Underwater robot adaptive stability control system based on multi-sensor fusion

Abstract

The invention relates to the technical field of underwater robot control, in particular to an underwater robot self-adaptive stability control system based on multi-sensor fusion, which comprises a multi-source information sensing module used for acquiring pose parameters, motion states and external environment flow field information of an underwater robot in real time through a multi-sensor fusion module; and the state feature fusion module is used for carrying out feature fusion on the pose parameters, the motion state and the environmental flow field information, and constructing a state vector conforming to a Markov decision process. According to the scheme, a neural control mechanism with an autonomous evolution capability is constructed through deep fusion of flow field inversion prediction and a near-end strategy optimization algorithm, advanced pre-judgment and accurate active compensation of unsteady high-current interference are realized, and the attitude stability and path following robustness of the robot under an extremely unknown sea condition are remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of underwater robot control technology, and in particular to an adaptive stability control system for underwater robots based on multi-sensor fusion. Background Technology

[0002] With the deepening of deep-sea resource development and precision underwater operations, the demand for high-precision stability control of underwater robots in complex dynamic flow fields is becoming increasingly urgent. Current mainstream solutions mostly employ multi-source information fusion sensing technology, combined with dynamic modeling and feedback control algorithms, to adjust power distribution in real time to maintain attitude when the robot is impacted by the flow field. To improve operational safety, researchers are committed to introducing more advanced predictive algorithms and control models, aiming to enable robots to capture the characteristics of the external environmental flow field in real time and achieve proactive suppression of complex fluid disturbances through refined compensation.

[0003] However, existing technologies are mostly based on passive feedback control, which suffers from lag in response to nonlinear strong turbulence and lacks robustness. In addition, control laws with fixed parameters are difficult to adapt to dynamically changing sea conditions, resulting in large attitude deviations of the robot when facing extreme disturbances and a lack of autonomous learning and calibration capabilities for complex flow fields. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides an adaptive stability control system for underwater robots based on multi-sensor fusion, which aims to improve the problems of slow response, limited accuracy, and poor sea state adaptability of traditional control systems.

[0005] This invention provides the following technical solution: an adaptive stabilization and control system for underwater robots based on multi-sensor fusion, comprising: The multi-source information sensing module is used to collect the underwater robot's pose parameters, motion status, and external environmental flow field information in real time through the multi-sensor fusion module; The state feature fusion module is used to fuse the pose parameters, motion state and environmental flow field information to construct a state vector that conforms to the Markov decision process. The intelligent strategy reasoning module is used to input the state vector into the strategy network constructed based on the near-end strategy optimization algorithm, and use the Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time. The thrust optimization allocation module is used by the main control unit to perform optimization solutions based on the attitude deviation compensation vector and the robot dynamics model, and generate control duty cycle signals for each thruster. The power drive execution module is used to generate a reverse thrust to counteract external water flow disturbances based on the control duty cycle signal of the thruster array, so as to realize dynamic adaptive correction of the deviation path or deflection attitude. The self-calibration control module combines the corrected pose feedback with a preset reward function, uses a Critic network to evaluate the value of the current control strategy, and fine-tunes the strategy network parameters in real time to complete the online autonomous calibration of the control law.

[0006] Preferably, in the multi-source information sensing module, the step of real-time acquisition of the underwater robot's pose parameters includes: The three-axis angular velocity and linear acceleration data of the robot are obtained by using a nine-axis inertial measurement unit, and the three-dimensional coordinate displacement vector of the robot relative to the seabed is calculated in real time by using a Doppler log. An extended Kalman filter algorithm is introduced to perform online denoising of the original high-frequency noise of the sensor, and a real-time coordinate transformation is performed between the geographic coordinate system and the body coordinate system to obtain absolute attitude angle data; By combining the hydrostatic pressure data output by the depth sensor, the robot's real-time vertical depth information is obtained by mapping the conversion model that correlates air pressure and water pressure. The three-axis angular velocity, linear acceleration, three-axis coordinate displacement vector, absolute attitude angle data, and real-time vertical depth information are aggregated to a unified time baseline.

[0007] Preferably, in the multi-source information sensing module, the step of acquiring the external environmental flow field information includes: Multiple pressure sensing arrays are pre-set on the surface of the robot shell to capture the impact pressure intensity of water flow from different directions in real time, and generate a dynamic pressure gradient field map of the robot surface. A simplified flow field inversion model based on computational fluid dynamics is established, and the dominant flow direction and velocity distribution of the current environmental flow field are identified by analyzing the pressure gradient distribution characteristics. Extract coherent structural characteristic parameters from the flow field and quantify the asymmetric disturbance torque generated by shear flow or eddies to provide a basis for feedforward compensation for subsequent attitude prediction.

[0008] Preferably, in the state feature fusion module, the step of performing feature fusion to construct a state vector conforming to a Markov decision process includes: Normalization processing is performed on the collected multi-source heterogeneous data, and the temporal motion trend features of the pose parameter sequence within the preset observation window are extracted using a one-dimensional convolution kernel. By using a fully connected mapping layer, unsteady flow field feature vectors are mapped to the latent space, and a comprehensive observation operator that includes the current state and environmental uncertainties is constructed. The time-domain motion trend features, environmental observation operators, and actuator action information from the previous moment are concatenated and spliced ​​to construct a high-dimensional continuous state feature vector that satisfies the Markov property.

[0009] Preferably, in the intelligent strategy reasoning module, the step of using an Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time includes: A deep neural network architecture containing residual connection modules is constructed, and the input state vector is subjected to multi-layer abstract mapping through nonlinear activation functions to extract key obstacle avoidance and stability features. The candidate action space is probabilistically modeled using a Gaussian policy distribution layer, and the mean and log standard deviation of the expected thrust for six-degree-of-freedom attitude correction are output. A residual control mechanism is introduced, which linearly superimposes the incremental compensation value output by the neural network with the traditional proportional-derivative control command to generate the final six-dimensional attitude deviation compensation vector.

[0010] Preferably, in the thrust optimization allocation module, the step of performing optimization by combining the robot dynamics model includes: A six-degree-of-freedom dynamic analytical equation for a robot, including the added mass effect and nonlinear hydrodynamic damping term, is established to accurately describe the dynamic response characteristics of the robot at different motion speeds. Construct a convex optimization mathematical model with the objective function of minimizing the output power of each thruster and with hard constraints of the upper limit of physical thrust and the rate of change of thrust; The convex optimization mathematical model is solved online using a sequential quadratic programming algorithm, which accurately decouples the global compensation requirements and allocates them to the target power output shaft of each actuator.

[0011] Preferably, in the power drive execution module, the step of dynamically adaptively correcting the offset path or deflection attitude includes: The allocated thrust commands are converted into pulse width modulation signals with dead zone compensation characteristics, which drive the H-bridge circuit of the thruster motor controller to perform duty cycle control at a preset frequency. The Hall encoder is used to provide real-time feedback on the actual rotational speed of the thruster, and a sliding mode variable structure control algorithm is introduced to quickly suppress load disturbances of the thruster, ensuring that the output thrust is consistent with the command requirements. The system monitors the convergence trajectory of deviations in various dimensions of the aircraft in real time. When the attitude overshoot exceeds the preset safety threshold, the dynamic limiting operator is triggered to perform secondary limiting protection on the thruster power.

[0012] Preferably, in the self-calibration control module, the step of evaluating the value of the current control strategy through the Critic network includes: Construct an evaluation model based on a value function and use a fully connected neural network to fit and evaluate the potential long-term return value of the current state vector; The residual between the actual benefit and the expected value after the current action is calculated using the temporal difference error algorithm, thereby quantifying the quality of the Actor network's output strategy. By combining the generalized dominance estimation method to smooth the multi-step sampling data, the dominance signal reflecting the magnitude of the improvement in control stability is extracted and used as the weight basis for the gradient update of the neural network.

[0013] Preferably, in the self-calibration control module, the step of completing the online autonomous calibration of the control law includes: The design incorporates a multi-objective composite reward function that considers attitude following accuracy, energy consumption, and control command smoothness, guiding the model towards high accuracy and low power consumption. The loss function is optimized using a near-end policy with a pruning operator. Gradient updates are performed on the weight parameters of the Actor network and the Critic network while ensuring that the policy update magnitude is within the trust region. The convergence curve of the reward function is monitored in real time. When the total deviation of the system within a preset number of iterations is lower than the minimum threshold, the current network weight parameters are automatically fixed to achieve self-optimization of the control strategy for a specific flow field environment.

[0014] The present invention has the following beneficial effects: 1. In this invention, the solution constructs a neural control mechanism with autonomous evolution capability by deeply integrating flow field inversion prediction and near-end strategy optimization algorithm, realizing the advanced prediction and accurate active compensation of unsteady strong current interference, which significantly enhances the robot's attitude stability and path following robustness under extreme unknown sea conditions.

[0015] 2. In this invention, by using a convex optimization thrust distribution strategy guided by a dynamic model, combined with residual control and sliding mode feedback mechanism, the adverse effects of nonlinear effects of the thruster and turbulent load on the stability control accuracy are eliminated, ensuring a fast and smooth response in the attitude correction process and significantly improving the attitude maintenance accuracy of underwater precision operations.

[0016] 3. In this invention, relying on the online self-calibration mechanism and multi-objective composite reward function, the power output distribution is automatically optimized while adaptively matching different working environments. This not only effectively reduces the overall power consumption of the system, but also reduces the mechanical fatigue of the actuator by suppressing control command oscillations, thus significantly extending the continuous working time and service life of the equipment. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of an adaptive stability control system for underwater robots based on multi-sensor fusion proposed in this invention. Detailed Implementation

[0018] The technical solutions in 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.

[0019] In embodiments of the present invention, the present invention provides an adaptive stability control system for underwater robots based on multi-sensor fusion, such as... Figure 1 As shown, it includes: The multi-source information sensing module is used to collect the underwater robot's pose parameters, motion status, and external environmental flow field information in real time through the multi-sensor fusion module; Furthermore, in the multi-source information sensing module, the steps for real-time acquisition of the underwater robot's pose parameters include: acquiring the three-axis angular velocity and linear acceleration data of the robot body using a nine-axis inertial measurement unit, and calculating the three-dimensional coordinate displacement vector of the robot relative to the ocean floor in real time using a Doppler log; introducing an extended Kalman filter algorithm to perform online denoising of the original high-frequency noise of the sensors, and performing real-time coordinate transformation between the geographic coordinate system and the body coordinate system to obtain absolute attitude angle data; combining the hydrostatic pressure data output by the depth sensor, and using a conversion model relating air pressure and water pressure to map and obtain the robot's real-time vertical depth information; and summarizing the three-axis angular velocity, linear acceleration, three-axis coordinate displacement vector, absolute attitude angle data, and real-time vertical depth information to a unified time baseline.

[0020] Furthermore, in the multi-source information sensing module, the steps for acquiring external environmental flow field information include: pre-setting multiple pressure sensing arrays on the surface of the robot shell to capture the water flow impact pressure intensity from different directions in real time, generating a dynamic pressure gradient field map of the body surface; establishing a flow field inversion model based on computational fluid dynamics, identifying the dominant flow direction and velocity distribution of the current environmental flow field by analyzing the pressure gradient distribution characteristics; extracting coherent structural characteristic parameters in the flow field, quantifying the asymmetric disturbance torque generated by shear flow or eddies, and providing a feedforward compensation basis for subsequent attitude prediction.

[0021] Specifically, during the multi-source sensing step, the system first initializes the hardware communication link, retrieves the raw register data from the nine-axis inertial measurement unit, and acquires the three-axis angular velocity components and three-axis acceleration components. To eliminate the influence of high-frequency mechanical vibration and sensor zero drift on attitude analysis, an extended Kalman filter algorithm is used to recursively estimate the motion state in real time. The prediction process of this algorithm is based on a nonlinear dynamic model, and its state vector is defined as... ,in These represent roll angle, pitch angle, and yaw angle, respectively. Let represent the angular velocities about the body axis, and represent the angular velocities respectively. The state transition equation is described as follows: ; In this formula, for The prior state estimate at time t. It is a nonlinear state transition function. This refers to process excitation noise.

[0022] The axial velocity vector relative to the ocean floor was obtained by synchronously acquiring feedback from a Doppler log. Using coordinate transformation matrix The transformation relationship between projecting the velocity in the body coordinate system to the global geographic coordinate system is expressed as follows: ; In the formula The velocity vector in the geographic coordinate system. The rotation matrix is ​​calculated based on the current real-time attitude angle and includes trigonometric function mapping relationships for the transformation from the body coordinate system to the geographic coordinate system.

[0023] When processing depth data, the depth sensor converts the sensed hydrostatic pressure signal into an electrical signal. The system then calls a conversion model that correlates air pressure and water pressure. By eliminating the influence of atmospheric pressure at sea level, the vertical pose of the underwater robot is accurately calculated. The depth calculation formula is as follows: ; in This indicates the robot's current real-time vertical depth. The total pressure measured by the pressure sensor. For the local sea-level atmospheric pressure, either preset or obtained in real time. The density of seawater, For gravitational acceleration, all sensing data are aggregated to a unified time baseline with microsecond-level precision according to a timestamp alignment algorithm, forming a synchronized pose state set.

[0024] For acquiring external environmental flow field information, the system reads pressure sensor array data distributed at different locations on the robot's shell, constructs a dynamic pressure gradient field map of the robot's surface, and estimates the local flow velocity using a simplified flow field inversion model based on computational fluid dynamics and Bernoulli's principle for any measuring point on the robot's surface. Its local flow velocity With pressure The relationship satisfies: ; in The total pressure at the station site, For the first The static pressure measured by each sensing unit is used to perform spatial integration on the pressure gradient distribution characteristics of the entire fuselage, thereby identifying the dominant flow direction and quantifying the asymmetric disturbance torque generated by the flow field on the fuselage. The calculation formula is as follows: ; In this formula, The output is the asymmetric disturbance torque result. To the robot's center of gravity to the first The distance vector of each pressure sensing unit. Let be the normal vector of the surface where the element is located. Given the effective force-bearing area corresponding to this unit, by extracting the coherent structural characteristic parameters in the flow field, the system can quantify the abstract water flow impact into a specific technical control boundary, thereby providing accurate physical input for subsequent adaptive stability control.

[0025] This step, through high-frequency denoising and spatiotemporal synchronization processing, enables deep perception of the underwater robot's motion state and external environmental disturbances, significantly improving the signal-to-noise ratio of the input data and the real-time accuracy of pose calculation, and providing reliable data support for attitude self-calibration in complex flow field environments.

[0026] The state feature fusion module is used to fuse pose parameters, motion state and environmental flow field information to construct a state vector that conforms to the Markov decision process. Furthermore, in the state feature fusion module, the steps of feature fusion to construct a state vector that conforms to the Markov decision process include: performing normalization processing on the collected multi-source heterogeneous data; using a one-dimensional convolutional kernel to extract the temporal motion trend features of the pose parameter sequence within a preset observation window; using a fully connected mapping layer to map the unsteady flow field feature vector to the latent space, constructing a comprehensive observation operator that includes the current state and environmental uncertainties; and concatenating and splicing the temporal motion trend features, the environmental observation operator, and the actuator action information from the previous moment to construct a high-dimensional continuous state feature vector that satisfies the Markov property.

[0027] Specifically, after acquiring multi-source heterogeneous data, the system first performs normalization preprocessing on pose parameters, motion state, and external environmental flow field information. Since the physical dimensions of different sensors differ, to eliminate the influence of range differences on feature extraction, a standard deviation normalization method is used to transform various types of input data. The calculation formula is as follows: ; in These are the normalized eigenvalues. The raw sensor data collected, This is the average value of this type of data over a preset observation period. This represents the corresponding standard deviation.

[0028] The processed data is divided into time-correlated feature sequences, providing standardized input for subsequent temporal feature extraction. For the normalized pose parameter sequence, the system introduces a one-dimensional convolution kernel to perform convolution operations within a preset sliding observation window to extract the motion trend features of the underwater robot under dynamic flow field impact. Let the sequence length within the observation window be , and the processing procedure of the one-dimensional convolution operator is as follows: ; In this formula, Indicates in The time-domain motion trend features extracted at each moment. The weights of the one-dimensional convolution kernel. For the normalized pose sequence within the window, The size of the convolution kernel. For bias terms, Representing a nonlinear activation function, this operator's sliding scan can effectively capture the acceleration trend of robot posture changes and the high-frequency pulsation characteristics caused by water flow disturbance.

[0029] For unsteady flow field eigenvectors, the system uses a fully connected mapping layer to project them into a high-dimensional latent space, constructing a comprehensive observation operator. This process aims to uncover the deep nonlinear mapping relationship between the water pressure gradient and the forces acting on the organism, and to characterize the uncertainty of the environment. The mapping process follows the following mathematical expression: ; In the formula For the generated integrated observation operator results, The input flow field feature vector, and These are the weight matrix and bias vector of the mapping layer, respectively.

[0030] Ultimately, the system will extract the temporal motion trend features. Integrated observation operator actuator action information from the previous moment By cascading and concatenating these features, a high-dimensional continuous state feature vector satisfying the Markov property is constructed: ; Here That is, the constructed current state vector. This represents the cascading concatenation operation of vectors, which fully characterizes the robot's current physical state and its interaction with the environment, ensuring the real-time nature and completeness of the decision-making process.

[0031] The intelligent policy reasoning module is used to input the state vector into the policy network built based on the near-end policy optimization algorithm, and use the Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time. Furthermore, in the intelligent strategy reasoning module, the step of using the Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time includes: constructing a deep neural network architecture containing a residual connection module, performing multi-layer abstract mapping on the input state vector through a nonlinear activation function, and extracting key obstacle avoidance and stability features; using a Gaussian policy distribution layer to probabilistically model the candidate action space, and outputting the expected thrust mean and logarithmic standard deviation for six-degree-of-freedom attitude correction; introducing a residual control mechanism, linearly superimposing the incremental compensation value output by the neural network with the traditional proportional-derivative control command to generate the final six-dimensional attitude deviation compensation vector.

[0032] Specifically, after obtaining the high-dimensional continuous state feature vector, the system feeds it as an input signal into a policy network constructed based on a proximal policy optimization algorithm. The Actor part of this policy network adopts a deep architecture containing multiple sets of residual connection modules based on the ResNet-50 architecture. The gradient vanishing problem during deep network training is solved through a skip connection mechanism, enhancing the ability to extract complex water flow interference features. The transformation process of each residual module is described as follows: ; In this formula, Indicates the first The input feature vector of the hidden layer, and These are the weight matrix and bias term for this layer, respectively. As a nonlinear activation function, after multiple layers of abstract mapping, the network can decouple key stability control features reflecting the robot's current offset degree and the intensity of water flow impact from the state vector.

[0033] The extracted features are transmitted to a Gaussian policy distribution layer for probabilistic modeling of the continuous action space. Addressing the six-DOF attitude correction requirements of the underwater robot, this distribution layer outputs the mean expected thrust and its corresponding log-standard deviation across six dimensions, thus constructing a multivariate normal distribution. Its probability density function is expressed as: ; in In the current state Take action below Probabilistic strategies, Let be the expected mean thrust vector output by the network. The covariance matrix is ​​generated from the logarithmic standard deviation. This represents the dimension of the action space.

[0034] To improve the system's stability boundary and reduce the search difficulty in the early stages of reinforcement learning, this implementation introduces a residual control mechanism. The system first uses a traditional proportional-derivative controller to calculate the basic control quantity based on the deviation between the attitude setpoint and the feedback value. Then, the incremental compensation value output by the Actor network is linearly superimposed with this basic command to generate the final six-dimensional attitude deviation compensation vector. The superposition logic is expressed as follows: ; In the above formula, The final output is a six-dimensional attitude deviation compensation vector result. and These are the proportional coefficient matrix and the differential coefficient matrix, respectively. and These are the robot's current posture deviation vector and its rate of change feedback value, respectively. This is the real-time incremental thrust compensation output calculated by the Actor network.

[0035] This synthetic command combines the stability of deterministic control with the flexibility of intelligent control, ensuring control continuity under variable flow fields. This step utilizes the nonlinear fitting capability of neural networks to achieve accurate prediction and compensation for complex flow field disturbances. The residual structure ensures smooth switching of control commands, effectively shortening the robot's attitude recovery time under strong flow impact and reducing static errors.

[0036] The thrust optimization allocation module is used by the main control unit to perform optimization solutions based on the attitude deviation compensation vector and the robot dynamics model, and generate control duty cycle signals for each thruster. Furthermore, in the thrust optimization allocation module, the steps for optimization in conjunction with the robot dynamics model include: establishing a six-degree-of-freedom dynamic analytical equation for the robot that includes additional mass effects and nonlinear hydrodynamic damping terms, to accurately describe the dynamic response characteristics of the robot at different motion speeds; constructing a convex optimization mathematical model with the minimization of the output power of each thruster as the objective function and the upper limit of physical thrust and the rate of change of thrust as hard constraints; and using a sequential quadratic programming algorithm to solve the convex optimization mathematical model online, accurately decoupling the global compensation requirements and allocating them to the target power output shaft of each actuator.

[0037] Specifically, after acquiring the six-dimensional attitude deviation compensation vector, the main control unit first establishes the six-degree-of-freedom dynamic analytical equations of the underwater robot to accurately describe the dynamic response of the robot in a complex flow field. These equations fully consider the additional mass effects generated by the fluid environment and the influence of nonlinear hydrodynamic damping terms on the motion characteristics. Their mathematical expression is as follows: ; In this formula, The system mass matrix is ​​composed of the superposition of the robot body mass matrix and the fluid-added mass matrix. This represents the nonlinear hydrodynamic damping matrix, which includes linear terms reflecting the low-speed motion characteristics and quadratic terms reflecting the high-speed turbulence characteristics. This represents the six-degree-of-freedom velocity vector of the robot in the body coordinate system; This represents the restoring torque vector generated by gravity and buoyancy; This is the input attitude deviation compensation vector.

[0038] To reduce system energy consumption while meeting compensation requirements, the main control unit constructed a constrained convex optimization mathematical model. This model takes minimizing the sum of squares of the output power of all thrusters as its objective function, and its optimization objective can be expressed as: ; in The result is the total power cost of the system to be optimized. The vector to be solved contains the magnitudes of the thrust of each thruster. It is a pre-defined weighted diagonal matrix used to balance the energy efficiency differences between different thrusters.

[0039] The model introduces multiple sets of constraints: ; ; ; in The thruster configuration matrix is ​​determined based on the thruster spatial layout and is used to decouple and map the thrust of each thruster into the resultant force of the six degrees of freedom of the body. and These represent the lower and upper physical limits of the thrust that the thruster can output; A preset hard constraint on the thrust change rate of the propeller is used to prevent electromagnetic shock or mechanical damage to the motor due to sudden changes in speed.

[0040] The system employs a sequential quadratic programming algorithm to solve the aforementioned convex optimization mathematical model online in a loop. Within each control cycle, the algorithm linearizes the nonlinear constraints and approximates the objective function quadratically, thereby searching for the optimal thrust distribution solution within the feasible region. After obtaining the optimal thrust vector, the main control unit uses the pre-stored thrust-duty cycle mapping curve to convert the target thrust value of each thruster into a specific control duty cycle signal. This process ensures the precise decoupling of global attitude compensation requirements and their allocation to the power output shaft of each actuator.

[0041] The power drive execution module is used to generate a reverse thrust to counteract external water flow disturbances based on the control duty cycle signal of the thruster array, so as to realize dynamic adaptive correction of deviation path or deflection attitude. Furthermore, in the power drive execution module, the steps for dynamically adaptively correcting the deviation path or yaw attitude include: converting the allocated thrust commands into pulse width modulation signals with dead-zone compensation characteristics to drive the H-bridge circuit of the thruster motor controller to perform duty cycle control at a preset frequency; using a Hall encoder to provide real-time feedback on the actual speed of the thruster, and introducing a sliding mode variable structure control algorithm to quickly suppress the load disturbance of the thruster, ensuring that the output thrust is consistent with the command requirements; monitoring the deviation convergence trajectory of each dimension of the body in real time, and triggering a dynamic limiting operator to perform secondary limiting protection on the thruster power when the attitude overshoot exceeds a preset safety threshold.

[0042] Specifically, the thruster control unit receives the target thrust commands for each thruster output from the aforementioned steps. Then, the system first performs the conversion calculation from thrust to duty cycle signal. Considering the static friction of the motor under low voltage and the dead zone characteristics of hydrodynamics, the system introduces a dead zone compensation algorithm to map the allocated thrust command into a pulse width modulation signal with dead zone cancellation characteristics, and outputs the duty cycle. The computational model is expressed as follows: ; in This is the final output duty cycle value. The input is the target thrust command data for a single thruster. The sign function is used to determine the forward and reverse direction of the motor. This is the preset minimum duty cycle offset used to overcome the initial dead zone of the thruster. It is the linear mapping ratio coefficient between thrust and duty cycle.

[0043] To ensure the accuracy of thrust output under complex turbulent loads, the system utilizes a Hall encoder integrated on the motor shaft to provide real-time feedback on the actual rotational speed of the thruster. To address the resistance fluctuations caused by external water flow on the spiral screw, the system introduces a sliding mode variable structure control algorithm to enhance the robustness of the control law. First, the rotational speed deviation is defined. ,in To establish a sliding mode switching surface based on the reference rotational speed calculated from the target thrust. The generated correction control increment Follow the law of exponential convergence: ; In this formula, The output result of the control quantity after correction by the sliding mode algorithm. For the defined sliding mode function values, For the input speed error data, To adjust the equivalent control gain of the switching intensity, To determine the convergence rate parameter, is the gain constant of the integral operator.

[0044] During the dynamic correction process, the system continuously monitors the body's deviation convergence trajectory in six degrees of freedom. When the sensor detects attitude overshoot... Exceeding the preset security alarm threshold At this time, the system automatically triggers the dynamic limiting operator to perform secondary limiting protection on the input power of the thruster to prevent mechanical damage. The command correction coefficient after limiting is... The calculation method is as follows: ; in The output power limiting factor result. The specific value of the current overshoot of the sensed body attitude. The preset safe deflection angle threshold, This is used to prevent the bias of extremely small positive numbers with a denominator of zero.

[0045] This coefficient directly affects the duty cycle generation process, preventing severe oscillations or structural overload caused by excessive correction of the robot by reducing the power output in real time.

[0046] The self-calibration control module combines the corrected pose feedback with a preset reward function, uses a Critic network to evaluate the value of the current control strategy, and fine-tunes the strategy network parameters in real time to complete the online autonomous calibration of the control law.

[0047] Furthermore, in the self-calibration control module, the step of evaluating the value of the current control strategy through the Critic network includes: constructing an evaluation model based on a value function; using a fully connected neural network to fit and evaluate the potential long-term return value of the current state vector; using a time-series difference error algorithm to calculate the residual between the actual return and the expected value after the current action is executed, quantifying the superiority or inferiority of the strategy output by the Actor network; and combining the generalized advantage estimation method to smooth the multi-step sampled data, extracting the advantage signal reflecting the magnitude of the improvement in control stability, and using it as the weight basis for the gradient update of the neural network.

[0048] Furthermore, in the self-calibration control module, the steps for online autonomous calibration of the control law include: designing a multi-objective composite reward function that includes attitude following accuracy, energy consumption, and control command smoothness to guide the model towards high accuracy and low power consumption; using a near-end policy optimization loss function with a pruning operator to perform gradient updates on the weight parameters of the Actor network and Critic network while ensuring that the policy update amplitude is within the trust region; and monitoring the convergence curve of the reward function in real time, automatically fixing the current network weight parameters when the total system deviation within a consecutive preset number of iterations is lower than the minimum threshold, thereby achieving self-optimization of the control policy for a specific flow field environment.

[0049] Specifically, after completing the initial attitude correction, the system enters the online self-calibration stage by constructing an evaluation model based on the value function. The Critic network adopts the same ResNet-50 architecture as the Actor network. It receives the current high-dimensional state vector and performs fitting operations through a deep architecture with multiple sets of residual connection modules, aiming to evaluate the potential long-term return value of the robot in a specific flow field environment.

[0050] To quantify the actual effectiveness of the Actor network's output strategy, the system introduces a temporal difference error algorithm to calculate the deviation value after the action is executed. Its mathematical definition is as follows: ; in The calculated time-series difference error results, The instant reward value obtained at the current moment. and These output the value assessments for the current state and the state at the next moment, respectively. This is a discount factor used to weigh future returns.

[0051] To suppress evaluation fluctuations caused by underwater random turbulence, the system combines the generalized dominance estimation method to smooth the continuously acquired multi-step sample data and extract the dominance signal reflecting the magnitude of the improvement in control stability. The calculation logic of this signal is as follows: ; in The generated generalized advantage estimation results are used as the weights for gradient updates in the neural network. This is the preset upper limit of the sampling time step. This is a smoothing parameter used to adjust the tradeoff between bias and variance.

[0052] In the online self-optimization stage of the control law, the system first calculates a multi-objective composite reward function based on task requirements. This function guides the model evolution by integrating pose deviation, thruster energy consumption, and command smoothness. Its quantitative expression is as follows: ; in The output is the composite reward value. The input is the six-DOF attitude following accuracy deviation data. The control command amplitude data characterizes the energy consumption loss of the thruster. This data reflects the change in command values ​​at adjacent time points to indicate the smoothness of control. These are the corresponding preset weighting coefficients.

[0053] Subsequently, the system uses a near-end strategy with a pruning operator to optimize the loss function and iteratively update the network parameters. This process ensures that the control law evolves smoothly within the trust region by limiting the probability ratio between the old and new strategies, preventing control oscillations under complex flow fields. The objective function is expressed as: ; in The calculated policy network loss function result, The ratio of the probability of the new strategy to the probability of the old strategy under the current action. This is a preset pruning threshold used to enforce constraints on the policy update magnitude.

[0054] The main control unit monitors the convergence curve of the reward function in real time. When the rate of change of the total deviation of the system within a consecutive preset number of iterations is lower than the minimum threshold, the system automatically executes the parameter solidification instruction and uses the current optimal network weight as the benchmark for stable control.

[0055] This step, through a closed-loop learning and evaluation mechanism, enables the control parameters to be autonomously calibrated for different sea conditions, significantly enhancing the system's robustness in unsteady flow fields and effectively optimizing the smoothness of power output while maintaining high-precision attitude stability.

[0056] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An adaptive stabilization and control system for underwater robots based on multi-sensor fusion, characterized in that, include: The multi-source information sensing module is used to collect the underwater robot's pose parameters, motion status, and external environmental flow field information in real time through the multi-sensor fusion module; The state feature fusion module is used to fuse the pose parameters, motion state and environmental flow field information to construct a state vector that conforms to the Markov decision process. The intelligent strategy reasoning module is used to input the state vector into the strategy network constructed based on the near-end strategy optimization algorithm, and use the Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time. The thrust optimization allocation module is used by the main control unit to perform optimization solutions based on the attitude deviation compensation vector and the robot dynamics model, and generate control duty cycle signals for each thruster. The power drive execution module is used to generate a reverse thrust to counteract external water flow disturbances based on the control duty cycle signal of the thruster array, so as to realize dynamic adaptive correction of the deviation path or deflection attitude. The self-calibration control module combines the corrected pose feedback with a preset reward function, uses a Critic network to evaluate the value of the current control strategy, and fine-tunes the strategy network parameters in real time to complete the online autonomous calibration of the control law.

2. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the multi-source information sensing module, the step of real-time acquisition of the underwater robot's pose parameters includes: The three-axis angular velocity and linear acceleration data of the robot are obtained by using a nine-axis inertial measurement unit, and the three-dimensional coordinate displacement vector of the robot relative to the seabed is calculated in real time by using a Doppler log. An extended Kalman filter algorithm is introduced to perform online denoising of the original high-frequency noise of the sensor, and a real-time coordinate transformation is performed between the geographic coordinate system and the body coordinate system to obtain absolute attitude angle data; By combining the hydrostatic pressure data output by the depth sensor, the robot's real-time vertical depth information is obtained by mapping the conversion model that correlates air pressure and water pressure. The three-axis angular velocity, linear acceleration, three-axis coordinate displacement vector, absolute attitude angle data, and real-time vertical depth information are aggregated to a unified time baseline.

3. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the multi-source information sensing module, the steps for acquiring the external environmental flow field information include: Multiple pressure sensing arrays are pre-set on the surface of the robot shell to capture the impact pressure intensity of water flow from different directions in real time, and generate a dynamic pressure gradient field map of the robot surface. A simplified flow field inversion model based on computational fluid dynamics is established, and the dominant flow direction and velocity distribution of the current environmental flow field are identified by analyzing the pressure gradient distribution characteristics. Extract coherent structural characteristic parameters from the flow field and quantify the asymmetric disturbance torque generated by shear flow or eddies to provide a basis for feedforward compensation for subsequent attitude prediction.

4. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the state feature fusion module, the step of performing feature fusion to construct a state vector conforming to a Markov decision process includes: Normalization processing is performed on the collected multi-source heterogeneous data, and the temporal motion trend features of the pose parameter sequence within the preset observation window are extracted using a one-dimensional convolution kernel. By using a fully connected mapping layer, unsteady flow field feature vectors are mapped to the latent space, and a comprehensive observation operator that includes the current state and environmental uncertainties is constructed. The time-domain motion trend features, environmental observation operators, and actuator action information from the previous moment are concatenated and spliced ​​to construct a high-dimensional continuous state feature vector that satisfies the Markov property.

5. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the intelligent strategy reasoning module, the step of using an Actor network to calculate the attitude deviation compensation vector for the current water flow impact in real time includes: A deep neural network architecture containing residual connection modules is constructed, and the input state vector is subjected to multi-layer abstract mapping through nonlinear activation functions to extract key obstacle avoidance and stability features. The candidate action space is probabilistically modeled using a Gaussian policy distribution layer, and the mean and log standard deviation of the expected thrust for six-degree-of-freedom attitude correction are output. A residual control mechanism is introduced, which linearly superimposes the incremental compensation value output by the neural network with the traditional proportional-derivative control command to generate the final six-dimensional attitude deviation compensation vector.

6. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the thrust optimization allocation module, the step of combining the robot dynamics model to perform optimization solution includes: A six-degree-of-freedom dynamic analytical equation for a robot, including the added mass effect and nonlinear hydrodynamic damping term, is established to accurately describe the dynamic response characteristics of the robot at different motion speeds. Construct a convex optimization mathematical model with the objective function of minimizing the output power of each thruster and with hard constraints of the upper limit of physical thrust and the rate of change of thrust; The convex optimization mathematical model is solved online using a sequential quadratic programming algorithm, which accurately decouples the global compensation requirements and allocates them to the target power output shaft of each actuator.

7. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the power drive execution module, the step of dynamically adaptively correcting the offset path or deflection attitude includes: The allocated thrust commands are converted into pulse width modulation signals with dead zone compensation characteristics, which drive the H-bridge circuit of the thruster motor controller to perform duty cycle control at a preset frequency. The Hall encoder is used to provide real-time feedback on the actual rotational speed of the thruster, and a sliding mode variable structure control algorithm is introduced to quickly suppress load disturbances of the thruster, ensuring that the output thrust is consistent with the command requirements. The system monitors the convergence trajectory of deviations in various dimensions of the aircraft in real time. When the attitude overshoot exceeds the preset safety threshold, the dynamic limiting operator is triggered to perform secondary limiting protection on the thruster power.

8. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the self-calibration control module, the step of evaluating the value of the current control strategy through the Critic network includes: Construct an evaluation model based on a value function and use a fully connected neural network to fit and evaluate the potential long-term return value of the current state vector; The residual between the actual benefit and the expected value after the current action is calculated using the temporal difference error algorithm, thereby quantifying the quality of the Actor network's output strategy. By combining the generalized dominance estimation method to smooth the multi-step sampling data, the dominance signal reflecting the magnitude of the improvement in control stability is extracted and used as the weight basis for the gradient update of the neural network.

9. The underwater robot adaptive stabilization and control system based on multi-sensor fusion according to claim 1, characterized in that, In the self-calibration control module, the steps for completing the online autonomous calibration of the control law include: The design incorporates a multi-objective composite reward function that considers attitude following accuracy, energy consumption, and control command smoothness, guiding the model towards high accuracy and low power consumption. The loss function is optimized using a near-end policy with a pruning operator. Gradient updates are performed on the weight parameters of the Actor network and the Critic network while ensuring that the policy update magnitude is within the trust region. The convergence curve of the reward function is monitored in real time. When the total deviation of the system within a preset number of iterations is lower than the minimum threshold, the current network weight parameters are automatically fixed to achieve self-optimization of the control strategy for a specific flow field environment.

Images