Intelligent detection and closed-loop correction method for deepwater foundation bed leveling defects

Abstract

The application provides a deep water base leveling defect intelligent detection and closed-loop correction method, which firstly constructs a base digital twin model based on communication and positioning technology to establish a collaborative mechanism of detection and correction equipment; secondly, a multi-modal detection is performed by an ROV robot to obtain base defect parameters, and after the three-dimensional characteristics of the defects are accurately identified by an artificial intelligence detection model, the defects are real-time mapped to the digital twin model, thereby driving the leveling device to move to the target; in the physical correction stage, a mechanical arm adjusts a vibrating leveling plate to perform compaction work, and an absolute compaction index is real-time fed back to a riprap system, and a riprap amount instruction is automatically issued according to the defect volume calculated by AI, and the riprap system performs adaptive closed-loop correction; finally, a secondary scanning verification is performed on the working area, and the model is iteratively optimized. The application completely solves the technical problems of disconnection between measurement and correction, easy to miss detection or excessive riprap in deep water base construction, and realizes seamless integration of intelligent sensing, accurate material supplement and dynamic compaction efficient work.

Description

Technical Field

[0001] This invention relates to the field of deep water foundation construction, and in particular to an intelligent detection and closed-loop correction method for leveling defects in deep water foundations. Background Technology

[0002] In the field of underwater engineering inspection and construction control, especially in the leveling of deep-water foundation beds at depths of 100 meters, it is a core process in the construction of major underwater infrastructure such as immersed tunnels and cross-sea channels. The flatness of the foundation bed directly affects the stress uniformity and long-term operational safety of the superstructure components. However, in deep-water environments of 100 meters, due to extremely low visibility, rapid currents, and the presence of a large amount of suspended sediment, the traditional detection of foundation bed defects (such as elevation deviations, pits, and protruding rocks) relies heavily on manual exploration by divers or scanning with a single sonar device. Manual exploration is not only highly subjective and extremely inefficient, but also faces extremely high safety risks in deep-water operations; while traditional sonar equipment often struggles to accurately distinguish between the foundation bed structure and silt deposits when facing strong underwater turbulence and irregular terrain, leading to frequent missed detections and misjudgments. Even more critically, in traditional construction methods, there is a huge time lag between the collection of test data and the physical correction by the leveling equipment. Due to the ever-changing water flow environment, this "test first, repair later" disconnect often causes the correction work to deviate from the actual defect state, which can easily lead to "over-cutting" or "inadequate filling" and make it difficult to meet the construction standards of modern ultra-fine flat foundation beds.

[0003] To overcome the limitations of traditional technologies, the industry has proposed several improvement schemes. For example, existing technology CN113494062A discloses an underwater rock-filled bed leveling device and its leveling method. This method primarily uses a multibeam echo sounder to perform three-dimensional feature scanning of the underwater bed, generating a three-dimensional model of the bed and dividing it into multiple printing layers on the Z-axis. Subsequently, a moving platform and a feeding pipe are controlled to perform leveling and rock-filling layer by layer. While this existing technology achieves a certain degree of integration between scanning and feeding, it relies solely on a multibeam echo sounder to obtain static elevation data, lacking an anti-interference mechanism for complex underwater suspended objects and strong flow fields. Furthermore, it does not incorporate artificial intelligence algorithms to intelligently identify and classify the severity of specific defects in the bed (such as pits and cracks), resulting in an overly mechanical rock-filling strategy that cannot achieve adaptive and precise matching. In addition, this scheme lacks real-time physical monitoring of the bed's compaction state during leveling, easily leading to the accumulation of loose material.

[0004] Another existing technology, CN121496930A, discloses an underwater subgrade leveling and replenishment device and construction method. It integrates a binocular camera, a leveling scraper, and a boulders on an underwater walking mechanism, attempting to identify protrusions and depressions through binocular vision and then scrape and fill them sequentially. However, this existing technology relies excessively on binocular optical vision for core perception. In deep-water, high-turbidity (muddy) environments, the optical camera is prone to large blind spots. This solution lacks multimodal fusion perception and turbidity-adaptive defogging enhancement algorithms, failing to guarantee robustness in extreme environments. Furthermore, its rigidly connected leveling scraper lacks flexible attitude adjustment buffering and dynamic pressure feedback when facing subgrade slopes or large, protruding rocks, making forced scraping prone to mechanical damage. In addition, deep-water currents generate significant hydrodynamic drag deformation on the cables connecting to the mother ship, and this existing technology does not propose a high-precision, flow-resistant flexible compensation and collaborative control strategy between the leveling vessel and underwater equipment. Summary of the Invention

[0005] The main objective of this invention is to provide an intelligent detection and closed-loop correction method for defects in deep water subgrade leveling. This application aims to solve the technical problems existing in the current deep water subgrade leveling construction, such as poor detection and identification accuracy due to strong turbulence and high turbidity, lag and disconnection in data transmission between detection and correction, lack of dynamic compaction feedback in the face of complex terrain, and easy mechanical damage under extreme working conditions.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for intelligent detection and closed-loop correction of defects in deep-water subgrade leveling, the method comprising: S1. Construct a digital twin model of the bed base based on communication and positioning technologies, calibrate the bed base boundary, and establish a collaborative mechanism between detection and correction equipment; The S2 and ROV robots, combined with robotic arms, adjust the vibration compaction detection of the vibratory plate and provide real-time feedback on the compaction degree. This feedback is then sent to the rock-throwing system, which automatically calculates the amount of rock to be thrown based on the defect volume detected by AI. The rock-throwing system then performs adaptive corrections. S3. ROV robots equipped with high-definition cameras and multiple sonars perform multimodal detection to obtain subgrade defect parameters, and after being identified by an artificial intelligence detection model, they are mapped to the subgrade digital twin model in real time, and the leveling device is driven to move according to the subgrade defect parameters. S4. Perform a secondary inspection scan on the area that has been leveled and riprap corrected to verify the closed loop, and iteratively optimize the artificial intelligence inspection model based on the operation data.

[0007] In the preferred embodiment, S1 includes building a base digital twin model based on a cloud server architecture, and deploying a data processing module and a 3D visualization service module on the server using Docker containerization technology; Nginx reverse proxy cluster is used to achieve network load balancing of multi-source heterogeneous data, and the received environmental status and device location data are persistently stored in PostgreSQL time series database; By combining the BeiDou positioning system to calibrate the bed boundary, a low-latency command interaction mechanism is established between the detection equipment and the correction equipment.

[0008] In the preferred embodiment, S3 includes deploying an ROV robot from a maintenance vessel to submerge in the water; The ROV robot uses the downward anti-current sonar among multiple sonars to obtain the Doppler ground velocity in real time, and uses the forward fixed-range sonar among multiple sonars to penetrate underwater suspended objects to locate abnormal areas, and coordinates with high-definition cameras to synchronously capture image data at a preset frame rate. Combining the U-INS underwater inertial navigation system, an error state Kalman filter algorithm based on flow field adaptation is used to ensure stable navigation trajectory; the specific algorithm implementation steps are as follows: Construct a state vector containing position error, velocity error, and attitude error. The system velocity output by U-INS and the ground velocity output by the downward anti-current sonar. The difference sequence is used as the observation benchmark; to resist high-frequency strong flow field interference in deep water environments, a value sequence based on the instantaneous velocity of the current deep water flow field is established. The adaptive measurement noise covariance matrix derived from the underwater robot's ground velocity: in The initial covariance matrix under calibration conditions. This is the environmental turbulence penalty coefficient derived from underwater hydrodynamic characteristics; Optimal Kalman gain is dynamically calculated using the covariance matrix: ; The measurement residuals are calculated using the optimal Kalman gain, and the trajectory drift integral error of the U-INS system is corrected in real time using a closed loop, outputting the optimal pose coordinates with anti-flow field interference characteristics. The acquired image data and environmental data containing the optimal pose coordinates are packaged and compressed using the feature extraction algorithm of the edge computing node. Utilizing the 5G communication link on the water, dynamic routing scheduling and network load balancing are performed via an Nginx reverse proxy cluster deployed on a cloud server, and the data is uploaded in real time in a high-concurrency mode to the backend interface of the digital twin model of the base bed deployed based on a Docker containerized architecture. After the received data stream is decompressed and multi-source spatiotemporal aligned by the backend service program, the structured optimal pose coordinates and environmental state parameters are persistently written to the PostgreSQL time series database, and the unstructured image data is pushed to the distributed object storage module. The digital twin model of the bed generates spatial control vectors based on the updated data, driving the leveling device connected to the leveling ship to move along the planned three-dimensional trajectory toward the target defect pose.

[0009] In the preferred embodiment, the training process of the artificial intelligence detection model is performed based on the PyTorch deep learning framework deployed on a cloud-based GPU-accelerated cluster, and the steps include: An initial dataset was constructed by collecting images of defects in underwater foundation beds from real engineering projects. Then, DCGAN generative adversarial network was used for feature enhancement and sample expansion. To constrain the texture and shadow generation quality of the generator in deep-water, high-turbidity environments, an adaptive generative adversarial loss function incorporating prior knowledge of underwater dark passages is constructed; the optimization formula for the generator's objective function is as follows: ; in, For the discriminator function, To augment the multimodal image matrix output by the generator, To derive the ambient light scattering coefficient using water flow data, A pre-defined a priori threshold for the dark channel to define underwater bedrock characteristics. and The nonlinear distribution hyperparameters are dynamically tuned based on real-time water turbidity feedback; an expanded sample library covering pits, protrusions and crack edges in extremely dark environments is iteratively generated through dynamic game between the generator and the discriminator. Using a Python script, the expanded sample library was divided into training, validation and test sets according to the proportions, and then input into the YOLOv5 detection model for iterative training. A channel attention mechanism module is embedded in the feature pyramid network of YOLOv5 to enhance feature extraction of underwater blurred edges; To improve the robustness of underwater defect size identification, a bounding box regression loss function for underwater defect morphology perception is constructed for the YOLOv5 model. This bounding box regression loss function combines depth and elevation outliers obtained from sonar to establish a three-dimensional spatial penalty mechanism. Its derivation formula is as follows:

[0010] in, To predict the spatial intersection-union ratio between the bounding box and the true bounding box, The square of the Euclidean distance between their centers. To cover the minimum diagonal distance of the closure regions of both, This is a penalty term for aspect ratio consistency in a two-dimensional plane. The scalar value represents the measured depth feature of the defects in the sonar point cloud feedback. For the pseudo-depth scalar of monocular reconstruction in visual networks, The confidence attenuation coefficient for acoustic-optical fusion is derived from the variance of multi-source heterogeneous data. The AdamW gradient descent optimization algorithm is used, combined with a cosine annealing learning rate dynamic decay strategy to optimize the network node weight matrix of the YOLOv5 detection model. The convergence status of the loss function is continuously monitored on the validation set until the average accuracy of all classes reaches a preset threshold. Training is then stopped and the lightweight model weight file reconstructed by the TensorRT tensor acceleration module is exported. This file is then packaged into a standard inference environment container image and pushed to edge computing nodes for deployment and execution.

[0011] In the preferred embodiment, when the deployed AI detection model receives image data and environmental data synchronously labeled with optimal pose coordinates by the underwater inertial navigation system in real time, the specific algorithm and computer module implementation steps include: A real-time vision processing pipeline is built using the TensorRT inference engine deployed on a cloud-based GPU acceleration node; upon receiving image data, the image processing module is invoked to execute turbidity-adaptive image enhancement and dehazing algorithms. To overcome backscattering interference caused by high turbidity in deep water and non-uniform light fields, a transmittance compensation defogging formula based on dynamic feedback of environmental parameters is constructed. The spatial transmittance distribution model is derived as follows: ; in, The color channel pixel values ​​of the input image. Estimate the global atmospheric illumination physical quantities. Adjust parameters for defogging level; The second-order Laplacian derivative of the image grayscale matrix is ​​used to extract high-frequency edge details in blurred underwater environments; NTU is the scalar reading of the water turbidity sensor synchronously transmitted in the environmental data. The high-frequency scattering compensation empirical coefficients are obtained by fitting multiple measured data; by introducing the Laplace operator and the turbidity logarithm term, the edge features of the bed rock are adaptively sharpened while eliminating underwater fogging. The dehazed and enhanced image matrix is ​​fed into the YOLOv5 detection model for feature extraction and bounding box regression; the forward propagation of the convolutional neural network is accelerated by the CUDA core, and the output includes the target confidence, defect classification probability and two-dimensional pixel coordinates; the detection results are spatiotemporally aligned and heterogeneous data fused with the corresponding area multibeam sonar point cloud data cached in the digital twin model of the bed bed. The system utilizes a 3D reconstruction and volume calculation module developed using C++ and the PCL library to accurately calculate the defect volume of the identified substrate defects. To eliminate errors caused by visual deformation and sonar noise, a visual attention-weighted 3D volume integral formula is constructed. ; in, Map the two-dimensional defect connected domain output by the YOLOv5 detection model to the projected integral region in a three-dimensional coordinate system; This is the actual elevation surface function measured by sonar point clouds. The standard flat datum elevation surface function is preset in the digital twin model of the subgrade bed; the area within parentheses on the right side of the formula is a visual attention confidence weight distribution map constructed based on the Sigmoid function, where... This refers to the spatial attention heatmap value output by the YOLOv5 feature pyramid layer. and These are the activation threshold and scaling factor, respectively. The final output includes the three-dimensional coordinates of the defect, the defect type, and the precise defect volume size calculated using the volume integral formula. The base bed defect parameters are then pushed to the control node via a microservice interface, providing a high-precision data source for calculating the amount of rock thrown by the intelligent rock-throwing system.

[0012] In the preferred embodiment, S2 includes a high-frequency vibrating block built into the vibrating plate for compaction, acquiring multi-point contact stress feedback signals through an integrated pressure sensor array, and transmitting compaction data to the rock-throwing system in real time via a controller local area network bus; the specific algorithm and software module implementation steps include: The embedded real-time operating system microcontroller deployed inside the leveling device is used as the main control node to read the analog voltage signal of the pressure sensor array at a sampling rate of 1000Hz, and convert it into a discrete digital signal sequence through its built-in ADC module. To filter out high-frequency mechanical vibrations and hydrodynamic noise generated by water flow from underwater equipment, a recursive least-squares digital filter based on an adaptive forgetting factor is constructed within the microcontroller; Let... This is the original pressure acquisition signal sequence. For the target smooth signal, the derivation of its recursive least squares digital filter parameter update algorithm is as follows: ; ; ; in, The input signal matrix, For the filter weight vector, Here is the gain matrix. It is the inverse correlation matrix. An adaptive forgetting factor; based on underwater vibration frequency characteristics... The dynamic stress is dynamically limited to the range [0.95, 0.99], and the smoothed effective dynamic contact stress sequence is output. The effective dynamic contact stress sequence of the multi-point pressure sensor is input into the compaction degree calculation software module; this module, combined with the real-time vibration acceleration characteristics of the vibrating plate, constructs a nonlinear soil dynamic impedance inversion model to calculate the absolute compaction degree index; the mathematical formula for the absolute compaction degree index is: ; in, This represents the total number of nodes in the pressure sensor array. For the first Smooth contact stress at each node, This is the local inclination angle of the contact surface at this node; and These represent the current real-time amplitude and angular frequency of the high-frequency vibrating block, respectively; the exponential term represents the decay function of energy dissipation due to soil plastic deformation. The empirical attenuation coefficient is pre-calibrated based on the subgrade soil type; the microcontroller will calculate the resulting... Scalar data is encapsulated into standard data frames according to application layer protocols and broadcast in real time to the control cabinet of the rock-throwing system via a redundant controller area network bus at a transmission frequency of 10Hz. Upon receiving the initial rock-throwing quantity command issued by the defect volume calculation module, the rock-throwing system begins precise replenishment operations and simultaneously parses the data on the controller area network bus. Data frame; compaction data for a continuous preset period Reaching or exceeding the target compaction threshold set by the system When its first derivative approaches zero, the control cabinet of the stone-throwing system triggers the closed-loop termination logic, automatically shuts off the stone-throwing valve through the actuator, and simultaneously issues a stop command to control the robotic arm to end the current pressing and adaptive angle adjustment actions.

[0013] In the preferred embodiment, the movement of the leveling device based on the base bed defect parameters includes: The dynamic positioning control software module deployed on the edge computing node of the leveling vessel subscribes to and receives in real time movement commands issued by the digital twin model of the base bed, which include the three-dimensional coordinates of the target defect, spatial attitude constraints and flow field prediction data. The software module is deployed on the edge computing node of the leveling vessel. It uses the MQTT Internet of Things communication protocol to subscribe to and receive movement commands in real time from the digital twin model of the base bed, which include the three-dimensional coordinates of the target defect, spatial attitude constraints and flow field prediction data. The control system parses movement commands and integrates the absolute coordinates output by the carrier phase differential technology system with the six-degree-of-freedom pose data output by the high-precision inertial measurement unit. A nonlinear model predictive control algorithm is used to calculate the optimal three-dimensional trajectory. To overcome the hydrodynamic deformation and nonlinear spatial position shift of the cable caused by the rapid currents at a depth of 100 meters, a flexible coupled spatial compensation formula based on ocean current vector feedforward is incorporated into the trajectory planning algorithm, deriving and generating the compensated high-precision tracking coordinates. ; in, The compensation three-dimensional coordinate vector that the leveling vessel actually needs to track; The three-dimensional coordinates of the target defect issued by the digital twin model; The length of the integration window for the model's prediction in the time domain; Let be the generalized mass inertia matrix of the underwater leveling system; Let be the Jacobian transformation matrix from the local coordinate system to the global coordinate system. To advance the control input torque; and These are the nonlinear Coriolis centripetal force matrix and the real-time ocean current velocity vector, respectively. Dynamically changing hydrodynamic damping dissipation matrix; This is to provide the real-time lowering length parameter for the connecting cable. The nonlinear flexible deformation attenuation coefficient is obtained by fitting historical underwater stress tensor data. To prevent extremely small positive numbers with a denominator of zero; The dynamic positioning control software module uses a sequential quadratic programming algorithm to perform rolling optimization of the motion trajectory, decomposing the optimal trajectory into thrust distribution commands for the multi-axis vector thrusters and dynamic take-up and release commands for the hydraulic winch, driving the leveling device connected to the leveling boat to move towards the defect location along the planned overshoot-free trajectory. After the leveling device reaches directly above the defect location, it calls the internal programmable logic controller to activate the hydraulic servo leveling subroutine. Combined with the closed-loop feedback of the high-precision depth gauge, it controls multiple sets of hydraulic cylinders to extend and retract to perform attitude decoupling and elevation fine adjustment, thereby performing precise physical scraping and compaction leveling operations on the defect location until the difference between the measured elevation and the target design converges within the safe tolerance range.

[0014] In the preferred embodiment, S4 includes controlling the ROV robot to initiate a secondary scanning and comparison operation on the area corrected by the leveling device and the stone-throwing system; and calling the cloud comparison module to calculate the difference in defect size before and after correction. If the difference is lower than the set standard, the completion status field of the corresponding coordinate will be updated using the database write command; The system backend automatically extracts global operation image data recorded in the PostgreSQL time series database according to a preset time period, and adjusts the hyperparameters of the recognition model through the gradient descent optimization algorithm to improve the defect recognition generalization ability under complex working conditions.

[0015] In the preferred embodiment, when the underwater flow velocity sensor detects in real time that the water flow velocity exceeds the safe operating threshold, the safety control module is automatically triggered to execute the anti-flow protection strategy. The specific algorithm and control steps include: The high-frequency sampling sequence of the flow velocity sensor is smoothed by a sliding window mean filtering algorithm to eliminate random high-frequency noise caused by underwater pulsating turbulence and output a real-time effective flow velocity scalar. When the effective flow velocity scalar continuously crosses the safe operation threshold for a preset time period, the ROV robot is forced to activate the dual sonar cooperative anti-current mode, which combines forward fixed distance and downward anti-current. In the dual sonar cooperative anti-current mode, the downward anti-current sonar locks onto the seabed topography to calculate the ROV robot's real-time seabed drift vector at high frequency, while the forward fixed distance sonar switches to a near-field high-resolution scanning mode to avoid suspended obstacles that move rapidly with the fluid. The ROV robot's underlying main control chip is based on the real-time seabed drift vector. It calls the linear quadratic regulator algorithm to calculate the optimal thrust compensation matrix to counteract the water flow thrust, and drives multiple vector thrusters to generate adaptive reverse thrust to maintain the original planned detection route without deviation. Meanwhile, the safety control module sends a fixed-point anchoring command to the leveling vessel through the communication network, activates the leveling vessel's dynamic positioning system, calculates the combined force of wind direction, wave and water flow data, controls the azimuth thruster combined with the anchoring system to fix the absolute attitude of the fuselage, and dynamically reduces the planned spacing of the underwater detection path according to the flow velocity ratio to ensure the data acquisition density in high flow velocity environments.

[0016] In the preferred embodiment, when the size of the stone block identified by the artificial intelligence detection model exceeds the mechanical safety cutting threshold, a protective operation mode is automatically executed to avoid equipment damage. The specific algorithm and mechanical control steps include: The 3D reconstruction module extracts the 3D point cloud bounding box of the protruding stone on the surface of the base bed based on the multimodal fusion data, and calculates the maximum geometric scalar of the stone along the principal inertial axis through the principal component analysis algorithm. When the maximum geometric scalar exceeds the set mechanical safety cutting threshold, the safety control module sends the highest priority hardware interrupt signal to the underlying actuator through the controller local area network bus, automatically intercepting and terminating the conventional rigid scraping command currently being executed by the leveling device. Then the leveling device automatically switches to a protective operation mode of first leveling and then compacting: first, the robotic arm is controlled to adjust the water-facing inclination angle of the vibrating leveling plate to the preset bulldozing angle, so as to limit the running speed and drive the leveling device to push the large stones to the adjacent pit and depression area around the foundation bed. During the pushing operation, the lateral resistance feedback value in the hydraulic circuit of the robotic arm is monitored in real time. When the lateral resistance feedback value drops sharply and falls below the set no-load resistance threshold, the system determines that the large-sized rock has detached from the push plate and fallen into the pit. The controller then issues a posture reset command, controlling the robotic arm to reset the tilt angle of the vibrating plate to a horizontal state, and activates the internal high-frequency vibration block to perform high-frequency compaction on the area. This avoids rigid cutting damage to the mechanical structure caused by large stones, while completing the flatness correction of the base bed defects in a closed loop.

[0017] This invention provides an intelligent detection and closed-loop correction method for defects in deep-water bed leveling. After implementation, by constructing a digital twin model of the bed based on communication and positioning technologies, seamless collaboration between underwater multimodal detection and surface leveling equipment is achieved, completely eliminating the time lag between detection and correction, and ensuring low-latency real-time correspondence between data transmission and physical execution. Utilizing an ROV robot equipped with sonar and a high-definition camera for acoustic-optical collaborative scanning, and integrating generative adversarial networks for sample expansion and transmittance compensation defogging algorithms, the system can still extract defect edge features with high clarity even in deep-water environments with high turbidity and turbulent sediment, greatly reducing the missed detection of tiny cracks and pits, and significantly improving the detection generalization ability and identification accuracy in complex deep-water environments.

[0018] This application innovatively utilizes a robotic arm to adjust the vibrating slab during physical correction operations. It also incorporates an adaptive forgetting factor filtering algorithm to handle multi-point contact stress, providing real-time feedback of a highly accurate absolute compaction index to the riprap system. This mechanism overcomes the reliance on elevation alone in traditional equipment, enabling the system to dynamically determine the opening and closing of the riprap valves based on the actual dynamic resistance of the subgrade soil. This fundamentally eliminates incomplete filling and excessive cutting, significantly improving the compaction density and overall bearing capacity of the subgrade. Simultaneously, the slab's omnidirectional adaptive adjustment capability allows it to perfectly conform to complex longitudinal slope terrain, ensuring smooth slab leveling operations.

[0019] This application demonstrates superior anti-interference and self-protection capabilities in the extreme ocean currents and geological conditions at depths of up to 100 meters. In terms of navigation and positioning, by introducing an error-state Kalman filter algorithm adapted to the flow field and nonlinear model predictive control based on ocean current vector feedforward, the system can automatically counteract trajectory drift caused by deep-water rapids and hydrodynamic deformation of the cable, ensuring accurate target tracking even under strong turbulence. In terms of mechanical protection, when the artificial intelligence model identifies giant boulders exceeding safety thresholds, the system can instantly intercept conventional commands and automatically switch to a protective operation mode of "flattening first, then compacting." This effectively avoids structural damage to expensive underwater equipment caused by rigid impacts while successfully filling adjacent pits, significantly extending equipment lifespan and reducing engineering maintenance costs. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is the overall flowchart of the closed-loop operation of this invention; Figure 2 This is a flowchart of the intelligent detection and data processing algorithm of the present invention; Figure 3 This is a flowchart of the compaction feedback and anomaly protection control of the present invention; Figure 4 This is a diagram of the ROV robot detection structure of the present invention; Figure 5 This is a measurement structure diagram of the ROV robot leveling device of the present invention; Figure 6 This is a diagram of the stone-throwing structure of the ROV robot leveling device measurement and post-measurement stone-throwing system of the present invention. Figure 7 This is a construction structure diagram of the leveling device of this invention working in conjunction with an ROV robot; Figure 8 This is a structural diagram of the secondary detection and scanning structure of the ROV robot of the present invention.

[0021] In the diagram: 1. Maintenance vessel; 2. ROV robot; 201. High-definition camera; 202. Robotic arm; 203. Vibration leveling plate; 204. Sonar; 3. Rock-throwing system; 4. Leveling vessel; 5. Leveling device. Detailed Implementation

[0022] Example 1 like Figure 1-8 As shown, a method for intelligent detection and closed-loop correction of defects in deep-water subgrade leveling is provided. The method includes: S1. Construct a digital twin model of the bed base based on communication and positioning technologies, calibrate the bed base boundary, and establish a collaborative mechanism between detection and correction equipment; S2, ROV robot 2 combined with robotic arm 202 adjusts the vibration compaction detection of vibrating plate 203 and provides real-time feedback on compaction degree, which is fed back to the rock-throwing system. The rock-throwing amount is automatically calculated based on the defect volume detected by AI, and the rock-throwing system 3 performs adaptive correction. S3. The ROV robot 2, equipped with a high-definition camera 201 and multiple sonars 204, performs multimodal detection to obtain the base bed defect parameters. After being identified by the artificial intelligence detection model, it is mapped to the base bed digital twin model in real time. The leveling device 5 is driven to move according to the base bed defect parameters. S4. Perform a secondary inspection scan on the area that has been leveled and riprap corrected to verify the closed loop, and iteratively optimize the artificial intelligence inspection model based on the operation data.

[0023] In the preferred embodiment, S1 includes building a base digital twin model based on a cloud server architecture, and deploying a data processing module and a 3D visualization service module on the server using Docker containerization technology; Nginx reverse proxy cluster is used to achieve network load balancing of multi-source heterogeneous data, and the received environmental status and device location data are persistently stored in PostgreSQL time series database; By combining the BeiDou positioning system to calibrate the bed boundary, a low-latency command interaction mechanism is established between the detection equipment and the correction equipment.

[0024] In specific implementations, building a digital twin model of the underwater bed based on a cloud server architecture is the core foundation for real-time monitoring and intelligent decision-making in deep-water construction operations. The system utilizes cloud servers for computing power, overcoming the physical limitations of underwater front-end computing resources. Based on this, data processing modules and 3D visualization service modules are deployed on the servers using Docker containerization technology. Docker containerization standardizes and packages the runtime environment and applications, achieving resource isolation and independent operation between different functional modules. This isolation mechanism ensures that when processing massive amounts of high-frequency underwater image data, it does not preempt and consume the computing memory of the 3D visualization service module. Its beneficial effects include significantly improving the system's fault tolerance, stability, and development and deployment efficiency, enabling the digital twin model of the underwater bed to maintain stable operation even when facing complex deep-water calculations and high-load rendering.

[0025] To handle the high-concurrency data streams generated by multimodal detection equipment, the system utilizes an Nginx reverse proxy cluster to achieve network load balancing for multi-source heterogeneous data. Since the underwater robot and leveling equipment simultaneously upload sonar point clouds, high-definition video streams, and various sensor status data, the Nginx reverse proxy cluster can dynamically and evenly distribute data access requests based on the real-time load status of each cloud node. To ensure globally optimal data link allocation, the system internally employs a response-aware adaptive dynamic weight allocation algorithm to determine the routing direction of multi-source heterogeneous data. The dynamic weight allocation calculation formula is as follows: ; In the formula, Representing the Dynamic weight allocation for each server node; Representing the The quantized value of the current average network response time of each server node; Representing the The current available computing resources of each server node; Representing the The baseline amount of total computing resources configured for each server node; and This represents the weight balance adjustment coefficients pre-defined by the system's network physical environment. Through the aforementioned dynamic weight allocation algorithm, the Nginx reverse proxy cluster can accurately distribute massive amounts of heterogeneous data to the optimal processing nodes. Its beneficial effect lies in completely eliminating the risk of network congestion and single-point failure during centralized data transmission, achieving efficient distribution and latency-free access of multi-source heterogeneous data within milliseconds.

[0026] At the underlying data storage level, the system persistently stores the received environmental status data and equipment location data in a PostgreSQL time-series database. Considering that deep-water bed leveling operations involve massive amounts of data that change dynamically over time, using a PostgreSQL time-series database allows for dedicated underlying index structure optimization for time-series data. The system adds precise nanosecond-level timestamps to every piece of water flow velocity, pressure data, and spatial positioning coordinate data collected by sensors before writing it into the database. This significantly improves the concurrent write throughput performance of temporal data and the retrieval speed of complex data based on time spans. This not only provides a continuous source of high-quality data for real-time rendering of the 3D scene of the digital twin model but also provides a highly structured, full-lifecycle historical operation sample for subsequent model optimization of artificial intelligence detection algorithms.

[0027] Regarding the establishment of spatial coordinate benchmarks, the system integrates the BeiDou Navigation Satellite System to calibrate the boundary of the subgrade. The surface support vessel and underwater equipment jointly receive high-precision timing and positioning signals from the BeiDou satellites, accurately mapping the absolute geographical location information of the construction area into the virtual three-dimensional space of the digital twin model of the subgrade. To eliminate spatial projection errors caused by the Earth's curvature in the vast construction area, the system internally constructs a high-precision transformation algorithm from absolute geographic coordinates to local spatial coordinates in the digital twin. The local spatial coordinate transformation calculation formula is as follows: ; In the formula, Represents the local three-dimensional spatial coordinate vector generated within the digital twin model of the bed base; This represents a three-dimensional spatial position vector containing latitude, longitude, and absolute elevation obtained through the BeiDou positioning system. Representatives combined local conditions The three-dimensional spatial rotation transformation matrix obtained from the calculation of gravity anomalies and physical parameters of the Earth ellipsoid; This represents the spatial three-dimensional translation compensation vector, which is shifted from the origin of the BeiDou geodetic coordinate system to the local relative origin of the digital twin model. The algorithm converts macroscopic absolute positioning data into microscopic operational coordinates at the engineering construction level with high precision. Its beneficial effect is ensuring absolute geometrical overlap between the virtual subgrade boundary and the actual seabed construction area, eliminating the potential for boundary-crossing construction or missed leveling caused by positioning signal drift from the algorithm's underlying layer.

[0028] Based on the aforementioned solid foundation of data flow and coordinate transformation, the system ultimately established a low-latency command interaction mechanism between the detection and correction equipment. Using a unified digital twin spatiotemporal reference, the cloud server can directly convert the three-dimensional coordinates of the subgrade defects detected in real-time by the underwater robot into spatial movement vector commands for the leveling device, and then transmit these commands to the underlying hardware actuators in real-time via a high-speed IoT communication protocol. The beneficial effect lies in completely breaking down the data silos between the measurement and correction stages in traditional underwater construction operations, constructing a closed-loop, synchronous, and real-time automated collaborative operation link. This ensures that the leveling device and the intelligent rock-throwing system can make instantaneous adaptive mechanical responses to the front-end detection results, thereby significantly improving the overall construction efficiency and final forming accuracy of deep-water subgrade leveling operations.

[0029] Example 2 S3 includes the deployment of ROV robot 2 by maintenance vessel 1 into the water; The ROV robot 2 uses the downward anti-current sonar in multiple sonars 204 to obtain the Doppler ground velocity in real time, and uses the forward fixed-range sonar in multiple sonars 204 to penetrate underwater suspended objects to locate abnormal areas, and coordinates with the high-definition camera 201 to synchronously capture image data at a preset frame rate. Combining the U-INS underwater inertial navigation system, an error state Kalman filter algorithm based on flow field adaptation is used to ensure stable navigation trajectory; the specific algorithm implementation steps are as follows: Construct a state vector containing position error, velocity error, and attitude error. The system velocity output by U-INS and the ground velocity output by the downward anti-current sonar. The difference sequence is used as the observation benchmark; to resist high-frequency strong flow field interference in deep water environments, a value sequence based on the instantaneous velocity of the current deep water flow field is established. The adaptive measurement noise covariance matrix derived from the underwater robot's ground velocity: in The initial covariance matrix under calibration conditions. This is the environmental turbulence penalty coefficient derived from underwater hydrodynamic characteristics; Optimal Kalman gain is dynamically calculated using the covariance matrix: ; The measurement residuals are calculated using the optimal Kalman gain, and the trajectory drift integral error of the U-INS system is corrected in real time using a closed loop, outputting the optimal pose coordinates with anti-flow field interference characteristics. The acquired image data and environmental data containing the optimal pose coordinates are packaged and compressed using the feature extraction algorithm of the edge computing node. Utilizing the 5G communication link on the water, dynamic routing scheduling and network load balancing are performed via an Nginx reverse proxy cluster deployed on a cloud server, and the data is uploaded in real time in a high-concurrency mode to the backend interface of the digital twin model of the base bed deployed based on a Docker containerized architecture. After the received data stream is decompressed and multi-source spatiotemporal aligned by the backend service program, the structured optimal pose coordinates and environmental state parameters are persistently written to the PostgreSQL time series database, and the unstructured image data is pushed to the distributed object storage module. The digital twin model of the bed generates a spatial control vector based on the updated data, which drives the leveling device 5 connected to the leveling vessel 4 to move along the planned three-dimensional trajectory toward the target defect pose.

[0030] In a specific implementation, the deployment of an ROV (Remotely Operated Vehicle) underwater from a maintenance vessel forms the physical basis for achieving high-precision detection of deep-water bedbeds. The ROV employs a multimodal fusion sensing architecture, utilizing downward-facing anti-current sonar to acquire Doppler velocity relative to the ground in real time, thus overcoming interference from underwater currents. Simultaneously, forward-facing fixed-range sonar penetrates the high concentration of suspended sediment in the deep-water environment to accurately locate abnormal areas on the bedbed surface, and coordinates with a high-definition camera to synchronously capture underwater image data at a preset frame rate. This acoustic-optical coordinated detection method effectively compensates for the inherent limitations of single optical vision in turbid water, significantly improving the quality of raw bottom-layer data acquisition and the all-weather capability for environmental perception under complex deep-water conditions.

[0031] To ensure the absolute stability of the underwater robot's trajectory in strong flow environments, the system deeply integrates an underwater inertial navigation system and innovatively employs an error-state Kalman filter algorithm based on flow field adaptation. The core of this algorithm lies in constructing a state vector that includes position error, velocity error, and attitude error. The system combines the system velocity output from the underwater inertial navigation system with the Doppler ground velocity output from the downward-facing sonar. The difference sequence between them serves as a reliable observation benchmark. This error state-space modeling method can transform nonlinear kinematic equations into linear error propagation equations, thereby greatly reducing the computational load on the microprocessor and improving the real-time response speed of trajectory calculation.

[0032] To address the frequent high-frequency, strong flow field interference in deep-water environments (up to 100 meters), the system establishes an adaptive measurement noise covariance matrix that can dynamically adjust according to the ambient flow velocity. The formula for calculating this covariance matrix is: ; In this mathematical model, This represents the measurement noise covariance matrix after adaptive adjustment to the environmental flow field. This represents the initial covariance matrix measured under static or extremely low flow rate calibration conditions. This represents the instantaneous velocity vector of the current deep-water flow field obtained from the environmental flow velocity sensor; This represents the actual ground velocity vector output by the downward-facing anti-current sonar; This represents the environmental turbulence penalty coefficient, derived and fitted from extensive underwater hydrodynamic experimental data. The physical meaning of this formula is that when the vector difference between the instantaneous underwater flow velocity and the robot's actual ground velocity increases sharply, it indicates that the underwater environment is currently experiencing severe turbulence, at which point the signal-to-noise ratio of the sonar echo will significantly decrease. By introducing an exponential function amplification mechanism with a natural constant as the base, the measurement noise covariance matrix increases sharply with flow field disturbances, thereby proactively reducing the reliability weight of the sonar measurement data at that moment at the algorithm level, effectively preventing trajectory divergence caused by sudden undercurrents.

[0033] After obtaining the adaptively adjusted covariance matrix, the system dynamically calculates the optimal Kalman gain through state prediction and measurement update steps. The formula for calculating the optimal Kalman gain is: ; In this formula, The optimal Kalman gain matrix represents the current time step; This represents the state prediction error covariance matrix extrapolated from the previous time step to the current time step; This represents the measurement observation matrix that maps the state space to the measurement space; This is the transpose of the measurement observation matrix; This is the adaptive measurement noise covariance matrix calculated above. By introducing an amplified... Optimal Kalman gain The gain automatically decreases, and when the system uses this gain to calculate the measurement residual, it relies more on internal kinematic predictions rather than heavily contaminated external sonar measurements. This real-time closed-loop correction mechanism thoroughly filters out trajectory drift integral errors, outputting optimal pose coordinates with extremely strong resistance to flow field interference, providing a solid spatial coordinate reference for subsequent accurate corrections.

[0034] At the data transmission and cloud interaction level, the system first performs data dimensionality reduction and compression packaging on the acquired high-definition image data and environmental data containing optimal pose coordinates using feature extraction algorithms built into edge computing nodes. The intervention of edge computing significantly reduces the bandwidth consumption of invalid background data. Subsequently, using the 5G communication link built on the water surface, dynamic routing scheduling and network load balancing are performed via an Nginx reverse proxy cluster deployed on cloud servers. The Nginx reverse proxy cluster can intelligently identify the real-time load rate of each cloud computing node, evenly distributing the massive heterogeneous data streams uploaded under high concurrency mode, avoiding single-point network congestion or server downtime, and uploading in real time to the backend interface of the base digital twin model deployed based on Docker containerization architecture with extremely high network throughput. Docker containerization technology ensures sandbox isolation and horizontal scaling of the complex backend environment, guaranteeing the absolute high availability of the system under massive concurrent requests.

[0035] In the final data processing and execution phase, the received data stream, after decompression and multi-source spatiotemporal alignment by the backend service program, is classified and stored using a categorized approach. The system persistently writes the structured optimal pose coordinates and environmental state parameters into a PostgreSQL time-series database. The PostgreSQL time-series database possesses excellent temporal data indexing and aggregation query capabilities, providing latency-free data support for the historical trajectory playback and trend prediction of the digital twin model. Simultaneously, the system pushes large volumes of unstructured image data to a distributed object storage module to optimize system storage costs. Based on the frequently updated underlying physical data in PostgreSQL, the bed-leveling digital twin model generates precise spatial control vectors in real time. These vectors can directly penetrate the underlying communication driver connected to the leveling device on the leveling vessel, ensuring it moves strictly along the planned three-dimensional trajectory towards the target defect pose. This complete link thoroughly establishes a closed-loop control channel from underwater intelligent sensing to surface decision-making and then to mechanical adaptive execution, achieving unmanned and highly precise deep-water bed-leveling operations.

[0036] In the preferred embodiment, the training process of the artificial intelligence detection model is performed based on the PyTorch deep learning framework deployed on a cloud-based GPU-accelerated cluster, and the steps include: An initial dataset was constructed by collecting images of defects in underwater foundation beds from real engineering projects. Then, DCGAN generative adversarial network was used for feature enhancement and sample expansion. To constrain the texture and shadow generation quality of the generator in deep-water, high-turbidity environments, an adaptive generative adversarial loss function incorporating prior knowledge of underwater dark passages is constructed; the optimization formula for the generator's objective function is as follows: ; in, For the discriminator function, To augment the multimodal image matrix output by the generator, To derive the ambient light scattering coefficient using water flow data, A pre-defined a priori threshold for the dark channel to define underwater bedrock characteristics. and The nonlinear distribution hyperparameters are dynamically tuned based on real-time water turbidity feedback; an expanded sample library covering pits, protrusions and crack edges in extremely dark environments is iteratively generated through dynamic game between the generator and the discriminator. Using a Python script, the expanded sample library was divided into training, validation and test sets according to the proportions, and then input into the YOLOv5 detection model for iterative training. A channel attention mechanism module is embedded in the feature pyramid network of YOLOv5 to enhance feature extraction of underwater blurred edges; To improve the robustness of underwater defect size identification, a bounding box regression loss function for underwater defect morphology perception is constructed for the YOLOv5 model. This bounding box regression loss function combines depth and elevation outliers obtained from sonar to establish a three-dimensional spatial penalty mechanism. Its derivation formula is as follows:

[0037] in, To predict the spatial intersection-union ratio between the bounding box and the true bounding box, The square of the Euclidean distance between their centers. To cover the minimum diagonal distance of the closure regions of both, This is a penalty term for aspect ratio consistency in a two-dimensional plane. The scalar value represents the measured depth feature of the defects in the sonar point cloud feedback. For the pseudo-depth scalar of monocular reconstruction in visual networks, The confidence attenuation coefficient for acoustic-optical fusion is derived from the variance of multi-source heterogeneous data. The AdamW gradient descent optimization algorithm is used, combined with a cosine annealing learning rate dynamic decay strategy to optimize the network node weight matrix of the YOLOv5 detection model. The convergence status of the loss function is continuously monitored on the validation set until the average accuracy of all classes reaches a preset threshold. Training is then stopped and the lightweight model weight file reconstructed by the TensorRT tensor acceleration module is exported. This file is then packaged into a standard inference environment container image and pushed to edge computing nodes for deployment and execution.

[0038] In a specific implementation, the training process of the AI ​​detection model is performed using the PyTorch deep learning framework deployed on a cloud-based GPU-accelerated cluster. The system first collects real underwater engineering bed defect images to construct an initial dataset, and then uses a DCGAN generative adversarial network for feature enhancement and sample expansion. The beneficial effect of this step is that it effectively overcomes the bottleneck of the extreme difficulty in obtaining real defect samples under deep-water conditions and the scarcity of data. By using adversarial generative techniques to fill the gap in long-tail defect data, it provides highly rich and feature-diverse multimodal data support for the subsequent training of the visual detection model, ensuring the generalization capability of the AI ​​detection system from the underlying data source.

[0039] To constrain the texture and shadow generation quality of the generator in deep-water, high-turbidity environments, the system constructs an adaptive generative adversarial loss function that incorporates prior knowledge of underwater dark passages. The optimization formula for the generator's objective function is as follows: ; In this formula, This represents the overall loss value of the generator; Represents sampling from the potential spatial distribution Calculation of expected value of noise data; For the discriminator function; To augment the multimodal image matrix output by the generator; The ambient light scattering coefficient was derived using water flow data; Preset a priori threshold for the dark channel to define underwater bedrock characteristics; and This is a nonlinearly distributed hyperparameter dynamically tuned based on real-time water turbidity feedback. The physical meaning of the absolute value term in the latter half of the formula is to forcibly approximate and align the dark channel features of the generated image to a priori threshold of real underwater bedrock. Through the constraints of the above mathematical model and the dynamic game between the generator and discriminator, an expanded sample library covering pits, protrusions, and crack edges in extremely dark environments can be iteratively generated. Its beneficial effect is that the generated artificial samples highly match the optical distortion characteristics of strong scattering and severe fogging in real deep water, eliminating the risk of feature extraction failure of the detection model under actual harsh water conditions from the source, and greatly improving the robustness of the model in dealing with underwater turbid environments.

[0040] After acquiring the expanded sample library, the system uses Python scripts to divide the data into training, validation, and test sets according to scientific proportions, and inputs them into the YOLOv5 detection model for iterative training. To further enhance the model's ability to extract features from blurred underwater edges, the system embeds a channel attention mechanism module into the YOLOv5 feature pyramid network. This module can adaptively redistribute the response weights of each feature channel, assigning higher activation weights to channels with key defect features. Its beneficial effects include significantly enhancing the model's sensitivity to capturing the edges of small defects against a low-contrast underwater background, effectively removing high-frequency visual noise interference from underwater suspended sediment, and significantly reducing the false negative rate of fine cracks and small pits in the substrate.

[0041] To improve the robustness of underwater defect size identification, the system constructs a bounding box regression loss function for underwater defect morphology perception based on the YOLOv5 model. This bounding box regression loss function, combined with depth and elevation outliers acquired by sonar, establishes a unique three-dimensional spatial penalty mechanism, the derivation formula of which is as follows: ; In this formula, The overall loss value represents the shape-aware bounding box regression. To predict the spatial intersection-union ratio between the bounding box and the true bounding box; To predict the square of the Euclidean distance between the center points of the bounding box and the true bounding box; The diagonal distance of the minimum closure region that covers both of the above; This is a consistency penalty term for aspect ratio in a two-dimensional plane, used to constrain the geometric proportion difference between the predicted bounding box and the ground truth bounding box; Scalar for the measured depth feature of defects in sonar point clouds for accurate feedback; For monocular reconstruction in visual networks, it is a pseudo-depth scalar; This is the confidence attenuation coefficient for acousto-optic fusion derived from the variance of multi-source heterogeneous data. The specific significance of introducing an exponential penalty term with a base of the natural constant is that when there is a large deviation between the visual monocular depth and the sonar measured depth, the loss value of this term will increase exponentially, thus forcing the visual network to closely align with the true sonar depth features during backpropagation. Its beneficial effect is that it completely breaks through the technical barrier of traditional two-dimensional vision's inability to accurately perceive underwater depth, achieving deep fusion of two-dimensional image features and three-dimensional physical elevation features, significantly improving the model's accuracy in assessing the size of bed bulges and pit volumes, and providing an absolutely reliable basis for subsequent rock-laying systems.

[0042] During the model optimization and deployment phase, the system employs the AdamW gradient descent optimization algorithm combined with a cosine annealing learning rate dynamic decay strategy to optimize the network node weight matrix of the YOLOv5 detection model. This optimization strategy enables the model to converge quickly with a large step size in the early stages of training, and then smoothly reduce the learning rate in the later stages to finely navigate local optima. The system continuously monitors the convergence status of the loss function on the validation set until the average accuracy of all classes reaches a preset strict threshold, at which point training stops. Finally, the system exports a lightweight model weight file reconstructed by the TensorRT tensor acceleration module, packages it into a standard inference environment container image, and pushes it to edge computing nodes for deployment and execution. The beneficial effects not only ensure extremely high detection and classification accuracy, but also significantly reduce the model's memory footprint and multiply the single-frame inference speed through tensor-level reconstruction and lightweight container packaging, ensuring that the underwater robot can still achieve high frame rate and extremely low latency real-time defect identification and command feedback on edge hardware platforms with limited underlying computing power.

[0043] In the preferred embodiment, when the deployed AI detection model receives image data and environmental data synchronously labeled with optimal pose coordinates by the underwater inertial navigation system in real time, the specific algorithm and computer module implementation steps include: A real-time vision processing pipeline is built using the TensorRT inference engine deployed on a cloud-based GPU acceleration node; upon receiving image data, the image processing module is invoked to execute turbidity-adaptive image enhancement and dehazing algorithms. To overcome backscattering interference caused by high turbidity in deep water and non-uniform light fields, a transmittance compensation defogging formula based on dynamic feedback of environmental parameters is constructed. The spatial transmittance distribution model is derived as follows: ; in, The color channel pixel values ​​of the input image. Estimate the global atmospheric illumination physical quantities. Adjust parameters for defogging level; The second-order Laplacian derivative of the image grayscale matrix is ​​used to extract high-frequency edge details in blurred underwater environments; NTU is the scalar reading of the water turbidity sensor synchronously transmitted in the environmental data. The high-frequency scattering compensation empirical coefficients are obtained by fitting multiple measured data; by introducing the Laplace operator and the turbidity logarithm term, the edge features of the bed rock are adaptively sharpened while eliminating underwater fogging. The dehazed and enhanced image matrix is ​​fed into the YOLOv5 detection model for feature extraction and bounding box regression; the forward propagation of the convolutional neural network is accelerated by the CUDA core, and the output includes the target confidence, defect classification probability and two-dimensional pixel coordinates; the detection results are spatiotemporally aligned and heterogeneous data fused with the corresponding area multibeam sonar point cloud data cached in the digital twin model of the bed bed. The system utilizes a 3D reconstruction and volume calculation module developed using C++ and the PCL library to perform accurate volume calculations for identified substrate defects, particularly pit types. To eliminate errors caused by visual deformation and sonar noise, a visual attention-weighted 3D volume integral formula is constructed. ; in, Map the two-dimensional defect connected domain output by the YOLOv5 detection model to the projected integral region in a three-dimensional coordinate system; This is the actual elevation surface function measured by sonar point clouds. The standard flat datum elevation surface function is preset in the digital twin model of the subgrade bed; the area within parentheses on the right side of the formula is a visual attention confidence weight distribution map constructed based on the Sigmoid function, where... This refers to the spatial attention heatmap value output by the YOLOv5 feature pyramid layer. and These are the activation threshold and scaling factor, respectively. The final output includes the three-dimensional coordinates of the defect, the defect type, and the precise defect volume size calculated using the volume integral formula. The base bed defect parameters are then pushed to the control node via a microservice interface, providing a high-precision data source for calculating the amount of rock thrown by the intelligent rock-throwing system.

[0044] In a specific implementation, when the deployed AI detection model receives image data and environmental data synchronously labeled with optimal pose coordinates by the underwater inertial navigation system in real time, the system first utilizes the TensorRT inference engine deployed on a cloud-based GPU acceleration node to construct a real-time visual processing pipeline. Upon receiving underwater image data, the system calls the image processing module to execute turbidity-adaptive image enhancement and dehazing algorithms. To overcome the severe backscattering interference caused by high turbidity in deep water and non-uniform light fields, the system constructs a transmittance compensation dehazing formula based on dynamic feedback of environmental parameters. Its spatial transmittance distribution model is derived as follows:

[0045] In this mathematical model, This represents the spatial transmittance distribution model after compensation and optimization. Represents the pixel values ​​of the color channels of the input image; Estimated values ​​representing global atmospheric illumination physical quantities; This represents the parameter for adjusting the degree of defogging; The second-order Laplacian derivative of the image grayscale matrix is ​​used to extract high-frequency edge details in a blurred underwater environment; NTU represents the real-time scalar reading of the water turbidity sensor that is synchronously transmitted in the environmental data. This represents the empirical coefficients for high-frequency scattering compensation derived from fitting multiple deep-water measurement data. By innovatively introducing the Laplacian operator and a logarithmic term based on water turbidity into the classic defogging model, this algorithm can adaptively sharpen the edge features of bedrock boulders according to changes in the surrounding sediment suspension concentration while eliminating underwater fogging. Its beneficial effects include significantly improving image resolution under harsh deep-water visibility conditions, avoiding the failure of traditional fixed-parameter defogging algorithms in dynamically turbid waters, and providing a high-quality pre-image source with an extremely high signal-to-noise ratio for subsequent target detection.

[0046] After image enhancement, the system feeds the dehazed and enhanced image matrix into the YOLOv5 detection model for feature extraction and bounding box regression. In this computational phase, the system utilizes the CUDA core to accelerate the forward propagation of the convolutional neural network, significantly reducing the time delay for edge inference and outputting accurate detection results including target confidence, defect classification probability, and 2D pixel coordinates. Subsequently, the system performs rigorous spatiotemporal alignment and heterogeneous data fusion with the corresponding region's multibeam sonar point cloud data cached in the digital twin model of the substrate bed. The beneficial effect of this step is that it fully integrates the extremely high resolution advantage of optical images in planar texture recognition with the high reliability advantage of multibeam sonar in penetrating turbid water to obtain 3D spatial structures. This multimodal heterogeneous data fusion completely overcomes the technical bottleneck of single sensors being easily limited or generating detection blind zones in deep-water environments, achieving absolutely accurate 3D localization of defects on the substrate bed surface.

[0047] After completing the spatiotemporal alignment of heterogeneous data, the system invokes a 3D reconstruction and volume calculation module developed using the C++ programming language and point cloud processing algorithm library to perform precise defect volume calculations on the identified substrate defects. This function is crucial for calculating the repair volume of pit-type defects. To completely eliminate the combined calculation errors caused by underwater nonlinear visual deformation and sonar multipath reflection noise, the system constructs an original visual attention-weighted 3D volume integral formula:

[0048] In this integral formula, This represents the final calculated precise defect volume size; The two-dimensional defect connected domain output by the YOLOv5 detection model is accurately mapped to the projected integral region in the three-dimensional coordinate system; The actual elevation surface function representing the high-density measurement of sonar point clouds; This represents the pre-defined standard flatness benchmark elevation surface function in the digital twin model of the subgrade bed. The absolute value of the difference between the two objectively reflects the actual physical deformation depth at that spatial coordinate point. The right-hand product term inside the integral sign of the formula is a visual attention confidence weight distribution map constructed based on the Sigmoid function, where... The spatial attention heatmap value represents the output of the YOLOv5 feature pyramid layer. This represents the attention activation threshold of the neural network; This represents a nonlinear scaling factor. The essence of this integral formula is to use the high-dimensional attention heatmap extracted by the deep learning model as a spatial weighting factor, dynamically adjusting and correcting the contribution ratio of each point cloud elevation difference in the total volume calculation. Its beneficial effect is that it can adaptively and strongly suppress the interference of blurred edge regions and sonar anomalies on volume integration, greatly improving the absolute accuracy and stability of the capacity measurement of pits and defects in the foundation bed under complex terrain.

[0049] Ultimately, the system outputs comprehensive subgrade defect parameters, including the three-dimensional coordinates of the defects, their specific classification types, and the precise volume size of the defects calculated rigorously using the aforementioned three-dimensional volume integral formula. These high-precision parameters, containing a wealth of physical information, are then pushed to the lower-level hardware execution control nodes via a lightweight microservice interface. The beneficial effect is that it provides a highly reliable data instruction source for the intelligent rock-laying system's automatic rock-laying quantity calculation and precise filling, ensuring that subsequent closed-loop correction operations can achieve extremely precise rock-laying at both fixed points and in fixed quantities. This fundamentally eliminates the material waste and under-laying or over-laying problems caused by fuzzy measurement data in traditional construction, significantly improving the automation level, operational efficiency, and first-time acceptance rate of the entire deep-water subgrade leveling construction.

[0050] Example 3 S2 includes a high-frequency vibrating block built into the vibrating plate 203 for compaction, which acquires multi-point contact stress feedback signals through an integrated pressure sensor array and transmits compaction data to the rock-throwing system 3 in real time via a controller local area network bus; its specific algorithm and software module implementation steps include: Using the embedded real-time operating system microcontroller deployed inside the leveling device 5 as the main control node, the analog voltage signal of the pressure sensor array is read at a sampling rate of 1000Hz and converted into a discrete digital signal sequence through its built-in ADC module. To filter out high-frequency mechanical vibrations and hydrodynamic noise generated by water flow from underwater equipment, a recursive least-squares digital filter based on an adaptive forgetting factor is constructed within the microcontroller; Let... This is the original pressure acquisition signal sequence. For the target smooth signal, the derivation of its recursive least squares digital filter parameter update algorithm is as follows: ; ; ; in, The input signal matrix, For the filter weight vector, Here is the gain matrix. It is the inverse correlation matrix. An adaptive forgetting factor; based on underwater vibration frequency characteristics... The dynamic stress is dynamically limited to the range [0.95, 0.99], and the smoothed effective dynamic contact stress sequence is output. The effective dynamic contact stress sequence of the multi-point pressure sensor is input into the compaction degree calculation software module; this module, combined with the real-time vibration acceleration characteristics of the vibrating plate 203, constructs a nonlinear soil dynamic impedance inversion model to calculate the absolute compaction degree index; the mathematical formula for the absolute compaction degree index is: ; in, This represents the total number of nodes in the pressure sensor array. For the first Smooth contact stress at each node, This is the local inclination angle of the contact surface at this node; and These represent the current real-time amplitude and angular frequency of the high-frequency vibrating block, respectively; the exponential term represents the decay function of energy dissipation due to soil plastic deformation. The empirical attenuation coefficient is pre-calibrated based on the subgrade soil type; the microcontroller will calculate the resulting... Scalar data is encapsulated into standard data frames according to application layer protocols and broadcast in real time to the control cabinet of the rock-throwing system 3 via a redundant controller area network bus at a transmission frequency of 10Hz. After receiving the initial command for the amount of rock to be thrown based on the defect volume calculation module, the rock-throwing system 3 begins precise replenishment operations and simultaneously parses the data on the controller area network bus. Data frame; compaction data for a continuous preset period Reaching or exceeding the target compaction threshold set by the system When its first derivative approaches zero, the control cabinet of the stone-throwing system 3 triggers the closed-loop termination logic, automatically cuts off the stone-throwing valve through the actuator, and simultaneously issues a stop command to control the robotic arm 202 to end the current pressing and adaptive angle adjustment actions.

[0051] In a specific implementation, for compaction operations performed by high-frequency vibrating blocks embedded in the vibrating slab, the system constructs a real-time monitoring and closed-loop control mechanism based on multi-point contact stress feedback. The underlying hardware uses an embedded real-time operating system microcontroller deployed within the slab as the master control node. It utilizes a sampling rate of up to 1 kHz to read the analog voltage signals generated by the pressure sensor array distributed on the slab in real time, and converts them into a high-precision discrete digital signal sequence through the microcontroller's built-in analog-to-digital converter module. The advantage of this high-frequency synchronous sampling mechanism is that it can completely capture the instantaneous dynamic stress change characteristics of the subgrade soil under high-frequency vibration impact. The embedded real-time operating system ensures an absolutely high-priority response to signal acquisition and processing tasks from the underlying software architecture, avoiding data loss or distortion due to system scheduling delays, and providing a reliable, high-fidelity raw data stream for subsequent accurate compaction degree calculation.

[0052] To completely filter out the high-frequency mechanical vibrations of the equipment and the complex hydrodynamic interference noise generated by water flow in deep-water environments, the system innovatively constructs a recursive least-squares digital filter based on an adaptive forgetting factor within the microcontroller. Let the original pressure acquisition signal sequence be... The target smoothing signal is The parameter update algorithm for this filter is derived as follows: ; ; ; In the matrix update formula above, The matrix representing the input signal at the current moment; The weight vector represents the dynamic adjustment of the filter; This represents the gain matrix used to control the error correction step size; This represents the inverse correlation matrix, reflecting the error covariance of historical data; This represents the adaptive forgetting factor. Based on the underwater vibration frequency characteristics, the system dynamically limits the adaptive forgetting factor to the range of 0.95 to 0.99. The beneficial effect of this algorithm is that, through matrix iteration using recursive least squares, it can remove hydrodynamic white noise and equipment harmonic interference mixed in the effective pressure signal in real time without introducing significant phase delay. The introduction of the adaptive forgetting factor allows the filter to dynamically adjust the dependence weight on historical data according to the severity of the underwater transient impact, thereby outputting an extremely smooth and physically meaningful effective dynamic contact stress sequence. .

[0053] After obtaining the smoothed stress sequence, the system inputs it into the compaction degree calculation software module. This module, combined with the real-time vibration acceleration characteristics of the vibrating slab, constructs a nonlinear soil dynamic impedance inversion model to accurately calculate the absolute compaction degree index. The mathematical model formula is as follows: ; In this formula, This represents the total number of valid nodes in the pressure sensor array. Representing the Smooth contact stress at each node after filtering; This represents the local dip angle of the contact surface at the node under complex terrain, and is used to project the stress vector to the effective compaction direction; and represent the current real-time physical amplitude and angular frequency of the high-frequency vibrating block, respectively; the exponential term on the right side of the formula represents the decay function of energy dissipation due to soil plastic deformation, where The empirical attenuation coefficient is pre-calibrated and extracted based on the subgrade soil type, while the integral part accumulates the rate of dramatic change in stress over time in real time. The beneficial effect of this algorithm is that it overcomes the technical blind spot of traditional methods that rely solely on a single pressure threshold to determine compaction. By introducing vibration energy normalization and an integral attenuation term reflecting the plastic dissipation of the subgrade soil, this model can accurately distinguish between the pseudo-high pressure generated by the probe touching a hard boulder and the real reaction force brought about by the overall compaction of the subgrade, thus providing the system with an absolutely objective global compaction quantification index unaffected by single geological abrupt changes.

[0054] At the data interaction and closed-loop execution level, the microcontroller encapsulates the calculated absolute compaction index scalar data into standard data frames according to the application layer communication protocol. This data is then broadcast in real-time to the control cabinet of the rock-throwing system via a redundant controller area network bus with strong electromagnetic interference resistance, transmitting ten times per second. After receiving the initial rock-throwing quantity command from the defect volume calculation module and commencing precise replenishment operations, the rock-throwing system synchronously and continuously parses the absolute compaction index data frames on the bus. The system internally employs a rigorous closed-loop termination logic condition: when the compaction data for a consecutive preset period reaches or exceeds the system's target compaction threshold, and the first derivative of the compaction data, i.e., the compaction growth rate, approaches zero, a termination response is triggered. The beneficial effect of this dual-judgment logic is that it ensures not only that the subgrade bed has reached the designed absolute density but also that the subgrade soil and rock have reached their physical compaction limit and no longer undergo plastic settlement. At this moment, the control cabinet automatically shuts off the rock-throwing valve instantly through the bottom actuator to stop the material feeding, and simultaneously issues a stop command to control the robotic arm to end its current pressing and adaptive angle adjustment actions. This complete closed-loop linkage mechanism completely eliminates the under-compaction and over-throwing phenomena commonly found in deep-water construction operations, significantly saves on crushed stone material costs, effectively avoids fatigue damage to mechanical parts due to excessive stress, and achieves the ultimate precision and fully automated control of underwater foundation leveling operations.

[0055] In the preferred embodiment, the movement of the leveling device 5 based on the base bed defect parameters includes: The dynamic positioning control software module deployed on the 4 edge computing nodes of the leveling vessel subscribes to and receives movement commands issued by the digital twin model of the base bed in real time through the MQTT Internet of Things communication protocol. The commands include the three-dimensional coordinates of the target defects, spatial attitude constraints and flow field prediction data. The control system parses movement commands and integrates the absolute coordinates output by the carrier phase differential technology system with the six-degree-of-freedom pose data output by the high-precision inertial measurement unit. A nonlinear model predictive control algorithm is used to calculate the optimal three-dimensional trajectory. To overcome the hydrodynamic deformation and nonlinear spatial position shift of the cable caused by the rapid currents at a depth of 100 meters, a flexible coupled spatial compensation formula based on ocean current vector feedforward is incorporated into the trajectory planning algorithm, deriving and generating the compensated high-precision tracking coordinates. ; in, The compensation three-dimensional coordinate vector that the leveling vessel 4 actually needs to track; The three-dimensional coordinates of the target defect issued by the digital twin model; The length of the integration window for the model's prediction in the time domain; Let be the generalized mass inertia matrix of the underwater leveling system; Let be the Jacobian transformation matrix from the local coordinate system to the global coordinate system. To advance the control input torque; and These are the nonlinear Coriolis centripetal force matrix and the real-time ocean current velocity vector, respectively. Dynamically changing hydrodynamic damping dissipation matrix; This is to provide the real-time lowering length parameter for the connecting cable. The nonlinear flexible deformation attenuation coefficient is obtained by fitting historical underwater stress tensor data. To prevent extremely small positive numbers with a denominator of zero; The dynamic positioning control software module uses a sequential quadratic programming algorithm to perform rolling optimization of the motion trajectory, decomposing the optimal trajectory into thrust distribution commands of the multi-axis vector thruster and dynamic take-up and release commands of the hydraulic winch, driving the leveling device 5 connected to the leveling boat 4 to move towards the defect location along the planned overshoot-free trajectory. After the leveling device 5 reaches directly above the defect location, it calls the internal programmable logic controller to activate the hydraulic servo leveling subroutine. Combined with the closed-loop feedback of the high-precision depth gauge, it controls multiple sets of hydraulic cylinders to extend and retract to perform attitude decoupling and elevation fine adjustment, thereby performing precise physical scraping and compaction leveling operations on the defect location until the difference between the measured elevation and the target design converges within the safe tolerance range.

[0056] In a specific implementation, driving the leveling device to move based on the parameters of the subgrade defect is the core physical execution step for achieving high-precision underwater correction. First, the dynamic positioning control software module deployed on the edge computing node of the leveling vessel employs a lightweight Internet of Things (IoT) communication protocol. Due to the limited network bandwidth and fluctuating signals in the offshore operating environment, this communication protocol ensures that the system can subscribe to and receive movement commands from the digital twin model of the subgrade at a high frequency and stably. These commands include the three-dimensional coordinates of the target defect, spatial attitude constraints, and flow field prediction data. Upon receiving the commands, the leveling vessel's control system deeply integrates the absolute geographic coordinates output by the carrier phase differential technology system with the six-degree-of-freedom pose data output by the high-precision inertial measurement unit. This multi-source sensor fusion mechanism constructs a drift-free absolute spatial coordinate reference for the entire underwater operating platform. Based on this coordinate reference, the system uses a nonlinear model predictive control algorithm to calculate the optimal three-dimensional motion trajectory. Its beneficial effect is that the nonlinear model predictive control algorithm has the ability to predict the future state of the system in advance. It can plan a smooth trajectory that avoids obstacles and conforms to the hydrodynamic characteristics in advance in the extremely challenging deep-water dynamic environment, which greatly reduces the ineffective movement and excessive energy consumption of the equipment.

[0057] To completely overcome the industry-wide technical challenge of hydrodynamic deformation and nonlinear spatial position displacement of cables caused by rapid currents at depths of 100 meters, the system incorporates a flexible coupled spatial compensation formula based on ocean current vector feedforward into the core of its trajectory planning algorithm. This formula derives and generates highly accurate tracking coordinates after precise compensation. The specific mathematical model is as follows: ; In this mathematical formula, This represents the calculated and compensated three-dimensional coordinate vector that the underlying control system of the leveling vessel actually needs to track. The three-dimensional coordinates of the target defect initially generated by the digital twin model; This represents the length of the integration window in the time domain for the nonlinear model prediction. The matrix within the curly braces inside the integral sign in the formula represents the rigid body dynamics model of the underwater leveling system. The generalized mass inertia matrix representing the underwater leveling system reflects the physical inertial resistance of the mechanical system against accelerated motion in a deep-water flow field. This represents the Jacobian transformation matrix from the local execution coordinate system to the global geodetic coordinate system; The control input torque represents the combined output of multiple vector thrusters; The matrix representing the nonlinear Coriolis centripetal force generated by the system's motion; This represents the hydrodynamic damping dissipation matrix, which is strictly based on the real-time ocean current velocity vector obtained from environmental sensors. It undergoes dynamic nonlinear changes. The exponential term following the dot product in the formula is the core of the entire algorithm: the cable flexibility compensation model. This parameter represents the real-time lowering length of the cable connecting the surface mother ship and the underwater execution equipment. This represents the nonlinear flexible deformation attenuation coefficient obtained by fitting a large amount of historical underwater stress tensor data. Representing a tiny positive constant, its purely mathematical function is to prevent the denominator from being zero in still water conditions (i.e., when the flow velocity is zero), thus avoiding meaningless singularity collapses in calculations. This formula perfectly decouples and feeds forward the complex nonlinear hydrodynamic equations underwater from the bow-shaped deformation effect of slender cables under the impact of water flow. Its beneficial effect is that the system can pre-calculate the spatial offset vector that needs to be resisted by counter-force before the turbulent current pushes the leveling device away from the target positioning point. This completely upgrades the passive hysteresis correction in traditional control to active pre-current compensation, ensuring that even in extremely harsh conditions such as hundreds of meters deep water and strong flow fields, the spatial positioning error of heavy equipment can still be locked within an extremely small centimeter range.

[0058] After generating high-precision tracking coordinates, the dynamic positioning control software module uses a sequential quadratic programming algorithm to perform rolling optimization on the entire motion trajectory. This algorithm can quickly find the globally optimal solution that minimizes trajectory tracking error and system energy consumption under strict boundary conditions that satisfy thrust limits and mechanical rotation angle constraints. Subsequently, the system precisely decomposes this globally optimal trajectory in the underlying control logic into independent thrust distribution commands for the multi-axis vector thrusters and dynamic deployment and retrieval commands for the hydraulic winch. The thrust commands control the surface hull to maintain its current-resistant positioning, while the winch commands control the dynamic follow-up compensation of the vertical descent depth. These two commands work in precise coordination to drive the leveling device to steadily move towards the underwater defect location along a pre-planned, overshoot-free, three-dimensional smooth trajectory. The beneficial effect is that it ensures that when large underwater machinery weighing tens of tons approaches a fragile gravel bed, it will not experience destructive pendulum-like spatial oscillations or coordinate overshoot sprints, achieving a smooth and flexible landing for heavy machinery and greatly protecting the already inspected and approved leveled area around the bed from secondary damage caused by violent hydrodynamic impacts.

[0059] Once the leveling device precisely reaches the target defect location, the control system immediately activates the hydraulic servo leveling subroutine via the internal programmable logic controller. At this point, a high-precision depth gauge installed at the bottom of the leveling device serves as the final physical feedback source for closed-loop control, feeding back the true physical elevation of the base bed to the microprocessor at an extremely high sampling frequency. Based on the measured elevation deviation data, the system controls the coordinated extension and retraction of multiple independent hydraulic cylinders via high-speed responsive hydraulic servo proportional valves. The independent differential extension and retraction of these multiple hydraulic cylinders perfectly isolates and decouples the pitch angle and roll attitude of the leveling device in three-dimensional space, completely filtering out high-frequency mechanical heave interference transmitted underwater by the cables due to surface wave fluctuations. After completing attitude decoupling and elevation fine-tuning, the leveling device chassis maintains an extremely stable, absolutely level attitude. The system then executes the final physical leveling and high-frequency vibration compaction leveling combined collaborative operation. This process continues in a closed loop iterative manner until the difference between the measured physical elevation of the base bed and the theoretical design of the digital twin model completely converges within the safety tolerance range required by engineering specifications. Its beneficial effect is that it completely solves the industry problem that traditional suspended underwater equipment cannot perform precise height-fixing operations due to wave coupling. Through the rigid decoupling of electromechanical-hydraulic integration and the deep control of the bottom closed loop, it achieves a perfect closed-loop acceptance of the 100-meter deep water foundation bed from high-precision perception detection to high-precision physical correction, giving the entire set of intelligent equipment the ultimate construction operation capability to cope with extremely fine and strict construction standards.

[0060] Example 4 S4 includes controlling the ROV robot 2 to initiate a secondary scanning and comparison operation on the area corrected by the leveling device 5 and the stone-throwing system 3; and calling the cloud comparison module to calculate the difference in defect size before and after correction. If the difference is lower than the set standard, the completion status field of the corresponding coordinate will be updated using the database write command; The system backend automatically extracts global operation image data recorded in the PostgreSQL time series database according to a preset time period, and adjusts the hyperparameters of the recognition model through the gradient descent optimization algorithm to improve the defect recognition generalization ability under complex working conditions.

[0061] In a specific implementation, initiating a secondary inspection scan of the corrected area is the final and crucial step in constructing a closed-loop control system for bed leveling. After the leveling device and the ballast system complete the physical correction, the control system immediately drives the ROV robot back to the work area for a secondary high-precision multimodal scan. After acquiring the secondary scan data, the system calls a comparison module deployed in the cloud to calculate the difference in defect size before and after correction. To accurately quantify the correction effect, the cloud comparison module constructs a calculation model based on the spatial residual of a three-dimensional point cloud. The formula for calculating the spatial residual is: ; In this mathematical formula, Represents the defect space residual volume before and after correction; The three-dimensional spatial integral domain representing the specific defect detected by the detection system; This represents the three-dimensional morphological geometry of the substrate surface constructed from the initial scan before correction; This represents the three-dimensional geometric function of the substrate surface obtained from a secondary scan after the correction operation. The difference between the two is taken, and the absolute value is then triple-integrated over the entire defect space domain to precisely calculate the absolute physical volume difference filled or removed by the correction action. The beneficial effect of this step is that it completely changes the technical drawbacks of traditional underwater construction, which involves blind operations and extreme difficulty in post-construction review. By leveraging the powerful computing capabilities of the cloud to perform real-time microscopic comparisons of massive point clouds, it provides an absolutely objective and quantitative mathematical basis for subsequent quality acceptance.

[0062] After calculating the difference in defect size and spatial residual volume before and after correction, the system compares them with the safety tolerance standards set in the engineering design. If the comparison result shows that the difference is lower than the set standard, it indicates that the flatness of the subgrade in that area has reached the extremely fine flatness standard. At this time, the system control center will immediately use the underlying database to write instructions to update the completion status field corresponding to the specific three-dimensional coordinates. This real-time update of the status will not only mark the completion with a specific highlighted color block on the panoramic visualization interface of the subgrade digital twin model, but more importantly, it will set the coordinates as an obstacle-avoided or completed area in the global path planning system. Its beneficial effect is that, through the millisecond-level time synchronization of the completion status field, the hidden dangers of repeated construction or missed leveling when multiple underwater devices are working simultaneously are completely eliminated, greatly improving the global scheduling efficiency and construction quality certainty of large-scale underwater subgrade leveling projects.

[0063] As the project progresses, the turbidity, light intensity, and ocean current characteristics of the underwater construction environment will undergo significant seasonal or sudden changes. To prevent the visual recognition model from experiencing a decline in generalization ability during long-term operations, the system automatically extracts global operation image data and corresponding manual review and fine-tuning instructions from the PostgreSQL time-series database at preset time intervals. Utilizing these massive amounts of challenging sample data accumulated from real-world complex working conditions, the system triggers a gradient descent optimization algorithm in the cloud to dynamically adjust the hyperparameters of the recognition model. The gradient descent iterative formula for optimizing the network node weights is as follows: ; In this optimization formula, This represents the model weight parameter matrix after this periodic iteration update; This represents the weight parameter matrix currently being used by the model in the current cycle. This represents the model learning rate, used to control the step size of each weight update; This represents the gradient vector of the partial derivatives of the loss function with respect to the current weight parameter matrix; This represents the total number of high-quality global job image samples extracted from a single batch of the PostgreSQL time-series database; The input is a measured image feature vector containing complex underwater noise; These are labels representing genuine defects that have been verified through real-world operations or manually corrected. This represents the predicted output feature vector of the current visual detection model for the input image; The cross-entropy loss function represents the relationship between the predicted output and the true label; This represents the regularization penalty coefficient, used to prevent the model from overfitting on new underwater image sample sets. The beneficial effect of this automated self-learning loop is that it endows the entire AI detection system with lifelong learning capabilities, allowing it to continuously evolve with increasing operational mileage. By continuously absorbing and digesting complex visual features under extreme conditions, the model's defect recognition generalization ability in harsh environments is substantially improved, completely overcoming the technical limitation of traditional static visual models that are prone to widespread failure when water quality changes drastically. This ensures absolute high availability and high recognition accuracy throughout the entire lifecycle of the deep-water bed intelligent detection system.

[0064] In the preferred embodiment, when the underwater flow velocity sensor detects in real time that the water flow velocity exceeds the safe operating threshold, the safety control module is automatically triggered to execute the anti-flow protection strategy. The specific algorithm and control steps include: The high-frequency sampling sequence of the flow velocity sensor is smoothed by a sliding window mean filtering algorithm to eliminate random high-frequency noise caused by underwater pulsating turbulence and output a real-time effective flow velocity scalar. When the effective flow velocity scalar continuously crosses the safe operation threshold for a preset time period, the ROV robot 2 is forced to activate the dual sonar cooperative anti-current mode of forward fixed distance and downward anti-current. In the dual sonar cooperative anti-current mode, the downward anti-current sonar locks the seabed topography to calculate the real-time seabed drift vector of the ROV robot 2 at high frequency, and the forward fixed distance sonar switches to the near-field high-resolution scanning mode to avoid suspended obstacles that move rapidly with the fluid. The underlying main control chip of ROV robot 2 is based on the real-time seabed drift vector. It calls the linear quadratic regulator algorithm to calculate the optimal thrust compensation matrix to counteract the water flow thrust, and drives multiple vector thrusters to generate adaptive reverse thrust to maintain the original planned detection route without deviation. Meanwhile, the safety control module sends a fixed-point anchoring command to the leveling vessel 4 through the communication network, activates the dynamic positioning system of the leveling vessel 4, calculates the combined force of wind direction, wave and water flow data, controls the azimuth thruster combined with the anchoring system to fix the absolute attitude of the fuselage, and dynamically reduces the planned spacing of the underwater detection path according to the flow velocity ratio to ensure the data acquisition density in high flow velocity environments.

[0065] In a specific implementation, when the underwater flow velocity sensor detects in real time that the water flow velocity exceeds the safe operating threshold in extreme conditions, the system automatically triggers the safety control module to execute the anti-flow protection strategy.

[0066] First, the system uses a sliding window mean filtering algorithm to smooth the high-frequency sampling sequence output by the flow velocity sensor. Pulsating turbulence in deep water environments generates a large amount of random high-frequency noise, causing drastic fluctuations in instantaneous flow velocity readings. The mathematical model of this filtering algorithm is as follows: ; In this formula, This represents the real-time effective flow velocity scalar output after filtering. This represents the total number of sampling points in the sliding window. This represents the raw high-frequency instantaneous flow velocity sequence collected by the flow velocity sensor; This represents the discrete sampling time interval of the flow velocity sensor. Using the above formula, the system calculates the arithmetic mean of the flow velocity within a specific time window in real time, thereby completely eliminating high-frequency data spikes caused by underwater turbulence. Its beneficial effect lies in its ability to extract the effective flow velocity scalar that truly represents the macroscopic flow intensity, effectively preventing frequent false triggering of anti-flow protection strategies due to instantaneous flow pulsations, and greatly improving the decision-making stability and robustness of the safety control system.

[0067] When the effective flow velocity scalar continuously crosses the safe operating threshold within a preset time period, the system determines that the current water area has entered a state of sustained strong current, and then forces the underwater robot to activate a dual-sonar cooperative current-fighting mode, combining forward ranging and downward current-fighting capabilities. In this dual-sonar cooperative current-fighting mode, the downward current-fighting sonar locks onto the absolutely static seabed topography by emitting high-frequency sound waves, and uses the Doppler frequency shift principle to calculate the underwater robot's real-time seabed drift vector at high frequency. At the same time, the forward ranging sonar instantly switches to a near-field high-resolution scanning mode. The beneficial effect is that the downward sonar provides the absolute motion reference frame required to combat undercurrents, while the forward sonar can accurately detect and avoid clumps of silt and suspended obstacles that move rapidly with the turbulent current. The physical coordination of the dual sonars fundamentally ensures that the underwater robot will neither lose its way nor be damaged by collisions in extremely turbulent and turbulent deep-water environments, giving the underwater equipment extremely high survivability and environmental awareness redundancy.

[0068] After acquiring the real-time seabed drift vector, the underwater robot's underlying main control chip calls a linear quadratic regulator algorithm to calculate the optimal thrust compensation matrix against the water flow thrust. The linear quadratic regulator algorithm solves for the optimal control law by minimizing the combined cost function of state error and control energy. Its core cost function formula is as follows: ; In this mathematical model, This represents the overall system performance index that needs to be minimized. This represents the system state matrix, which contains the real-time seabed drift vector and its rate of change. The thrust control matrix represents the output of multiple vector thrusters of an underwater robot; The state error weight matrix is ​​used to severely punish the amount of spatial drift of the robot that deviates from the predetermined path. The control energy weight matrix is ​​used to limit the peak power output of the thrusters to save energy. The underlying main control chip obtains the optimal feedback gain matrix by solving the algebraic Riccati equation, thereby driving multiple vector thrusters to generate adaptive reverse thrust. Its advantage lies in finding a perfect mathematical balance between maintaining the original planned detection path without deviation and minimizing battery energy consumption. This avoids the thrust saturation and fuselage oscillation phenomena that are prone to occur in conventional proportional-integral-derivative control when dealing with strong currents, achieving flexible and extremely precise dynamic hovering and trajectory tracking.

[0069] Meanwhile, the extremely high underwater current speeds are often accompanied by harsh sea conditions on the surface. The safety control module sends a fixed-point anchoring command to the leveling vessel on the surface via the communication network. Upon receiving the command, the leveling vessel immediately activates its dynamic positioning system. The system accurately calculates the environmental force vector currently acting on the hull by integrating high-frequency wind direction data, wave undulation data, and measured water flow velocity data. Subsequently, the control system directs the azimuth thruster in conjunction with the physical anchoring system to generate a counteracting torque equal in magnitude but opposite in direction to the environmental force, thereby firmly fixing the leveling vessel's hull attitude to the construction area. The beneficial effect is that the absolute stillness of the surface mother ship eliminates the enormous drag force transmitted to the underwater robot via the umbilical cable due to the hull drifting with the current, providing a rock-solid physical anchor point for the stable operation of the underwater equipment.

[0070] Furthermore, to counteract the dilution effect of strong current on scan quality, the system dynamically reduces the planned spacing of the underwater detection path according to the current velocity. The dynamic spacing adjustment formula is as follows: ; In this formula, This represents the actual underwater detection path spacing after dynamic scaling calculation; This represents the initial detection path spacing pre-planned under still water conditions; This represents the real-time effective flow velocity scalar of the aforementioned filter output; The flow rate represents the safe operating threshold set by the system. This represents the velocity spacing attenuation coefficient derived from the sonar beam coverage characteristics. When the flow velocity exceeds a safe threshold, the path spacing decreases exponentially and smoothly. Its beneficial effect is that strong currents can cause unavoidable yaw angle sideslip during underwater robot tracking, resulting in gaps in the sonar scanning trajectory. By forcibly reducing the detection path spacing, the physical overlap between adjacent scanning strips is significantly increased, ensuring the density and integrity of the 3D point cloud data acquisition of the substrate bed under harsh high-velocity environments, and completely eliminating geological blind spots under extreme sea conditions.

[0071] In the preferred embodiment, when the size of the stone block identified by the artificial intelligence detection model exceeds the mechanical safety cutting threshold, a protective operation mode is automatically executed to avoid equipment damage. The specific algorithm and mechanical control steps include: The 3D reconstruction module extracts the 3D point cloud bounding box of the protruding stone on the surface of the base bed based on the multimodal fusion data, and calculates the maximum geometric scalar of the stone along the principal inertial axis through the principal component analysis algorithm. When the maximum geometric scalar exceeds the set mechanical safety cutting threshold, the safety control module sends the highest priority hardware interrupt signal to the underlying actuator through the controller local area network bus, automatically intercepting and terminating the conventional rigid scraping command currently being executed by the leveling device 5. Then the leveling device 5 automatically switches to a protective operation mode of first leveling and then compacting: first, the mechanical arm 202 is controlled to adjust the water-facing inclination angle of the vibrating leveling plate 203 to the preset bulldozing angle, so as to limit the running speed and drive the leveling device 5 to push the large-sized stones to the adjacent pit and depression area around the foundation bed. During the pushing operation, the lateral resistance feedback value in the hydraulic circuit of the robotic arm 202 is monitored in real time. When the lateral resistance feedback value drops sharply and falls below the set no-load resistance threshold, the system determines that the large-sized stone has detached from the push plate and fallen into the pit. The controller then issues a posture reset command, controlling the robotic arm 202 to reset the tilt angle of the vibrating plate 203 to a horizontal state, and activates the internal high-frequency vibration block to perform high-frequency compaction on the area, thereby avoiding rigid cutting damage to the mechanical structure by large stones while completing the flatness correction of the base bed defects in a closed loop.

[0072] In a specific implementation, for extreme working conditions where the artificial intelligence detection model identifies abnormally large boulders, the system constructs a highly rigorous protective operating mode. First, the 3D reconstruction module uses multimodal fusion data to extract the 3D point cloud bounding boxes of the protruding boulders on the subgrade surface. To accurately extract physical dimensions with engineering guidance significance from the highly irregular boulder geometry, the system employs principal component analysis to calculate the maximum geometric scalar of the boulder along its principal axis of inertia. The calculation model for the maximum geometric scalar is as follows: ; In this mathematical model, This represents the maximum geometric scalar dimension of the stone along the principal axis of inertia, obtained through principal component analysis. The total number of valid discrete point clouds within the bounding box of the extracted 3D point cloud of the protruding rock; Represents the first in the enclosure box A spatial three-dimensional coordinate vector of a point cloud; The first principal component eigenvector, corresponding to the largest eigenvalue, represents the spatially longest geometric principal axis of the boulder after eigenvalue decomposition of the point cloud covariance matrix. This vector physically represents the longest possible geometric principal axis of the boulder. By projecting all point cloud coordinate vectors onto this first principal component eigenvector and calculating the projection span, the system can obtain the longest physical contour dimension of the boulder with extreme accuracy. Its advantage lies in completely eliminating the size assessment errors caused by the shooting angle and optical distortion in two-dimensional vision, achieving absolutely accurate quantification of the actual cutting cross-section of irregular underwater boulders, and providing impeccable data support for determining whether a mechanical protection mechanism is triggered.

[0073] When the calculated maximum geometric scalar value exceeds the system's set mechanical safety cutting threshold, the safety control module instantly sends a highest-priority hardware interrupt signal to the underlying actuator via the controller area network bus. Due to the immense inertia of underwater heavy machinery, conventional software polling mechanisms suffer from fatal communication delays. The highest-priority hardware interrupt signal, however, can directly penetrate the operating system's task scheduling layer, forcibly cutting off the servo drive's enable circuit within microseconds, thereby automatically intercepting and terminating the leveling device's currently executing conventional rigid leveling command. Its beneficial effect is that it averts a catastrophic collision between the cutting edge and the oversized boulder, absolutely ensuring the structural integrity of the robotic arm's core hydraulic cylinder and the vibrating leveling plate, fundamentally preventing catastrophic damage to underwater equipment caused by reckless construction.

[0074] After successfully intercepting the dangerous command, the system automatically switches the leveling device into a protective operation mode of first leveling and then compacting. The system first controls the robotic arm to finely adjust the inclination angle of the vibratory leveling plate's water-facing side, rotating it to the preset bulldozing angle. Once the bulldozing angle is reached, the system, by limiting the operating speed of the hydraulic motor, flexibly drives the leveling device to slowly push large boulders into adjacent depressions around the foundation bed. Its beneficial effect lies in cleverly transforming the originally highly destructive rigid cutting resistance into constructive horizontal thrust. This not only perfectly resolves the mechanical obstruction crisis caused by the boulders but also directly utilizes the discarded protruding boulders to fill the surrounding depressions, achieving on-site balance and turning waste into treasure in underwater earthmoving within a micro-area, greatly saving the operating costs and time of the mother ship dropping additional crushed stone materials.

[0075] During the slow movement of large boulders, the system needs to accurately determine whether the boulder has landed in the target crater. Since the boulder is completely obscured by the push plate, both optical vision and sonar have detection blind spots. Therefore, the system innovatively introduces a purely physical hydraulic feedback monitoring mechanism. The microcontroller monitors the lateral resistance feedback value in the robotic arm's hydraulic circuit at high frequency in real time. The mathematical logic criteria for its state determination are as follows: ; ; In the above logical decision equation, This represents the lateral pushing resistance feedback value that is collected in real time and filtered and converted in the hydraulic circuit of the robotic arm; The no-load resistance threshold is pre-set by the system based on water flow resistance and internal mechanical friction; its first derivative term represents the transient rate of change of lateral resistance over time. This represents the threshold for the step descent rate at which resistance experiences a precipitous drop. When the real-time resistance value is lower than the no-load resistance threshold, and its descent rate exceeds the step descent rate threshold, the system, from a physical and mechanical perspective, definitively determines that the large rock has lost support, completely detached from the push plate, and successfully fallen into the deep pit. Its beneficial effect is that it endows underwater equipment with a tactile sensing capability completely unaffected by the poor visibility in deep water. Through the transient physical changes in hydraulic pressure, it achieves blind-state precise confirmation of operational progress, ensuring zero-error switching of complex operational state machines.

[0076] After the system determines that a boulder has fallen into a depression, the controller immediately issues a posture reset command, controlling the robotic arm to quickly and smoothly restore the tilt angle of the vibrating slab from a bulldozing angle to a horizontal state. Next, the system activates the high-frequency vibrating blocks built into the slab at full power, performing high-frequency compaction on the area where the boulder was filled. The beneficial effect is that the high-frequency vibration liquefies the loose gravel around the boulder, causing it to tightly wrap around and wed into the gaps between the boulder and completely eliminate the potential for voids caused by the boulder filling. This series of smooth, closed-loop controls, while perfectly avoiding mechanical rigid cutting damage, successfully corrected the flatness and compacted the extreme defects in the subgrade to a high standard, demonstrating the extremely high self-healing and flexible construction capabilities of the entire underwater intelligent system when encountering extremely harsh geological conditions.

[0077] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A method for intelligent detection and closed-loop correction of defects in deep-water subgrade leveling, characterized by: The method includes: S1. Construct a digital twin model of the bed base based on communication and positioning technologies, calibrate the bed base boundary, and establish a collaborative mechanism between detection and correction equipment; S2, ROV robot (2) combined with robotic arm (202) adjusts the vibration compaction detection of vibrating plate (203) and provides real-time feedback on compaction degree, which is fed back to the stone throwing system. The stone throwing system (3) automatically calculates the amount of stone thrown based on the defect volume detected by AI and performs adaptive correction. S3. Using an ROV robot (2) equipped with a high-definition camera (201) and multiple sonars (204), multimodal detection is performed to obtain the parameters of the base bed defects. After being identified by an artificial intelligence detection model, the parameters are mapped to the digital twin model of the base bed in real time. The leveling device (5) is driven to move according to the parameters of the base bed defects. S4. Perform a secondary inspection scan on the area that has been leveled and riprap corrected to verify the closed loop, and iteratively optimize the artificial intelligence inspection model based on the operation data.

2. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 1, characterized in that: S1 includes building a base digital twin model based on a cloud server architecture, and deploying data processing modules and 3D visualization service modules on the server through Docker containerization technology; Nginx reverse proxy cluster is used to achieve network load balancing of multi-source heterogeneous data, and the received environmental status and device location data are persistently stored in PostgreSQL time series database; By combining the BeiDou positioning system to calibrate the bed boundary, a low-latency command interaction mechanism is established between the detection equipment and the correction equipment.

3. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 2, characterized in that: S3 This includes deploying ROV robots (2) into the water from a maintenance vessel (1); The ROV robot (2) uses the downward anti-current sonar in multiple sonars (204) to obtain the Doppler ground velocity in real time, uses the forward fixed-range sonar in multiple sonars (204) to penetrate underwater suspended objects to locate abnormal areas, and coordinates with the high-definition camera (201) to synchronously capture image data at a preset frame rate. Combining the U-INS underwater inertial navigation system, an error state Kalman filter algorithm based on flow field adaptation is used to ensure stable navigation trajectory; the specific algorithm implementation steps are as follows: Construct a state vector containing position error, velocity error, and attitude error. The system velocity output by U-INS and the ground velocity output by the downward anti-current sonar. The difference sequence is used as the observation benchmark; to resist high-frequency strong flow field interference in deep water environments, a value sequence based on the instantaneous velocity of the current deep water flow field is established. The adaptive measurement noise covariance matrix derived from the underwater robot's ground velocity: in The initial covariance matrix under calibration conditions. This is the environmental turbulence penalty coefficient derived from underwater hydrodynamic characteristics; Optimal Kalman gain is dynamically calculated using the covariance matrix: ； The measurement residuals are calculated using the optimal Kalman gain, and the trajectory drift integral error of the U-INS system is corrected in real time using a closed loop, outputting the optimal pose coordinates with anti-flow field interference characteristics. The acquired image data and environmental data containing the optimal pose coordinates are packaged and compressed using the feature extraction algorithm of the edge computing node. Utilizing the 5G communication link on the water, dynamic routing scheduling and network load balancing are performed via an Nginx reverse proxy cluster deployed on a cloud server, and the data is uploaded in real time in a high-concurrency mode to the backend interface of the digital twin model of the base bed deployed based on a Docker containerized architecture. After the received data stream is decompressed and multi-source spatiotemporal aligned by the backend service program, the structured optimal pose coordinates and environmental state parameters are persistently written into the PostgreSQL time series database, and the unstructured image data is pushed to the distributed object storage module. The digital twin model of the bed generates a spatial control vector based on the updated data, and drives the leveling device (5) connected to the leveling vessel (4) to move along the planned three-dimensional trajectory toward the target defect pose.

4. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 3, characterized in that: The training process of the AI ​​detection model is performed using the PyTorch deep learning framework deployed on a cloud-based GPU-accelerated cluster. The steps include: An initial dataset was constructed by collecting images of defects in underwater foundation beds from real engineering projects. Then, DCGAN generative adversarial network was used for feature enhancement and sample expansion. To constrain the texture and shadow generation quality of the generator in deep-water, high-turbidity environments, an adaptive generative adversarial loss function incorporating prior knowledge of underwater dark passages is constructed; the optimization formula for the generator's objective function is as follows: ； in, For the discriminator function, To augment the multimodal image matrix output by the generator, To derive the ambient light scattering coefficient using water flow data, A pre-defined a priori threshold for the dark channel to define underwater bedrock characteristics. and The nonlinear distribution hyperparameters are dynamically tuned based on real-time water turbidity feedback; an expanded sample library covering pits, protrusions and crack edges in extremely dark environments is iteratively generated through dynamic game between the generator and the discriminator. Using a Python script, the expanded sample library was divided into training, validation and test sets according to the proportions, and then input into the YOLOv5 detection model for iterative training. A channel attention mechanism module is embedded in the feature pyramid network of YOLOv5 to enhance feature extraction of underwater blurred edges; To improve the robustness of underwater defect size identification, a bounding box regression loss function for underwater defect morphology perception is constructed for the YOLOv5 model. This bounding box regression loss function combines depth and elevation outliers obtained from sonar to establish a three-dimensional spatial penalty mechanism. Its derivation formula is as follows: in, To predict the spatial intersection-union ratio between the bounding box and the true bounding box, The square of the Euclidean distance between their centers. To cover the minimum diagonal distance of the closure regions of both, This is a penalty term for aspect ratio consistency in a two-dimensional plane. The scalar value represents the measured depth feature of the defects in the sonar point cloud feedback. For the pseudo-depth scalar of monocular reconstruction in visual networks, The confidence attenuation coefficient for acoustic-optical fusion is derived from the variance of multi-source heterogeneous data. The AdamW gradient descent optimization algorithm is used, combined with a cosine annealing learning rate dynamic decay strategy to optimize the network node weight matrix of the YOLOv5 detection model. The convergence status of the loss function is continuously monitored on the validation set until the average accuracy of all classes reaches a preset threshold. Training is then stopped and the lightweight model weight file reconstructed by the TensorRT tensor acceleration module is exported. This file is then packaged into a standard inference environment container image and pushed to edge computing nodes for deployment and execution.

5. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 4, characterized in that: The deployed AI detection model receives image data and environmental data with optimal pose coordinates synchronously marked by an underwater inertial navigation system in real time. The specific algorithm and computer module implementation steps include: A real-time vision processing pipeline is built using the TensorRT inference engine deployed on a cloud-based GPU acceleration node; upon receiving image data, the image processing module is invoked to execute turbidity-adaptive image enhancement and dehazing algorithms. To overcome backscattering interference caused by high turbidity in deep water and non-uniform light fields, a transmittance compensation defogging formula based on dynamic feedback of environmental parameters is constructed. The spatial transmittance distribution model is derived as follows: ； in, The color channel pixel values ​​of the input image. Estimate the global atmospheric illumination physical quantities. Adjust parameters for defogging level; The second-order Laplacian derivative of the image grayscale matrix is ​​used to extract high-frequency edge details in blurred underwater environments; NTU is the scalar reading of the water turbidity sensor synchronously transmitted in the environmental data. The high-frequency scattering compensation empirical coefficients are obtained by fitting multiple measured data; by introducing the Laplace operator and the turbidity logarithm term, the edge features of the bed rock are adaptively sharpened while eliminating underwater fogging. The dehazed and enhanced image matrix is ​​fed into the YOLOv5 detection model for feature extraction and bounding box regression; the forward propagation of the convolutional neural network is accelerated by the CUDA core, and the output includes the target confidence, defect classification probability and two-dimensional pixel coordinates; the detection results are spatiotemporally aligned and heterogeneous data fused with the corresponding area multibeam sonar point cloud data cached in the digital twin model of the bed bed. The system utilizes a 3D reconstruction and volume calculation module developed using C++ and the PCL library to perform precise volume calculations on identified substrate defects (especially pitted types). To eliminate errors caused by visual deformation and sonar noise, a visual attention-weighted 3D volume integral formula is constructed. ； in, Map the two-dimensional defect connected domain output by the YOLOv5 detection model to the projected integral region in a three-dimensional coordinate system; This is the actual elevation surface function measured by sonar point clouds. The standard flat datum elevation surface function is preset in the digital twin model of the subgrade bed; the area within parentheses on the right side of the formula is a visual attention confidence weight distribution map constructed based on the Sigmoid function, where... This refers to the spatial attention heatmap value output by the YOLOv5 feature pyramid layer. and These are the activation threshold and scaling factor, respectively. The final output includes the three-dimensional coordinates of the defect, the defect type, and the precise defect volume size calculated using the volume integral formula. The base bed defect parameters are then pushed to the control node via a microservice interface, providing a high-precision data source for calculating the amount of rock thrown by the intelligent rock-throwing system.

6. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 1, characterized in that: S2 The process includes compaction operations using a high-frequency vibrating block integrated into the vibrating plate (203), acquiring multi-point contact stress feedback signals through an integrated pressure sensor array, and transmitting compaction data to the rock-throwing system (3) in real time via a controller local area network bus; the specific algorithm and software module implementation steps include: Using the embedded real-time operating system microcontroller deployed inside the leveling device (5) as the main control node, the analog voltage signal of the pressure sensor array is read at a sampling rate of 1000Hz and converted into a discrete digital signal sequence through its own ADC module. To filter out high-frequency mechanical vibrations and hydrodynamic noise generated by water flow from underwater equipment, a recursive least-squares digital filter based on an adaptive forgetting factor is constructed within the microcontroller; Let... This is the original pressure acquisition signal sequence. For the target smooth signal, the derivation of its recursive least squares digital filter parameter update algorithm is as follows: ； ； ； in, The input signal matrix, For the filter weight vector, Here is the gain matrix. It is the inverse correlation matrix. An adaptive forgetting factor; based on underwater vibration frequency characteristics... The dynamic stress is dynamically limited to the range [0.95, 0.99], and the smoothed effective dynamic contact stress sequence is output. The effective dynamic contact stress sequence of the multi-point pressure sensor is input into the compaction degree calculation software module; this module, combined with the real-time vibration acceleration characteristics of the vibrating plate (203), constructs a nonlinear soil dynamic impedance inversion model to calculate the absolute compaction degree index; the mathematical formula for the absolute compaction degree index is: ； in, This represents the total number of nodes in the pressure sensor array. For the first Smooth contact stress at each node, This is the local inclination angle of the contact surface at this node; and These represent the current real-time amplitude and angular frequency of the high-frequency vibrating block, respectively; the exponential term represents the decay function of energy dissipation due to soil plastic deformation. The empirical attenuation coefficient is pre-calibrated based on the subgrade soil type; the microcontroller will calculate the resulting... Scalar data is encapsulated into standard data frames according to the application layer protocol and broadcast in real time to the control cabinet of the rock-throwing system (3) via the redundant controller local area network bus at a transmission frequency of 10Hz; After receiving the initial command for the amount of rock thrown based on the defect volume calculation module, the rock-throwing system (3) starts the precise replenishment operation and simultaneously parses the data on the controller local area network bus. Data frame; compaction data for a continuous preset period Reaching or exceeding the target compaction threshold set by the system When its first derivative approaches zero, the control cabinet of the stone-throwing system (3) triggers the closed-loop termination logic, automatically cuts off the stone-throwing valve through the actuator, and simultaneously issues a stop command to control the robotic arm (202) to end the current pressing and adaptive angle adjustment action.

7. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 6, characterized in that: The movement of the leveling device (5) driven by the base bed defect parameters includes, The dynamic positioning control software module deployed on the edge computing node of the leveling vessel (4) subscribes to and receives in real time movement commands issued by the digital twin model of the bed bed, which include the three-dimensional coordinates of the target defect, spatial attitude constraints and flow field prediction data. The control system parses movement commands and integrates the absolute coordinates output by the carrier phase differential technology system with the six-degree-of-freedom pose data output by the high-precision inertial measurement unit. A nonlinear model predictive control algorithm is used to calculate the optimal three-dimensional trajectory. To overcome the hydrodynamic deformation and nonlinear spatial position shift of the cable caused by the rapid currents at a depth of 100 meters, a flexible coupled spatial compensation formula based on ocean current vector feedforward is incorporated into the trajectory planning algorithm, deriving and generating the compensated high-precision tracking coordinates. ； in, The actual compensation three-dimensional coordinate vector that needs to be tracked for leveling the ship (4); The three-dimensional coordinates of the target defect issued by the digital twin model; The length of the integration window for the model's prediction in the time domain; Let be the generalized mass inertia matrix of the underwater leveling system; Let be the Jacobian transformation matrix from the local coordinate system to the global coordinate system. To advance the control input torque; and These are the nonlinear Coriolis centripetal force matrix and the real-time ocean current velocity vector, respectively. Dynamically changing hydrodynamic damping dissipation matrix; This is to provide the real-time lowering length parameter for the connecting cable. The nonlinear flexible deformation attenuation coefficient is obtained by fitting historical underwater stress tensor data. To prevent extremely small positive numbers with a denominator of zero; The dynamic positioning control software module uses a sequential quadratic programming algorithm to perform rolling optimization of the motion trajectory, decomposes the optimal trajectory into the thrust distribution command of the multi-axis vector thruster and the dynamic take-up and release command of the hydraulic winch, and drives the leveling device (5) connected to the leveling boat (4) to move towards the defect position along the planned non-overshoot trajectory. After the leveling device (5) reaches directly above the defect location, it calls the internal programmable logic controller to activate the hydraulic servo leveling subroutine. Combined with the closed-loop feedback of the high-precision depth gauge, it controls multiple sets of hydraulic cylinders to extend and retract to perform attitude decoupling and elevation fine adjustment, thereby performing precise physical scraping and compaction leveling operations on the defect location until the difference between the measured elevation and the target design converges within the safe tolerance range.

8. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 7, characterized in that: S4 includes controlling the ROV robot (2) to initiate a secondary scanning and comparison operation on the area corrected by the leveling device (5) and the stone-throwing system (3); The cloud-based comparison module is invoked to calculate the difference in defect size before and after correction. If the difference is lower than the set standard, the completion status field of the corresponding coordinate will be updated using the database write command; The system backend automatically extracts global operation image data recorded in the PostgreSQL time series database according to a preset time period, and adjusts the hyperparameters of the recognition model through the gradient descent optimization algorithm to improve the defect recognition generalization ability under complex working conditions.

9. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 1, characterized in that: When an underwater flow velocity sensor detects that the water flow velocity exceeds the safe operating threshold in real time, the safety control module is automatically triggered to execute an anti-flow protection strategy. The specific algorithm and control steps include: The high-frequency sampling sequence of the flow velocity sensor is smoothed by a sliding window mean filtering algorithm to eliminate random high-frequency noise caused by underwater pulsating turbulence and output a real-time effective flow velocity scalar. When the effective flow velocity scalar continuously crosses the safe operation threshold for a preset time period, the ROV robot (2) is forced to activate the dual sonar cooperative anti-flow mode of forward fixed distance and downward anti-flow. In the dual-sonar cooperative anti-current mode, the downward anti-current sonar locks the seabed topography to calculate the real-time seabed drift vector of the ROV robot (2) at high frequency, and the forward fixed-range sonar switches to the near-field high-resolution scanning mode to avoid suspended obstacles that move rapidly with the fluid. The underlying main control chip of the ROV robot (2) is based on the real-time seabed drift vector calculated. It calls the linear quadratic regulator algorithm to calculate the optimal thrust compensation matrix to counteract the water flow thrust, and drives multiple vector thrusters to generate adaptive reverse thrust to maintain the original planned detection route from deviating. At the same time, the safety control module sends a fixed-point anchoring command to the leveling vessel (4) through the communication network, activates the dynamic positioning system of the leveling vessel (4), calculates the combined force of wind direction, wave and water flow data, controls the azimuth thruster combined with the anchoring system to fix the absolute attitude of the fuselage, and dynamically reduces the planned spacing of the underwater detection path according to the flow velocity ratio to ensure the data acquisition density in a high flow velocity environment.

10. The intelligent detection and closed-loop correction method for deep-water foundation leveling defects according to claim 9, characterized in that: When the AI ​​detection model identifies a rock mass larger than the mechanical safety cutting threshold, a protective operating mode is automatically executed to prevent equipment damage. The specific algorithm and mechanical control steps include: The 3D reconstruction module extracts the 3D point cloud bounding box of the protruding stone on the surface of the base bed based on the multimodal fusion data, and calculates the maximum geometric scalar of the stone along the principal inertial axis through the principal component analysis algorithm. When the maximum geometric scalar size is greater than the set mechanical safety cutting threshold, the safety control module sends the highest priority hardware interrupt signal to the underlying actuator through the controller local area network bus, automatically intercepting and terminating the conventional rigid scraping command currently being executed by the leveling device (5). Then the leveling device (5) is automatically switched to enter the protective operation mode of first leveling and then compacting: first, the mechanical arm (202) is controlled to adjust the water-facing inclination angle of the vibrating leveling plate (203) to the preset bulldozing angle, so as to limit the running speed and drive the leveling device (5) to push the large-sized stones to the adjacent pit and depression area around the foundation bed. During the pushing operation, the lateral resistance feedback value in the hydraulic circuit of the robotic arm (202) is monitored in real time. When the lateral resistance feedback value drops sharply and falls below the set no-load resistance threshold, the system determines that the large-sized stone has fallen off the push plate and into the pit. The controller then issues a posture reset command, controlling the robotic arm (202) to reset the tilt angle of the vibrating plate (203) to a horizontal state, and activates the internal high-frequency vibration block to perform high-frequency compaction operation on the area, thereby avoiding rigid cutting damage to the mechanical structure by large stones while completing the flatness correction of the base bed defects in a closed loop.

Images