A method and system for intelligent inspection of a water conservancy inspection robot dog based on a Beidou AI

Abstract

The application provides a water conservancy inspection robot dog intelligent inspection method and system based on Beidou AI, the method first constructs a multi-dimensional digital twin model of water conservancy facilities, and uses a knowledge graph and a graph neural network to dynamically predict risk transmission effects, and forms a risk potential field. Based on the potential field, the optimal path of the inspection robot dog is planned, which takes into account the information income and energy consumption. In the inspection process, the cognitive accident is found by comparing the real-time data with the expected value of the model, and the path re-planning is triggered in real time to investigate the abnormal point. The abnormal data with Beidou high-precision positioning information are fed back to the graph neural network for iterative training, and a closed-loop correction is formed. The application realizes the effect of improving the accuracy of water conservancy inspection risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent water conservancy inspection technology, specifically relating to an intelligent inspection method and system for a water conservancy inspection robot dog based on Beidou AI. Background Technology

[0002] As a critical infrastructure safeguarding national economic development and the safety of people's lives and property, the stability and reliability of water conservancy facilities are of paramount importance. For a long time, the safe operation and maintenance of water conservancy facilities has relied primarily on regular on-site inspections and fixed sensor networks deployed at key locations. This traditional inspection model not only requires a significant investment of manpower and resources, resulting in high labor intensity and low efficiency, but also, due to fixed inspection cycles and limited coverage, makes it difficult to comprehensively and promptly grasp the overall health status of the facilities. Furthermore, it often leads to delayed responses to some gradual or sudden safety hazards, lacking effective proactive risk warning and prevention capabilities.

[0003] With the expansion of water conservancy facilities and the increase in their structural complexity, their safety risks are often not caused by a single factor, but rather are the result of multiple factors intertwined and dynamically coupled, such as the aging of the facilities themselves, changes in surrounding geological and hydrological conditions, and extreme weather events. Existing technologies have significant shortcomings in analyzing the intrinsic connections and cumulative effects among these complex risk factors, leading to insufficient accuracy in risk assessment. Consequently, the allocation of inspection resources and the planning of inspection tasks lack sufficient scientific basis, making it difficult to achieve focused and efficient responses. Therefore, establishing an intelligent inspection system capable of dynamically assessing overall risks, intelligently planning inspection tasks, and rapidly responding to sudden anomalies is a pressing technical problem that needs to be solved in the field of water conservancy safety operation and maintenance. Summary of the Invention

[0004] This invention provides a method and system for intelligent inspection of water conservancy facilities using a robot dog based on BeiDou AI, in order to solve the above-mentioned technical problems.

[0005] In a first aspect, the present invention provides an intelligent inspection method for a water conservancy inspection robot dog based on BeiDou AI, the method comprising the following steps:

[0006] Construct a multidimensional digital twin model of water conservancy facilities, and combine historical data and facility structure information of water conservancy facilities to assign initial risk weights to the multidimensional digital twin model to form an initial risk potential energy field;

[0007] Knowledge graphs and graph neural networks are used to predict the transmission and accumulation effects of environmental risks around water conservancy facilities in the facility structure, and the initial risk potential field is dynamically updated to the risk potential field.

[0008] Based on the risk potential energy field and combined with the inspection information benefits and inspection energy consumption costs of the inspection robot dog, the optimal inspection path is generated for the inspection robot dog.

[0009] The robot dog is controlled to perform inspections according to the optimal inspection path, and real-time environmental data around the robot dog is collected during the inspection process.

[0010] The real-time environmental data is compared with the environmental expectation value extracted from the risk potential field. When a cognitive surprise occurs as defined by the difference between the real-time environmental data and the environmental expectation value, the online replanning of the inspection path is triggered to conduct local inspection of the anomaly point where the cognitive surprise occurs.

[0011] The inspection robot dog acquires abnormal environmental data and BeiDou positioning information at abnormal points, and feeds the abnormal environmental data and BeiDou positioning information back to the graph neural network for iterative training in order to achieve closed-loop correction of the graph neural network.

[0012] Optionally, constructing a multi-dimensional digital twin model of the water conservancy facility, and assigning initial risk weights to the multi-dimensional digital twin model by combining historical data and facility structure information to form an initial risk potential energy field includes the following steps:

[0013] The original three-dimensional point cloud data and texture data of water conservancy facilities were collected using UAV oblique photography and ground three-dimensional laser scanning technology;

[0014] The original three-dimensional point cloud data is registered with geographic coordinates using BeiDou differential positioning technology to generate a multi-dimensional digital twin model with absolute geographic accuracy.

[0015] Structured risk information is extracted from design drawings, historical damage reports, and geological exploration data of water conservancy facilities;

[0016] Structured risk information is mapped to corresponding positions on a multidimensional digital twin model and quantified into initial risk weights;

[0017] An initial risk potential field covering the entire water conservancy facility is generated based on the initial risk weights at all locations.

[0018] Optionally, using knowledge graphs and graph neural networks to predict the transmission and accumulation effects of environmental risks around water conservancy facilities in the facility structure, and dynamically updating the initial risk potential field to a risk potential field includes the following steps:

[0019] A water conservancy knowledge graph is constructed by taking the components of water conservancy facilities as facility nodes and the physical or logical influence relationships between facility components as node edges.

[0020] Real-time environmental impact factors, including rainfall forecasts, upstream water levels, and soil moisture, are periodically obtained from external data sources.

[0021] Real-time environmental impact factors are loaded as input variables into the corresponding facility nodes of the water resources knowledge graph;

[0022] Using graph neural networks to perform reasoning on a water conservancy knowledge graph, we calculate the cumulative effect of risk energy caused by real-time environmental impact factors after transmission and aggregation along node edges.

[0023] The updated risk potential field is output based on the calculation results of the cumulative effect.

[0024] Optionally, generating the optimal inspection path for the inspection robot dog based on the risk potential energy field and combining the inspection information benefits and inspection energy consumption costs includes the following steps:

[0025] The updated risk potential energy field is transformed into an inspection information entropy field, where areas with higher risk potential energy have higher inspection information benefits.

[0026] The multidimensional digital twin model is divided into a path planning grid, and the corresponding inspection energy consumption cost is calculated for each feasible edge in the path planning grid.

[0027] Define the path evaluation function by using the benefits of inspection information and the energy cost of inspection as variables.

[0028] Given a total energy consumption budget constraint, the optimal inspection path is found by searching using a path optimization algorithm and a path evaluation function.

[0029] Optionally, calculating the corresponding inspection energy cost for each feasible edge in the path planning grid includes the following steps:

[0030] Extract the slope parameters and surface material type of each feasible edge in the path planning grid from the multidimensional digital twin model;

[0031] The energy consumption required to overcome the slope parameters by gravity is calculated based on the dynamic model of the inspection robot dog.

[0032] Based on the type of surface material, the corresponding friction coefficient is retrieved from the energy consumption database, and the friction energy consumption required for the inspection robot dog to overcome friction is calculated.

[0033] The basic travel energy consumption is calculated by combining the average travel energy consumption of the inspection robot dog and the length of the feasible edge.

[0034] The energy consumption cost of inspection is obtained by adding the energy consumption of overcoming gravity, the energy consumption of overcoming friction, and the energy consumption of basic travel.

[0035] Optionally, when a cognitive surprise occurs as defined by the difference between real-time environmental data and expected environmental values, online replanning of the inspection path is triggered to conduct local investigation of the anomaly points where the cognitive surprise occurred, including the following steps:

[0036] When the inspection robot dog acquires real-time environmental data, the current location corresponding to the real-time environmental data is determined by the inspection robot dog's Beidou positioning.

[0037] Extract the expected risk value of the current location from the risk potential field, and construct the prior belief distribution of the risk state of the current location based on the expected risk value;

[0038] Real-time environmental data is modeled as an observation likelihood function. Bayesian inference is applied, and the posterior belief distribution of the current location risk state is calculated by combining the prior belief distribution with the observation likelihood function.

[0039] The cognitive surprise value caused by the difference between real-time environmental data and environmental expectations is quantified by calculating the Kolb-Leibler divergence between the posterior belief distribution and the prior belief distribution.

[0040] When the cognitive error value exceeds the preset error threshold, the current position is designated as the abnormal point where the cognitive error occurred, and the online replanning of the inspection robot dog's inspection route is triggered.

[0041] Optionally, triggering online replanning of the inspection robot dog's inspection route includes the following steps:

[0042] Delineate a local exploration area centered on the anomaly point;

[0043] A local risk potential energy field is generated within a local survey area using real-time environmental data.

[0044] The local risk potential energy field is transformed into a local inspection information entropy field, and the local energy consumption cost of the inspection robot dog in the local inspection information entropy field is calculated.

[0045] By combining the local inspection information entropy field and local energy consumption cost, and using the optimal path algorithm, a local inspection sub-path for the inspection robot dog is generated for the local inspection area.

[0046] Optionally, after conducting a localized investigation of the anomalies where cognitive surprises occur, the following steps may also be included:

[0047] Acquire local environmental data collected by the inspection robot dog during local reconnaissance sub-paths;

[0048] Analyze local environmental data and quantify the risk severity level of anomalies;

[0049] Determine whether the severity level of the risk exceeds the preset global impact threshold;

[0050] If the global impact threshold is not exceeded, the inspection robot dog is controlled to return to the breakpoint of the original optimal inspection path and continue to execute the task.

[0051] If the global impact threshold is exceeded, the remaining part of the original optimal inspection path is discarded, the global risk potential field is updated based on the risk severity level and local environmental data, and the adaptive inspection path of the inspection robot dog is regenerated from the anomaly point location.

[0052] Optionally, the abnormal environmental data and BeiDou positioning information at the abnormal points are obtained by the inspection robot dog, and the abnormal environmental data and BeiDou positioning information are fed back to the graph neural network for iterative training to achieve closed-loop correction of the graph neural network. The steps include the following:

[0053] The inspection robot dog obtains abnormal environmental data and BeiDou positioning information at abnormal points.

[0054] Abnormal environmental data, BeiDou positioning information, and real-time environmental impact factors are packaged into structured risk samples;

[0055] Assign a label to each risk sample, consisting of the risk type and severity level;

[0056] Labeled risk samples are stored in a pre-defined historical risk database to expand the training dataset of the graph neural network.

[0057] The retraining process of the graph neural network is triggered when the preset iteration cycle is reached or the number of accumulated risk samples reaches a threshold.

[0058] The original graph neural network is replaced with the retrained graph neural network to complete the closed-loop correction of the risk potential field assessment.

[0059] Secondly, the present invention also provides a water conservancy inspection robot dog intelligent inspection system based on Beidou AI, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the water conservancy inspection robot dog intelligent inspection method based on Beidou AI as described in the first aspect.

[0060] The beneficial effects of this invention are:

[0061] This invention constructs a multi-dimensional digital twin model of water conservancy facilities and integrates knowledge graphs and graph neural networks to achieve dynamic, accurate prediction and quantitative assessment of risks to these facilities, transforming traditional passive response inspections into proactive, predictive maintenance. Based on the optimal inspection path generated from the risk potential field, it comprehensively considers the information benefits and energy costs of inspections, significantly improving inspection efficiency and reducing operating costs while ensuring priority is given to key risk points. The cognitive accident triggering mechanism and online path replanning capability during the inspection process endow the inspection system with a high degree of autonomy and emergency response speed, enabling timely detection and investigation of localized sudden anomalies that are easily overlooked in traditional inspection models. More importantly, through closed-loop feedback correction of BeiDou positioning information and anomaly data, the system can continuously learn and iterate, continuously improving the accuracy of the risk model and the scientific nature of inspection decisions, ultimately forming a complete predictive, adaptive, and highly efficient intelligent inspection system to comprehensively ensure the safe and stable operation of water conservancy facilities. Attached Figure Description

[0062] Figure 1 This is a flowchart illustrating one embodiment of the intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI in this application.

[0063] Figure 2 This is a schematic diagram illustrating the process of predicting the transmission and cumulative effects of environmental risks around water conservancy facilities in the facility structure using knowledge graphs and graph neural networks in one embodiment of this application. Detailed Implementation

[0064] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0065] The terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0066] Figure 1This is a flowchart illustrating an intelligent inspection method for water conservancy inspection using a robot dog based on BeiDou AI, as shown in one embodiment. It should be understood that, although... Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps. For example Figure 1 As shown, the intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI disclosed in this invention specifically includes the following steps:

[0067] S101. Construct a multi-dimensional digital twin model of water conservancy facilities, and combine the historical data and structural information of water conservancy facilities to assign initial risk weights to the multi-dimensional digital twin model to form an initial risk potential energy field.

[0068] To accurately reproduce the physical structure and operational status of the water conservancy facilities, firstly, UAV oblique photography and ground-based 3D laser scanning technologies were used to collect 3D point clouds and high-definition textures of the facilities from multiple angles, both aerial and ground-based, constructing a high-fidelity basic model. Subsequently, using BeiDou differential positioning technology, each point in the model was assigned precise real-world geographic coordinates, forming a multi-dimensional digital twin with absolute accuracy down to the meter or even centimeter level. Based on this, by reviewing and analyzing the facility's design drawings, historical damage reports, and geological exploration data, structured risk information such as the historical location of cracks, weak points in materials, and areas of foundation settlement was extracted. This information was mapped and quantified into initial risk weights, which were then attached to the corresponding spatial locations in the digital twin model. Ultimately, the initial risk weights for all locations collectively constitute an initial risk potential energy field covering the entire facility. , where x represents spatial location. The effect of this process is to generate a static, historical, and inherently attribute-based risk map of the facility, providing a solid benchmark for subsequent dynamic risk assessments.

[0069] S102. Use knowledge graphs and graph neural networks to predict the transmission and accumulation effects of environmental risks around water conservancy facilities in the facility structure, and dynamically update the initial risk potential field to the risk potential field.

[0070] To ensure that risk assessments dynamically reflect current environmental changes, it is necessary to predict the transmission and cumulative effects of external environmental risks within the facilities. This process begins by constructing a hydraulic knowledge graph, abstracting facility components such as dams, gates, and dikes as nodes in the graph, and defining the physical support or logical influence relationships between them, such as water flow conduction and stress transfer, as edges connecting the nodes. Next, the system periodically retrieves real-time environmental impact factors from external data sources such as meteorology and hydrology, including rainfall forecasts for the next few hours, real-time upstream water levels, and surrounding soil moisture. These impact factors... This information will be loaded as a variable onto the corresponding facility node in the knowledge graph. Then, a graph neural network (GNN) is used to perform inference calculations on the entire knowledge graph G, simulating how the risk energy brought about by these environmental impact factors is transmitted, superimposed, and accumulated along the edges between nodes. This calculation process can be represented as the cumulative risk amount. Ultimately, the calculated cumulative risk effect is superimposed on the initial risk potential field to dynamically update and generate a risk potential field reflecting the current overall risk situation. The effect is to transform risk assessment from static analysis into dynamic prediction that can anticipate the chain reactions triggered by environmental changes.

[0071] S103. Based on the risk potential energy field and combined with the inspection information benefits and inspection energy consumption costs of the inspection robot dog, the optimal inspection path is generated for the inspection robot dog.

[0072] To plan the most efficient inspection route, the goal is to maximize the information gain within a limited energy budget. This process transforms the dynamically updated risk potential energy field into an inspection information entropy field. The core idea is that areas with higher risk potential energy have greater uncertainty, but also higher information value or gains from inspection. Simultaneously, based on a high-precision multi-dimensional digital twin model, the entire inspection area is divided into a fine-grained path planning grid. For each passable edge in the grid, its corresponding inspection energy cost must be precisely calculated. This comprehensively considers the slope and surface material information provided by the digital twin model, as well as the robot dog's own dynamic model. Then, a path evaluation function is defined, aiming to find a path that maximizes the ratio of total information gain to total energy cost. Under the constraint of the robot dog's total energy consumption budget, through A... Path optimization algorithms, such as algorithms or genetic algorithms, are used to solve the evaluation function, ultimately generating an optimal inspection path. The advantage of this step is that it abandons the traditional fixed or random inspection mode and achieves intelligent path planning with optimal resources for high-risk areas.

[0073] S104. Control the inspection robot dog to perform inspections according to the optimal inspection path, and collect real-time environmental data around the inspection robot dog during the inspection process.

[0074] Once the optimal inspection path is generated, the system enters the autonomous execution phase in the physical world. The central control system sends this path data, containing a series of precise three-dimensional coordinates and corresponding action commands, to the inspection robot. Upon receiving the commands, the robot activates its autonomous navigation and control system. This system deeply integrates multi-source sensor information, including a BeiDou high-precision positioning module, an inertial measurement unit (IMU), lidar, and a visual camera, to achieve accurate tracking of the planned path and agile adaptation to complex terrain. During its journey along the path, various payloads carried by the inspection robot, such as a high-definition visible light camera, an infrared thermal imager, a gas sensor, and a microphone, continuously collect real-time environmental data, such as images of facility surfaces, temperature distribution, and air composition. All collected data is tagged with real-time, high-precision time and space labels provided by the BeiDou module. This step transforms the intelligent planning in the virtual space into reliable autonomous action and precise data collection in the physical world, providing a realistic and credible on-site input for subsequent risk verification and model closure.

[0075] S105. Compare the real-time environmental data with the environmental expectation value extracted from the risk potential field. When a cognitive surprise occurs as defined by the difference between the real-time environmental data and the environmental expectation value, online replanning of the inspection path is triggered to conduct local investigation of the anomaly point where the cognitive surprise occurs.

[0076] To address unexpected emergencies during the inspection process, an online path replanning mechanism based on cognitive surprise was introduced. When the inspection robot moves to a certain location and collects real-time environmental data, the system first uses its BeiDou positioning information to extract the expected environmental value for that location from the risk potential field. This expected value constitutes the prior belief distribution of the risk state at that point. Next, the newly acquired real-time environmental data is modeled as an observation likelihood function. Through Bayesian inference, combined with prior beliefs and observational data, the updated perception of the current location's risk state, i.e., the posterior belief distribution, is calculated. To quantify the degree of difference between old and new knowledge, the Kolbec-Leibler divergence was used. The system calculates a cognitive surprise value. When this value exceeds a preset surprise threshold, a cognitive surprise is identified and marked as an anomaly. At this point, the system automatically pauses the original inspection task, triggers online path replanning, and instructs the robot to conduct a more detailed local investigation of the anomaly. The effect is to give the inspection system the ability to react on the spot, enabling it to proactively discover and focus on potential risks that the model failed to predict.

[0077] S106. Obtain abnormal environmental data and BeiDou positioning information at abnormal points through the inspection robot dog, and feed the abnormal environmental data and BeiDou positioning information back to the graph neural network for iterative training in order to achieve closed-loop correction of the graph neural network.

[0078] When an anomaly is detected during inspection and a local investigation is completed, the entire system enters a closed-loop correction phase of learning and optimization. The inspection robot collects detailed environmental data at the anomaly point, along with its precise BeiDou positioning information and the external environmental influencing factors at that time, and packages it into a structured risk sample. This sample is assigned a risk type and severity level label determined by professionals or auxiliary algorithms. Subsequently, this labeled new sample is stored in the historical risk database, thereby expanding the training dataset of the graph neural network. When the number of accumulated new risk samples reaches a certain threshold, or when a preset iteration cycle is reached, the system automatically triggers a retraining process for the graph neural network. In retraining, the parameters of the network model are updated by minimizing the loss function on the entire dataset including the new samples. After training, this graph neural network model, which has learned new knowledge and has stronger performance, will replace the old model. The ultimate effect of this process is that the system can learn from each unexpected event it discovers, continuously improving its prediction accuracy of risk transmission and cumulative effects, realizing continuous self-evolution and closed-loop correction of the model, making the inspection increasingly intelligent.

[0079] In one implementation, constructing a multidimensional digital twin model of the water conservancy facility and assigning initial risk weights to the multidimensional digital twin model by combining historical data and facility structure information to form an initial risk potential energy field includes the following steps:

[0080] The original three-dimensional point cloud data and texture data of water conservancy facilities were collected using UAV oblique photography and ground three-dimensional laser scanning technology;

[0081] The original three-dimensional point cloud data is registered with geographic coordinates using BeiDou differential positioning technology to generate a multi-dimensional digital twin model with absolute geographic accuracy.

[0082] Structured risk information is extracted from design drawings, historical damage reports, and geological exploration data of water conservancy facilities;

[0083] Structured risk information is mapped to corresponding positions on a multidimensional digital twin model and quantified into initial risk weights;

[0084] An initial risk potential field covering the entire water conservancy facility is generated based on the initial risk weights at all locations.

[0085] In this embodiment, a drone equipped with an oblique photography camera flies over the facility from multiple angles with high overlap, capturing high-definition texture images of the top and facades from all directions. Simultaneously, a 3D laser scanner is deployed in key ground areas and blind spots of the drone's view, emitting laser pulses and receiving echoes to acquire hundreds of millions of 3D coordinate points on the facility surface at extremely high point frequencies, forming raw point cloud data. These two technologies complement each other: the drone oblique photography ensures the realism and integrity of the model's texture, while the ground laser scanning guarantees the geometric accuracy of key areas. To accurately map the isolated raw point cloud data to the real world, BeiDou differential positioning technology is used for geographic coordinate registration.

[0086] In practice, a BeiDou reference station is deployed at a known precise coordinate point near the water conservancy facility. This reference station continuously receives satellite signals and calculates error correction information. The rover carried by the UAV and ground scanning equipment synchronously receives the satellite signals and the differential correction signal emitted by the reference station, thereby calculating its own absolute geographic coordinates with centimeter-level precision in real time. By using these high-precision coordinates as control points, coordinate transformation is performed on the entire collected point cloud data, that is, the relative coordinate system of the model is transformed to a national or globally unified geodetic coordinate system. This coordinate transformation process can be represented as follows: ,in These are the local coordinates of the original point cloud, after being rotated by a rotation matrix. Translation vector After transformation, its absolute geographic coordinates are obtained. The final result is the generation of a multi-dimensional digital twin model, in which each point has real latitude, longitude, and elevation, achieving precise spatial unification between the virtual model and the physical entity.

[0087] After obtaining a precise geometric model, it is necessary to inject inherent risk attribute information into it. This step is achieved through in-depth analysis of the design drawings of the water conservancy facilities, historical operation and maintenance records, defect inspection reports, and preliminary geological exploration data. Information such as material strength, structural connection methods, and design loads can be extracted from the design drawings; the specific locations and severity of past leaks, cracks, and settlements can be identified from historical defect reports; and geological exploration data reveals the soil and rock types, fault distribution, and groundwater conditions of the foundation. With the assistance of expert interpretation and natural language processing technology, these unstructured descriptions scattered in text and charts are transformed into structured risk entries, such as "Dam Block 3 of XX Dam Section, a transverse crack was discovered in 2018, with a length of 2m and a depth of 0.1m." Next, the extracted structured risk information is precisely spatially correlated with the three-dimensional digital twin model and quantified as initial risk weights. For each risk entry in the knowledge base, its corresponding three-dimensional spatial location or region is marked on the digital twin model based on its geographical location description. After completing the spatial mapping, the qualitative risk description needs to be converted into quantitative values. Therefore, a multi-factor evaluation function can be established. This function takes into account the risk type. (e.g., cracks, leaks) and historical severity level Factors such as minor or severe risk are used to output a standardized initial risk weight for position i. Its value range is typically between 0 and 1. For example, an area with a history of severe leakage will have a significantly higher weight than an area that only has theoretical weaknesses in its design.

[0088] Finally, based on the initial risk weights at each discrete location on the model, a continuous initial risk potential field covering the entire water conservancy facility is generated. Since risk information typically exists only at specific points or regions, spatial interpolation algorithms are needed to extend it to obtain a global risk view. For example, the inverse distance weighting (IDW) method can be applied to determine the initial risk potential of any unweighted location x on the model surface. It is calculated by weighting and averaging multiple known risk points in its vicinity. The core idea is that the degree to which a point is affected by known risk points is inversely proportional to its distance from them. The calculation formula can be expressed as:

[0089]

[0090] in is the distance from location x to the known risk point i, and k is a power parameter of the distance. The final result of this process is to form an intuitive, continuous, static risk distribution map, which integrates the facility's design vulnerabilities and historical defects. High-potential areas represent inherently high-risk areas that need to be focused on, providing a basic risk map for subsequent inspection planning.

[0091] In one implementation, using knowledge graphs and graph neural networks to predict the transmission and accumulation effects of environmental risks around water conservancy facilities on the facility structure, and dynamically updating the initial risk potential field to a new risk potential field, includes the following steps:

[0092] A water conservancy knowledge graph is constructed by taking the components of water conservancy facilities as facility nodes and the physical or logical influence relationships between facility components as node edges.

[0093] Real-time environmental impact factors, including rainfall forecasts, upstream water levels, and soil moisture, are periodically obtained from external data sources.

[0094] Real-time environmental impact factors are loaded as input variables into the corresponding facility nodes of the water resources knowledge graph;

[0095] Using graph neural networks to perform reasoning on a water conservancy knowledge graph, we calculate the cumulative effect of risk energy caused by real-time environmental impact factors after transmission and aggregation along node edges.

[0096] The updated risk potential field is output based on the calculation results of the cumulative effect.

[0097] In this implementation, key physical components such as reservoir dams, spillways, and water pipelines are abstracted as facility nodes in the graph, while the physical connections, hydraulic relationships, or stress transfers between components are defined as edges connecting the nodes. For example, there is a pressure-bearing relationship between the upstream reservoir and the dam, and a flow-discharging relationship between the dam and the spillway. Each node and edge can be assigned attributes, such as the node's material, design life, and the edge's connection strength or maximum flow velocity. In this way, the massive and complex water conservancy project is transformed into a graph structure composed of nodes and edges that can be understood and processed by machines. To ensure the timeliness of risk assessment, it is necessary to periodically capture environmental changes that may affect facility safety from external data sources. By accessing application programming interfaces (APIs) of public service platforms such as meteorology and hydrology or proprietary monitoring networks, a series of key environmental impact factors can be automatically and in near real-time acquired. These factors mainly include 24-hour rainfall forecasts, real-time water level data of upstream rivers, and soil moisture monitoring values ​​around the dam. The acquired data, such as rainfall intensity and upstream water level, are combined into a feature vector representing the current external environmental state. This process is typically performed at a fixed frequency, such as once per hour, to ensure the freshness of the input data.

[0098] The acquired environmental impact factors must be accurately applied to their corresponding positions in the knowledge graph to initiate subsequent transmission analysis. This step loads the components of the external environmental feature vector onto the facility nodes directly associated with them in the water resources knowledge graph, serving as initial input features or state updates for these nodes. For example, upstream water level values ​​are loaded onto nodes representing the dam's upstream face and the reservoir area, while heavy rainfall forecasts primarily affect nodes related to watershed catchment and slope stability. Specifically, this involves updating the initial feature vectors of the corresponding nodes with the values ​​of these environmental factors. This step effectively maps external environmental events to specific points within the facilities, transforming abstract environmental pressure into initial risk incentives for specific nodes in the graph structure. A graph neural network model is then used to perform inference on the constructed water resources knowledge graph to calculate risk transmission and accumulation. The core mechanism of the graph neural network is to update node states by simulating information transfer between nodes.

[0099] In each round of computation, each node transmits its own risk information to all its connected neighbors and also receives risk information from all its neighbors. Then, each node aggregates the received information and updates its current state based on its state from the previous round. After multiple iterations, the risk energy propagates along the edges of the graph throughout the network. A simplified single-layer update rule can be represented as follows: ,in This refers to the risk state of node j in the new round. It is the set of its neighboring nodes. and It is the weight matrix learned by the model. It is an activation function. The effectiveness of this step is to quantitatively predict the distribution of risk among the various components of the facility under the current environmental input, revealing potential risk accumulation points.

[0100] Finally, based on the cumulative effect calculated by the graph neural network, a fully updated and dynamic risk potential field is output. The final output of the graph neural network is the cumulative risk value of each facility node in the knowledge graph. These discrete node risk values ​​need to be transformed back into a continuous risk field covering the entire three-dimensional space. By mapping the risk value of each node back to its physical location in the multidimensional digital twin model and using spatial interpolation methods, a dynamic risk increment field driven by environmental factors can be generated. This dynamic incremental field is combined with the original initial risk potential field. By performing weighted fusion, the final dynamic risk potential field is obtained. ,in It is a weighting coefficient that balances historical and real-time risks. The final result is a dynamic risk map that comprehensively considers the inherent defects of the facility and the current environmental pressures, enabling more precise guidance on the allocation of inspection resources.

[0101] In one implementation, generating the optimal inspection path for the inspection robot dog based on the risk potential energy field and combining the inspection information benefits and inspection energy consumption costs includes the following steps:

[0102] The updated risk potential energy field is transformed into an inspection information entropy field, where areas with higher risk potential energy have higher inspection information benefits.

[0103] The multidimensional digital twin model is divided into a path planning grid, and the corresponding inspection energy consumption cost is calculated for each feasible edge in the path planning grid.

[0104] Define the path evaluation function by using the benefits of inspection information and the energy cost of inspection as variables.

[0105] Given a total energy consumption budget constraint, the optimal inspection path is found by searching using a path optimization algorithm and a path evaluation function.

[0106] In this embodiment, the updated dynamic risk potential energy field is transformed into an inspection information entropy field. The core principle of this transformation is that the higher the risk potential energy of a region, the greater the uncertainty of its future state. Therefore, the amount of information, i.e., the information gain, that can be obtained by inspecting the real-time data of that region is also higher. One implementation method is to establish a mapping function to represent the dynamic risk potential energy value at each position x. Directly converted into corresponding information benefit value For example, a simple linear proportional relationship can be used for quantification: ,in This is a constant gain coefficient used to adjust the magnitude of information gain. By performing this operation on the entire risk potential field, an information gain map covering the entire area is generated. To enable the path planning algorithm to be computed in a structured environment, the continuous multidimensional digital twin model needs to be discretized into a path planning grid. This process divides the traversable surface area of ​​the entire hydraulic facility, such as the dam crest, inspection road, and slope platform, into a series of interconnected grid cells or nodes.

[0107] These nodes represent the possible locations where the inspection robot dog might stop, while the edges connecting adjacent nodes represent feasible movement paths. The key is to accurately calculate the corresponding inspection energy cost for each feasible edge j. Based on terrain data extracted from a digital twin model, the energy consumption required to overcome the slope of the road section, the energy consumption due to friction caused by different ground materials (such as concrete, grass, and dirt roads), and the basic travel distance energy consumption can be calculated. This step abstracts the complex real-world physical environment into a weighted graph, where nodes represent locations and edge weights represent energy costs. This transforms the physical movement constraints and energy consumption of the robot dog into a mathematical problem that can be analyzed and optimized in a computer. To find the optimal choice among numerous feasible paths, a unified path evaluation function must be defined, which comprehensively measures the benefits and costs of the inspection path. The goal of this function is to achieve the best balance between the two contradictory variables of inspection benefits and energy costs. An ideal path should consume as little energy as possible while passing through areas with high information benefits. An effective implementation method is to define a path evaluation function. Let be the ratio of the total information gain to the total energy cost of path p, i.e., the inspection efficiency. For a path p consisting of multiple nodes and edges, its evaluation function can be expressed as: ,in It is the information gain brought by node i along the path, and It is the energy cost of edge j contained in the path.

[0108] Given a total energy consumption budget constraint for a single inspection robot dog task, the goal is to find the inspection path with the highest score using a path optimization algorithm. This is a typical resource-constrained path planning problem. Due to the extremely large number of possible path combinations, it is difficult to find the optimal solution through exhaustive search. Therefore, heuristic search algorithms or metaheuristic algorithms are usually employed. For example, an improved greedy algorithm can be used, starting from the initial point and selecting the next node that brings the highest marginal utility (i.e., the additional information gain per unit of energy consumption) at each step, while ensuring that the total energy consumption does not exceed the budget B. Throughout the search process, the following constraints must always be satisfied: ,in This is the final selected optimal path. The ultimate result of this step is to output a specific and executable optimal inspection path that not only considers the highest risk points but also takes into account the energy consumption cost of reaching these points, ensuring that the inspection value is maximized within the limited operating time.

[0109] In one implementation method, calculating the corresponding inspection energy cost for each feasible edge in the path planning grid includes the following steps:

[0110] Extract the slope parameters and surface material type of each feasible edge in the path planning grid from the multidimensional digital twin model;

[0111] The energy consumption required to overcome the slope parameters by gravity is calculated based on the dynamic model of the inspection robot dog.

[0112] Based on the type of surface material, the corresponding friction coefficient is retrieved from the energy consumption database, and the friction energy consumption required for the inspection robot dog to overcome friction is calculated.

[0113] The basic travel energy consumption is calculated by combining the average travel energy consumption of the inspection robot dog and the length of the feasible edge.

[0114] The energy consumption cost of inspection is obtained by adding the energy consumption of overcoming gravity, the energy consumption of overcoming friction, and the energy consumption of basic travel.

[0115] In this embodiment, to accurately calculate the energy consumption of the inspection robot dog on different road sections, key physical environment parameters first need to be extracted from a high-precision multi-dimensional digital twin model. For each feasible edge in the path planning grid, by querying the three-dimensional coordinates of its starting and ending points, the length and vertical height difference of the road section can be calculated, thereby obtaining accurate slope parameters. Simultaneously, the surface material attribute layer stored in the digital twin model, endowed with high-definition drone textures and manual annotations, can be queried to determine the material category of the road segment, such as concrete, grass, or bare soil. This step effectively imbues the abstract path mesh with concrete physical properties, directly linking each virtual path segment to real-world terrain features (such as steepness and surface roughness), providing foundational data for subsequent energy consumption calculations based on the physics model. After obtaining the slope parameters of the path, the next step is to calculate the work done by the inspection robot dog to overcome gravity, i.e., the energy consumption to overcome gravity. This energy consumption calculation is based on the inspection robot dog's own dynamic model, where the most critical parameter is its total mass m. According to basic physics principles, when the robot dog moves along a slope of... When moving upwards along the slope, a gravitational force must be overcome, acting downwards along the slope. This energy consumption is calculated by multiplying this gravitational force by the distance L traveled along the slope. Therefore, for a feasible edge, the energy consumption for overcoming gravity is... It can be represented as Where g is the acceleration due to gravity. If the robot dog is moving downhill, the slope angle is... A negative value indicates that energy consumption is negative, theoretically meaning that potential energy can be recovered or that travel is less strenuous. This step quantifies the core impact of terrain undulations on energy consumption, enabling route planning to instinctively avoid unnecessary uphill climbs.

[0116] Besides gravity, friction is another major source of energy consumption, and its magnitude is closely related to the surface material. Based on the surface material type obtained in the first step, the system will query a pre-established energy consumption database to obtain the corresponding coefficient of friction for that material. This database was built through experimental testing of the robot dog on various real-world road surfaces. The magnitude of friction is equal to the product of the coefficient of friction and the normal force perpendicular to the contact surface. On a slope of... On the inclined plane, the normal force is another component of gravity. Therefore, the energy required to overcome friction is... The formula is derived by multiplying the frictional force by the path length. Even under ideal conditions with no slope and no friction, the inspection robot still consumes energy to maintain its operation and drive its motor. This energy consumption is called basic travel energy consumption. It mainly includes the inherent power consumption of the robot's internal computing unit, sensors, communication module, and drive system when traveling at an average speed on a level surface. This parameter is usually obtained through calibration experiments and is defined as the basic energy consumption value per unit distance. Therefore, for a feasible edge of length L, its basic travel energy consumption is... It can be calculated using a simple linear relationship: This step ensures that the total path length is always a component of energy cost, preventing the algorithm from choosing an excessively winding and long path, even if it is flat, in some cases.

[0117] Finally, by summing up the energy consumption calculations for each item in the previous steps, we can obtain the total energy cost of the inspection robot dog traversing the feasible edge. This total cost is a comprehensive reflection of all the energy expenditure required by the robot dog on this road segment, taking into account the terrain challenges, road resistance, and travel distance. The calculation formula is a direct summation: By performing this complete computation process on every feasible edge in the path planning grid, a global path graph with energy consumption weights is ultimately obtained. The final effect of this step is to provide a crucial, quantitative decision basis for subsequent optimal path planning algorithms, namely the cost of each segment of the journey, thereby ensuring that the final generated inspection path is physically feasible and energy efficient.

[0118] In one implementation, triggering online replanning of the inspection path when a cognitive unexpected event occurs, defined by the difference between real-time environmental data and expected environmental values, to conduct local investigation of the anomaly points where the cognitive unexpected event occurred, includes the following steps:

[0119] When the inspection robot dog acquires real-time environmental data, the current location corresponding to the real-time environmental data is determined by the inspection robot dog's Beidou positioning.

[0120] Extract the expected risk value of the current location from the risk potential field, and construct the prior belief distribution of the risk state of the current location based on the expected risk value;

[0121] Real-time environmental data is modeled as an observation likelihood function. Bayesian inference is applied, and the posterior belief distribution of the current location risk state is calculated by combining the prior belief distribution with the observation likelihood function.

[0122] The cognitive surprise value caused by the difference between real-time environmental data and environmental expectations is quantified by calculating the Kolb-Leibler divergence between the posterior belief distribution and the prior belief distribution.

[0123] When the cognitive error value exceeds the preset error threshold, the current position is designated as the abnormal point where the cognitive error occurred, and the online replanning of the inspection robot dog's inspection route is triggered.

[0124] In this implementation, while the inspection robot dog moves along a predetermined path and collects real-time environmental data using its sensor suite, its onboard BeiDou high-precision positioning module assigns a geographic coordinate accurate to the centimeter level to each frame of data. This coordinate is the sole link between the data and space. Whether it's a water stain captured by a high-definition camera or a temperature anomaly detected by an infrared thermal imager, it will be immediately labeled with a three-dimensional spatial tag containing longitude, latitude, and elevation information. The effect of this step is to ensure that every piece of on-site data collected in the physical world can find an accurate mapping point in its corresponding multi-dimensional digital twin model, achieving spatiotemporal synchronization between real observation and virtual model, which is the basis for subsequent differential analysis. With the precise location of the real-time data, the expected risk value of that location can be extracted from the dynamic risk potential energy field and transformed into a probabilistic expression.

[0125] Specifically, using the geographical coordinates obtained in the previous step, in the risk potential energy field A risk prediction value can be obtained by querying the query. This value represents the most likely risk state at that point, inferred by the model based on historical data and the environment. However, any model prediction contains uncertainty, so it cannot be used as an absolute value. Therefore, a prior belief distribution is constructed based on this expected value, which can usually be assumed to be a normal distribution. Where R is the risk state variable, and the variance is... This represents the degree of uncertainty the model has regarding its own predictions. Next, the real-time environmental data collected by the inspection robot is also described using a probabilistic model and combined with prior beliefs to update the understanding of the current location's risk status. The collected real-time data, such as temperature readings or crack widths... Because sensors inherently possess measurement errors, these errors are also modeled as a probability distribution, namely the observation likelihood function, for example... ,in This represents the measurement uncertainty of the sensor.

[0126] Then, applying the core rules of Bayesian inference, the prior beliefs are multiplied by the observed likelihood to obtain the posterior belief distribution. It is proportional to The benefit of this step is that it integrates information from both model predictions and real observations, generating a more accurate and updated perception of the risk state. To quantify the new information brought by the real observation data—that is, the degree of cognitive surprise—it is necessary to calculate the magnitude of the change in the posterior belief distribution relative to the prior belief distribution. The Kolb-Leibler divergence (KL divergence) is an ideal tool for measuring the difference between two probability distributions. This is achieved by calculating the KL divergence value from the prior distribution to the posterior distribution. This yields a non-negative value that precisely quantifies the extent to which observed data compels changes in the system's perception. The formula for its calculation is: If the observed data closely matches the expectations, then the posterior and prior distributions are very close. The value approaches 0; conversely, if an observation deviates significantly from expectations, The value will increase significantly.

[0127] Finally, a decision is made based on the calculated cognitive surprise value to determine whether the current task needs to be interrupted to deal with the unexpected situation. The calculated cognitive surprise value is compared with a pre-set surprise threshold derived from expert experience or historical data statistics. If the cognitive surprise value is greater than the surprise threshold, it is determined that a significant cognitive surprise has occurred at the current location, meaning that the actual situation here is far beyond the model's prediction and there is a potential unknown risk. At this time, the location will be automatically marked as an anomaly by the system, and an interruption signal will be immediately triggered to suspend the original inspection plan, start the online replanning program of the inspection path, and instruct the inspection robot to conduct a more detailed local survey of the anomaly. The ultimate effect of this step is to endow the inspection system with intelligent alertness and adaptive capabilities, enabling it to proactively identify and focus on real surprises beyond the model's prediction, thereby avoiding repeated and inefficient inspections of known risks.

[0128] In one implementation, triggering online replanning of the inspection robot dog's inspection route includes the following steps:

[0129] Delineate a local exploration area centered on the anomaly point;

[0130] A local risk potential energy field is generated within a local survey area using real-time environmental data.

[0131] The local risk potential energy field is transformed into a local inspection information entropy field, and the local energy consumption cost of the inspection robot dog in the local inspection information entropy field is calculated.

[0132] By combining the local inspection information entropy field and local energy consumption cost, and using the optimal path algorithm, a local inspection sub-path for the inspection robot dog is generated for the local inspection area.

[0133] In this implementation, when an anomaly is triggered, a local survey area needs to be delineated spatially, centered on the anomaly, for targeted investigation. This area is not fixed but dynamically adjusted based on the magnitude of the perceived unexpectedness; that is, the higher the degree of unexpectedness, the larger the delineated survey radius or range, ensuring complete coverage of the potential problem's spatial impact. In practice, a region with a radius of [missing information] is generated, using the anomaly's BeiDou positioning coordinates as the geometric center. spherical or circular area , where the radius It is an increasing function of the unexpected value. The effect of this step is to transform the global, aimless search problem into a local, well-defined, detailed exploration problem, which greatly focuses computing resources and the attention of the inspection robot.

[0134] Within the designated local survey area, a more refined and timely local risk potential energy field needs to be generated. This process no longer relies on global, model-based predictions, but is entirely driven by real-time environmental data acquired through rapid scanning of the area by a robotic inspection dog. The robot dog utilizes its onboard sensors, such as LiDAR or structured light cameras, to perform a rapid 3D scan and data acquisition of the area, obtaining high-density point cloud or image data. After processing, it is directly mapped to a local risk value through a preset local risk assessment function. Here, y is any point within the local area. The result of this step is the creation of a high-resolution local risk map based on firsthand field data. This map reveals subtle anomalies that the global model fails to capture, providing the most reliable basis for subsequent exploration route planning. To plan the most efficient local exploration route, the newly generated local risk potential energy field needs to be transformed into a decision space that includes benefits and costs. Similar to global planning, the local risk potential energy field is first transformed into a local inspection information entropy field through a mapping function. Points with higher risk have higher information benefits. At the same time, the grid data corresponding to this local area in the multidimensional digital twin model is retrieved, and the local energy consumption cost of each feasible edge within the area is recalculated. This calculation process is based on the same principle as the global energy consumption calculation, but it operates on a smaller, more refined grid, also considering factors such as slope and surface material.

[0135] Finally, combining the calculated local inspection information entropy field and local energy consumption cost, an optimal path algorithm is used to generate a local reconnaissance sub-path specifically for this area for the inspection robot dog. The goal of this algorithm is to access the highest information-gaining points within this small area with minimal energy cost. It can be modeled as a local variation of the traveling salesman problem, solving for the shortest loop or path that effectively covers all high-risk points. This sub-path is generated as a series of dense 3D coordinate points, guiding the inspection robot dog to perform a series of precise actions such as moving, hovering, and adjusting sensor angles to conduct comprehensive, blind-spot-free data collection of the anomaly area. The ultimate result of this step is the output of an immediately executable and efficient deep reconnaissance plan, ensuring that the investigation of unknown anomalies is systematic and thorough, rather than random.

[0136] In one embodiment, after conducting a local investigation of the anomaly where cognitive surprise occurred, the following steps are also included:

[0137] Acquire local environmental data collected by the inspection robot dog during local reconnaissance sub-paths;

[0138] Analyze local environmental data and quantify the risk severity level of anomalies;

[0139] Determine whether the severity level of the risk exceeds the preset global impact threshold;

[0140] If the global impact threshold is not exceeded, the inspection robot dog is controlled to return to the breakpoint of the original optimal inspection path and continue to execute the task.

[0141] If the global impact threshold is exceeded, the remaining part of the original optimal inspection path is discarded, the global risk potential field is updated based on the risk severity level and local environmental data, and the adaptive inspection path of the inspection robot dog is regenerated from the anomaly point location.

[0142] In this implementation, after the inspection robot dog completes its local reconnaissance sub-path for the anomaly point, it first needs to acquire and integrate all the high-density, multimodal local environmental data collected during this process. This includes detailed 3D point clouds generated by lidar, image sequences captured by high-definition cameras, and temperature distribution maps recorded by infrared thermal imagers. This data is uniformly aggregated and bound to precise BeiDou positioning information to form a comprehensive dataset about the anomaly area. This dataset differs from the sparse data previously collected along the global path; it provides unprecedented detail in depicting the anomaly, forming the basis for in-depth analysis and accurate assessment. After acquiring detailed local environmental data, it needs to be processed using specialized analysis algorithms to quantify the actual risk severity level of the anomaly point. This process involves extracting key features from the dataset. For example, for a crack, the algorithm automatically calculates its length, maximum width, and depth from the 3D point cloud; for a leak, it analyzes its affected area and temperature difference. Then, these extracted multidimensional features are... Inputting this into a pre-trained risk assessment model or a multi-criteria decision function will produce a standardized risk severity rating. This rating By weighting and combining various features, the complex on-site situation is transformed into a single, intuitive numerical value.

[0143] The next step is to compare the quantified risk severity level with a key decision threshold to determine the nature of the anomaly. This preset global impact threshold... It is not a fixed value; it is set by hydraulic engineering experts based on the importance of the facility, its structural characteristics, and safety regulations. It represents the critical point at which a localized problem may threaten the safety or stable operation of the entire facility structure. For example, the severity level of a tiny surface crack may be far below this threshold, while a penetrating, continuously expanding crack may far exceed it. The judgment process involves a direct numerical comparison: if... If the anomaly is localized and non-urgent, then it is considered a localized, non-urgent problem; otherwise, if If so, it is considered to have potential global impact.

[0144] If the assessment indicates that the severity of the anomaly does not exceed the global impact threshold, it means that while the issue needs to be recorded and monitored, it does not pose an urgent threat and is insufficient to alter the overall inspection strategy. In this case, the control system will instruct the inspection robot to terminate its local reconnaissance task. Subsequently, the system will plan the most efficient path from its current location back to the interruption point of the original optimal inspection path. Once the robot reaches this interruption point, it will seamlessly resume the remaining portion of the suspended original inspection task. This step ensures the continuity and integrity of the inspection task, allowing it to return to the main task with minimal time and energy costs after handling a localized emergency, thus ensuring that the overall inspection objective is not significantly affected.

[0145] Conversely, if the risk severity level exceeds the global impact threshold, it indicates the discovery of a major hidden danger, rendering the original inspection plan, based on outdated risk perception, inapplicable. In this case, the system immediately discards all unexecuted portions of the original optimal inspection path. Next, using the newly quantified high-risk severity level and detailed local environmental data, the global risk potential field is forcibly updated, forming a new, significant high-risk peak at the anomaly location. Then, using this anomaly point as the new starting point, and based on the updated global risk map and remaining energy budget, the optimal path planning algorithm is restarted to generate a completely new, adaptive inspection path. The ultimate effect of this step is the dynamic and intelligent reshaping of the inspection strategy, ensuring that limited inspection resources can be immediately reallocated, prioritizing the most severe and newly discovered risks.

[0146] In one implementation, the abnormal environmental data and BeiDou positioning information at the abnormal points are obtained by the inspection robot dog, and the abnormal environmental data and BeiDou positioning information are fed back to the graph neural network for iterative training to achieve closed-loop correction of the graph neural network. The steps include the following:

[0147] The inspection robot dog obtains abnormal environmental data and BeiDou positioning information at abnormal points.

[0148] Abnormal environmental data, BeiDou positioning information, and real-time environmental impact factors are packaged into structured risk samples;

[0149] Assign a label to each risk sample, consisting of the risk type and severity level;

[0150] Labeled risk samples are stored in a pre-defined historical risk database to expand the training dataset of the graph neural network.

[0151] The retraining process of the graph neural network is triggered when the preset iteration cycle is reached or the number of accumulated risk samples reaches a threshold.

[0152] The original graph neural network is replaced with the retrained graph neural network to complete the closed-loop correction of the risk potential field assessment.

[0153] In this implementation, after completing the local in-depth investigation of the anomaly point, to enable the model's self-learning and evolution, the discovery first needs to be solidified into a learnable experience. The inspection robot dog will package all the key environmental data it collects at the anomaly point, such as 3D point clouds describing crack morphology and infrared thermal images reflecting leakage, along with the precise BeiDou positioning information of that point, into a raw data package. This data package is an objective, first-hand record of this unexpected event. To enable the machine learning model to understand this event, the raw data package needs to be processed into a structured, complete sample containing cause and effect. Specifically, the environmental data and BeiDou positioning information of the anomaly point are integrated with the global real-time environmental impact factors that triggered the anomaly, such as the rainfall and upstream water level at that time. This forms a structured risk sample S containing environmental input, spatial location, and anomaly result. This sample completely describes what kind of anomaly occurred under what environmental conditions, at what specific location, and in what way. To enable the model to perform supervised learning, this structured risk sample must be assigned a correct answer, i.e., a label.

[0154] This label is typically assigned by water conservancy experts after reviewing abnormal data in the background, or by an auxiliary artificial intelligence classification model followed by expert verification. The label Y is usually a combination of two parts: risk type (e.g., crack, settlement, leakage) and quantified risk severity level. Through this process, an unlabeled sample S becomes a complete sample pair (S, Y) with the correct answer, suitable for training. The effect of this step is to provide a clear learning objective for the machine learning model: to accurately predict the label Y when the model encounters inputs similar to S in the future. A single training sample is insufficient to improve the overall performance of the model; it must be incorporated into a continuously growing knowledge base. New risk samples with labels are stored in a historical risk database specifically for this purpose. This database acts as the long-term memory of the entire system; each discovery and labeling of a new anomaly adds a valuable record to this database. This process can be represented as a training dataset. Expansion: The benefit of this step is the creation of a dynamic and continuously enriched training dataset that reflects the latest state changes of the facility and all known risk patterns, providing a constant source of nourishment for the continuous evolution of the model.

[0155] To ensure the model's performance remains up-to-date while avoiding unnecessary computational overhead, the retraining process is triggered by specific conditions. The system employs two triggering mechanisms: one is a time period, such as retraining every month regardless of data volume; the other is a data volume threshold, where retraining is initiated when the number of newly accumulated labeled risk samples reaches a preset value, such as 100. Once triggered, the system accesses the entire updated historical risk database and adjusts the internal parameters of the graph neural network using the backpropagation algorithm, aiming to minimize the model's prediction error across the entire dataset. This step ensures that the model's iterative updates are efficient and meaningful, accumulating sufficient new knowledge before systematic learning and integration. Finally, when the retraining process is complete, a new graph neural network model with optimized parameters and stronger predictive capabilities is generated. At this point, the system performs a crucial replacement operation: replacing the old model currently in the online service with this newly trained model. This deployment process ensures that the system's core risk assessment engine is always the latest and most intelligent version. Thus, the entire closed loop from anomaly detection and analysis to model upgrade is completed. The ultimate result is that the system learns from an accident and achieves self-correction. When it conducts the risk potential field assessment next time, it will be able to predict similar situations more accurately, thereby transforming future cognitive accidents into known risk expectations and realizing the continuous evolution of the entire inspection intelligence system.

[0156] This invention also discloses a water conservancy inspection robot dog intelligent inspection system based on Beidou AI, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the water conservancy inspection robot dog intelligent inspection method based on Beidou AI as described in any of the above embodiments.

[0157] The processor can be a central processing unit (CPU). Of course, depending on the actual use, it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc., and this application does not limit it.

[0158] The memory can be an internal storage unit of a computer device, such as a hard disk or RAM, or an external storage device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD), or flash memory card (FC) provided on the computer device. Furthermore, the memory can be a combination of internal storage units and external storage devices of a computer device. The memory is used to store computer programs and other programs and data required by the computer device. The memory can also be used to temporarily store data that has been output or will be output. This application does not limit this.

[0159] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0160] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A water conservancy inspection robot dog intelligent inspection method based on Beidou AI, characterized in that, Includes the following steps: Construct a multidimensional digital twin model of water conservancy facilities, and combine historical data and facility structure information of water conservancy facilities to assign initial risk weights to the multidimensional digital twin model to form an initial risk potential energy field; A water conservancy knowledge graph is constructed by taking the components of water conservancy facilities as facility nodes and the physical or logical influence relationships between facility components as node edges. Real-time environmental impact factors, including rainfall forecasts, upstream water levels, and soil moisture, are periodically obtained from external data sources. Real-time environmental impact factors are loaded as input variables into the corresponding facility nodes of the water resources knowledge graph; Using graph neural networks to perform reasoning on a water conservancy knowledge graph, we calculate the cumulative effect of risk energy caused by real-time environmental impact factors after transmission and aggregation along node edges. The updated risk potential field is output based on the calculation results of the cumulative effect; Based on the risk potential energy field and combined with the inspection information benefits and inspection energy consumption costs of the inspection robot dog, the optimal inspection path is generated for the inspection robot dog. The robot dog is controlled to perform inspections according to the optimal inspection path, and real-time environmental data around the robot dog is collected during the inspection process. The real-time environmental data is compared with the environmental expectation value extracted from the risk potential field. When a cognitive surprise occurs as defined by the difference between the real-time environmental data and the environmental expectation value, the online replanning of the inspection path is triggered to conduct local inspection of the anomaly point where the cognitive surprise occurs. The inspection robot dog acquires abnormal environmental data and BeiDou positioning information at abnormal points, and feeds the abnormal environmental data and BeiDou positioning information back to the graph neural network for iterative training in order to achieve closed-loop correction of the graph neural network.

2. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 1, characterized in that, The process of constructing a multidimensional digital twin model of water conservancy facilities and assigning initial risk weights to the multidimensional digital twin model by combining historical data and facility structure information to form an initial risk potential energy field includes the following steps: The original three-dimensional point cloud data and texture data of water conservancy facilities were collected using UAV oblique photography and ground three-dimensional laser scanning technology; The original three-dimensional point cloud data is registered with geographic coordinates using BeiDou differential positioning technology to generate a multi-dimensional digital twin model with absolute geographic accuracy. Structured risk information is extracted from design drawings, historical damage reports, and geological exploration data of water conservancy facilities; Structured risk information is mapped to corresponding positions on a multidimensional digital twin model and quantified into initial risk weights; An initial risk potential field covering the entire water conservancy facility is generated based on the initial risk weights at all locations.

3. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to any one of claims 1 and 2, characterized in that, The process of generating the optimal inspection path for the inspection robot dog based on the risk potential energy field and combining the inspection information benefits and inspection energy consumption costs includes the following steps: The updated risk potential energy field is transformed into an inspection information entropy field, where areas with higher risk potential energy have higher inspection information benefits. The multidimensional digital twin model is divided into a path planning grid, and the corresponding inspection energy consumption cost is calculated for each feasible edge in the path planning grid. Define the path evaluation function by using the benefits of inspection information and the energy cost of inspection as variables. Given a total energy consumption budget constraint, the optimal inspection path is found by searching using a path optimization algorithm and a path evaluation function.

4. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 3, characterized in that, The calculation of the inspection energy consumption cost for each feasible edge in the path planning grid includes the following steps: Extract the slope parameters and surface material type of each feasible edge in the path planning grid from the multidimensional digital twin model; The energy consumption required to overcome the slope parameters by gravity is calculated based on the dynamic model of the inspection robot dog. Based on the type of surface material, the corresponding friction coefficient is retrieved from the energy consumption database, and the friction energy consumption required for the inspection robot dog to overcome friction is calculated. The basic travel energy consumption is calculated by combining the average travel energy consumption of the inspection robot dog and the length of the feasible edge. The energy consumption cost of inspection is obtained by adding the energy consumption of overcoming gravity, the energy consumption of overcoming friction, and the energy consumption of basic travel.

5. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 1, characterized in that, The step of triggering online replanning of the inspection path when a cognitive unexpected event occurs, defined by the difference between real-time environmental data and expected environmental values, to conduct local investigation of the anomaly points where the cognitive unexpected event occurred, includes the following steps: When the inspection robot dog acquires real-time environmental data, the current location corresponding to the real-time environmental data is determined by the inspection robot dog's Beidou positioning. Extract the expected risk value of the current location from the risk potential field, and construct the prior belief distribution of the risk state of the current location based on the expected risk value; Real-time environmental data is modeled as an observation likelihood function. Bayesian inference is applied, and the posterior belief distribution of the current location risk state is calculated by combining the prior belief distribution with the observation likelihood function. The cognitive surprise value caused by the difference between real-time environmental data and environmental expectations is quantified by calculating the Kolb-Leibler divergence between the posterior belief distribution and the prior belief distribution. When the cognitive error value exceeds the preset error threshold, the current position is designated as the abnormal point where the cognitive error occurred, and the online replanning of the inspection robot dog's inspection route is triggered.

6. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 5, characterized in that, The online replanning of the inspection route for the robot dog includes the following steps: Delineate a local exploration area centered on the anomaly point; A local risk potential energy field is generated within a local survey area using real-time environmental data. The local risk potential energy field is transformed into a local inspection information entropy field, and the local energy consumption cost of the inspection robot dog in the local inspection information entropy field is calculated. By combining the local inspection information entropy field and local energy consumption cost, and using the optimal path algorithm, a local inspection sub-path for the inspection robot dog is generated for the local inspection area.

7. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 6, characterized in that, After conducting a localized investigation of the anomalies that revealed unexpected cognitive events, the following steps are also included: Acquire local environmental data collected by the inspection robot dog during local reconnaissance sub-paths; Analyze local environmental data and quantify the risk severity level of anomalies; Determine whether the severity level of the risk exceeds the preset global impact threshold; If the global impact threshold is not exceeded, the inspection robot dog is controlled to return to the breakpoint of the original optimal inspection path and continue to execute the task. If the global impact threshold is exceeded, the remaining part of the original optimal inspection path is discarded, the global risk potential field is updated based on the risk severity level and local environmental data, and the adaptive inspection path of the inspection robot dog is regenerated from the anomaly point location.

8. The intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI according to claim 1, characterized in that, The process of acquiring abnormal environmental data and BeiDou positioning information at abnormal points using a patrol robot dog, and then feeding this data back to the graph neural network for iterative training to achieve closed-loop correction, includes the following steps: The inspection robot dog obtains abnormal environmental data and BeiDou positioning information at abnormal points. Abnormal environmental data, BeiDou positioning information, and real-time environmental impact factors are packaged into structured risk samples; Assign a label to each risk sample, consisting of the risk type and severity level; Labeled risk samples are stored in a pre-defined historical risk database to expand the training dataset of the graph neural network. The retraining process of the graph neural network is triggered when the preset iteration cycle is reached or the number of accumulated risk samples reaches a threshold. The original graph neural network is replaced with the retrained graph neural network to complete the closed-loop correction of the risk potential field assessment.

9. A water conservancy inspection robot dog intelligent inspection system based on Beidou AI, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the intelligent inspection method for water conservancy inspection robot dogs based on Beidou AI as described in any one of claims 1 to 8.

Images