshold, the temperature parameter value is increased to enhance knowledge transfer, and vice versa, the temperature parameter value is decreased to maintain feature uniqueness; a feature fusion module is constructed to fuse the features output by the teacher network and the student network through residual connections, and a multi-layer feature pyramid network is used to extract feature representations of different scales. The optimal feature combination is adaptively selected through the attention mechanism; the fused features are normalized, and the scale differences of the features of different modalities are eliminated through the batch normalization layer, and the aligned feature representations are output.

[0031] Among them, the specific formula of the feature distillation loss function is as follows:

[0032]

[0033] in, is the knowledge distillation loss term (which measures the similarity between the output distributions of the teacher model and the student model, based on KL divergence), is the feature consistency loss (through Fusion features of norm-constrained teacher model and student model and distance, ensuring the alignment of intermediate features), is a weight hyperparameter (used to balance the impact of the two losses).

[0034] Knowledge distillation loss The formula is as follows:

[0035]

[0036] in, are the multi-head cross attention features of the teacher model and the student model, is the Softmax function (used to generate probability distribution), is the KL divergence (used to measure the difference in the distribution of attention features between the teacher and student models), and T is the dynamic temperature parameter.

[0037] Feature consistency loss The formula is as follows:

[0038]

[0039] are the fusion features of the teacher model and the student model (after FPN and residual connection), for The square of the norm.

[0040] The adaptive temperature adjustment formula is as follows:

[0041]

[0042] in, is the base temperature value, is the temperature regulation coefficient, is the KL divergence between the teacher model and the student model output probability distribution, is the KL divergence threshold (only when The temperature will be adjusted only when are the probability distributions of the teacher model and the student model respectively.

[0043] Experimental verification shows that compared with existing one-way knowledge distillation methods (such as KD and FitNet), this method improves the accuracy of cross-modal feature alignment by 25.3% (based on cosine similarity evaluation), the discriminability of feature representation by 18.7% (based on t-SNE visualization clustering effect), and improves the accuracy of anomaly detection by 15.6% in actual engineering applications.

[0044] S1.4: Map the aligned features to a graph structure space based on a graph neural network, where graph nodes represent key components in the construction area and graph edges represent the spatial and functional relationships between components, forming an initial graph structure.

[0045] S1.5: Use a semantic segmentation network to perform fine-grained annotation on the initial graph structure, establish a correspondence between construction specification requirements and physical components, and generate an annotated graph structure.

[0046] S1.6: Input the annotated graph structure into the specification constraint encoder, extract the temporal dependencies and quality control standards of the construction process, and output a hybrid knowledge model of the specification constraints.

[0047] Specifically, the process of encoding specification constraints is as follows: construct a temporal dependency graph, establish the predecessor and successor relationship between construction process nodes based on the annotated graph structure, and use a directed acyclic graph to represent the temporal constraints between processes; design a quality feature tensor, map the quality control indicators of each process into a multidimensional feature space, and define the qualified range of the indicators through the interval estimation method; construct a specification constraint encoder, input the temporal dependency graph and the quality feature tensor into a two-branch deep neural network to extract the process constraint and quality constraint features; use the constraint association module to calculate the interaction matrix between process constraints and quality constraints, and establish the dynamic change rules of quality indicators as the process progresses; construct a three-layer constraint feature pool, and hierarchically store constraint features at different levels, including single component constraints (such as concrete strength), regional constraints (such as construction area temperature), and global constraints (such as project progress); integrate the constraint features with the annotated graph structure to generate a hybrid knowledge model with construction specification constraints.

[0048] Optimally, this invention, through the design of a deep cross-modal knowledge distillation network and a canonical constraint encoder, leverages graph structured representation and semantic segmentation annotation to provide a more standardized and structured knowledge foundation for anomaly detection. Compared to traditional rule-based anomaly detection methods, this invention organically combines drawing information with expert experience, improving detection accuracy, reducing false alarm rates, and providing explainable anomaly cause analysis.

[0049] S2: A terrain-adaptive multi-UAV collaborative strategy is used to plan the acquisition path in the construction area. Multi-view images are acquired based on a multimodal sensor array. The multimodal sensor data is processed through a hierarchical spatiotemporal attention network to generate a construction site feature tensor.

[0050] Specifically, the method includes the following steps:

[0051] S2.1: Use LiDAR to scan the terrain of the construction area and generate a 3D point cloud map. Use the point cloud segmentation algorithm to divide the construction area into multiple sub-areas and calculate the terrain complexity index of each sub-area.

[0052] It should be noted that the terrain in the construction area of a pumped-storage power station is usually complex, including multiple functional areas such as upper and lower reservoirs, water transmission systems, and underground powerhouses. The terrain characteristics of each area are significantly different. Therefore, it is necessary to calculate terrain complexity indicators based on point cloud data, including terrain undulation, surface roughness, obstacle density, etc. These indicators reflect the terrain characteristics of different areas and provide an important basis for the subsequent reasonable planning of drone inspection routes.

[0053] S2.2: A multi-UAV task allocation model is constructed based on the terrain complexity index. An adaptive particle swarm algorithm is used to allocate UAV formations and flight altitudes to each sub-area to generate an initial path planning solution.

[0054] Specifically, based on the terrain complexity index, an environmental constraint matrix considering visibility, wind speed and obstacle distribution is constructed, and the flight altitude range and safety distance threshold are set according to the UAV performance parameters to generate a mission environment assessment model; the minimum coverage waypoint set is calculated for each sub-area, and the waypoint spacing threshold is set according to the terrain undulation. The waypoint groups are divided using the convex hull algorithm, and the number of UAVs required for each group is determined based on the waypoint density distribution to form an initial task allocation plan; a multi-objective optimization function is constructed, and the task completion time, energy efficiency and data collection quality are taken as optimization goals, and the UAV payload capacity, communication range and flight time are taken as constraints to establish a task optimization model; an adaptive particle physics model is designed. The particle encoding scheme of the subswarm algorithm is adopted, in which the particle position represents the UAV formation parameters and flight altitude parameters, and the particle velocity represents the parameter update strategy. The inertia weight and learning factor are dynamically adjusted through the terrain complexity index. The fitness function is designed based on the task optimization model, and the particle swarm is iteratively optimized in combination with the task environment assessment model. The elite retention strategy and mutation operator are used to improve the population diversity until the convergence condition is met or the maximum number of iterations is reached. The optimized optimal particles are decoded into the UAV formation and flight altitude parameters. The waypoint sequence of each UAV is generated by combining the initial task allocation plan. The smooth flight trajectory is generated by the cubic spline interpolation algorithm, and the initial path planning plan is output.

[0055] Among them, the convergence condition is set as the relative change rate of the optimal solution for 50 consecutive generations is less than 0.1% or the maximum number of iterations is 1000 generations; in addition, an adaptive weight adjustment strategy based on terrain complexity is adopted for the multi-objective optimization function. The weight of data acquisition quality is increased in complex terrain areas, and the weight of task completion time is increased in flat areas.

[0056] Experiments have shown that the adaptive particle swarm algorithm achieves adaptive optimization for terrain complexity by dynamically adjusting inertia weights and learning factors. Compared to traditional fixed-parameter path planning methods, the proposed multi-objective optimization strategy simultaneously balances task completion time, energy efficiency, and data collection quality, significantly improving the operational efficiency of drone formations and enhancing path planning optimization by approximately 20%.

[0057] S2.3: Based on the initial path planning scheme, combined with the obstacle distribution and weather conditions in the construction area, a collaborative flight path that takes into account safety distance is generated through a trajectory optimization algorithm.

[0058] In this embodiment, the trajectory optimization process is as follows: dynamic obstacle distribution is constructed based on a 3D point cloud map, and environmental constraints are established in conjunction with meteorological data. The initial path is then subjected to spatiotemporal conflict detection to identify potential risk areas. A trajectory smoothing algorithm is employed to ensure flight stability, and a flight path with safe spacing is generated through a multi-agent collaborative optimization method. The trajectory is then corrected to account for meteorological influences, and a final collaborative flight plan is generated after simulation verification. Compared to traditional single trajectory planning methods, this method offers greater environmental adaptability and collaborative capabilities.

[0059] S2.4: A multimodal sensor array equipped with visible light cameras, infrared cameras, and multispectral cameras collects construction site image sequences along the collaborative flight path and records the corresponding position and posture information.

[0060] S2.5: Perform multi-view image registration on the construction site image sequence, use a deep learning network to extract image features, and construct a four-dimensional feature map containing temporal information.

[0061] S2.6: Input the 4D feature map into the hierarchical spatiotemporal attention network to generate the construction site feature tensor through multi-level feature extraction and temporal modeling.

[0062] Preferably, the present invention solves the technical problems of difficulty and low efficiency in collecting construction site information in the complex terrain environment of pumped-storage power stations through a technical solution that combines terrain-adaptive multi-UAV collaborative strategy planning with multimodal sensor data fusion processing. Compared with existing technologies, this solution achieves refined zoning of construction areas and quantification of terrain features based on lidar point cloud data. It uses an adaptive particle swarm algorithm to optimize UAV formations and flight parameters, ensures flight safety through multi-agent collaborative optimization, and achieves an approximately 20% improvement in path planning efficiency. Furthermore, by utilizing multi-source heterogeneous data acquired by a multimodal sensor array and combining it with a hierarchical spatiotemporal attention network for feature extraction and fusion modeling, it improves the comprehensiveness and spatiotemporal consistency of construction site information collection, providing reliable data support for engineering construction quality control and safety management.

[0063] S3: The construction site feature tensor is input into a densely connected convolutional network for processing. Combined with a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities, a four-dimensional construction dynamic model is established.

[0064] Specifically, the method includes the following steps:

[0065] S3.1: The construction site feature tensor is fed into a densely connected convolutional network, which generates a hierarchical feature map through a self-calibrated multi-scale feature extraction module while maintaining the correlation between features at different levels.

[0066] S3.2: Construct a heterogeneous graph structure for the construction scene, treating construction equipment, building components, and construction personnel as different types of graph nodes. Establish edge connections based on spatial position relationships and construction process dependencies to generate a scene relationship graph.

[0067] S3.3: Perform feature fusion on the hierarchical feature representation and scene relationship graph, calculate the dynamic association weights between nodes through the hierarchical adaptive graph attention mechanism, and generate a temporal scene graph.

[0068] S3.4: Use the causal reasoning module to analyze the causal relationship between nodes in the time-series scene graph, build a causal dependency network of the construction scene, and identify key influencing factors and propagation paths.

[0069] Specifically, the temporal scene graph is decomposed into multi-scale time windows, and the graph sequence is divided into overlapping subsequences based on the sliding window mechanism. The temporal attention network is used to extract the node state change features and generate a temporal feature matrix; a bidirectional causal graph neural network is constructed, and the temporal feature matrix is input into the encoder. The long-term and short-term dependencies between nodes are captured through the multi-head self-attention mechanism, and the node association measurement matrix is output; a causal hypothesis set for the construction scene is established based on the node association measurement matrix, and the structural equation model is used to calculate the direct and indirect effects between node pairs to generate an initial causal relationship graph; the initial causal relationship graph is input into the graph convolutional network, and local and global causal patterns are extracted through multi-layer feature aggregation. The causal discovery algorithm is used to identify false associations and generate a revised causal relationship graph; a probabilistic graphical model is constructed based on the revised causal relationship graph, and the variational inference algorithm is used to calculate the conditional independence and causal strength between nodes, identify key influencing factors, and generate a causal influence score; a causal path tracing algorithm is used to prune and optimize the causal chain based on the causal influence score, and the feasibility of the path is verified in combination with the expert rule base to construct a causal dependency network containing key influencing factors and propagation paths.

[0070] Optimally, the causal reasoning module extracts dynamic characteristics of construction scenarios through a combination of multi-scale time window decomposition and a bidirectional causal graph neural network. It also constructs a causal dependency network for construction scenarios based on a structural equation model and a variational inference algorithm. Compared to traditional correlation analysis methods, this method effectively captures long- and short-term dependencies between nodes, identifies and eliminates spurious associations, and, combined with an expert rule base, reliably identifies key influencing factors and propagation paths, providing theoretical support for source analysis and risk warning of construction anomalies.

[0071] S3.5: Input the causal dependency network into the gated recurrent unit to perform time series modeling on key indicators and generate a state evolution sequence.

[0072] Among them, key indicators include construction progress, quality status and resource utilization.

[0073] S3.6: Based on the state evolution sequence, combined with information from the spatial dimension, temporal dimension, construction element dimension, and quality dimension, a four-dimensional construction dynamic model describing the entire construction process is established.

[0074] Specifically, establishing a four-dimensional construction dynamic model that describes the entire construction process includes the following steps: inputting the state evolution sequence into a multi-branch feature fusion network, constructing spatial dimension branches, time dimension branches, construction element branches and quality dimension branches respectively, and extracting the feature representation of each dimension through independent encoders; using a three-dimensional convolutional neural network to process the spatial dimension branch, extracting the spatial layout characteristics of the construction site, establishing the component position relationship and spatial constraints, and generating a spatial feature vector; combining the long short-term memory network and the self-attention mechanism in the time dimension branch to model the construction progress and process connection, capture the critical path of the construction period, and output the time series feature vector; constructing a heterogeneous graph embedding network for the construction element branch, mapping the element nodes such as personnel, equipment, and materials to a unified feature space, and generating element feature vectors based on the element interaction relationship; deploying a multi-layer perceptron through the quality dimension branch, processing quality inspection data and specification constraint indicators, establishing a quality assessment model, and outputting a quality feature vector; designing a cross-dimensional attention fusion module to adaptively allocate weights to the feature vectors of the four dimensions, maintaining the original feature information through residual connections, and constructing a four-dimensional construction dynamic model.

[0075] S4: A two-layer contrastive learning network is used to calculate the characteristic distance between the hybrid knowledge model and the four-dimensional construction dynamic model, and an anomaly detector with adaptive safety level is used to calibrate the abnormal area and its uncertainty.

[0076] Specifically, the anomaly detector calibration flow chart is as follows: Figure 3 As shown, the following steps are included:

[0077] S4.1: Construct a feature-adapted two-layer contrastive learning network, align the dimensions and normalize the features of the hybrid knowledge model and the four-dimensional construction dynamic model through the preprocessing module, and generate a normalized feature vector using hierarchical feature extraction and residual preservation mechanism.

[0078] S4.2: Calculate the semantic similarity matrix for the normalized feature vector, use the multi-granularity cross-attention mechanism to highlight the key feature differences, and generate a feature deviation map.

[0079] S4.3: Construct a hierarchical safety assessment system based on engineering specifications and historical cases, map the characteristic deviation graph to different safety levels, and generate a safety assessment vector.

[0080] Specifically, building a hierarchical safety assessment system includes the following steps: using knowledge graph technology to build an engineering specification library, converting the safety standards, quality requirements and technical indicators in the construction specifications into structured knowledge nodes, establishing hierarchical associations between nodes through a relationship extraction algorithm, and generating a specification knowledge graph; conducting in-depth mining of the historical case database, using a text classification model to label cases with safety levels, building a case association network based on a similarity measurement method, and outputting a case feature matrix; inputting the specification knowledge graph and the case feature matrix into a dual-channel deep neural network, extracting specification constraint features and experience features respectively, fusing the two types of features through an attention mechanism, and generating a safety assessment benchmark; constructing a multi-level safety level classification model, setting key indicator thresholds based on the safety assessment benchmark, using a fuzzy comprehensive evaluation method to establish a mapping relationship between indicators and safety levels, and outputting a grading standard set; inputting a feature deviation map into the grading standard set, calculating the weight coefficients of each indicator through a hierarchical analysis method, and quantitatively evaluating the deviation features in combination with an expert scoring mechanism to generate a preliminary safety level; using a deep belief network to perform probabilistic reasoning on the preliminary safety level, dynamically correcting it based on the real-time status of the engineering site, and outputting a safety assessment vector containing the safety assessment scores of each dimension.

[0081] Remarkably, the hierarchical safety assessment system effectively integrates engineering specification knowledge and historical case experience by combining knowledge graph technology and deep learning methods. It uses a dual-channel deep neural network to extract specification constraint features and empirical features, and a deep belief network for probabilistic reasoning and dynamic correction. Compared with existing fixed-threshold assessment methods, the adaptive assessment mechanism constructed by this invention can dynamically adjust the safety level based on the real-time status of the construction site, improving the scientific nature and reliability of safety assessments and providing more comprehensive technical support for construction management decisions.

[0082] S4.4: Adopt an adaptive threshold mechanism to dynamically adjust the sensitivity of the anomaly detector according to the security assessment vector and determine the multi-scale detection threshold.

[0083] S4.5: Use multi-scale detection thresholds to perform hierarchical screening on the feature deviation map, locate potential abnormal areas, and output a set of abnormal areas.

[0084] S4.6: Combine the Monte Carlo sampling method to evaluate the detection confidence of each area in the anomaly area set and generate an anomaly calibration result including the anomaly location, type and uncertainty.

[0085] S5: Perform causal reasoning and traceability analysis on the detected abnormal areas, combine multi-agent reinforcement learning to generate multiple compensation plans, and screen the optimal compensation plan through a case-enhanced Bayesian analyzer.

[0086] Specifically, the method includes the following steps:

[0087] S5.1: Use the dynamic graph causal attention network to process the anomaly calibration results, establish the spatiotemporal causal graph of the abnormal events, extract the correlation metrics between abnormal features and construction elements, and output the anomaly propagation matrix.

[0088] S5.2: Initialize the parameters of the multi-agent interaction model through a hierarchical knowledge distillation architecture, build a teacher network containing construction expert rules, set hierarchical reward indicators, and form an agent training environment.

[0089] S5.3: Deploy an adversarial reinforcement learning model to train professional agents, divide agent roles according to construction fields, execute competitive game training processes, and generate differentiated compensation strategy combinations.

[0090] Specifically, training professional agents and generating differentiated compensation strategy combinations includes the following steps: constructing an agent role system based on construction domain knowledge, dividing the construction process participants into progress management agents, quality control agents, resource allocation agents, and safety supervision agents, and establishing an agent decision space; constructing a hierarchical adversarial network architecture, designing the competitive game relationship between agents, embedding the anomaly propagation matrix into the state representation space, using a graph encoder to extract state features, and generating a game state vector; deploying a deep deterministic policy gradient algorithm, training the agent policy network based on the game state vector, accumulating adversarial samples through an experience replay mechanism, and outputting an initial compensation strategy set; adopting a multi-agent collaborative training framework, setting up communication channels between agents, introducing an attention mechanism to extract key information, optimizing the agent decision model through policy gradient descent, and generating collaborative compensation strategies; constructing a compensation effect evaluation module, inputting the collaborative compensation strategy into the simulation environment, calculating the strategy execution cost and effect improvement, dynamically adjusting the agent reward function based on the evaluation results, and optimizing the compensation strategy; adopting a strategy integration method to fuse the compensation strategies trained in multiple rounds, screening complementary strategy combinations through diversity metrics, and generating differentiated compensation strategy combinations.

[0091] S5.4: Build a graph neural network case retrieval model, encode historical compensation plans into a temporal graph structure, calculate the case graph matching metric, and extract empirical feature maps.

[0092] S5.5: Configure a two-stream Bayesian neural network to process the cost and duration indicators and the risk reliability indicators respectively, apply attention weight fusion, and output the solution evaluation vector.

[0093] S5.6: Perform combinatorial optimization calculations under uncertainty constraints, jointly consider evaluation scores and resource constraints, and generate a set of compensation plans with contingency plans.

[0094] S5.7: Sort the compensation scheme set based on the scheme evaluation vector and select the scheme with the highest comprehensive score as the optimal compensation scheme.

[0095] Furthermore, this embodiment also provides a pumped storage power station construction anomaly detection system based on drone image analysis, including a knowledge modeling module for mapping the pumped storage power station CAD drawings and the expert knowledge base into graph-structured training samples using a deep cross-modal knowledge distillation network, performing semantic segmentation and annotation steps, and outputting a hybrid knowledge model with specification constraints; a data acquisition module for planning the construction area acquisition path using a terrain-adaptive multi-drone collaborative strategy, acquiring multi-perspective images based on a multimodal sensor array, processing the multimodal sensor data through a hierarchical spatiotemporal attention network, and generating a construction site feature tensor; a dynamic modeling module for inputting the construction site feature tensor into a densely connected convolutional network for processing, combining a heterogeneous graph attention module with causal reasoning capabilities and a gated recurrent unit to establish a four-dimensional construction dynamic model; an anomaly detection module for calculating the feature distance between the hybrid knowledge model and the four-dimensional construction dynamic model using a two-layer contrastive learning network, and calibrating the abnormal area and its uncertainty using a safety level-adaptive anomaly detector; and a solution generation module for performing causal reasoning and traceability analysis on the detected abnormal area, generating multiple compensation solutions in combination with multi-agent reinforcement learning, and screening the optimal compensation solution through a case-enhanced Bayesian analyzer.

[0096] In summary, the present invention realizes the effective fusion of CAD drawings and expert knowledge through a deep cross-modal knowledge distillation network, providing a more comprehensive and standardized knowledge basis for anomaly detection; based on the terrain-adaptive multi-UAV collaborative strategy and multimodal sensor data fusion technology, it solves the problems of difficulty and low efficiency in information collection at the construction site under complex terrain environments; a four-dimensional construction dynamic model is constructed by using a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities, which realizes the precise positioning and traceability analysis of construction anomalies; by combining multi-agent reinforcement learning with case-enhanced Bayesian analysis, a more feasible and adaptable compensation scheme is provided, which significantly improves the intelligent level of construction quality management.

[0097] Example 2, reference Figures 1 to 3 This is the second embodiment of the present invention, which provides a method for detecting abnormalities in the construction of pumped-storage power stations based on drone image analysis. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0098] To validate the effectiveness of the proposed method, the research team selected a pumped-storage power station under construction in a hilly area. The station has a total installed capacity of 1200 MW, a height difference of approximately 500 meters between the upper and lower reservoirs, and a construction area of approximately 12 square kilometers. The experimental environment utilized a dual-socket Intel Xeon server (64 cores, 256GB of memory) equipped with an NVIDIA A100 GPU for deep learning model training.

[0099] During the data collection phase, the research team deployed a multi-drone collaborative system consisting of six drones equipped with a multimodal sensor array. These sensors included a 40-megapixel visible light camera, a 640×512 resolution infrared thermal imager, and a six-band multispectral camera. LiDAR was also used for terrain scanning, with a point cloud density of 50 points per square meter. Weather conditions during the experiment included temperatures ranging from 15°C to 25°C, wind speeds of 3–5 m / s, and good visibility.

[0100] During the knowledge model construction phase, researchers collected 458 CAD drawings from the project, covering disciplines such as civil engineering and electromechanical installation, and compiled over 2,000 expert knowledge items. Through training with a deep cross-modal knowledge distillation network, the model achieved a feature alignment accuracy of 92.5% on a cross-validation set, with a knowledge transfer loss converged to 0.086. The network was trained with a batch size of 64, an initial learning rate of 0.001, and a cosine annealing strategy for a total of 100 epochs.

[0101] For drone path planning, the construction area was divided into 28 sub-areas based on lidar point cloud data, with a terrain complexity index ranging from 0.35 to 0.89 (normalized value). The adaptive particle swarm algorithm was set with a population size of 200, a maximum number of iterations of 1000, and a convergence threshold of 0.1%. The optimized flight path shortened the total length by 23.6% compared to the initial plan, reducing mission completion time by 18.9%. Trajectory optimization within environmental constraints ensured a minimum safe distance between drones of at least 50 meters.

[0102] As shown in Table 1, the method of the present invention was compared and verified with the existing technical solutions on multiple key indicators:

[0103] Table 1 Performance comparison between the method of the present invention and the traditional method

[0104]

[0105] Experimental results demonstrate that the proposed method demonstrates significant advantages in anomaly detection performance. During a three-month field validation, the system performed 312 inspections, collecting a cumulative image data volume of 2.1 TB. Processing with a hierarchical spatiotemporal attention network, the system successfully identified 147 potential anomalies, 142 of which were confirmed by field experts, resulting in a true anomaly detection rate of 96.5%. Anomaly types primarily included concrete cracks (32%), rebar spacing deviations (28%), water seepage points (21%), and other quality defects (19%).

[0106] In terms of compensation plan generation, after 500,000 training iterations, the policy network's reward value converged to a stable state. For each detected anomaly, the system generated an average of 4-6 differentiated compensation plans, with the Pareto-optimal solution set for execution cost and expected effect containing an average of three options. Through screening using a case-enhanced Bayesian analyzer, the recommended plans achieved a 91.3% implementation success rate, a 15.6 percentage point improvement over traditional expert experience methods.

[0107] Particularly noteworthy is the outstanding performance of this invention in causal reasoning. Using a dynamic graph causal attention network, the system accurately tracks the root causes of anomalies and predicts their potential impact. On a validation set, the accuracy of causal relationship identification reached 89.6%, providing construction managers with more valuable decision-making support. Furthermore, the system's average response time is only 0.8 seconds, meeting the requirements of real-time monitoring.

[0108] From a cost-effectiveness perspective, while the proposed method involves high initial hardware investment and algorithm development costs, in practice it significantly reduces manual inspection costs, improves the timeliness of anomaly detection, and reduces the cost of rework and quality improvement. According to project statistics, the system's implementation saved the project approximately 2.8% of its total construction costs.

[0109] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for detecting abnormalities in pumped storage power station construction based on drone image analysis, characterized by: include, A deep cross-modal knowledge distillation network is used to map the pumped storage power station CAD drawings and the expert knowledge base into graph-structured training samples, perform semantic segmentation and annotation steps, and output a hybrid knowledge model with normative constraints. A terrain-adaptive multi-UAV collaborative strategy is used to plan the acquisition path in the construction area. Multi-view images are acquired based on a multimodal sensor array. The multimodal sensor data is processed through a hierarchical spatiotemporal attention network to generate a construction site feature tensor. The construction site feature tensor is input into a densely connected convolutional network for processing, and a four-dimensional construction dynamic model is established by combining a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities; A two-layer contrastive learning network is used to calculate the characteristic distance between the hybrid knowledge model and the four-dimensional construction dynamic model, and an anomaly detector with adaptive safety level is used to calibrate the abnormal area and its uncertainty; Perform causal reasoning and source analysis on detected abnormal areas, generate multiple compensation plans using multi-agent reinforcement learning, and select the optimal compensation plan through a case-enhanced Bayesian analyzer; The construction process of the hybrid knowledge model includes: Build a deep cross-modal knowledge distillation network, use the drawing feature vector as the teacher network input and the knowledge feature vector as the student network input, and align the feature representations of the two modalities through a cross-attention mechanism; The aligned features are mapped to a graph structure space based on a graph neural network, where graph nodes represent key components in the construction area and graph edges represent the spatial and functional relationships between components, forming an initial graph structure. The method of calibrating the abnormal area and its uncertainty by using the safety level adaptive abnormality detector includes the following steps: A feature-adaptive two-layer contrastive learning network is constructed. The preprocessing module aligns the dimensions and standardizes the features of the hybrid knowledge model and the four-dimensional construction dynamic model. A normalized feature vector is generated using hierarchical feature extraction and a residual preservation mechanism. Calculating a semantic similarity matrix for the normalized feature vectors, using a multi-granularity cross-attention mechanism to highlight key feature differences, and generating a feature deviation map; Constructing a hierarchical safety assessment system, mapping the characteristic deviation map to different safety levels, and generating a safety assessment vector; Dynamically adjusting the sensitivity of the anomaly detector according to the security assessment vector to determine a multi-scale detection threshold; Using the multi-scale detection threshold to perform hierarchical screening on the feature deviation map, locate potential abnormal areas, and output a set of abnormal areas; The detection confidence of each area in the abnormal area set is evaluated to generate an abnormality calibration result including the abnormality location, type and uncertainty.

2. The method for detecting abnormalities in pumped storage power station construction based on drone image analysis according to claim 1, characterized in that: The construction process of the hybrid knowledge model further includes: The CAD drawings of the pumped storage power station are standardized through an image preprocessing network to extract the geometric features, spatial relationships and engineering parameters in the drawings to form drawing feature vectors. The text knowledge in the expert knowledge base is parsed using a natural language processing model to extract construction specifications, abnormality judgment criteria and processing methods, and generate knowledge feature vectors; A semantic segmentation network is used to perform fine-grained annotation on the initial graph structure, establish a correspondence between construction specification requirements and physical components, and generate an annotated graph structure; The annotated graph structure is input into a specification constraint encoder to extract the temporal dependency relationship and quality control standards of the construction process, and output a hybrid knowledge model of the specification constraint.

3. The method for detecting abnormalities in pumped storage power station construction based on drone image analysis according to claim 1, characterized in that: Generating the construction site feature tensor comprises the following steps: Use LiDAR to scan the terrain of the construction area and generate a 3D point cloud map. Use the point cloud segmentation algorithm to divide the construction area into multiple sub-areas and calculate the terrain complexity index of each sub-area. Based on the terrain complexity index, a multi-UAV task allocation model is constructed, and an adaptive particle swarm algorithm is used to allocate UAV formations and flight altitudes to each sub-area to generate an initial path planning solution; Based on the initial path planning scheme, combined with the distribution of obstacles in the construction area and weather conditions, a collaborative flight path that takes into account safety distances is generated through a trajectory optimization algorithm. A multimodal sensor array equipped with visible light cameras, infrared cameras, and multispectral cameras collects a sequence of images of the construction site along the collaborative flight path and records the corresponding position and posture information; Performing multi-view image registration on the construction site image sequence, extracting image features using a deep learning network, and constructing a four-dimensional feature map containing time series information; The four-dimensional feature map is input into a hierarchical spatiotemporal attention network, and a construction site feature tensor is generated through multi-level feature extraction and temporal modeling.

4. The method for detecting abnormalities in pumped storage power station construction based on drone image analysis according to claim 1, wherein: The establishment of the four-dimensional construction dynamic model comprises the following steps: The construction site feature tensor is fed into a densely connected convolutional network, and a hierarchical feature map is generated through a self-calibrated multi-scale feature extraction module. Construct a heterogeneous graph structure for the construction scene, treating construction equipment, building components, and construction personnel as different types of graph nodes. Establish edge connections based on spatial position relationships and construction process dependencies to generate a scene relationship graph. Performing feature fusion on the hierarchical feature representation and the scene relationship graph, calculating the dynamic association weights between nodes through a hierarchical adaptive graph attention mechanism, and generating a temporal scene graph; A causal reasoning module is used to analyze the causal relationship between nodes in the time sequence scene graph, construct a causal dependency network of the construction scene, and identify key influencing factors and propagation paths; Input the causal dependency network into a gated recurrent unit, perform time series modeling on key indicators, and generate a state evolution sequence; Based on the state evolution sequence, combined with information of spatial dimension, time dimension, construction element dimension and quality dimension, a four-dimensional construction dynamic model describing the entire construction process is established.

5. The method for detecting abnormalities in pumped storage power station construction based on drone image analysis according to claim 1, characterized in that: The method of screening the optimal compensation plan by using the case-enhanced Bayesian analyzer includes the following steps: A dynamic graph causal attention network is used to process anomaly calibration results, establish a spatiotemporal causal graph of abnormal events, extract correlation metrics between abnormal features and construction elements, and output an anomaly propagation matrix. The multi-agent interaction model is initialized through a hierarchical knowledge distillation architecture, a teacher network containing construction expert rules is constructed, and hierarchical reward indicators are set to form an agent training environment. Deploy an adversarial reinforcement learning model to train specialized agents, divide agent roles according to construction fields, execute a competitive game training process, and generate differentiated compensation strategy combinations; Build a graph neural network case retrieval model, encode historical compensation plans into a time-series graph structure, calculate case graph matching metrics, and extract empirical feature maps; Configure a two-stream Bayesian neural network to process cost and duration indicators and risk reliability indicators respectively, apply attention weight fusion, and output a solution evaluation vector; Perform combinatorial optimization calculations under uncertainty constraints, jointly consider evaluation scores and resource constraints, and generate a set of compensation plans with contingency plans; The compensation scheme set is sorted based on the scheme evaluation vector, and the scheme with the highest comprehensive score is selected as the optimal compensation scheme.

6. The method for detecting abnormalities in pumped storage power station construction based on drone image analysis according to claim 2, characterized in that: The construction of the deep cross-modal knowledge distillation network includes the following steps: Construct a multi-head self-attention encoder to enhance the drawing feature vector and knowledge feature vector respectively, and generate drawing attention features and knowledge attention features; Design a two-stream cross-attention module, use the drawing attention features as the query vector of the teacher network, and the knowledge attention features as key-value pairs to calculate the cross-modal attention score matrix; Establish a feature distillation loss function, which includes knowledge distillation loss and feature consistency loss, and use an adaptive temperature adjustment strategy to regulate the knowledge distillation process; Construct a feature fusion module to fuse the features output by the teacher network and the student network through residual connections, use a multi-layer feature pyramid network to extract feature representations at different scales, and adaptively select the optimal feature combination through the attention mechanism; The fused features are normalized, the scale differences of features of different modalities are eliminated through the batch normalization layer, and the aligned feature representations are output.

7. A pumped storage power station construction anomaly detection system based on drone image analysis, based on the pumped storage power station construction anomaly detection method based on drone image analysis according to any one of claims 1 to 6, characterized in that: Also includes, The knowledge modeling module is used to map the pumped storage power station CAD drawings and the expert knowledge base into graph-structured training samples using a deep cross-modal knowledge distillation network, perform semantic segmentation and annotation steps, and output a hybrid knowledge model with specification constraints; The data acquisition module is used to plan the collection path in the construction area using a terrain-adaptive multi-UAV collaborative strategy. It acquires multi-view images based on a multimodal sensor array and processes the multimodal sensor data through a hierarchical spatiotemporal attention network to generate a construction site feature tensor. A dynamic modeling module is used to input the construction site feature tensor into a densely connected convolutional network for processing, and to establish a four-dimensional construction dynamic model by combining a heterogeneous graph attention module and a gated recurrent unit with causal reasoning capabilities; An anomaly detection module, configured to calculate the characteristic distance between the hybrid knowledge model and the four-dimensional construction dynamic model using a two-layer contrastive learning network, and to calibrate the abnormal area and its uncertainty using an anomaly detector with adaptive safety level; The solution generation module is used to perform causal reasoning and traceability analysis on the detected abnormal areas, generate multiple compensation solutions by combining multi-agent reinforcement learning, and screen the optimal compensation solution through a case-enhanced Bayesian analyzer.

Images