An underwater robot dynamic cooperative control method and system fusing multi-source data

Abstract

The application discloses a kind of underwater robot dynamic cooperative control method and system of fusion multi-source data, it is related to general control or regulating system field.In the method, control inspection unmanned plane obtains the laser point cloud and multispectral data of reservoir dam and water table, constructs centimeter level global space datum;Abnormal detection algorithm based on Mahalanobis distance is used to identify abnormal water surface area, and generate cooperative exploration task package;After receiving task package, underwater robot realizes centimeter level accurate positioning by permanent magnet marker array and visual positioning, and carries out surface observation on underwater dam face;Water quality sampling unmanned plane automatically collects stratified water sample based on vertical water quality profile data;Collect multi-element heterogeneous data, construct causal network model by principal component analysis, bayesian inference and granger causality test, and generate comprehensive diagnosis report.The application greatly improves positioning accuracy and diagnosis efficiency, and reduces operation and maintenance cost.

Description

Technical Field

[0001] This invention relates to the field of general control or regulation systems, specifically to a dynamic cooperative control method and system for underwater robots that integrates multi-source data. Background Technology

[0002] As a critical national water conservancy infrastructure, the structural safety of reservoir dams and the quality of the reservoir's aquatic ecosystem directly affect downstream public safety, drinking water safety, and the sustainable use of water resources. However, current reservoir dam operation and maintenance practices still face several prominent technical bottlenecks:

[0003] First, underwater inspection positioning accuracy is insufficient. Traditionally, underwater structural exploration relies on underwater robots. However, since GPS signals are unusable underwater, robots can only rely on inertial navigation systems and Doppler logs for dead reckoning. This method inevitably introduces cumulative errors, which, in practice, have been shown to reach 0.1%-0.5% of the distance traveled per hour. When inspecting reservoirs and dams hundreds of meters long, this error means that structural defects such as cracks and leaks discovered cannot be accurately correlated with design drawings or BIM models, creating significant difficulties for subsequent maintenance positioning and construction.

[0004] Secondly, tracing the source of aquatic ecological anomalies is difficult. Current water quality monitoring methods mostly combine fixed shore-based monitoring stations with periodic manual boat sampling. When fixed monitoring stations detect abnormalities in parameters such as temperature, chlorophyll, and dissolved oxygen in a certain area of ​​water, it is difficult to quickly and accurately determine whether the anomaly is caused by underwater structural defects. In other words, current technology is severely lacking in the ability to accurately and quickly trace the source of underwater problems.

[0005] Furthermore, the existing unmanned systems suffer from low levels of collaboration. Although drones, underwater robots, and other equipment are widely used, their collaboration largely remains at the level of simple command transmission and independent data collection, lacking a truly intelligent collaborative mechanism. The data collected by different platforms are inconsistent in format and spatiotemporal reference, requiring significant manual intervention for data alignment and comprehensive analysis, which severely restricts operational efficiency and automation levels.

[0006] Therefore, there is an urgent need for an innovative technology and system that can achieve high-precision underwater positioning, intelligent air-water integration, and causal correlation analysis of structural defects and ecological anomalies. Summary of the Invention

[0007] The purpose of this invention is to provide a dynamic collaborative control method and system for underwater robots that integrates multi-source data. By establishing a unified spatial benchmark and data fusion mechanism, it enables correlation analysis and accurate diagnosis of dam structural status and aquatic ecological environment, thereby overcoming the aforementioned problems in the prior art.

[0008] To achieve the objective, the present invention provides the following technical solution:

[0009] A dynamic cooperative control method for underwater robots that integrates multi-source data includes the following steps:

[0010] Step S1: Control the inspection drone to fly along the preset route of the reservoir dam and adjacent waters to acquire laser point cloud and multispectral data of the reservoir dam and water surface; register the laser point cloud with the BIM model of the reservoir dam to construct a global spatial reference.

[0011] Step S2: Analyze the multispectral data using an anomaly detection algorithm based on Mahalanobis distance to identify abnormal water surface areas; automatically generate a collaborative exploration task package containing the geographical location, range, type, and priority of the abnormal water surface areas;

[0012] Step S3: After receiving the collaborative exploration task package, the underwater robot autonomously navigates to the abnormal water surface area specified in the collaborative exploration task package; when the underwater robot arrives at the abnormal water surface area, the underwater dam surface image corresponding to the abnormal water surface area captured by the underwater robot is visually matched and pose calculated with the reference image of the BIM model to achieve centimeter-level precise positioning; under the guidance of the centimeter-level precise positioning information, the underwater robot is controlled to conduct close-up observation of the underwater dam surface in the abnormal water surface area, and the defects found are photographed and measured from multiple angles, recording their precise position, length and width;

[0013] Step S4: The water quality sampling drone receives vertical water quality profile data from the underwater robot in real time, automatically identifies the characteristic depth based on preset trigger rules, and collects water samples at that characteristic depth.

[0014] Step S5: Collect the multivariate heterogeneous data from Steps S1 to S4, and use principal component analysis to reduce data dimensionality and extract features; combine Bayesian inference and Granger causality test methods to construct a causal network model between structural defect parameters and aquatic ecological anomaly parameters, quantitatively identify key correlation factors, and automatically generate a comprehensive diagnostic report.

[0015] Furthermore, in step S1, the laser point cloud is registered with the BIM model of the reservoir dam to construct a global spatial reference, which specifically includes the following sub-steps: Step S1.1: Perform point cloud calculation and denoising on the laser point cloud;

[0016] Step S1.2: Select several common feature points in the laser point cloud and BIM model; calculate the coordinate transformation matrix using these common feature points to initially align the laser point cloud and BIM model; Step S1.3: Find the nearest point pair between the triangular mesh surfaces of the laser point cloud and BIM model, and iteratively calculate the optimal rigid body transformation using the least squares method to minimize the root mean square error between the two point sets, and finally establish a centimeter-level global spatial benchmark.

[0017] Further, in step S2, an anomaly detection algorithm based on Mahalanobis distance is used to analyze the multispectral data and identify abnormal water surface areas. Specifically, this includes the following sub-steps: Step S2.1: Aerial triangulation and orthophoto stitching are performed on the original multispectral images obtained in step S1 to generate reflectance images for each band; Step S2.2: For the thermal infrared band, the radiative transfer equation method is used, combined with real-time meteorological data, to invert and obtain the absolute temperature distribution map of the water surface; using the reflectance of the red and near-infrared bands, a variant algorithm of the normalized difference vegetation index is used to invert and generate a chlorophyll a concentration distribution map; using the reflectance ratio of the blue and red light bands, an empirical or semi-empirical model is used to invert and obtain a turbidity distribution map; Step S2.3: The three distribution maps are read, and for each pixel with the same geographical location... The values ​​of its three bands are used to construct a three-dimensional feature vector, thus obtaining a three-dimensional vector set representing the entire inspection area. A set of pixels is randomly selected from the three-dimensional vector set as a training sample, and its mean vector and covariance matrix are calculated to construct a statistical model for characterizing the characteristics of normal water areas. The mean vector represents the average level of temperature, chlorophyll, and turbidity in the entire area, and the covariance matrix represents the relationship between these three parameters. Each pixel in the image is traversed, and its Mahalanobis distance to the distribution defined by the statistical model is calculated. Pixels whose Mahalanobis distance exceeds the preset value are judged as abnormal pixels. Step S2.4: DBSCAN spatial clustering is performed on the abnormal pixels to obtain abnormal water surface areas. Detailed information of the abnormal water surface areas is extracted from the datasets of steps S2.1 to S2.3 and packaged into a task package.

[0018] Furthermore, in step S3, multiple permanent magnet markers are pre-deployed on the underwater wall of the dam using a topology structure combining equally spaced grids and key point encryption, forming a permanent magnet marker array; step S3 specifically includes the following sub-steps: Step S3.1: The underwater robot is deployed from the mother ship, and its initial position is given by the mother ship's GNSS; after entering the water, the underwater robot mainly relies on the inertial navigation system and Doppler log for combined navigation; Step S3.2: When the inertial navigation system instructs the underwater robot to navigate to a preset distance from the previous calibration point and approach a... When pre-placed magnetic markers are installed, the triaxial magnetometer array begins to monitor the regular changes in the strength and direction of the spatial magnetic field, acquiring a magnetic field vector map. The inertial navigation system matches the real-time measured magnetic field vector map with a pre-stored magnetic field grid map. Through a least squares algorithm, it uniquely determines that the underwater robot is currently located near a permanent magnet marker. When the match is successful, the navigation system uses the precise coordinates of the permanent magnet marker recorded in the BIM model as a high-confidence observation to provide feedback correction to the system state of the inertial navigation system, thereby suppressing the cumulative error of the inertial navigation system.

[0019] Furthermore, step S3 specifically includes the following sub-steps: Step S3.3: When the underwater robot arrives at the abnormal water surface area, the visual-assisted navigation system is activated, and the camera captures the underwater dam surface image corresponding to the abnormal water surface area, and extracts multiple key points and their multi-dimensional feature descriptors; these multi-dimensional feature descriptors are matched with the reference image feature library pre-rendered in the area by the BIM model, and the coordinates and attitude of the current underwater robot are calculated by the EPnP algorithm to achieve centimeter-level precise positioning; Step S3.4: Under the guidance of the centimeter-level precise positioning information, the operator controls the underwater robot to conduct close-up observation of the underwater dam surface in the abnormal water surface area, and takes multi-angle pictures and measurements of the discovered defects through the camera and imaging sonar, recording its precise position, length, width and other information.

[0020] Furthermore, in step S4, the feature depth is automatically identified based on a preset triggering rule. Specifically, the vertical water quality profile data is analyzed, and when a depth layer with an absolute temperature gradient value ≥ 8.0 ℃ / m or a chlorophyll concentration mutation rate exceeding 30% is detected, the depth layer is determined as the feature depth that needs to be sampled.

[0021] Furthermore, step S5 includes the following sub-steps: Step S5.1: Collect the laser point cloud, abnormal water surface area information, underwater structural defect parameters and location information, water quality profile data and stratified sampling laboratory analysis results from steps S1 to S4, and construct a structure-ecology association data table under a unified spatiotemporal reference.

[0022] Step S5.2: After standardizing the structure-ecology association data table, perform principal component analysis to select principal components with eigenvalues ​​greater than 1 and cumulative variance contribution rates ≥ 85%. Analyze the principal component loading matrix, defining the original variables with absolute loading values ​​greater than 0.7 as core representation variables, and assigning physical meaning labels to the corresponding principal components based on the commonalities of these core representation variables. Step S5.3: Based on the principal components obtained in Step S5.2, and the structural semantic information and historical maintenance records extracted from the BIM model of the reservoir dam, initialize a Bayesian network structure. Input the observational evidence from the structure-ecology association data table into this Bayesian network structure, and quantitatively assess the probability of different structural defects leading to specific ecological anomalies by calculating posterior probabilities. The process begins with identifying key correlation factors. For these initially correlated variables, time-series data is extracted and Granger causality tests are performed. When the p-value is less than 0.05, a statistically significant causal relationship is determined, thus establishing the causal direction. A quantitative structural defect-aquatic ecological anomaly causal network model is constructed by integrating the probabilistic correlation strength from the Bayesian inference with the causal direction from the Granger causality test, thus completing the quantitative identification of key correlation factors. Step S5.4: Based on the analysis results of the causal network model, a comprehensive diagnostic report is automatically generated. This report includes at least the precise location of the defect in the BIM model, a causal mechanism analysis, a list of quantitatively identified key correlation factors, and targeted maintenance strategy recommendations.

[0023] A dynamic collaborative control system for underwater robots that integrates multi-source data, used to implement any of the above-mentioned dynamic collaborative control methods for underwater robots that integrate multi-source data, includes: an inspection drone for executing steps S1 and S2; an underwater robot for executing step S3; a water quality sampling drone for executing step S4; and a ground control and intelligent analysis center for collaboratively controlling the inspection drone, the underwater robot, and the water quality sampling drone, and executing step S5.

[0024] Furthermore, it also includes a permanent magnet marker array for underwater robot navigation correction. The permanent magnet marker array consists of several permanent magnet markers fixed to the underwater wall of the dam, and the spatial coordinates of each permanent magnet marker are recorded in the BIM model.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] Firstly, this invention establishes a global spatial benchmark by highly calibrating the laser point cloud acquired by the inspection drone with the BIM model, and combines it with a pre-deployed permanent magnet marker array for magnetic field matching and positioning, as well as underwater robot pose calculation based on visual features and the EPnP algorithm. This forms a multi-source fusion positioning correction chain, reducing the absolute positioning error of the underwater robot from the meter level of traditional inertial navigation to the centimeter level, providing a reliable spatial benchmark guarantee for the precise maintenance of dam defects.

[0027] Secondly, this invention automatically identifies and generates collaborative exploration task packages through a multispectral data anomaly detection algorithm based on Mahalanobis distance, driving underwater robots and water quality sampling drones to perform dynamic collaborative operations. It also utilizes the triggering rules of vertical water quality profile data to achieve adaptive stratified sampling, shortening the traditional anomaly investigation process that relies on manual labor and takes several weeks to complete within a few hours, resulting in a significant improvement in overall efficiency.

[0028] Thirdly, this invention employs principal component analysis for data dimensionality reduction and feature extraction, and integrates Bayesian inference and Granger causality testing methods to construct a quantitative causal network model between structural defect parameters and aquatic ecological anomaly parameters. This model can quantitatively identify key correlation factors and analyze the causal mechanisms, achieving a leap from phenomenon description to causal diagnosis. This allows for the identification of early hidden dangers that are difficult to detect with the naked eye, and promotes the transformation of operation and maintenance mode towards early warning maintenance.

[0029] Fourth, through the automated collaborative operation and intelligent analysis and diagnosis of the above-mentioned air-space-water multi-platform, the system significantly reduces the need for high-risk manual diving exploration and ship operations, reduces the intensity of manual intervention and safety risks, and practical applications show that the overall operation and maintenance costs can be significantly reduced. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a system structure block diagram of the present invention. Detailed Implementation

[0032] Specific embodiments of the present invention will now be described with reference to the accompanying drawings. Many details are described below to provide a comprehensive understanding of the invention; however, those skilled in the art will be able to implement the invention without these details.

[0033] like Figure 1 and Figure 2 As shown, a dynamic cooperative control method for underwater robots that integrates multi-source data includes the following steps:

[0034] Step S1: Control the inspection drone to fly along the preset route of the reservoir dam and adjacent waters to acquire laser point cloud and multispectral data of the reservoir dam and water surface; register the laser point cloud with the BIM model of the reservoir dam to construct a global spatial reference with centimeter-level accuracy.

[0035] In one specific embodiment, step S1, flying along the preset route of the reservoir dam and adjacent waters, specifically involves: using the dam axis as a reference, extending 200 meters upstream of the reservoir to plan a bow-shaped flight route.

[0036] In step S1, the laser point cloud is registered with the BIM model of the reservoir dam to construct a global spatial reference with centimeter-level accuracy. This step includes the following sub-steps:

[0037] Step S1.1: Perform point cloud calculation and denoising on the laser point cloud.

[0038] Step S1.2: Select several common feature points in the laser point cloud and BIM model; calculate the coordinate transformation matrix using these common feature points, and perform preliminary alignment of the laser point cloud and BIM model.

[0039] Step S1.3: Find the nearest point pair between the laser point cloud and the triangular mesh surface of the BIM model, and calculate the optimal rigid body transformation iteratively using the least squares method to minimize the root mean square error between the two point sets, and finally establish a centimeter-level global spatial benchmark.

[0040] Step S2: The multispectral data is analyzed using an anomaly detection algorithm based on Mahalanobis distance to identify abnormal water surface areas; a collaborative exploration task package containing the geographical location, range, type, and priority of the abnormal water surface areas is automatically generated.

[0041] In one specific embodiment, step S2 involves using an anomaly detection algorithm based on Mahalanobis distance to analyze the multispectral data and identify abnormal water surface areas. This includes the following sub-steps:

[0042] Step S2.1: Perform aerial triangulation and orthophoto stitching on the multispectral raw images obtained in step S1 to generate reflectance images for each band, and use the FLAASH model for atmospheric correction to convert the apparent reflectance into the true reflectance of the land surface.

[0043] Step S2.2: For the thermal infrared band, the absolute temperature distribution map of the water surface is obtained by using the radiative transfer equation method and combining real-time meteorological data; the chlorophyll a concentration distribution map is generated by using the reflectance of the red edge and near-infrared bands through a variant algorithm of the normalized difference vegetation index; the turbidity distribution map is obtained by using the reflectance ratio of the blue light (e.g., 450nm) and red light (e.g., 650nm) bands according to empirical or semi-empirical models (e.g., the Nechad algorithm).

[0044] Step S2.3: Read the three distribution maps. For each pixel with the same geographical location, construct a three-dimensional feature vector from the values ​​of its three bands, thus obtaining a three-dimensional vector set representing the entire inspection area. Randomly select a set of pixels from the three-dimensional vector set as training samples, calculate its mean vector and covariance matrix, and construct a statistical model to characterize the features of normal water bodies. The mean vector represents the average level of temperature, chlorophyll, and turbidity in the entire area, and the covariance matrix represents the relationship between these three parameters. Traverse each pixel in the image and calculate its Mahalanobis distance to the distribution defined by the statistical model. Pixels whose Mahalanobis distance exceeds the preset value are judged as abnormal pixels.

[0045] Step S2.4: Perform DBSCAN spatial clustering on the abnormal pixels to obtain the abnormal water surface region; extract the detailed information of the abnormal water surface region from the dataset of Steps S2.1 to S2.3 and package it into a task package.

[0046] Step S3: After receiving the collaborative exploration task package, the underwater robot autonomously navigates to the abnormal water surface area specified in the collaborative exploration task package. When the underwater robot arrives at the abnormal water surface area, its camera captures an image of the underwater dam surface corresponding to the abnormal water surface area. The underwater dam surface image is then matched with the reference image of the BIM model for visual feature matching and pose calculation to achieve centimeter-level precise positioning. Guided by the centimeter-level precise positioning information, the underwater robot is controlled to conduct close-up observation of the underwater dam surface in the abnormal water surface area. The discovered defects are photographed and measured from multiple angles, and their precise location, length, width, and other information are recorded.

[0047] In one specific embodiment, multiple permanent magnet markers are pre-deployed on the underwater wall of the dam using a topology structure that combines equally spaced grids with key point encryption, forming a permanent magnet marker array. Step S3 specifically includes the following sub-steps:

[0048] Step S3.1: The underwater robot is deployed from the mother ship, and its initial position is determined by the mother ship's GNSS. After entering the water, the underwater robot relies on a combined navigation system and a Doppler log for navigation. However, due to the scaling factor error of the inertial navigation system (approximately 0.2%) and the zero-bias instability of the gyroscope in the Doppler log, the position error accumulates at a rate of approximately 0.3% of the travel distance per hour. Therefore, this invention further utilizes a permanent magnet marker matrix to perform absolute magnetic field grid calibration on the underwater robot's position.

[0049] Step S3.2: When the inertial navigation system instructs the underwater robot to navigate to a preset distance from the previous calibration point and approach a pre-placed magnetic marker, the triaxial magnetometer array begins to monitor the regular changes in the strength and direction of the spatial magnetic field and acquire a magnetic field vector map. The inertial navigation system matches the real-time measured magnetic field vector map with a pre-stored magnetic field grid map. Through the least squares algorithm, it uniquely determines that the underwater robot is currently located near a permanent magnet marker. When the match is successful, the navigation system uses the precise coordinates of the permanent magnet marker recorded in the BIM model as a high-confidence observation to perform feedback correction on the system state of the inertial navigation system, thereby suppressing the cumulative error of the inertial navigation system.

[0050] Step S3.3: When the underwater robot arrives at the abnormal water surface area, the magnetic marker positioning method used in step S3.2 is used to periodically correct the inertial navigation system error. Simultaneously, the visual-assisted navigation system is activated for precise positioning. Specifically, the visual-assisted navigation system involves: capturing images of the underwater dam surface corresponding to the abnormal water surface area using a camera, and extracting multiple key points and their multi-dimensional feature descriptors; matching these multi-dimensional feature descriptors with a pre-rendered reference image feature library generated from the BIM model in that area; and using the EPnP algorithm to calculate the current coordinates and attitude of the underwater robot, achieving centimeter-level precise positioning.

[0051] Step S3.4: Guided by centimeter-level precise positioning information, the operator controls the underwater robot to conduct close-up observation of the underwater dam surface in the abnormal water surface area, and uses cameras and imaging sonar to take pictures and measure the defects found from multiple angles, recording their precise location, length and width and other information.

[0052] Step S4: The water quality sampling drone receives vertical water quality profile data from the underwater robot in real time via a wireless network, automatically identifies the characteristic depth based on preset trigger rules, and collects water samples at that characteristic depth.

[0053] In one specific embodiment, in step S4, the feature depth is automatically identified based on a preset triggering rule. Specifically, the vertical water quality profile data is analyzed, and when a depth layer with an absolute temperature gradient ≥ 8.0 ℃ / m or a chlorophyll concentration mutation rate exceeding 30% is detected, that depth layer is determined as the feature depth requiring sampling. The aforementioned chlorophyll concentration mutation rate = |(C2-C1)| / C1×100%, where C1 and C2 are the concentration values ​​of adjacent depth layers.

[0054] Step S5: Collect the multivariate heterogeneous data from Steps S1 to S4, and use principal component analysis to reduce data dimensionality and extract features. Combine Bayesian inference and Granger causality tests to construct a causal network model between structural defect parameters and aquatic ecological anomaly parameters, quantitatively identify key correlation factors, and automatically generate a comprehensive diagnostic report including defect location, causal analysis, and maintenance recommendations. This includes, but is not limited to, laser point clouds, spectral images, water quality profile data, and structural defect parameters.

[0055] In one specific embodiment, step S5 includes the following sub-steps:

[0056] Step S5.1: Collect the laser point cloud, abnormal water surface area information, underwater structural defect parameters and location information, water quality profile data and stratified sampling laboratory analysis results from steps S1 to S4, and construct a structure-ecology correlation data table under a unified spatiotemporal reference. The rows represent different observation samples, and the column variables include structural defect parameters and aquatic ecological anomaly parameters.

[0057] Step S5.2: After standardizing the structure-ecology association data table, perform principal component analysis and select principal components with eigenvalues ​​greater than 1 and cumulative variance contribution rate ≥ 85%. By analyzing the principal component loading matrix, define the original variables with absolute loading values ​​greater than 0.7 as core characterization variables, and assign physical meaning labels to the corresponding principal components based on the commonalities of this group of core characterization variables.

[0058] Step S5.3: Based on the principal components with physical meaning labels obtained in step S5.2, and the structural semantic information and historical maintenance records extracted from the BIM model of the reservoir dam, initialize a Bayesian network structure, and input the observation evidence from the structure-ecology association data table into the Bayesian network structure. By calculating the posterior probability, quantitatively assess the probability that different structural defects will lead to specific ecological anomalies, and preliminarily identify key association factors.

[0059] For variables that are initially associated, their time series data are extracted and Granger causality tests are performed. When the p-value of the test result is less than the threshold (e.g., 0.05), a statistically significant causal relationship is determined, thereby determining the causal direction.

[0060] By integrating the probabilistic correlation strength of the Bayesian inference with the causal direction of the Granger causality test, a quantitative causal network model of structural defects and aquatic ecological anomalies is constructed to complete the quantitative identification of key correlation factors.

[0061] Step S5.4: Based on the analysis results of the causal network model, automatically generate a comprehensive diagnostic report. The report includes at least the precise location of the defect in the BIM model, the causal mechanism analysis based on causal relationships, a list of quantitatively identified key related factors, and targeted maintenance strategy recommendations.

[0062] like Figure 1 and Figure 2 As shown, an underwater robot dynamic cooperative control system that integrates multi-source data is used to implement the above-mentioned underwater robot dynamic cooperative control method that integrates multi-source data. The system includes:

[0063] Inspection drone: includes lidar, multispectral camera, high-precision GNSS / IMU integrated navigation system and first wireless communication unit, used to perform steps S1 and S2.

[0064] An underwater robot is used to perform step S3.

[0065] A water quality sampling drone is used to perform step S4.

[0066] The ground control and intelligent analysis center is used to coordinate the control of the inspection drone, underwater robot and water quality sampling drone, and to execute step S5.

[0067] In addition, the system includes an array of permanent magnet markers for underwater robot navigation correction. This array consists of several permanent magnet markers fixed to the underwater wall of the dam, and the spatial coordinates of each permanent magnet marker are recorded in the BIM model. The permanent magnet markers are fixed to the underwater wall of the dam either embedded or anchored.

[0068] Example 1

[0069] Taking the periodic comprehensive inspection of a large concrete gravity dam as an example, this paper fully demonstrates the entire implementation process of this system and method.

[0070] (I) System Configuration

[0071] 1. Inspection drone: DJI Matrice 350RTK platform, equipped with Zenmuse L1 LiDAR and Zenmuse P1 multispectral camera.

[0072] 2. Underwater Robot: A custom-designed Observation-Class ROV with a frame size of 1.5m (length) × 1.0m (width) × 0.8m (height). It employs six brushless DC thrusters, enabling six degrees of freedom maneuverability. Equipped with: 1) a three-axis magnetometer array consisting of 12 sensors; 2) a 1080P high-definition low-light camera; 3) a SoundMetrics ARIS Explorer 1800 imaging sonar as an underwater acoustic beacon receiver; and 4) a Kearfott T24INS / DVL integrated navigation system.

[0073] 3. Water sampling drone: The DRAGONFLY 2 water sampling drone (model M400-FC100) from Guangzhou Feichuang Intelligent Technology Co., Ltd. was selected.

[0074] 4. Ground Control and Intelligent Analysis Center: This is a high-performance mobile workstation that runs mission planning software, data management software, and the core analysis algorithm of this invention.

[0075] 5. Permanent Magnet Marker Array: The permanent magnet markers are N52 grade neodymium iron boron permanent magnets, measuring 50mm × 50mm × 25mm, with a surface magnetic field strength of 0.8±0.1T. Each permanent magnet marker is encapsulated in a laser-welded 316 stainless steel waterproof shell. Multiple permanent magnet markers are deployed according to the topology on key structural parts such as the underwater wall of the dam and the gate slot, thus forming a permanent magnet marker array. The topology deployment combines an equidistant grid (base layer) with a non-uniform topology based on key structural points (application layer). Specifically, the equidistant grid (base layer) is deployed on the main underwater wall of the dam with 80m × 80m grid nodes. Key point densification (application layer): The density is increased to 20m × 20m around all transverse joints of the dam section, within 5 meters upstream and downstream of the spillway gate slot, and around the drainage gallery outlet. After the deployment of each permanent magnet marker, the three-dimensional coordinates (accuracy: ±2mm) of its center point are entered into the BIM model.

[0076] (II) Implementation of the Method

[0077] Step S1: Global Spatial Benchmark Construction and Preliminary Screening

[0078] Mission Planning and Flight: At the ground control and intelligent analysis center, operators open the DJIPilot2 software and import the boundary KML file of the reservoir dam. Using the dam's axis as a reference, a bow-shaped flight path is planned, extending 200 meters upstream. Specific parameters are set as follows: flight altitude 80 meters relative to the dam crest, flight speed 5 meters per second, forward overlap rate 85%, and lateral overlap rate 25%. This setting ensures sufficient density of the laser point cloud and enough stereo pairs for modeling of the multispectral imagery. Subsequently, the UAV automatically flies along the preset flight path, simultaneously acquiring laser point cloud and raw multispectral imagery via lidar and multispectral camera. Exposure points for both lidar and multispectral camera are triggered by the UAV's high-precision GNSS / IMU system, ensuring that each data point carries accurate latitude, longitude, and altitude (WGS84UTM coordinate system) and attitude information (pitch, roll, yaw).

[0079] Point cloud data processing and registration: After the flight, the ground control and intelligent analysis center received the laser point cloud and imported it into ContextCapture software. First, the laser point cloud was processed for point cloud analysis and denoising, removing obvious aircraft noise and abnormal drift points. Then, several common feature points were selected from the laser point cloud and the BIM model. In this inspection, four corners of the dam crest, two lamppost bases, and the center of the floodgate opening and closing mechanism were selected.

[0080] Using these common feature points, a preliminary coordinate transformation matrix was calculated to initially align the laser point cloud and the BIM model. Then, the closest point pairs were found between the triangular mesh surfaces of the laser point cloud and the BIM model, and the optimal rigid body transformation (rotation matrix R and translation vector T) was iteratively calculated using the least squares method to minimize the root mean square error between the two point sets, ultimately establishing a centimeter-level global spatial benchmark. The iteration termination condition was set as follows: the change in root mean square error between two consecutive iterations was less than 1 cm, or the maximum number of iterations reached 100. The final registration report of this inspection showed a plane mean square error of 2.5 cm and an elevation mean square error of 4.1 cm, successfully establishing a centimeter-level global spatial benchmark.

[0081] Step S2: Intelligent identification of water surface anomalies and dynamic task generation

[0082] Multispectral data preprocessing and distribution map generation: First, the raw multispectral images acquired by the multispectral camera are subjected to aerial triangulation and orthophoto stitching in Pix4Dmatic to generate reflectance images for each band. Then, for the thermal infrared band, atmospheric and emissivity corrections are performed using the radiative transfer equation method in ENVI software, combined with real-time meteorological data, to invert and obtain the absolute temperature distribution map of the water surface. Using the reflectance of the red-edge (730nm) and near-infrared (840nm) bands, a variant algorithm of the normalized difference vegetation index is used to invert and generate the chlorophyll a concentration distribution map. Using the reflectance ratio of the blue (450nm) and red (650nm) bands, an empirical or semi-empirical model (such as the Nechad algorithm) is used to invert and obtain the turbidity distribution map.

[0083] Multi-parameter joint anomaly detection based on Mahalanobis distance:

[0084] Read three distribution maps, and for each pixel with the same geographical location, construct a three-dimensional feature vector x from the values ​​of its three bands. i =[temp i ,chl i ,turb i ]; where temp i chl i and turb i (representing the temperature, chlorophyll a concentration, and turbidity at point i, respectively), resulting in a set of three-dimensional feature vectors representing the entire inspection area;

[0085] A large number of 10,000 pixels were randomly selected from the entire area as training samples, representing the normal water state of the reservoir. The mean vector μ of the set of three-dimensional feature vectors of these training samples was calculated. temp ,μ chl ,μ turb And the covariance matrix Σ. Where, μ temp μ chl μ turb These represent the average levels of temperature, chlorophyll, and turbidity across the entire region, respectively.

[0086] In this inspection, the mean vector μ of the three-dimensional feature vector set of the training samples is [20.2, 13.5, 12.3], and the covariance matrix Σ is:

[0087]

[0088] Among them, Σ[1,2]=Σ[2,1]=2.16, which is the covariance between temperature and chlorophyll a concentration, with a correlation coefficient ρ=2.16 / √(1.21×9.61)≈0.63, showing a positive correlation, indicating that an increase in temperature is usually accompanied by an increase in chlorophyll a concentration. Σ[1,3]=Σ[3,1]=-0.85, which is the covariance between temperature and turbidity, with a correlation coefficient ρ=-0.85 / √(1.21×2.25)≈-0.52, showing a negative correlation, indicating that water bodies with higher temperatures tend to have lower turbidity. Σ[2,3]=Σ[3,2]=1.23, which is the covariance between chlorophyll a concentration and turbidity, with a correlation coefficient ρ=1.23 / √(9.61×2.25)≈0.27, showing a positive correlation, indicating that algal growth may slightly increase the turbidity of the water body.

[0089] traversing each pixel x in the reflectance image pixel by pixel i And independently calculate its Mahalanobis distance to normal water. Mahalanobis distance D M (x i The formula for calculating D is: M (x i )=√[(x i μ) T Σ -1 (x i Under the assumption that the three-dimensional eigenvectors follow a multivariate normal distribution, the squared Mahalanobis distance (D) is... M (xi)) 2 It follows a chi-square distribution with 3 degrees of freedom: χ 2 3. Set a 95% confidence interval and look up the χ² value in the distribution table. 2 3. Distribution table to obtain P(χ²) 2 (3 ≤ 7.815) = 0.95. Therefore, the Mahalanobis distance D M (x i Pixels with a value greater than √7.815 ≈ 2.796 are considered abnormal pixels. The set 95% confidence interval can be dynamically adjusted based on historical data, or an adaptive threshold mechanism can be introduced.

[0090] Clustering and Task Packet Generation: The DBSCAN clustering algorithm (parameters: eps=5 meters, min_samples=10) is used to spatially cluster anomalous pixels, forming an anomalous water surface region, and generating a corresponding collaborative exploration task packet for this region. This collaborative exploration task packet is pushed to the underwater robot control unit on the mother ship in real time.

[0091] During this inspection, based on the pixel-by-pixel anomaly detection results, the DBSCAN clustering algorithm was used to spatially cluster all pixels marked as anomalies. These anomaly pixels collectively formed an abnormal water surface region above the micro-cracks. This is because the cracks cause reservoir water to seep into the dam body, forming a local seepage field and altering the heat conduction path of the water, thus creating an abnormal temperature region on the corresponding water surface. The mean vector of this abnormal water surface region is [18.5, 15.2, 12.5]. Specifically, the average absolute water temperature is 18.5 ℃, approximately 1.7 ℃ lower than the surrounding normal water temperature of 20.2 ℃; the average chlorophyll a concentration is 15.2 μg / L, slightly higher than the normal value of 13.5 μg / L; and the average turbidity is 12.5 NTU, showing no significant anomalies. The calculated average Mahalanobis distance is 3.45, far exceeding the threshold of 2.796.

[0092] Step S3: Underwater centimeter-level precise positioning and close-range exploration

[0093] Initial navigation and INS / DVL calculation: The underwater robot enters the water from the mother ship's launch point (coordinates known). Initial position and attitude are provided by the mother ship's GNSS (Global Navigation Satellite System) and attitude indicator. After entry, navigation is performed using a combined inertial navigation system and a Doppler log (INS / DVL integrated navigation system). The inertial navigation system provides the ground velocity, which is fused with the angular rate and acceleration from the Doppler log in an extended Kalman filter to output the optimal navigation solution.

[0094] Absolute calibration of the magnetic field grid: When the inertial navigation system instructs the underwater robot to navigate to a preset distance (e.g., 60 m) from the previous calibration point and approach a permanent magnet marker, the triaxial magnetometer array begins to monitor the regular changes in the strength and direction of the spatial magnetic field, acquiring a magnetic field vector map. The inertial navigation system performs least-squares matching between the real-time measured magnetic field vector map and the pre-stored magnetic field map, calculating that the most likely position of the underwater robot differs from the theoretical position of the permanent magnet marker by only (Δx=0.15 m, Δy=0.08 m, Δz=0.05 m).

[0095] The absolute position (X3+Δx, Y3+Δy, Z3+Δz) with uncertainty is used as an observation update and input into the extended Kalman filter. The filter compares the predicted position of the inertial navigation system with this observed position to calculate estimates of the current position error, velocity error, attitude error, and sensor bias error of the inertial navigation system. The filter then uses these error estimates to correct the entire system state vector of the inertial navigation system, thereby correcting the navigation path to the correct trajectory.

[0096] Visual-assisted centimeter-level repositioning:

[0097] After the underwater robot reaches the abnormal water surface area, it activates its high-definition camera to capture images of the corresponding underwater dam surface. The contrast of the underwater dam surface image is enhanced using the CLAHE algorithm; approximately 520 scale-invariant, rotation-invariant keypoints and their 128-dimensional descriptors are extracted from the underwater dam surface image.

[0098] These 128-dimensional feature descriptors were matched with the feature library of the pre-rendered benchmark image of the BIM model in the area, initially yielding 162 point pairs. 1000 RANSAC mismatch elimination iterations were performed, randomly selecting 4 point pairs each time to calculate the homography matrix, ultimately finding an optimal set containing 135 interior points, eliminating 27 exterior points. Using these 135 interior points, the transformation matrix from the camera coordinate system to the world coordinate system was solved using the EPnP algorithm, thus achieving centimeter-level precise positioning of the underwater robot. The results show that the underwater robot's positional accuracy relative to the target dam surface is ±2.1 cm, and its attitude accuracy is ±0.3 degrees.

[0099] Guided by centimeter-level precise positioning information, operators controlled an underwater robot to conduct close-up observations of the underwater dam surface in areas with abnormal water surfaces. Using cameras and imaging sonar, they captured and measured the discovered defects from multiple angles, recording their precise location, length, width, and other information. During this inspection, a minute crack measuring 1.2 meters in length and with a maximum width of 2.5 millimeters was discovered.

[0100] Step S4: Adaptive stratified water quality sampling and validation

[0101] Data Reception and Feature Depth Recognition: The water quality sampling drone receives vertical water quality profile data relayed by the underwater robot via underwater acoustic communication. Analyzing this data, the system identifies depths with an absolute temperature gradient ≥ 8.0 ℃ / m or a chlorophyll concentration mutation rate exceeding 30% as characteristic depths. During this inspection, the temperature gradient analysis revealed a sharp drop in temperature from 15.2 ℃ to 11.2 ℃ between 15.5 and 16.0 meters, a gradient of 4.0 ℃ / 0.5m = 8.0 ℃ / m, or 0.8 ℃ / 10cm, reaching the threshold of 8.0 ℃ / m. Based on this, the system determined 15.75 meters (the center of the thermocline) as the characteristic depth.

[0102] Precise stratified sampling: The water quality sampling drone flies over the target water area and begins to lower the sampler.

[0103] When the sampler reaches 15.70 meters, the water quality sampling drone begins collecting water samples. After sampling is completed, the water quality sampling drone retrieves the sampler.

[0104] Step S5: Multi-source data causal association analysis and diagnostic report generation

[0105] Multi-source heterogeneous data aggregation and structuring: The Ground Control and Intelligent Analysis Center aggregates raw and processed data from the following sources:

[0106] From step S1: Registered BIM model and laser point cloud.

[0107] From step S2: Boundary, type (temperature / chlorophyll), and confidence level of the abnormal water surface area.

[0108] From step S3: the three-dimensional coordinates of the defect, the defect type (e.g., crack), the defect size (length, width, depth), visual image, and sonar image.

[0109] All data were integrated into a structure-ecology association data table, with behavioral observation samples listed as variables. These variables included crack length, crack width, crack depth, distance to a specific component in the BIM model, absolute water surface temperature, chlorophyll a concentration, turbidity, temperature at the characteristic depth, and Cl at the characteristic depth. - SO4 concentration and characteristic depth 2- Concentration, etc.

[0110] To conduct principal component analysis and causality tests, the Ground Control and Intelligent Analysis Center collected data from this inspection and historical inspections, obtaining ≥100 sets of observation samples. A structure-ecology correlation data table was constructed for modeling. The table below serves as a demonstration, showing only a portion of the observation samples (where observation sample 1 is from this inspection).

[0111]

[0112] Principal component analysis is used for data dimensionality reduction and feature extraction.

[0113] Perform Z-score standardization on the data.

[0114] Principal component analysis was performed on 10 samples, yielding eigenvalues ​​of [5.8, 3.2, 1.5, 0.9, ...]. The first three principal components with eigenvalues ​​greater than 1 were selected, and their cumulative variance contribution rate was 88%.

[0115] Observation of the load matrix: By analyzing the principal component load matrix, the original variables with an absolute load value greater than 0.7 were defined as core characterization variables, and the corresponding principal components were assigned physical meaning labels based on the commonalities of this group of core characterization variables; In this inspection, principal component 1 was found to have significant differences in crack width, Cl - Concentration, SO4 2-Principal component 2 has a high loading on concentration (>0.85). This core set of variables collectively reflects the chemical migration characteristics caused by leakage, therefore, its physical meaning is labeled as a leakage-chemical corrosion factor. Principal component 2 has high loadings on surface absolute temperature and temperature at a depth of 15.7m (>0.82). These variables collectively characterize the thermodynamic anomalies of the water body, therefore, it is labeled as a thermodynamic factor. Principal component 3 has a high loading on chlorophyll a (loading >0.88). This variable directly indicates the level of biological activity in the water body, therefore, it is labeled as a bioactivity factor.

[0116] Constructing causal network models and quantitatively identifying key association factors:

[0117] a. Bayesian inference:

[0118] In this embodiment, the Bayesian network adopts a "three-node, three-state" Naive Bayesian network, and the topology and variable definitions are shown in the table below:

[0119]

[0120] The edge structure is fixed as C→I and C→T, prohibiting reverse or cyclic dependencies, which conforms to the physical assumption that "structural defects drive ecological anomalies".

[0121] The prior probability P(C) is directly referenced from the dam's maintenance database over the past 5 years: P(C=None)=0.62, P(C=Slight)=0.27, P(C=Significant)=0.11.

[0122] The Conditional Probability Table (CPT) was obtained by combining expert knowledge (from two rounds of Delphi evaluation by three senior engineers (hydraulic / structural / water environment)) with statistical frequencies, as shown in the table below:

[0123] The initial frequency was obtained from 10 sets of samples, and three structural / water environment experts were invited to perform two rounds of correction on the frequency using the Delphi method. The final CPT is shown in the table below and is stored in the ground control and intelligent analysis center for automatic loading by the system before each inspection.

[0124]

[0125] Posterior probability calculation: Observed evidence E={I=yes, T=yes}, calculated for each state θ∈{no, slight, significant}:

[0126]

[0127] Substituting the above table and prior information, and then normalizing, the results are as follows:

[0128]

[0129] Conclusion: Under the condition of observing both chemical and temperature anomalies, the posterior probability of "significant structural defects" is 65%, which is much higher than that of other states, satisfying the criteria for determining key correlation factors. Therefore, significant structural defects can be determined as the most likely key correlation factor causing ecological anomalies.

[0130] b. Granger causality test:

[0131] The time series of micro-strain of crack opening monitored by the underwater robot at the crack for 1 hour and the time series of water temperature at 0.5 meters above were extracted (both at 1-minute intervals, totaling 361 data points). The first 50 data points are shown in the table below.

[0132]

[0133] Using water temperature as the dependent variable and crack aperture as the independent variable, we examine whether crack aperture is a Granger cause of water temperature changes. The lag order is set to 2.

[0134] Result: The null hypothesis that crack aperture was not caused by water temperature was corrected, with a p-value of 0.0028.

[0135] Conclusion: At a 99% confidence level, the null hypothesis is rejected, confirming that the change in crack aperture is a Granger cause of the water temperature change.

[0136] Causal network model: Integrating the two, the final conclusion is that structural defects (key correlation factor) caused the chemical and temperature anomalies of the water body with a high probability (65%) and high statistical confidence (p value < 0.01).

[0137] Automatically generate comprehensive diagnostic reports:

[0138] The core contents of the comprehensive diagnosis during this inspection include:

[0139] Title: Inspection and Diagnosis Report of a Large Reservoir Dam on [Date] 2025

[0140] Abstract: A 1.2m long crack was found in the flood discharge gate section (BIM coordinates: X=505105.123, Y=343205.456, Z=-15.75). Quantitative analysis confirmed that it was the root cause of the abnormal temperature and ion concentration in the water above.

[0141] Defect details: Includes crack images, dimensions, and precise coordinates.

[0142] Causal analysis: Display PCA results diagram, Bayesian network diagram, and Granger test results table.

[0143] Repair recommendations: It is recommended to perform epoxy resin pressure grouting on the crack during the dry season and conduct a detailed investigation of the surrounding area.

[0144] The above are merely specific embodiments of the present invention, but the design concept of the present invention is not limited thereto. Any non-substantial modifications made to the present invention using this concept shall be considered as infringing upon the protection scope of the present invention.

Claims

1. A dynamic cooperative control method for underwater robots that integrates multi-source data, characterized in that, Includes the following steps: Step S1: Control the inspection drone to fly along the preset route of the reservoir dam and adjacent waters to acquire laser point cloud and multispectral data of the reservoir dam and water surface; register the laser point cloud with the BIM model of the reservoir dam to construct a global spatial reference. Step S2: Analyze the multispectral data using an anomaly detection algorithm based on Mahalanobis distance to identify abnormal water surface areas; automatically generate a collaborative exploration task package containing the geographical location, range, type, and priority of the abnormal water surface areas; Step S3: After receiving the collaborative exploration task package, the underwater robot autonomously navigates to the abnormal water surface area specified in the collaborative exploration task package; when the underwater robot arrives at the abnormal water surface area, the underwater dam surface image corresponding to the abnormal water surface area captured by the underwater robot is visually matched and pose calculated with the reference image of the BIM model to achieve centimeter-level precise positioning; under the guidance of the centimeter-level precise positioning information, the underwater robot is controlled to conduct close-up observation of the underwater dam surface in the abnormal water surface area, and the defects found are photographed and measured from multiple angles, recording their precise position, length and width; Step S4: The water quality sampling drone receives vertical water quality profile data from the underwater robot in real time, automatically identifies the characteristic depth based on preset trigger rules, and collects water samples at that characteristic depth. Step S5: Collect the multivariate heterogeneous data from Steps S1 to S4, and use principal component analysis to reduce data dimensionality and extract features; combine Bayesian inference and Granger causality test methods to construct a causal network model between structural defect parameters and aquatic ecological anomaly parameters, quantitatively identify key correlation factors, and automatically generate a comprehensive diagnostic report. Step S3 specifically includes the following sub-steps: Step S3.1: Deploy the underwater robot from the mother ship, and the mother ship's GNSS will provide the initial position; after the underwater robot enters the water, it mainly relies on the inertial navigation system and the Doppler log for combined navigation; Step S3.2: When the inertial navigation system instructs the underwater robot to navigate to a preset distance from the previous calibration point and approach a pre-placed magnetic marker, the three-axis magnetometer array begins to monitor the regular changes in the strength and direction of the spatial magnetic field and acquire a magnetic field vector map; the inertial navigation system matches the real-time measured magnetic field vector map with a pre-stored magnetic field grid map; through the least squares algorithm, it uniquely determines that the underwater robot is currently located near a permanent magnet marker; When a match is successful, the navigation system uses the precise coordinates of the permanent magnet mark recorded in the BIM model as a high-confidence observation to provide feedback correction to the system state of the inertial navigation system, thereby suppressing the cumulative error of the inertial navigation system. Step S3.3: When the underwater robot arrives at the abnormal water surface area, the vision-assisted navigation system is activated. The camera captures the underwater dam surface image corresponding to the abnormal water surface area and extracts multiple key points and their multi-dimensional feature descriptors. These multi-dimensional feature descriptors are matched with the reference image feature library pre-rendered in the area by the BIM model, and the coordinates and attitude of the current underwater robot are calculated by the EPnP algorithm to achieve centimeter-level precise positioning. Step S3.4: Guided by centimeter-level precise positioning information, the operator controls the underwater robot to conduct close-up observation of the underwater dam surface in the abnormal water surface area, and uses cameras and imaging sonar to take pictures and measure the defects found from multiple angles, recording their precise location, length and width information.

2. The underwater robot dynamic cooperative control method according to claim 1, characterized in that, In step S1, the laser point cloud is registered with the BIM model of the reservoir dam to construct a global spatial reference, which specifically includes the following sub-steps: Step S1.1: Perform point cloud calculation and denoising on the laser point cloud; Step S1.2: Select several common feature points in the laser point cloud and BIM model; calculate the coordinate transformation matrix using these common feature points, and perform preliminary alignment of the laser point cloud and BIM model; Step S1.3: Find the nearest point pair between the laser point cloud and the triangular mesh surface of the BIM model, and calculate the optimal rigid body transformation iteratively using the least squares method to minimize the root mean square error between the two point sets, and finally establish a centimeter-level global spatial benchmark.

3. The underwater robot dynamic cooperative control method according to claim 1, characterized in that, In step S2, an anomaly detection algorithm based on Mahalanobis distance is used to analyze the multispectral data and identify abnormal water surface areas. This specifically includes the following sub-steps: Step S2.1: Perform aerial triangulation and orthophoto stitching on the multispectral raw image obtained in step S1 to generate reflectance images for each band; Step S2.2: For the thermal infrared band, the absolute temperature distribution map of the water surface is obtained by using the radiative transfer equation method and combining real-time meteorological data; the chlorophyll a concentration distribution map is generated by using the reflectance of the red and near-infrared bands through a variant algorithm of the normalized differential vegetation index; the turbidity distribution map is obtained by using the reflectance ratio of the blue and red bands according to empirical or semi-empirical models. Step S2.3: Read the three distribution maps. For each pixel with the same geographical location, construct a three-dimensional feature vector from the values ​​of its three bands, thus obtaining a three-dimensional vector set representing the entire inspection area. Randomly select a set of pixels from the three-dimensional vector set as training samples, calculate its mean vector and covariance matrix, and construct a statistical model to characterize the features of normal water bodies. The mean vector represents the average level of temperature, chlorophyll, and turbidity in the entire area, and the covariance matrix represents the relationship between these three parameters. Traverse each pixel in the image and calculate its Mahalanobis distance to the distribution defined by the statistical model. Pixels whose Mahalanobis distance exceeds the preset value are judged as abnormal pixels. Step S2.4: Perform DBSCAN spatial clustering on the abnormal pixels to obtain the abnormal water surface region; Extract detailed information about the abnormal water surface area from the datasets from steps S2.1 to S2.3 and package it into a task package.

4. The underwater robot dynamic cooperative control method according to claim 1, characterized in that, In step S4, the feature depth is automatically identified based on the preset triggering rules. Specifically, the vertical water quality profile data is analyzed, and when a depth layer with an absolute temperature gradient value ≥ 8.0 ℃ / m or a chlorophyll concentration mutation rate exceeding 30% is detected, the depth layer is determined as the feature depth that needs to be sampled.

5. The underwater robot dynamic cooperative control method according to claim 1, characterized in that, Step S5 includes the following sub-steps: Step S5.1: Collect the laser point cloud, abnormal water surface area information, underwater structural defect parameters and location information, water quality profile data and stratified sampling laboratory analysis results from steps S1 to S4, and construct a structure-ecology correlation data table under a unified spatiotemporal reference. Step S5.2: After standardizing the structure-ecology association data table, perform principal component analysis and select principal components with eigenvalues ​​greater than 1 and cumulative variance contribution rate ≥ 85%; by analyzing the principal component loading matrix, define the original variables with absolute loading values ​​greater than 0.7 as core characterization variables, and assign physical meaning labels to the corresponding principal components based on the commonalities of this group of core characterization variables. Step S5.3: Based on the principal components obtained in Step S5.2, and the structural semantic information and historical maintenance records extracted from the BIM model of the reservoir dam, a Bayesian network structure is initialized. Observational evidence from the structure-ecology association data table is input into this Bayesian network structure. By calculating the posterior probability, the probability of different structural defects leading to ecological anomalies is quantitatively assessed, and key association factors are initially identified. For the variables with preliminary associations, their time series data are extracted, and Granger causality tests are performed. When the p-value of the test result is less than 0.05, a statistically significant causal relationship is determined, thereby determining the causal direction. By integrating the probabilistic association strength of the Bayesian inference and the causal direction of the Granger causality test, a quantitative structural defect-aquatic ecological anomaly causal network model is constructed to complete the quantitative identification of key association factors. Step S5.4: Based on the analysis results of the causal network model, automatically generate a comprehensive diagnostic report. The report includes at least the precise location of the defect in the BIM model, the causal mechanism analysis based on causal relationships, a list of quantitatively identified key related factors, and targeted maintenance strategy recommendations.

6. A dynamic cooperative control system for underwater robots that integrates multi-source data, used to implement the dynamic cooperative control method for underwater robots that integrates multi-source data as described in any one of claims 1-5, characterized in that, include: An inspection drone is used to perform steps S1 and S2. An underwater robot is used to perform step S3; A water sampling drone is used to perform step S4; The ground control and intelligent analysis center is used to coordinate the control of the inspection drone, underwater robot and water quality sampling drone, and to execute step S5; It also includes a permanent magnet marker array for underwater robot navigation correction, the permanent magnet marker array consisting of several permanent magnet markers fixed to the underwater wall of the dam, and the spatial coordinates of each permanent magnet marker are recorded in the BIM model.

Images