A deep learning-based adaptive sediment repair robot and a control method thereof

Abstract

The application discloses a kind of adaptive bottom mud repair robot based on deep learning and control method thereof, comprising the following steps: obtaining multi-modal perception data, the multi-modal perception data includes underwater topography data, pollution field distribution data and robot posture data;Based on the multi-modal perception data, a three-dimensional dynamic digital twin model of the bottom mud environment is constructed;An reinforcement learning model is constructed, and the state space is initialized by the three-dimensional dynamic digital twin model for model training;The optimal action parameters are output by the trained reinforcement learning model;According to the optimal action parameters, perform the bottom mud repair work;The application constructs multi-dimensional state space based on deep learning identification, and realizes the accurate control of the robot by combining the reinforcement learning model, which can more targetedly carry out decontamination operation.

Description

Technical Field

[0001] This invention relates to the field of underwater environment remediation technology, and more specifically to a deep learning-based adaptive sediment remediation robot and its control method. Background Technology

[0002] With the increasing demand for water pollution control, sediment remediation, as an important means of improving the aquatic ecological environment, faces complex unstructured operating environments in engineering practice. The sediment environment usually has significant spatial heterogeneity and temporal dynamics. Under the influence of multiple factors such as water flow disturbance, sediment diffusion, and biological activity, its pollutants exhibit obvious spatial gradient differences and continuous diffusion characteristics.

[0003] Actual monitoring has revealed that areas with high concentrations of pollution tend to have higher diffusion rates, meaning that pollutants migrate more rapidly to surrounding low-concentration areas per unit time, leading to a continuous expansion of the pollution range. If high-concentration areas are not prioritized for treatment, the spread of pollution will further reduce overall remediation efficiency, requiring subsequent remediation work to cover a larger area, thus significantly increasing operation time and energy costs.

[0004] Current sediment remediation operations mostly rely on pre-planned path planning or uniform coverage strategies, failing to effectively combine changes in pollution concentration gradients and diffusion dynamics for differentiated scheduling.

[0005] Furthermore, due to noise interference and time delay effects in underwater environmental sensing, local extreme points in polluted areas may change rapidly over time, making it difficult for fixed strategies to continuously and effectively track pollution hotspots.

[0006] Therefore, how to adapt to the highly time-varying underwater environment and achieve precise operational control for bottom sediment remediation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide a deep learning-based adaptive mud remediation robot control method to overcome or at least partially solve the above problems.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A deep learning-based adaptive sediment remediation robot control method includes the following steps: Acquire multimodal perception data, including underwater topography data, pollution field distribution data, and robot posture data; A three-dimensional dynamic digital twin model of the sediment environment is constructed based on the multimodal sensing data; A reinforcement learning model is constructed, and the state space is initialized through the three-dimensional dynamic digital twin model for model training; the trained reinforcement learning model is used to output the optimal action parameters. Perform sediment remediation work according to the optimal action parameters.

[0010] Preferably, the steps for constructing a three-dimensional dynamic digital twin model of the sediment environment include: Preprocessing and spatiotemporal registration are performed based on the multimodal sensing data; A three-dimensional mesh model of the sediment was constructed based on the topographic curvature matrix. A virtual underwater environment is generated by superimposing a dynamic pollution field onto the three-dimensional mesh model based on the pollutant concentration gradient tensor. The dynamic pollution field includes a scalar field and a vector field. The scalar field is used to bind concentration values ​​at each mesh point, and the vector field is used to calculate the direction of concentration change and diffusion rate at each location based on the concentration gradient. The robot's real-time pose is calculated based on the attitude quaternion and placed in the virtual underwater environment to form a three-dimensional dynamic digital twin model of the bottom sediment environment.

[0011] Preferably, the steps further include: Based on the three-dimensional dynamic digital twin model, environmental deviations are identified. When the environmental deviations exceed a preset threshold, the optimal action parameters currently used to control the sediment remediation work are updated.

[0012] The specific aspects of identifying environmental deviations include: The direction of pollutant diffusion is determined by the vector field of the dynamic pollution field in the three-dimensional dynamic digital twin model, and the high-pollution grid area and the ordinary sediment grid area are divided according to the scalar field concentration value. Based on the regional division results, corresponding data acquisition frequencies are set for each grid region to increase the data acquisition frequency of high-pollution grids along the direction of pollution diffusion vector, and / or reduce the data acquisition frequency of ordinary sediment grid regions. The collected real-world perception data is compared with the simulation data of the corresponding grid in the three-dimensional dynamic digital twin model to obtain the regional deviation value, which is used as the environmental deviation.

[0013] Preferably, the initialization of the state space through the three-dimensional dynamic digital twin model includes the following steps: Train a deep learning model to identify pollutants, generate pollution categories, pollution levels and corresponding spatial locations, and construct a pollution removal list; Structured data is extracted based on the digital twin model and combined with the pollutant removal list to form the state space.

[0014] Preferably, the reward function used when training the reinforcement learning model is:

[0015] in, The amount of pollutants removed is used to characterize the actual contribution of robot actions to pollutant reduction; This is the target approximation term, used to characterize the approximation effect of the action command on the target contaminated area; This is the heading stability term, used to characterize the impact of motion commands on the robot's underwater attitude stability; This is a pollution priority item, used to guide consideration of prioritizing highly polluted areas; This is a regional coverage term, used to incentivize traversal of uncovered, heavily polluted areas; This is an energy consumption penalty term used to constrain the energy consumption of action commands; This is a penalty item for collision risk. , , , , , and The penalty coefficient corresponds to different reward items.

[0016] Preferably, the action space of the reinforcement learning model includes: the robot displacement heading vector, the robot arm end effector trajectory, and the sediment remediation drug injection parameters.

[0017] A deep learning-based adaptive sediment remediation robot includes: A multimodal perception module is used to collect multimodal data, including underwater terrain data, pollution concentration distribution data, and robot posture data. The embedded processing module is used to perform spatiotemporal registration and fusion based on the multimodal data to construct a three-dimensional dynamic digital twin model of the sediment environment; based on the three-dimensional dynamic digital twin model, a multi-dimensional state space is constructed, and optimal action parameters are generated through a trained reinforcement learning model; A dynamic actuator is used to perform corresponding actions based on the optimal action parameters.

[0018] Preferably, the multimodal sensing module includes a three-dimensional sonar array, a turbidity sensor array, an inertial measurement unit, and a preprocessing unit; The three-dimensional sonar array is used to acquire sonar point cloud data, and the sonar point cloud data is converted into a terrain network model through point cloud registration, and feature extraction is performed to obtain a terrain curvature feature matrix. The turbidity sensor array is used to acquire turbidity data at different locations; The inertial measurement unit is used to acquire the robot's attitude data; The preprocessing unit is used to convert the sonar point cloud data into a terrain network model through point cloud registration, and to extract features to obtain a terrain curvature feature matrix as the terrain feature. This is used to perform Kalman filtering on turbidity data from different locations, and through correlation fusion, to generate a pollutant concentration gradient tensor as the pollutant concentration distribution feature; The data is used to collect the robot's roll angle, pitch angle, and yaw angle. The collected angle data is then converted into attitude quaternions using an attitude conversion algorithm, which are used as the attitude parameters.

[0019] Preferably, the dynamic actuator includes a robotic arm, a vector propulsion system, and a repair drug injector; The vector propulsion system is mounted on the robotic arm, and the repair drug injector is fixed to the end of the robotic arm; The robot's motion trajectory is controlled by path planning and adjustment of thrust magnitude and direction. The pose of the robotic arm is controlled by joint angle adjustment, thereby achieving the working positioning of the drug repair syringe.

[0020] As can be seen from the above technical solutions, compared with the prior art, this invention discloses an adaptive sediment remediation robot system and its control method based on deep learning. It introduces a differentiated priority scheduling mechanism based on pollution concentration gradient and diffusion rate during the sediment remediation process. By prioritizing the treatment of highly polluted areas, control can be completed in the early stages of further pollution diffusion, thereby reducing the overall remediation range growth rate, improving pollution reduction efficiency per unit time, and reducing repetitive work and resource waste. This invention uses deep learning to identify and construct a multi-dimensional state space, combined with a reinforcement learning model, to achieve precise robot control, enabling more targeted decontamination operations. Furthermore, this invention constructs a digital twin model for underwater environmental deviation identification and deep learning model trial and error, enabling rapid acquisition of underwater environmental changes and dynamic strategy adjustments, thus better adapting to highly time-varying underwater operating environments. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of an adaptive sediment remediation robot control method based on deep learning provided in an embodiment of the present invention; Figure 2 is a schematic diagram illustrating the performance verification effect of the repair operation compared with the traditional control method in the embodiment of the present invention; Figure 3 This is a schematic diagram of an adaptive sediment remediation robot based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1 like Figure 1 An adaptive sediment remediation robot control method based on deep learning includes the following steps: S1: Acquire multimodal perception data, which includes underwater terrain data, pollution field distribution data, and robot posture data; S2: Construct a three-dimensional dynamic digital twin model of the sediment environment based on the multimodal sensing data; S3: Construct a reinforcement learning model, initialize the state space through the three-dimensional dynamic digital twin model, and train the model; use the trained reinforcement learning model to output the optimal action parameters; S4: Perform the sediment remediation work according to the optimal action parameters.

[0025] The following is a detailed explanation of each step in this embodiment: S1 is the data acquisition step. For underwater environments with high turbidity, no light, and complex topography, a collaborative sensing approach using 3D sonar, a turbidity sensor array, and an IMU is employed. The 3D sonar collects underwater topographic point clouds and quantifies underwater topographic features to obtain topographic data. The turbidity sensor array collects turbidity information at different locations around the robot to reflect the distribution and changes of pollutants in the interface between the water and sediment. Pollution distribution data is collected in real time using an IMU to capture robot posture data.

[0026] Specifically, after each data point is collected, it undergoes front-end preprocessing. For the terrain point cloud acquired by sonar, coordinate unification, noise suppression, and feature extraction are performed to form terrain data that can characterize the changes in the bottom sediment surface. For the data collected by the turbidity sensor, filtering and smoothing are performed by combining data from adjacent time points and adjacent locations to reduce the impact of instantaneous fluctuations on the judgment of polluted areas and to form concentration feature data that characterizes the changes in pollutant distribution. For the attitude information output by the inertial measurement unit, attitude calculation and data correction are performed to obtain robot attitude data that can reflect the current motion state of the robot.

[0027] S2 is the construction step of the digital twin model. The present invention dynamically displays the real underwater environment and the real working situation of the robot through the constructed twin model, and monitors the time-varying environment based on the real-time update of the digital twin model, so as to update the control strategy in a timely manner.

[0028] Specifically, the steps for constructing a three-dimensional dynamic digital twin model of the sediment environment include: S21: Preprocess the multimodal sensing data and perform spatiotemporal registration. In this embodiment, an embedded unit is used to receive the collected data in S1, and the collected multi-source data is timestamped and then spatiotemporally registered. Data with different sampling frequencies is unified through data alignment or resampling, and the independent coordinate systems of each data are globally normalized.

[0029] For example, using the global coordinate system of the actual operating water area as the reference coordinate system, the discrete terrain points are mapped to the global coordinates in the reference coordinate system using the terrain curvature matrix to generate a basic spatial network; the pollutant gradient tensor is bound to the same spatial grid grid by grid; the local coordinate system of the robot body is converted to the global reference coordinate system to achieve alignment of the robot's virtual and real poses.

[0030] S22: Construct a three-dimensional mesh model of the bottom sediment based on the topographic curvature matrix; generate a regular three-dimensional network, such as a voxel network or a triangular network, from the discrete points of the topographic curvature matrix according to spatial resolution; calculate the elevation, slope, and concavity / convexity of each grid point using curvature numerical values ​​to reconstruct the underwater topographic undulations; identify curvature abrupt changes based on thresholds, and mark obstacles such as obstacles, deep pits, and protruding areas in the grid based on the curvature abrupt change areas to form a static three-dimensional mesh model of the bottom sediment.

[0031] S23: A dynamic pollution field is superimposed on the 3D mesh model based on the pollutant concentration gradient tensor to generate a virtual underwater environment. The dynamic pollution field includes a scalar field and a vector field. The scalar field is used to bind concentration values ​​at each mesh point, and the vector field is used to calculate the direction of concentration change and diffusion rate at each location based on the concentration gradient. Isosurfaces can be generated from the superimposed scalar field, and the visualization effect can be improved by setting different colors or transparency.

[0032] S24: The robot's real-time pose is calculated using attitude quaternions and placed in the virtual underwater environment to form a three-dimensional dynamic digital twin model of the sediment environment. Based on the robot's original attitude data, the robot's rotation state is calculated in real time using attitude quaternions, sequentially decomposing the device's pitch angle, roll angle, and yaw angle to completely restore the robot's real-time rotation attitude. The calculated quaternion attitude parameters are converted into corresponding spatial attitude information to determine the robot's orientation and tilt state. The robot's real-time spatial positioning data is read simultaneously to obtain the device's basic coordinate position in the scene. The robot's local attitude data is fused and matched with global positioning data to complete the conversion from the robot's local coordinate system to the scene's global coordinate system. Step-by-step calculation and correction are performed to finally obtain the robot's accurate and complete real-time position and attitude pose data in the global space.

[0033] S25: Based on the three-dimensional dynamic digital twin model, identify environmental deviations. When the environmental deviations exceed a preset threshold, update the optimal action parameters currently used to control the sediment remediation work.

[0034] Specifically, identifying environmental deviations includes: The direction of pollutant diffusion is determined by the vector field of the dynamic pollution field in the three-dimensional dynamic digital twin model, and the high-pollution grid area and the ordinary sediment grid area are divided according to the scalar field concentration value. Based on the regional division results, corresponding data acquisition frequencies are set for each grid region to increase the data acquisition frequency of high-pollution grids along the direction of pollution diffusion vector, and / or reduce the data acquisition frequency of ordinary sediment grid regions. The collected real-world perception data is compared with the simulation data of the corresponding grid in the three-dimensional dynamic digital twin model to obtain the regional deviation value, which is used as the environmental deviation.

[0035] This implementation relies on a three-dimensional dynamic digital twin model to differentiate scalar field concentration and vector field diffusion direction to achieve grid-based differentiated sampling control. On the one hand, it increases the sensing and acquisition frequency for high-pollution grids arranged along the pollution diffusion vector, enabling high-density capture of dynamic disturbance details such as pollutant migration, concentration abrupt changes, and intensified diffusion, avoiding the omission of high-risk pollution evolution characteristics due to sparse sampling. On the other hand, it reduces the acquisition frequency for low-risk ordinary sediment grids, significantly reducing the massive amount of redundant data generated by meaningless sampling of multimodal sensor arrays, and alleviating the computational load of real-time data fusion and virtual-real comparison of embedded processing modules. Ultimately, only the difference between virtual and real data of high-pollution grids obtained by encrypted sampling is used as the core environmental deviation for judgment. This not only accurately locks the key environmental disturbances caused by pollution diffusion and ensures priority updates and remediation control parameters for high-pollution areas, but also filters out invalid deviation interference caused by small topographic and concentration fluctuations in ordinary sediment areas, avoiding operational oscillations caused by frequent triggering of action parameter updates by robots, and simultaneously taking into account the accuracy of dynamic monitoring of sediment pollution and the real-time computing efficiency of embedded devices.

[0036] In another implementation, the pollutant field grid data stored in the digital twin model at the previous moment and the current moment are called up, and the pollutant concentration, diffusion direction, and diffusion coverage of all grid points are compared point by point. The concentration difference, diffusion angle offset, and coverage area change of the same grid at different moments are calculated to quantify the real-time change deviation of the pollutant field. The deviation values ​​of all grids are statistically summarized to obtain the overall environmental deviation value. The calculated real-time environmental deviation is compared with a fixed threshold preset by the system. If the deviation value is less than or equal to the threshold, it is determined that the current model environmental state is stable and the existing remediation action parameters remain unchanged. If the deviation value is greater than the preset threshold, it is determined that the current pollutant diffusion state has changed significantly and the original remediation parameters are no longer suitable for the current grid environmental state. Based on the pollutant distribution and diffusion changes after the current grid is updated, the optimal action parameters of the robot remediation operation bound to the model are directly and forcibly updated to make the remediation operation parameters adapt to the latest pollutant field grid state.

[0037] This implementation method compares the global pollution grid data point by point before and after the twin model, which can fully capture global environmental changes such as the diffusion and migration of pollutants in the sediment and the emergence of sudden pollution blobs. It uses a fixed threshold to uniformly judge the overall fluctuation of the pollution field. Once the overall pollution diffusion state changes significantly, the robot's remediation action parameters are forcibly updated synchronously. This allows it to adapt to the overall evolution trend of the pollution field in the first instance. It avoids the problem of mismatch between the original operation path, reagent dosage and other parameters and the latest pollution distribution from the pollution source level, and ensures that the sediment remediation operation in the whole area always matches the real-time diffusion and evolution law of pollutants.

[0038] In another implementation, motion parameters can be updated in real time based on robot execution deviations. Pre-set standard robot working pose, standard travel trajectory, and standard working point data are retrieved from the digital twin model. The actual robot pose, actual travel trajectory, and actual stationary grid points in real-time are then matched and read from the model. The standard execution data and actual execution data are compared point-by-point to calculate the robot's heading offset angle, trajectory offset distance, and working point offset error. Single-frame offset data are summarized to obtain the robot's real-time motion execution deviation value, and continuous offset data across multiple frames are statistically analyzed to determine the persistent offset state. The calculated motion execution deviation is compared with a system-preset motion error threshold. If the deviation is within the threshold range, the robot's motion execution is considered accurate, the work process is without deviation, and the current optimal motion parameters are maintained. If the deviation exceeds the preset threshold, the robot's actual work action is considered to deviate from the standard work instruction, indicating an execution error. Based on the trajectory offset direction, offset distance, and attitude deviation magnitude recorded by the model, the robot's travel trajectory, working positioning points, and attitude calibration parameters are specifically corrected. The optimal repair motion parameters bound to the model are updated in real time to correct the robot's work execution deviation and ensure that underwater sediment remediation work is executed accurately according to the standard process.

[0039] This implementation focuses on the precision of robot body operation. It relies on the standard operation benchmark data stored in digital twins to perform point-by-point matching and comparison with real-time operation data, quantifying multi-dimensional execution offset errors in heading, trajectory, and operation point. It also combines continuous multi-frame data to judge persistent offset faults, which is different from the logic of adjusting parameters solely based on changes in environmental pollution. When robot hardware execution causes drift or underwater turbulence to cause the trajectory to deviate from the standard instructions, it can specifically correct the relevant motion parameters of propulsion, robotic arm positioning, and attitude calibration, correcting the operation deviation at the execution level in real time. This ensures that the repair point is accurately aligned with the target pollution grid, eliminating repair omissions and reagent waste caused by positioning offset from the equipment execution end, and stabilizing the standardized repair operation process.

[0040] In another implementation, a reinforcement learning reward value deviation verification is performed based on a digital twin model. The optimal action parameters for the repair operation are updated based on the difference between the actual reward and the theoretical reward. Relying on the digital twin 3D mesh model and the reinforcement learning operation strategy architecture, the definition standards for two types of reward values ​​are clarified. The theoretical reward is the standard reward value calculated under offline simulation and preset environmental conditions by the reinforcement learning model. It is the expected gain value of the model based on an ideal pollution diffusion state and error-free robot action output. The actual reward is the real-time environmental state feedback reward from the digital twin, which is the real gain value calculated in real time based on the current actual pollution distribution, real-time diffusion changes, and the robot's actual operation actions. Within each round of action iteration in the model, the single-step theoretical reward value output by reinforcement learning is retrieved. Simultaneously, based on the real-time updated pollution field mesh data and robot operation completion status of the twin model, the actual reward value corresponding to the current round of actions is calculated synchronously. Environmental rewards; the difference between the theoretical reward and the actual reward for the same operation iteration cycle is calculated to quantify the reward deviation; the reward deviation is compared with the system's preset reward threshold; if the actual reward value is close to the theoretical reward and the deviation is within the threshold range, it is determined that the current reinforcement learning strategy parameters are adapted to the underwater environment and the operation benefit is normal, and the original model parameters remain unchanged; if the actual reward value is significantly lower than the theoretical reward and the deviation exceeds the preset threshold, it indicates that the current reinforcement learning strategy cannot adapt to real-time pollution diffusion changes, and the strategy action benefit is invalid; based on the magnitude of the reward deviation, the reinforcement learning strategy network parameters and robot repair operation action parameters are updated in reverse iteration, the strategy output logic is corrected, and the model action benefit is made to fit the real underwater bottom sediment environment conditions.

[0041] This implementation constructs parameter update criteria from the perspective of reinforcement learning strategy returns, distinguishing between the theoretical rewards of ideal offline simulations and the actual rewards of twin real-time environmental feedback. The difference between the two quantifies the degree of adaptation between the current strategy and the real underwater working conditions, moving away from the traditional approach of relying solely on surface data of the environment and equipment hardware to determine the timing of updates. When the actual remediation returns are far lower than theoretical expectations, the strategy network and robot operation parameters are iteratively optimized in reverse, and the reinforcement learning output action logic is actively corrected. This addresses the core issues of strategy failure under dynamic pollution changes and the failure of priority treatment in highly polluted areas from the bottom layer of algorithm decision-making, continuously optimizing the model decision logic to adapt to the complex dynamic underwater sediment environment.

[0042] S3 is the step for constructing and training the reinforcement learning model and outputting the optimal action. Based on the constructed real-time synchronously updated three-dimensional digital twin environment of the sediment, it completes the initialization of the state space and iterative training of the reinforcement learning model. It uses the policy model after training convergence to output the optimal action parameters for sediment restoration that are adapted to the underwater dynamic environment in real time.

[0043] In this step, considering the characteristic that underwater sediment remediation is a continuous control operation, the Deep Deterministic Policy Gradient (DDPG) reinforcement learning algorithm is selected to construct a dual-network training architecture. Closed-loop iterative training is completed based on a digital twin simulation environment to form a standardized state-action decision-making capability.

[0044] First, the basic model structure of DDPG is constructed. This invention adopts an Actor-Critic dual-network architecture. The Actor network uses a 4-layer MLP, and the Critic network also uses a 4-layer MLP. The activation function for both is ReLU, and the output layer uses tanh (Actor) and linear activation (Critic). The Actor network first outputs initial action commands, which are then pre-simulated and collision detected in the digital twin model. Process deviation data is recorded simultaneously and fed back to the Critic network. The Critic network evaluates the action adaptability based on multi-dimensional data, generating a quantitative feedback signal containing optimization directions. The Actor network iteratively adjusts the action parameters based on the feedback signal. After multiple iterations of optimization, the output action commands meet the requirements of collision-free, redundancy-free, and sediment remediation operations, ultimately forming a multi-actuator collaborative control command.

[0045] Next, the basic space definition is completed. The state space is based on the environmental data updated in real time by the digital twin model. It defines the basic environmental state, including the undulation features of the bottom mud terrain, the distribution features of obstacles, the real-time posture and pose of the robot, and the mask of the repaired area.

[0046] Specifically, for the undulating features of the bottom mud terrain, the mean and variance of the 8×8 terrain curvature grid of the twin model are extracted as the terrain undulation quantification features; for the obstacle distribution features, the number of obstacle marker points in the current grid and the shortest distance between the obstacle and the robot center are counted as the obstacle distribution status; for the robot's real-time 3D pose, the pose quaternions and global coordinates calculated by the twin model are directly read to form the robot pose state variables; for the mask features of the repaired area, a binary mask matrix of the same size as the working grid can be generated, with the repaired grid assigned a value of 1 and the unrepaired grid assigned a value of 0, to avoid repetitive work.

[0047] Furthermore, in order to achieve more precise targeted remediation, this invention uses a lightweight deep learning model to identify pollutants and classify them, as well as to determine the degree of pollution. The model output is then used to augment the dimensions of the state space.

[0048] Specifically, an improved MobileNetV3 lightweight detection model can be used. The input of this model is a global environmental feature visualization map generated by mapping the digital twin 3D mesh model at the current moment, including the sediment topographic mesh feature map rendered by the twin model, the dynamic pollution field scalar concentration heat map, and the pollutant vector diffusion direction features. Figure 3The model, which is based on the twin-derived feature image, adds a coordinate attention enhancement module and a DAFA-DBN domain adaptive branch on the basis of MobileNetV3-Large. Relying on multi-scale feature extraction, location information enhancement and multi-task reasoning, it generates three accurate output results in parallel: pollutant category, pollutant two-dimensional pixel coordinates and pollution levels 1 to 5.

[0049] The main body of the model uses MobileNetV3-Large as the backbone network and utilizes the built-in depthwise separable convolutional structure to compress the number of model parameters, adapting to the real-time inference requirements of embedded devices. At the same time, a coordinate attention module is embedded in the feature pyramid Neck layer to encode the horizontal and vertical orientation information of the feature map, making up for the defects of weak features of small pollutant targets and easy loss of position information in complex underwater environments. The detection output end is based on the SSD single-step detection branch to build a multi-task prediction head. In addition to the original target classification and bounding box regression branches, an additional pollution level regression branch is added to complete the end-to-end inference output and realize accurate pollutant perception based on the structured environment of digital twin.

[0050] Considering the significant environmental distribution shift between the laboratory clean water source domain and the real complex underwater target domain, this paper embeds a DAFA domain adaptive feature alignment sub-network and a DBN domain batch normalization layer into the feature output segment of the aforementioned basic detection network to improve the model's cross-domain generalization ability. The two together constitute the model's cross-domain adaptive module: the DAFA sub-network weakens the feature bias caused by scene environment differences through feature distribution alignment and domain adversarial learning, realizing the transfer of source domain labeled knowledge to unlabeled real underwater scenes; the DBN layer corrects the feature statistical shift of different water environments through an adaptive normalization strategy, further improving the model's domain adaptability.

[0051] The model input uniformly uses underwater pollutant feature maps derived from digital twins as training samples. The dataset labels include three types of information: pollutant category, target bounding box coordinates, and pollution level. The model is trained under supervision using a multi-task joint loss function. The total loss is composed of a weighted fusion of classification loss, bounding box regression loss, pollution level regression loss, and domain adversarial loss, which can simultaneously drive the model to complete pollutant category discrimination, precise target localization, and quantitative assessment of pollution level.

[0052] The model training employs a two-stage strategy: the first stage involves supervised training on 5000 labeled images collected in the source domain, i.e., the laboratory's clean water environment. The labels include three categories of annotation information: pollutant category (heavy metal deposits, organic pollutant blobs, and microplastic particles), bounding box coordinates (x, y, w, h), and pollution level (1-5). The total loss function is:

[0053] in, For classification loss, FocalLoss (α=0.25, γ=2.0) is used to alleviate the problems of class imbalance and low proportion of small target samples in underwater datasets. For the bounding box regression loss, CIoULoss is adopted to fully consider the target distance, aspect ratio and overlap deviation, thereby improving the positioning accuracy of deformable pollutants; To account for pollution level loss, SmoothL1Loss is used for regression constraints to improve the stability of level prediction. For domain adversarial loss, binary cross-entropy loss is adopted to reduce the difference in environmental distribution between the source domain and the target domain; , , and The weights for the corresponding loss terms are set to 1.0, 5.0, 0.5, and 0.1, respectively. The optimizer used is AdamW with an initial learning rate of 0.001, which decays to 0.00001 with cosine annealing. The training batch size is 16, and the training lasts for 300 epochs.

[0054] In the second stage, domain adaptive fine-tuning is performed on the target domain (2000 unlabeled images collected in the real underwater environment). Only the parameters of the DAFA adaptive sub-network and the DBN domain normalization layer are updated, while the parameters of the backbone network and the detection head are frozen. Fine-tuning is performed for 100 epochs until the domain classifier loss converges.

[0055] To adapt to real-time inference on embedded platforms, multi-dimensional lightweight optimizations were implemented on the trained model. L1 norm channel pruning was used to remove redundant channels to reduce the number of parameters. The native SiLU activation function of the network was uniformly replaced with the more computationally efficient Leaky-ReLU. 8-bit integer quantization was performed based on the TensorRT framework to further compress the model size and improve the forward inference speed. The optimized model was deployed on an embedded Jetson platform, enabling end-to-end inference on real-time acquired images and synchronously outputting pollutant categories, detection box coordinates, and pollution levels, providing reliable input for coordinate registration and path planning.

[0056] After detection, the two-dimensional pixel coordinates of the detected pollutants are converted into planar coordinates in the robot's local coordinate system using a pre-calibrated hand-eye calibration matrix and perspective transformation matrix. These coordinates are then mapped to the global coordinate system using twin terrain mesh elevation data to obtain the three-dimensional spatial coordinates of the pollutant target. Simultaneously, the extreme values ​​of the pollutant concentration gradient tensor at the current mesh location are read and matched with the pollution level Li (levels 1-5) output by deep learning to calculate the corresponding priority weights. Finally, the number of effective pollutant targets detected in a single frame, the three-dimensional coordinate offset of the pollutant target, the normalized value of the pollution level, and the priority weights are concatenated into a fixed-dimensional pollution feature vector. This pollution feature vector is then fused with the aforementioned basic environmental feature vector to generate a complete state vector containing multi-dimensional features of terrain, obstacles, robot pose, repair mask, and pollutants. This state vector serves as the standard input for each iteration of the reinforcement learning model, enabling the model to observe the location of underwater pollutant targets, the severity of pollution, and operational priorities in real time, thus achieving targeted removal decisions for bottom sediment pollutants.

[0057] The definition of the motion space specifically includes the thrust vector motion of the thruster, the fine-tuning motion of the robotic arm end effector, and the parameters for the removal of contaminants.

[0058] Among them, the thrust vector motion parameter of the thruster is a two-dimensional continuous motion quantity, which corresponds to the horizontal forward and backward and left and right propulsion thrust of the robot, and the thrust value is constrained within a specific range to control the horizontal movement and heading deflection of the robot, so as to achieve precise approach to the polluted area. The fine-tuning motion parameters at the end of the robotic arm are four-dimensional continuous motion quantities, including the three-dimensional coordinate offset of the end of the robotic arm and the binarized injection trigger signal. The coordinate offset is quantitatively constrained and used to fine-tune the robotic arm's fit to the bottom mud contamination operation point. Pollutant removal is a one-dimensional continuous removal intensity action quantity, with the numerical normalization constraint in the range of [0,1]. It corresponds to the sediment pollutant removal rate, agent spraying dosage, and disturbance removal amplitude. The larger the value, the stronger the pollutant removal operation at the current point, which directly corresponds to the core task of sediment pollution removal and realizes differentiated cleaning operations in areas with different pollution concentrations.

[0059] Reward function:

[0060] in, This refers to the amount of pollutants removed. For the target approximation term, For the heading stability term, As a pollution priority item, For regional coverage items, As an energy consumption penalty item, This is a penalty item for collision risk. , , , , , and The penalty coefficient corresponds to different reward items.

[0061] The reward function is calculated based on the action instructions output by the reinforcement learning model and the feedback data from the digital twin environment, where: The amount of pollutant removed is calculated from the difference results of the pollution concentration field in the digital twin model within a continuous time step. Specifically, it is the difference between the average pollution concentration of the grid at the current moment before the action is performed and the average pollution concentration of the grid at the next moment after the action is performed. It is used to characterize the actual contribution of the robot's action to the reduction of pollutants, and the value is normalized to the interval [0,1].

[0062] The target approximation term first calculates the Euclidean distance *d* between the spatial coordinates of the robot's end effector and the coordinates of the target contamination point after the robot performs its action. Then, a mapping formula is used to convert this distance into a dimensionless positive reward; the mapping formula is:

[0063] In the formula, This is the distance attenuation coefficient, used to reflect the approximation effect of the motion command on the target contaminated area. This item is set to 0 when the robotic arm exceeds its maximum extension stroke.

[0064] The heading stability term is obtained by converting the attitude quaternions collected by the robot's inertial measurement unit into Euler angles, and then taking the absolute deviation value of the heading angle and performing a negative exponential mapping. The smaller the deviation, the higher the reward. This term is used to characterize the impact of action commands on the robot's underwater attitude stability.

[0065] The pollution priority item combines the pollutant concentration amplitude of the action area grid and the pollution diffusion rate output of the dynamic pollution field vector field to perform a joint index normalization mapping to obtain the basic priority weight, and superimposes the unvisited area markers of the digital twin model for weighted calculation; the faster the pollution diffusion and the higher the concentration, the greater the reward for this item, which specifically guides the reinforcement learning strategy to prioritize the remediation of areas with high pollution diffusion rates.

[0066] The regional coverage reward is based on the statistics of the 3D mesh mask of the polluted area updated after the action is executed. When any mesh changes from an unrepaired state to a qualified repair state, a fixed value is accumulated and all are dimensionless. This is used to incentivize the traversal of the polluted meshes in the uncovered heavily polluted area.

[0067] The energy consumption penalty term is obtained by summing the square integrals of the thrust output of the actuator and the torque of each joint of the robotic arm within a single time step. A sliding window is used to limit the maximum cumulative penalty, which is used to constrain the energy consumption of a single action command.

[0068] The collision risk penalty is calculated to determine the minimum safe distance between the robot body and underwater obstacles or protruding terrain in the mud. Two levels of safety thresholds are set: a mild penalty is applied when the distance is below the first-level warning threshold, and a severe penalty is applied when the distance is below the second-level danger threshold. This is used to layer constraints on the underwater operation safety of action commands.

[0069] In this embodiment, during training, the robot interacts in real time within a digital twin environment to obtain its current state, executes continuous actions, and calculates a seven-component composite reward function to obtain a single-step immediate reward, forming time-series training samples and storing them in the experience replay pool. During the training phase, samples are randomly sampled in batches from the replay pool. First, the Critic network is input with the current state and corresponding actions to fit the current Q value. The target Q value is then calculated by combining the discount factor and the target network output. The value loss is calculated using the mean squared error loss function, and the gradient is backpropagated to complete the iterative update of the Critic network parameters, enabling the network to accurately fit the long-term reward value of actions under different pollution scenarios. Subsequently, the Critic network... Maximizing the output Q-value is used as the optimization objective. An Actor policy network loss function is constructed, and policy gradient backpropagation is performed. The Actor network parameters are iteratively updated to continuously optimize the continuous action output strategy of robot heading propulsion, robotic arm fine-tuning, and drug removal intensity. After each iteration, the target Actor and target Critic network parameters are slowly synchronized through soft updates to ensure training stability. Through multiple rounds of iterative iteration, the network loss is continuously minimized and the gradient weights are optimized until the model converges. Ultimately, the reinforcement learning model has the decision-making ability to adaptively output the optimal repair action based on terrain obstacles, robot pose, pollution level, and diffusion characteristics.

[0070] Specifically, this algorithm includes two independent loss functions for updating the parameters of the two networks. The Critic value network uses mean squared error loss, calculated as follows:

[0071] in, The current Critic network predicts action value. The target Critic network predicts value. For single-step compound rewards, The return discount factor can be fixed at 0.99.

[0072] The Actor policy network aims to maximize the cumulative Q-value, employing policy gradient loss, calculated as follows:

[0073] The Actor policy network optimizes continuous action policies by minimizing the negative expected Q-value.

[0074] Parameter updates are divided into online network gradient updates and target network soft updates. Both the online Actor and Critic networks use the Adam optimizer to perform gradient backpropagation iterations. The learning rate of the Actor network is set to... The learning rate of the Critic network is set to The target network employs a soft update mechanism to achieve smooth iteration, with the update formula being:

[0075] In the formula, the soft update coefficient , for online network parameters, These are the target network parameters.

[0076] To further address the issues of asynchronous actions and significant susceptibility to external disturbances in collaborative operations involving multiple actuators, this invention introduces collaborative control. It analyzes the action parameters output by the reinforcement learning model, combines the deviation prediction of the digital twin model to perform deviation correction, and generates the final control command based on the corrected action parameters.

[0077] This step addresses the issues of asynchronous actions and susceptibility to external disturbances that can easily occur when multiple actuators work together. It uses a unified embedded processing module to complete instruction parsing and signal distribution, and employs a dual closed-loop control and cross-coupling compensation strategy to suppress interference caused by sudden changes in the sediment environment. At the same time, it integrates real-time sensing data throughout the process to form a closed-loop feedback, ensuring that actions such as robot movement, robotic arm positioning, and drug injection are coordinated, consistent, accurate, and reliable.

[0078] For example, multi-agency collaboration includes: After receiving the thrust vector command, the vector thruster maps the desired resultant force and resultant torque to the independent thrust value of each thruster through a thrust distribution algorithm, dynamically adjusting the output of each thruster. Simultaneously, it reads the attitude data fed back by the inertial measurement unit in real time, calculating the current roll and pitch angles. When the attitude deviation exceeds the preset stability range, the attitude controller calculates an additional corrective thrust based on the angle deviation, superimposing it on the original thrust vector command to form a closed-loop attitude stabilization, ensuring the robot moves smoothly along the planned path.

[0079] After receiving the end-effector coordinate command, the six-DOF robotic arm calculates the target angles of each joint through inverse kinematics, and then drives the joints to reach the target angles through closed-loop position control of the servo motors. At the same time, it reads the terrain curvature feature matrix generated in step S1 to determine the terrain slope at the current location of the robotic arm base. If the terrain slope is large, a gravity compensation term is automatically added in the inverse kinematics solution, and the approach posture of the end effector is finely adjusted to ensure that the axis of the end effector is always perpendicular to the bottom mud surface, thereby accurately aligning with the core pollution area.

[0080] Independent injection control: The injection rate, total flow rate, and needle extension length are determined according to the type and concentration level of contamination. Needle positioning and drug delivery are executed independently through position closed loop and flow closed loop.

[0081] S4 is a dynamic execution step, which moves the robot to the contaminated area, injects the agent through the precision repair module, and feeds back the real-time operation data to the digital twin model of the sediment environment.

[0082] This step employs a closed-loop working mode of simultaneous execution and learning. Addressing the dynamic changes in pollutant distribution and sediment topography during operations, it continuously updates the digital twin model in real time by transmitting sensor data. Furthermore, it utilizes the online fine-tuning capabilities of reinforcement learning algorithms to dynamically correct path planning and injection parameters, constructing an adaptive closed loop of perception-decision-execution-re-perception. This effectively addresses the uncertainties of the unstructured underwater environment. The expected outcome is stable and efficient pollutant removal, significantly reducing reagent consumption while ensuring remediation effectiveness. Leveraging the real-time synchronization capabilities of the digital twin model, it enables rapid perception and early warning of environmental conditions and deviations. Compared to traditional remediation methods, this significantly improves operational accuracy, reduces overall costs, and makes the entire remediation process more stable, economical, and intelligent.

[0083] In this embodiment, the digital twin model can provide multifaceted visualization effects during the control process. For example, in the digital twin model, the point cloud data collected in real time by sonar is converted into triangular mesh terrain using a Poisson surface reconstruction algorithm and continuously updated at a frequency of 50Hz, thus showing the changes in terrain undulations. Then, based on the previously obtained pollutant concentration gradient tensor, a heatmap texture is applied to the surface of the terrain mesh, with higher concentrations resulting in darker colors (gradient from red to purple). A semi-transparent isosurface is also overlaid to visually demonstrate the trend of pollution diffusion. The robot's own pose synchronization relies on receiving attitude quaternions from the IMU and position data from the thruster encoders to drive the robot model, which includes a six-degree-of-freedom robotic arm, to move in real time within the Unity scene, maintaining a refresh rate of 50Hz. This data is pushed from the Jetson platform to the Unity engine in real time via the ROS-TCP-Connector protocol. Each time new data is received, only incremental updates are performed, without rebuilding the entire scene, making the computational burden relatively controllable.

[0084] During collaborative control, once a lateral offset exceeding 0.1 meters or a terrain deviation is detected, the model highlights the deviation trajectory between the robot's actual position and the planned path, and marks the suggested correction direction with a red dashed line. Operators or the automated decision-making module can quickly confirm the correction plan. After correction, closed-loop verification is performed. The model simultaneously displays a comparison overlay of the planned path and the actual trajectory. If the average tracking error exceeds 0.05 meters for 10 consecutive cycles, it automatically marks the area with excessive deviation and issues a warning. In addition, when the system dynamically adjusts the priority of pollution targets based on measured concentrations, the model also updates the color depth and priority number of the polluted areas in real time, making high-pollution areas readily apparent and facilitating the checking of whether the priority ranking is reasonable.

[0085] To further illustrate the technical effects of the present invention, after completing the repair work in the preset area, the present invention conducts performance evaluation and application verification to verify the repair effect and system performance, and confirm that the actual problem has been solved.

[0086] This invention conducts performance evaluation through a complete process: scenario setup → environmental perception → modeling and planning → collaborative operation → data monitoring → effect detection → comparative analysis, as detailed below: Typical scenario validation: Two standard pollution test scenarios were established for oil spills and heavy metals, with unified sediment conditions and initial pollution concentrations. In the emergency oil spill scenario, the system operates according to a closed-loop process of environmental modeling, path generation, and reagent injection. It completes 3D modeling of the sediment environment within 5 minutes, generates the optimal remediation path in 30 seconds, and completes precise reagent injection in 15 seconds, ultimately achieving an oil spill removal rate of 94.3%. In the heavy metal pollution remediation scenario, the system adaptively matches injection parameters according to pollutant category, stably adding reagents at a rate of 30 μL / s, reducing lead ion concentration by 73%, and demonstrating significantly better remediation results than traditional processes.

[0087] Performance Comparison: In a complex sediment environment with sludge mobility of 35% and a pollutant gradient >50 NTU / m, the system was compared with traditional fixed injection devices and open-loop controlled underwater robots using a controlled variable method with the same site, pollution concentration, reagent dosage, and operation time. Traditional devices rely on static preset paths and manual positioning, lacking real-time environmental perception and dynamic planning capabilities. Their robotic arms use open-loop control with low positioning accuracy, and the injection process lacks closed-loop flow control, making them susceptible to sediment disturbances. These traditional devices suffer from drawbacks such as blind operation, numerous ineffective actions, significant reagent waste, and poor environmental adaptability. The results show that the system's remediation efficiency increased from 12.7 m³ / h to 18.3 m³ / h (a 44% increase), the planar positioning error decreased from 5.2 mm to 1.8 mm, the reagent flow rate standard deviation decreased from 4.5 μL to 0.8 μL, and the pollutant removal rate increased from 68.7% to 92.5%, demonstrating comprehensive improvements in operational efficiency, control accuracy, and remediation effect.

[0088] Industrial Application Validation: In a real-world lake sediment remediation project, long-term validation was conducted following the process of site survey → work zoning → system deployment → continuous operation → periodic sampling inspection → cost accounting. The system operated continuously and stably for 30 days, with an average daily processing capacity of 439 m³, a pollutant removal rate consistently >92%, and maintenance costs reduced by 41% compared to traditional processes. This fully validated the system's practicality and reliability in real-world engineering scenarios, fundamentally addressing the core pain points of traditional remediation devices: low precision, poor efficiency, and insufficient adaptability to complex environments. The above performance validation process, test conditions, and quantitative results are as follows: Figure 2 As shown.

[0089] Example 2 like Figure 3 Based on the same inventive concept, this invention discloses an adaptive sediment remediation robot based on deep learning, comprising: The multimodal perception module is used to collect multimodal data, including underwater terrain data, pollution concentration distribution data, and robot posture data. The embedded processing module is used to perform spatiotemporal registration and fusion based on multimodal data to construct a three-dimensional dynamic digital twin model of the sediment environment; a multi-dimensional state space is constructed based on the three-dimensional dynamic digital twin model, and optimal action parameters are generated through a trained reinforcement learning model. A dynamic actuator is used to perform corresponding actions based on optimal motion parameters.

[0090] In one embodiment, the multimodal sensing module includes a three-dimensional sonar array, a turbidity sensor array, an inertial measurement unit, and a preprocessing unit. A three-dimensional sonar array is used to acquire sonar point cloud data, and the sonar point cloud data is transformed into a terrain network model through point cloud registration, and feature extraction is performed to obtain the terrain curvature feature matrix. Turbidity sensor arrays are used to acquire turbidity data at different locations; The inertial measurement unit is used to acquire the robot's attitude data; The preprocessing unit is used to convert sonar point cloud data into a terrain network model through point cloud registration, and to extract features to obtain a terrain curvature feature matrix as terrain features. This is used to perform Kalman filtering on turbidity data from different locations, and through correlation fusion, to generate a pollutant concentration gradient tensor as a feature of pollutant concentration distribution. This data is used to collect the robot's roll, pitch, and yaw angles. The collected angle data is then converted into attitude quaternions using an attitude conversion algorithm, which serve as attitude parameters.

[0091] Furthermore, the dynamic actuators include a robotic arm, a vector propulsion system, and a repair drug injector; The vector propulsion system is mounted on the robotic arm, and the repair drug injector is fixed to the end of the robotic arm; The robot's motion trajectory is controlled by path planning and adjustment of thrust magnitude and direction. The pose of the robotic arm is controlled by joint angle adjustment, thereby achieving the working positioning of the drug repair syringe.

[0092] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0093] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A deep learning-based adaptive sediment remediation robot control method, characterized in that, Includes the following steps: Acquire multimodal perception data, including underwater topography data, pollution field distribution data, and robot posture data; A three-dimensional dynamic digital twin model of the sediment environment is constructed based on the multimodal sensing data; A reinforcement learning model is constructed, and the state space is initialized through the three-dimensional dynamic digital twin model for model training; the trained reinforcement learning model is used to output the optimal action parameters. Perform sediment remediation work according to the optimal action parameters.

2. The adaptive sediment remediation robot control method based on deep learning according to claim 1, characterized in that, The steps for constructing a three-dimensional dynamic digital twin model of the sediment environment include: Preprocessing and spatiotemporal registration are performed based on the multimodal sensing data; A three-dimensional mesh model of the sediment was constructed based on the topographic curvature matrix. A virtual underwater environment is generated by superimposing a dynamic pollution field onto the three-dimensional mesh model based on the pollutant concentration gradient tensor. The dynamic pollution field includes a scalar field and a vector field. The scalar field is used to bind concentration values ​​at each mesh point, and the vector field is used to calculate the direction of concentration change and diffusion rate at each location based on the concentration gradient. The robot's real-time pose is calculated based on the attitude quaternion and placed in the virtual underwater environment to form a three-dimensional dynamic digital twin model of the bottom sediment environment.

3. The adaptive sediment remediation robot control method based on deep learning according to claim 1, characterized in that, The steps also include: Based on the three-dimensional dynamic digital twin model, environmental deviations are identified. When the environmental deviations exceed a preset threshold, the optimal action parameters currently used to control the sediment remediation work are updated.

4. The adaptive sediment remediation robot control method based on deep learning according to claim 3, characterized in that, The specific environmental deviations identified include: The direction of pollutant diffusion is determined by the vector field of the dynamic pollution field in the three-dimensional dynamic digital twin model, and the high-pollution grid area and the ordinary sediment grid area are divided according to the scalar field concentration value. Based on the regional division results, corresponding data acquisition frequencies are set for each grid region to increase the data acquisition frequency of high-pollution grids along the direction of pollution diffusion vector, and / or reduce the data acquisition frequency of ordinary sediment grid regions. The collected real-world perception data is compared with the simulation data of the corresponding grid in the three-dimensional dynamic digital twin model to obtain the regional deviation value, which is used as the environmental deviation.

5. The adaptive sediment remediation robot control method based on deep learning according to claim 1, characterized in that, Initializing the state space using the aforementioned three-dimensional dynamic digital twin model includes the following steps: Train a deep learning model to identify pollutants, generate pollution categories, pollution levels and corresponding spatial locations, and construct a pollution removal list; Structured data is extracted based on the digital twin model and combined with the pollutant removal list to form the state space.

6. A deep learning-based adaptive sediment remediation robot control method according to claim 1 or 5, characterized in that, The reward function used when training the reinforcement learning model is: in, The amount of pollutants removed is used to characterize the actual contribution of robot actions to pollutant reduction; This is the target approximation term, used to characterize the approximation effect of the action command on the target contaminated area; This is the heading stability term, used to characterize the impact of motion commands on the robot's underwater attitude stability; This is a pollution priority item, used to guide consideration of prioritizing highly polluted areas; This is a regional coverage term, used to incentivize traversal of uncovered, heavily polluted areas; This is an energy consumption penalty term used to constrain the energy consumption of action commands; This is a penalty item for collision risk. , , , , , and The penalty coefficient corresponds to different reward items.

7. The adaptive sediment remediation robot control method based on deep learning according to claim 6, characterized in that, The action space of the reinforcement learning model includes: the robot's displacement heading vector, the robot's end effector trajectory, and the parameters for injecting the sediment remediation drug.

8. A deep learning-based adaptive sediment remediation robot, characterized in that, include: A multimodal perception module is used to collect multimodal data, including underwater terrain data, pollution concentration distribution data, and robot posture data. The embedded processing module is used to perform spatiotemporal registration and fusion based on the multimodal data to construct a three-dimensional dynamic digital twin model of the sediment environment; based on the three-dimensional dynamic digital twin model, a multi-dimensional state space is constructed, and optimal action parameters are generated through a trained reinforcement learning model; A dynamic actuator is used to perform corresponding actions based on the optimal action parameters.

9. The deep learning-based adaptive sediment remediation robot according to claim 8, characterized in that, The multimodal sensing module includes a three-dimensional sonar array, a turbidity sensor array, an inertial measurement unit, and a preprocessing unit. The three-dimensional sonar array is used to acquire sonar point cloud data, and the sonar point cloud data is converted into a terrain network model through point cloud registration, and feature extraction is performed to obtain a terrain curvature feature matrix. The turbidity sensor array is used to acquire turbidity data at different locations; The inertial measurement unit is used to acquire the robot's attitude data; The preprocessing unit is used to convert the sonar point cloud data into a terrain network model through point cloud registration, and to extract features to obtain a terrain curvature feature matrix as the terrain feature. This is used to perform Kalman filtering on turbidity data from different locations, and through correlation fusion, to generate a pollutant concentration gradient tensor as the pollutant concentration distribution feature; The data is used to collect the robot's roll angle, pitch angle, and yaw angle. The collected angle data is then converted into attitude quaternions using an attitude conversion algorithm, which are used as the attitude parameters.

10. The deep learning-based adaptive sediment remediation robot according to claim 8, characterized in that, The dynamic actuator includes a robotic arm, a vector propulsion system, and a repair drug injector; The vector propulsion system is mounted on the robotic arm, and the repair drug injector is fixed to the end of the robotic arm; The robot's motion trajectory is controlled by path planning and adjustment of thrust magnitude and direction. The pose of the robotic arm is controlled by joint angle adjustment, thereby achieving the working positioning of the drug repair syringe.

Images