An underwater robot intelligent operation method and system based on digital twinning

Abstract

The application discloses an underwater robot intelligent operation method and system based on digital twinning, which comprises the following steps: collecting multi-modal sensor data of the underwater robot during operation, and jointly performing online inversion and correction on model parameters in the digital twinning environment; using a physical information neural network to constrain reinforcement learning operation strategy and deploying the operation strategy on the underwater robot; triggering switching of a three-layer progressive safety architecture composed of an operation strategy, a model predictive controller and an impedance controller by using uncertainty measurement; performing interpretable analysis and task planning on high-level natural language instructions; during task execution, performing real-time abnormal diagnosis and root cause analysis; feeding back interaction data and abnormal patterns to the digital twinning environment and the knowledge graph to drive model correction, operation strategy optimization and knowledge updating, and realizing continuous co-evolution of the digital twinning environment and the operation strategy. The application significantly improves the reliability, safety and autonomy of underwater operation of the underwater robot.

Description

Technical Field

[0001] This invention relates to the field of intelligent underwater robot operation technology, and in particular to an intelligent underwater robot operation method and system based on digital twins. Background Technology

[0002] Underwater robots (AUVs / ROVs) are core equipment for deep-sea exploration and facility maintenance, but the extreme environments of the deep sea, such as high pressure, turbulence, and attenuation of optical / acoustic channels, pose severe challenges to their intelligent perception, autonomous decision-making, and stable control. Currently, the industry generally adopts a technical approach combining digital twins and simulation training, hoping to transfer strategies trained in virtual environments to real robots. However, this approach faces a serious problem of discrepancies between simulation and reality in practice: traditional digital twin environments are mostly built based on idealized or static physical parameters, and their fluid dynamics and sensor noise models cannot accurately reflect the complex, time-varying, and nonlinear physical processes in real underwater environments (such as transient eddies and drastic changes in optical imaging quality). This inherent bias at the model level leads to a sharp drop in performance after deployment of strategies that perform well in simulation, often due to inaccurate dynamic response or distorted sensory input, creating a migration gap from virtual to reality.

[0003] To address this challenge, numerous technological innovations have emerged in recent years, but each has its own limitations. In the area of ​​digital twin environment construction and simulation training, patent CN117408082A (Underwater robot simulation method, system, and medium based on digital twin technology) proposes constructing a virtual model and calculating the "underwater scene influence coefficient" and "operation level coefficient," optimizing operations through preset thresholds. This method focuses on operator skill training and evaluation, but its twin environment parameters are fixed, lacking a mechanism for dynamic correction based on feedback from the real robot, and cannot close the loop to resolve the virtual-real discrepancy caused by model drift. In the area of ​​perception and specific task control, patent CN114862904B (A continuous target tracking method using a twin network for underwater robots) utilizes a twin network for underwater target tracking, judging the tracking status through similarity scoring and thresholds. This approach focuses on the robustness of visual perception tasks, but its tracking strategy is independent of the robot's overall operation planning and control. It fails to integrate perceptual uncertainty into high-level decision-making and safety control loops, making it difficult to guarantee system-level safety in complex tasks. Regarding swarm collaboration and high-level decision-making, patent CN116700299A (An AUV Swarm Control System and Method Based on Digital Twin) achieves dynamic task allocation and route planning by constructing a swarm digital twin. Patent CN119717649A (An AUV Motion Control Method and System Based on LLM Human-Machine Interaction Framework) introduces a Large Language Model (LLM) to parse user commands and generate high-level decisions to drive the lower-level controller. These methods improve the system's planning capabilities and human-machine interaction level, but their decision-making process relies on pre-set models or semantic understanding, lacking quantitative evaluation of the confidence level of the lower-level control strategy in real-world environments. Furthermore, they do not embed underwater physical laws as hard constraints into the strategy learning process, potentially leading to high-risk commands that violate physical principles or exceed actuator capabilities when encountering unmodeled interference or unexpected situations.

[0004] In summary, while existing technologies have made progress in some aspects, they generally suffer from systemic defects such as static and rigid digital twin models, a disconnect between policy training and physical reality, a lack of safety adaptive mechanisms to perceive uncertainty under sparse feedback, and a lack of physically interpretable knowledge to guide task planning. Most solutions develop independently, failing to achieve coordinated closed-loop evolution of the twin environment, perception modules, control strategies, and task planners, making it difficult to continuously guarantee the reliability, safety, and autonomy of intelligent operations in real, complex underwater environments. Therefore, there is an urgent need for a solution that can dynamically compensate for differences between virtual and real environments and achieve coordinated evolution of the entire system's intelligent agents. Summary of the Invention

[0005] This invention provides an intelligent operation method and system for underwater robots based on digital twins, which systematically solves problems such as simulation-reality differences, weak cross-domain generalization ability of strategies, insufficient safety adaptive ability under sparse feedback, and poor interpretability of task planning in complex underwater environments.

[0006] The technical solution of the present invention is as follows: A digital twin-based intelligent operation method for underwater robots includes the following steps: (1) Construct a digital twin environment for the underwater robot; during the operation, collect multimodal sensor data of the underwater robot simultaneously, and perform joint online inversion and correction of the computational fluid dynamics transient flow field model parameters and the perception degradation model parameters in the digital twin environment; (2) In the modified digital twin environment, a physical information neural network is used as the operation strategy network constraint reinforcement learning operation strategy and deployed on the underwater robot; (3) Based on the received sparse success and failure feedback signals, update the posterior distribution of the situation parameters, and trigger the switching of the three-layer progressive safety architecture consisting of the operation strategy, model prediction controller and impedance controller with uncertainty measure. (4) Based on the knowledge graph, interpretable parsing and task planning are performed on high-level natural language instructions; during task execution, anomaly diagnosis and root cause analysis are achieved by comparing the sensor data with the prediction results of the physical forward model associated in the knowledge graph. Interactive data and anomaly patterns are fed back to the digital twin environment and knowledge graph to drive model correction, operational strategy optimization, and knowledge updates, thereby enabling the continuous co-evolution of the digital twin environment and operational strategies.

[0007] This invention addresses the combined challenges of highly dynamic fluid disturbances, strong perception degradation, low-bandwidth communication, and high-frequency model mismatch in underwater environments. It constructs a complete perception-decision-execution-learning closed loop that enables the co-evolution of a digital twin environment and robot operation strategies. Specifically, the high-fidelity digital twin environment output in step (1) online evolution provides the foundation for training the operation strategy network; the real-world interaction data generated in step (3) safe deployment and the specific abnormal patterns diagnosed in step (4) task understanding are structured and fed back to the digital twin environment and knowledge graph, thereby driving model correction, strategy optimization, and knowledge updates. This enables the entire system to continuously co-evolve in complex underwater environments, solving problems such as simulation-reality discrepancies, weak cross-domain strategy generalization ability, insufficient safety adaptive capability under sparse feedback, and poor task planning interpretability in complex underwater environments.

[0008] Preferably, in step (1), the multimodal sensor data includes six-dimensional force / torque data, sonar raw data, and optical image data; the perception degradation model includes an acoustic attenuation model, an optical scattering model, and a turbulence disturbance model.

[0009] Preferably, step (1) includes: (1-1) Construct a digital twin environment that includes a computational fluid dynamics transient flow field model and a perceptual degradation model; (1-2) During the operation, six-dimensional force / torque data and multi-modal sensor data of the underwater robot under fluid disturbance are collected simultaneously. Based on the joint optimization framework of extended Kalman filtering and maximum likelihood estimation, the parameters of the computational fluid dynamics transient flow field model and the parameters of the perception degradation model are constructed into a joint state vector. The continuous online inversion and correction of the model parameters are realized through Bayesian update, so that the virtual environment dynamically approximates the real underwater complex physical field.

[0010] Preferably, the joint state vector ,in To calculate the parameters of the transient flow field model in fluid dynamics, For parameters of the perceived degradation model; Bayesian update satisfies: ; in, Formal representation in , Likelihood probability under given conditions; Indicates that given model parameters and current state Under these conditions, sensor data were observed. The likelihood probability; Represents the time from the initial moment to the current moment. Multimodal sensor data sequences; Computational Fluid Dynamics Transient Flow Field Model Output and six-dimensional force / torque data Residual drive Gradient update; Perceptual degradation model Optimization through perceptual consistency loss .

[0011] Preferably, in step (2), the operation strategy network inputs multimodal sensor data. Field-invariant features Output of the job policy network ; Joint objective function of training job policy network for: ; in, For the loss of the task master, To counteract domain adaptation loss, For feature distribution alignment loss, For physical consistency loss, , , To balance hyperparameters; ; This is a fluid-structure interaction model. For the fluid velocity vector, This represents a tensor representing physical quantities related to fluid dynamics (such as momentum or force density). The formula above describes the physical conservation laws of fluids.

[0012] Preferably, in step (3), the uncertainty measure , Represents the scenario parameter vector. Represents the scenario parameter vector The posterior covariance matrix, Represents the trace of a matrix; The switching rules for the three-tier progressive security architecture are as follows: (i) When At that time, execute the work strategy. ; (ii) When Then, solve the model predictive controller constraint optimization: ; in, and Predict the controller weight matrix for the model; As a reference trajectory, the system's state at the next moment. From the current state and control input Calculated using a computational fluid dynamics transient flow field model; This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The system state at any given moment; This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The control input to be solved at each moment; (iii) When When the impedance controller is activated. ,in and These represent the stiffness and damping matrices of the impedance controller, and the threshold value, respectively. , It learns and updates online by using historical fault data.

[0013] Preferably, the knowledge graph integrates knowledge from fields such as underwater robot structure, corrosion characteristics, and the effects of bioattachment.

[0014] Preferably, in step (4), the task planning process is formalized as a skill sequence optimization problem under knowledge graph constraints, and the task planning is generated through the following optimization objectives: ; in, For skill sequences; The knowledge graph for underwater operations consists of nodes containing entities such as equipment components, tools, and environmental parameters, and edges containing relationships such as spatial constraints, preconditions, and physical effects. for exist The logical rationality score is calculated using a graph neural network to determine the path matching degree and constraint satisfaction degree. The current environmental context includes equipment status, ocean current speed, and visibility data acquired in real time through underwater robot sensors; To conduct a feasibility assessment; and These are weighting coefficients, calibrated using domain expert rules or historical data, used to balance the weights of logical rationality and execution feasibility.

[0015] Preferably, the anomaly diagnosis in step (4) includes: 1) Residual calculation and threshold alarm: The residual calculation is performed between the real sensor data and the prediction results of the physical forward model associated with the knowledge graph. When the residual exceeds the preset threshold, an anomaly alarm is triggered; 2) Graph structure reasoning and root cause backtracking: Based on the graph structure reasoning of the knowledge graph, backtracking is performed along the associated path of equipment-component-material-corrosion characteristics-environmental parameters to locate the root cause of the anomaly; 3) Structured storage and feedback update: The diagnosed anomaly pattern is structured and stored and fed back to the knowledge graph to update the node attributes and relationship weights.

[0016] Based on the same inventive concept, the present invention also provides an intelligent underwater robot operation system based on digital twins, for performing the method, including: The co-evolution module of the digital twin environment performs joint online inversion and correction of the parameters of the computational fluid dynamics transient flow field model and the parameters of the perception degradation model in the digital twin environment based on the multimodal sensor data of the underwater robot collected synchronously during the operation. The task strategy training module, in the modified digital twin environment, uses a physical information neural network as the task strategy network constraint reinforcement learning task strategy and deploys it on the underwater robot. The safety adaptive module updates the posterior distribution of the scenario parameters based on the received sparse success and failure feedback signals, and triggers the switching of a three-layer progressive safety architecture consisting of operation strategy, model prediction controller and impedance controller with uncertainty metric. The task understanding module performs interpretable parsing and task planning for high-level natural language instructions based on knowledge graphs; during task execution, it performs real-time anomaly diagnosis and root cause analysis. When the system starts up, a digital twin environment is constructed. When the underwater robot performs its tasks, the interactive data generated by the safety adaptation module and the abnormal patterns diagnosed by the task understanding module are fed back to the twin environment co-evolution module and the task understanding module to drive model correction, strategy optimization and knowledge update.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The digital twin closed-loop evolution method of the present invention, through online Bayesian parameter correction based on real operation data, enables the simulation environment to dynamically approximate the complex and ever-changing underwater physical field, fundamentally reducing the difference between simulation and reality, and providing a high-fidelity foundation for strategy training and testing that continuously approximates reality. (2) The physical model-guided strategy training method of the present invention forces the agent to learn essential characteristics consistent with underwater dynamics through physical information neural network constraints and joint optimization, effectively overcoming the non-physical behavior generated by traditional methods when facing strong fluid-structure coupling and perception degradation, and significantly improving cross-domain generalization ability and decision interpretability. (3) The uncertainty-driven safety adaptive method of the present invention, under the condition of sparse feedback in the deep sea, realizes the smooth and autonomous switching of the three-layer architecture of intelligent strategy, model predictive control and impedance control based on the online quantification of strategy confidence, breaks through the rigid bottleneck of traditional fixed rules in dealing with sudden disturbances, and ensures the active safety and reliability degradation of complex operations. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an intelligent operation method for underwater robots based on digital twins. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0020] This invention provides an intelligent operation method for underwater robots. Addressing the combined challenges of highly dynamic fluid disturbances, strong perception degradation, low-bandwidth communication, and high-frequency model mismatch in underwater environments, this method constructs a cohesive closed loop that enables the co-evolution of the digital twin environment and the robot's operation strategy. Specifically, it includes the following four collaborative operation steps: First, online co-evolution of the twin environment addresses fluid disturbance and model mismatch. A multi-physics digital twin environment is constructed, incorporating a computational fluid dynamics (CFD) transient flow field model and simulating underwater acoustic and optical sensing degradation. During real robot operations, its dynamic response under fluid disturbance and multi-modal sensor data are simultaneously collected. Based on this, the fluid dynamic parameters and sensor degradation model parameters in the twin environment are jointly inverted and corrected online, enabling the virtual environment to dynamically approximate the complex physical field of real underwater environments. The core correction mechanism employs a joint optimization framework based on extended Kalman filtering and maximum likelihood estimation, achieving continuous optimization of model parameters through Bayesian updates.

[0021] Second, training for physical consistency of strategies across domain differences. In the aforementioned online evolutionary high-fidelity twin environment, a domain-invariant reinforcement learning method constrained by Physical Information Neural Network (PINN) is used to train the operational strategy. This method introduces a differentiable underwater physics forward model as a constraint, forcing the policy network to extract essential features consistent with underwater bulk-fluid interaction dynamics from simulation data collected by multimodal sensors (such as force sensors, cameras, sonar, etc.) and processed by a degradation model. The training process uses a joint objective function that integrates the task master loss, adversarial domain adaptation loss, feature distribution alignment loss, and physical consistency loss for optimization, ensuring that the learned strategy has both generalization ability across simulation and real domains and conforms to physical laws.

[0022] Third, secure online adaptive design for sparse-latency communication. The trained lightweight policy is deployed on a real underwater robot. Under the constraints of low bandwidth and high latency communication in the deep sea, the robot estimates the confidence level of the policy under the current unknown perturbation based on the received sparse success and failure feedback signals. Using this confidence level as a criterion, the robot automatically and smoothly switches between three progressive safety architectures consisting of a lightweight policy, a model predictive controller (MPC), and an impedance controller, thereby maximizing intelligent performance while ensuring absolute safety in the event of sudden anomalies.

[0023] Fourth, knowledge-enhanced task understanding of underwater equipment characteristics. Based on a knowledge graph integrating knowledge of underwater equipment structure, corrosion characteristics, and the effects of biofouling, interpretable parsing and task planning are performed on high-level natural language commands. The planning process is formalized as a skill sequence optimization problem constrained by the knowledge graph. During execution, anomaly diagnosis and root cause analysis are achieved by comparing the prediction results of mathematical models associated with real sensor data and the knowledge graph, which can simulate the physical responses (such as stress, strain, corrosion rate, and sealing performance) of specific underwater equipment (such as valves, welds, and submarine cables) under specific environmental conditions (such as specific pressure, salinity, and flow rate). This model is obtained by instantiating and parameterizing the equipment's physical properties, material parameters, environmental conditions, and their interactions (such as Hooke's Law, corrosion kinetic equations, and fluid resistance formulas) defined in the knowledge graph. Its inputs are equipment status and environmental observations, and its outputs are predicted values ​​of key physical quantities.

[0024] The above four steps constitute a complete closed loop of perception-decision-execution-learning. The high-fidelity environment output from the online evolution step provides the foundation for policy training; the real-world interaction data generated by the safe deployment step (including robot state sequences, action sequences during policy execution, success / failure reward signals obtained from sparse communication, safety architecture switching records, and control commands generated by MPC or impedance controllers, etc.) and the specific anomaly patterns diagnosed by the task understanding step are structured and fed back to the digital twin environment and knowledge base, thereby driving model correction, policy optimization, and knowledge updates. This is specifically achieved through the following process: 1) Data collection and structuring: Format the interactive data and diagnosed anomaly patterns (including anomaly type, associated components, root cause path, timestamp, etc.) into standard logs; 2) Twin environment update: Utilizing the state-action-dynamic response sequence in the interactive data, the parameters of the CFD model and the perceptual degradation model are continuously inverted and corrected online through the Bayesian joint optimization framework described in step 1; 3) Policy optimization: Fine-tune the deployed policy online using new interaction data (especially success and failure signals) (e.g., offline reinforcement learning), or retrain the policy in an evolved twin environment; 4) Knowledge Graph Update: New, verified anomaly patterns and their root causes are added to the knowledge graph as new nodes and edges, or the attributes of existing nodes (e.g., component failure rates) and relationships (e.g., the weight of "causing" relationships) are updated. This enables the entire system to continuously co-evolve in complex underwater environments. Furthermore, this invention also provides a system for implementing the above method. This system includes: a twin environment co-evolution module for performing online co-evolution of fluid disturbances and model mismatch; a policy physical consistency training module for performing domain-invariant reinforcement learning training guided by a physical model; a safety online adaptive module for performing three-layer safety switching control based on uncertainty estimation; and a knowledge-enhanced task understanding module for performing interpretable task planning and anomaly diagnosis guided by the knowledge graph.

[0025] Upon system startup, an underwater digital twin environment is first constructed, incorporating a transient flow field model and a multimodal perception degradation model. When the real underwater robot performs its tasks, the system dynamically evolves through the following closed-loop collaborative mechanism: Step 1: Real-time calibration of the twin environment The robot's sensors collect six-dimensional force / torque data under fluid disturbances in real time. and raw data from sonar and optics Based on the Bayesian joint optimization framework, fluid parameters are... With perceived degradation parameters Construct as a joint state vector Cross-physics inversion is achieved through extended Kalman filter (EKF) prediction and maximum likelihood estimation (MLE) correction: ; in, Formal representation in , Likelihood probability under given conditions; This represents the vector of model parameters to be estimated. Indicates that given model parameters and current state Under these conditions, sensor data were observed. The likelihood probability; Represents the time from the initial moment to the current moment. Multimodal sensor data sequences.

[0026] Among them, the differentiable CFD model generates the flow field function. Its output is the same as Residual drive Gradient update; acoustic / optical degradation model Then optimize through perceptual consistency loss This process allows the twin environment to continuously approximate the real physical field, laying the foundation for strategy training.

[0027] Step 2: Training the physical constraint strategy In the evolved high-fidelity environment, a Physical Information Neural Network (PINN) constrained reinforcement learning strategy is employed. : Policy network input degradation data Field-invariant features ,function It is a parameter of A neural network encoder is used to extract feature representations from raw sensor data that are independent of specific sensor domains (such as simulation or reality). Represents encoder network A set of trainable weight parameters. A fusion task main loss is employed. Combating losses Feature distribution alignment loss and loss of physical consistency Joint objective function: ; in, For the action output of the policy network (such as thruster commands). This is a fluid-structure interaction model, describing the characteristics... Apply control to the indicated state At that time, the robot interacts with the fluid field (using physical quantities). (This indicates) the dynamic effects produced. This represents the local fluid velocity vector. , , To balance hyperparameters, physics terms As a hard constraint embedding, it ensures that the policy behavior strictly conforms to the laws of hydrodynamics. Training requires... The data in the strategy training phase mainly comes from the high-fidelity digital twin environment simulation generated after step 1, and is superimposed with the corresponding perception degradation model to simulate real underwater degradation perception; in the strategy fine-tuning phase, it comes from the data collected by the real robot.

[0028] After the policy network is trained, it is deployed on the onboard computing unit of a real underwater robot for online real-time decision-making, and then enters the safe online adaptive loop in step 3.

[0029] Step 3: Confidence-driven secure handover During the deployment phase, the robot receives sparse success and failure signals via underwater acoustic communication. (A binary signal indicating the success or failure of a task segment, determined by the surface control center or autonomously). The posterior distribution of contextual parameters is updated based on variational Bayesian inference. The specific steps are as follows: Assuming a posterior distribution Following a certain parameterized family of distributions (such as a Gaussian distribution), by maximizing the lower bound of the evidence ( This can be used to optimize the parameters of the distribution.

[0030] Defined as:

[0031] in, As a prior distribution, Let be the likelihood function. Represents the log-likelihood. Optimization is performed using gradient descent. The parameters are set to approximate the true posterior distribution. KL divergence is used to measure the difference between two probability distributions. and The nonnegative scalar of the difference between them. This represents the expected value of a mathematical expression within parentheses. ,exist obey Given the distribution, find its average value. Represents a vector of scenario parameters (ocean current speed, visibility, etc.), and measures uncertainty. Triggering a three-layer control switch, among which This indicates finding the trace of a matrix: (i) Lightweight strategy layer: when ,implement ; (ii) MPC layer: when Solving constraint optimization at time: ; in, and This is the weight matrix of the MPC controller. As a reference trajectory, the system's state at the next moment. From the current state and control input It was calculated using a CFD model. This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The system state at a given moment. This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The control input to be solved at each moment; (iii) Impedance control layer: when Enabled at time For safety, this formula is the core equation of impedance control, used to describe the relationship between the required output force / torque and position / velocity when a robot's end effector (such as a robotic arm) comes into contact with the environment. Threshold , It learns and updates online by using historical fault data. and The stiffness and damping matrix of the impedance controller. This indicates the actual position (or orientation) of the robot's end effector. This indicates the actual speed (or angular velocity) of the end effector. This indicates the desired position (or orientation) of the robot's end effector.

[0032] Step 4: Knowledge Enhancement Task Execution High-level task instructions via knowledge graph Analysis into skill sequence Task planning is generated through physically interpretable optimization objectives: ; in, Skill sequences refer to the processing of high-level natural language instructions (such as "inspect the subsea pipeline") through a knowledge graph ( After parsing and planning, an ordered, atomized chain of executable actions (such as "approaching the target - identifying the interface - performing grasping - placing") is generated, which acts as a bridge connecting the high-level task semantics and the low-level robot control. The knowledge graph for underwater operations consists of nodes containing entities such as equipment components, tools, and environmental parameters, and edges containing relationships such as spatial constraints, preconditions, and physical effects. To measure logical rationality, a sequence is represented. exist The logical rationality score is calculated using a graph neural network to determine the path matching degree and constraint satisfaction degree. O It indicates the current environmental context, including equipment status, ocean current speed, and visibility data obtained in real time through underwater robot sensors; Assess the feasibility of implementation; and These are weighting coefficients, calibrated using domain expert rules or historical data, used to balance the weights of logical rationality and execution feasibility.

[0033] Therefore, the core process of the above method is: planning, decomposing into skill sequences, scheduling and handing them over to an adaptive controller for execution, monitoring throughout the process and making safe switching based on uncertainty feedback, and adjusting or replanning online when necessary. This forms a closed-loop autonomous execution system from high-level semantic instructions to low-level physical actions.

[0034] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent operation of underwater robots based on digital twins, characterized in that, include: (1) Constructing a digital twin environment for underwater robots; During the operation, multimodal sensor data of the underwater robot are collected simultaneously, and the parameters of the computational fluid dynamics transient flow field model and the perception degradation model in the digital twin environment are jointly inverted and corrected online. (2) In the modified digital twin environment, a physical information neural network is used as the operation strategy network constraint reinforcement learning operation strategy and deployed on the underwater robot; (3) Based on the received sparse success and failure feedback signals, update the posterior distribution of the situation parameters, and trigger the switching of the three-layer progressive safety architecture consisting of the operation strategy, model prediction controller and impedance controller with uncertainty measure. (4) Based on knowledge graphs, perform interpretable parsing and task planning for high-level natural language instructions; During task execution, anomaly diagnosis and root cause analysis are achieved by comparing sensor data with the prediction results of the physical forward model associated with the knowledge graph. Interactive data and anomaly patterns are fed back to the digital twin environment and knowledge graph to drive model correction, operational strategy optimization, and knowledge updates, thereby enabling the continuous co-evolution of the digital twin environment and operational strategies.

2. The intelligent underwater robot operation method based on digital twin according to claim 1, characterized in that, In step (1), the multimodal sensor data includes six-dimensional force / torque data, sonar raw data, and optical image data; the perception degradation model includes an acoustic attenuation model, an optical scattering model, and a turbulence disturbance model.

3. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, Step (1) includes: (1-1) Construct a digital twin environment that includes a computational fluid dynamics transient flow field model and a perceptual degradation model; (1-2) During the operation, six-dimensional force / torque data and multi-modal sensor data of the underwater robot under fluid disturbance are collected simultaneously. Based on the joint optimization framework of extended Kalman filtering and maximum likelihood estimation, the parameters of the computational fluid dynamics transient flow field model and the parameters of the perception degradation model are constructed into a joint state vector. The continuous online inversion and correction of the model parameters are realized through Bayesian update, so that the virtual environment dynamically approximates the real underwater complex physical field.

4. The intelligent underwater robot operation method based on digital twin according to claim 3, characterized in that, The joint state vector ,in To calculate the parameters of the transient flow field model in fluid dynamics, For parameters of the perceived degradation model; Bayesian update satisfies: ； in, Formal representation in , Likelihood probability under given conditions; Indicates that given model parameters and current state Under these conditions, sensor data were observed. The likelihood probability; Represents the time from the initial moment to the current moment. Multimodal sensor data sequences; Computational Fluid Dynamics Transient Flow Field Model Output and six-dimensional force / torque data Residual drive Gradient update; Perceptual degradation model Optimization through perceptual consistency loss .

5. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, In step (2), the operation strategy network inputs multimodal sensor data. Field-invariant features Output of the job policy network ; Joint objective function of training job policy network for: ； in, For the loss of the task master, To counteract domain adaptation loss, For feature distribution alignment loss, For physical consistency loss, , , To balance hyperparameters; ； This is a fluid-structure interaction model. For the fluid velocity vector, Tensors representing physical quantities related to fluid dynamics.

6. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, In step (3), uncertainty measurement , Represents the scenario parameter vector. Represents the scenario parameter vector The posterior covariance matrix, Represents the trace of a matrix; The switching rules for the three-tier progressive security architecture are as follows: (i) When At that time, execute the work strategy. ; (ii) When Then, solve the model predictive controller constraint optimization: ； in, and Predict the controller weight matrix for the model; For reference trajectory; This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The system state at any given moment; This indicates that within the prediction time domain, from the current time... Initially, predicting the future... The control input to be solved at each moment; (iii) When When the impedance controller is activated. ,in and The stiffness and damping matrices of the impedance controller, and the threshold are respectively. , It learns and updates online by using historical fault data.

7. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, The knowledge graph integrates domain knowledge, including underwater robot structure, corrosion characteristics, and the effects of bioattachment.

8. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, In step (4), the task planning process is formalized as a skill sequence optimization problem under knowledge graph constraints. The task planning is generated through the following optimization objectives: ； in, For skill sequences; A knowledge graph for underwater operations; for exist The logical rationality score; For the current environment context; To conduct a feasibility assessment; and These are the weighting coefficients.

9. The intelligent operation method for underwater robots based on digital twins according to claim 1, characterized in that, The anomaly diagnosis in step (4) includes: calculating the residual between the real sensor data and the prediction results of the physical forward model associated with the knowledge graph; triggering an anomaly alarm when the residual exceeds a preset threshold; tracing back along the associated path of equipment-component-material-corrosion characteristics-environmental parameters based on the graph structure reasoning of the knowledge graph to locate the root cause of the anomaly; and storing the diagnosed anomaly pattern in a structured manner and feeding it back to the knowledge graph to update the node attributes and relationship weights.

10. An intelligent underwater robot operation system based on digital twins, characterized in that, For performing the method according to any one of claims 1-9, comprising: The co-evolution module of the digital twin environment performs joint online inversion and correction of the parameters of the computational fluid dynamics transient flow field model and the parameters of the perception degradation model in the digital twin environment based on the multimodal sensor data of the underwater robot collected synchronously during the operation. The task strategy training module, in the modified digital twin environment, uses a physical information neural network as the task strategy network constraint reinforcement learning task strategy and deploys it on the underwater robot. The safety adaptive module updates the posterior distribution of the scenario parameters based on the received sparse success and failure feedback signals, and triggers the switching of a three-layer progressive safety architecture consisting of operation strategy, model prediction controller and impedance controller with uncertainty metric. The task understanding module performs interpretable parsing and task planning for high-level natural language instructions based on knowledge graphs; during task execution, it performs real-time anomaly diagnosis and root cause analysis. When the system starts up, a digital twin environment is constructed. When the underwater robot performs its tasks, the interactive data generated by the safety adaptation module and the abnormal patterns diagnosed by the task understanding module are fed back to the twin environment co-evolution module and the task understanding module to drive model correction, strategy optimization and knowledge update.

Images