Fault-tolerant control method, program and device for underwater robot based on physical information residual network and dynamic control allocation, and storage medium
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN ENGINEERING UNIVERSITY SANYA NANHAI INNOVATION & DEVELOPMENT BASE
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-03
Smart Images
Figure CN122086106B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of underwater robot control methods, specifically to an underwater robot fault-tolerant control method, program, device, and storage medium based on physical information residual networks and dynamic control allocation. Background Technology
[0002] As marine development tasks become more long-term and complex, underwater robots (AUVs / ROVs) often need to operate continuously for extended periods in deep-sea environments with harsh sea conditions and numerous obstacles. In practical applications, the robot's propulsion system is the component with the highest failure rate, easily leading to reduced thrust efficiency or even complete failure due to seawater corrosion, marine organism attachment, entanglement with seaweed, or aging circuitry.
[0003] Currently, the underlying control schemes for underwater robots typically employ either a "PID control + fixed control allocation" approach or a "traditional adaptive control" approach. PID control with fixed control allocation relies heavily on an accurate linearized model and a pre-defined thrust allocation matrix. This matrix assumes all thrusters are healthy and their thrust coefficients are constant. If a thruster fails (e.g., thruster #3's thrust becomes zero), the torque calculated by the controller cannot be correctly executed, resulting in a huge yaw torque that causes the robot to become instantly unstable, rotate violently, or even crash. While traditional adaptive control schemes utilize adaptive laws to estimate unknown parameters, they typically assume the model structure is known and the fault is linear. However, the hydrodynamic coefficients (added mass, damping) of underwater robots vary nonlinearly with the flow field, and faults are often accompanied by strong unstructured disturbances. Traditional parameter adaptive control has slow convergence speed and is prone to "parameter drift," making it difficult to recover attitude within seconds of a sudden fault. Existing technologies face the challenges of "lagging fault perception" and "rigid control reconfiguration": simple feedback control can only passively resist errors and cannot actively identify "who is broken and how much is broken"; while fixed control allocation logic lacks redundancy adjustment capabilities and cannot tap the potential of remaining healthy thrusters to maintain the control objective when some actuators fail. Therefore, there is an urgent need for an intelligent fault-tolerant control scheme that can integrate physical mechanisms and data-driven approaches to quickly identify fault models online and reconfigure thrust allocation in real time. Summary of the Invention
[0004] The purpose of this invention is to increase the underwater robot's ability to perceive faults by constructing a coupled architecture of Physical Information Residual Network (PINN-Residual) and Dynamic Control Allocation (DCA) without relying on external fault detection sensors, thereby achieving "soft perception" of the thruster's health status and "soft reconstruction" of the control law.
[0005] This invention provides a fault-tolerant control method for underwater robots based on physical information residual networks and dynamic control allocation, comprising the following steps:
[0006] Based on the dynamic equations, a generalized dynamic model of the underwater robot under fault conditions is constructed by introducing the thruster health diagonal matrix and the external lumped disturbance term.
[0007] Construct a physical information neural network; the physical information neural network adopts a fully connected neural network, and its input is the current state. The output is an estimate of the total system disturbance. ;
[0008] Based on the design of a composite control law using a non-singular terminal sliding surface, the virtual total control force required to resist faults is calculated. The control law includes an equivalent control term and a robust switching term, and the total system disturbance output by the physical information neural network is introduced.
[0009] The control allocation is transformed into a constrained quadratic programming optimization problem with full-degree-of-freedom control. A thruster health-aware weighting mechanism and slack variables are introduced into the optimization objective function, and the thrust of each thruster is redistributed through the optimization algorithm.
[0010] When the underwater robot's propulsion system malfunctions, a graded degradation strategy is determined based on the slack variables in the optimization objective function. The graded degradation strategy is as follows: if the L2 norm of the slack variable is less than or equal to a set threshold, the full degree of freedom control is maintained to allocate thrust to each thruster; if the L2 norm of the slack variable is greater than the set threshold, non-critical degrees of freedom are abandoned, depth and heading degree of freedom control is maintained, and the thrust of each thruster is reassigned.
[0011] Furthermore, the dynamic equations of the generalized dynamic model of the underwater robot under fault conditions are as follows:
[0012]
[0013] in, The inertia matrix includes the added mass; The Coriolis force matrix; The nominal damping matrix; This is the restoring force term; For the thruster configuration matrix; , For the first Health status of each thruster; For the input commands of each thruster; This is external lumped interference.
[0014] Furthermore, the total system disturbance is the sum of the fault residual and the disturbance, including thruster physical faults, external lumped disturbances, and model uncertainties;
[0015] .
[0016] Furthermore, the Newton-Euler equations are introduced into the physical information neural network, and the prior dynamics model of the underwater robot is embedded in it. The weight coefficients between the two terms are dynamically and adaptively adjusted by using a composite loss function of physical equation constraints and data equation constraints. By minimizing the composite loss function in backpropagation, the parameters of the neural network are forced to be updated adaptively.
[0017] The loss function of the physical information neural network is:
[0018]
[0019]
[0020] in, , and These are the velocity, acceleration, and pose data measured in real time by the sensors, respectively. This is the output vector of the physical information residual network; An adaptive penalty coefficient; For the set of all trainable parameters of the physical information residual network The gradient operator.
[0021] Furthermore, a non-singular terminal sliding mode control strategy is adopted, and a non-singular terminal sliding mode surface containing the fractional power of the error is designed. ;
[0022]
[0023]
[0024] in, For tracking error; , , ; ; .
[0025] Furthermore, the control law for:
[0026]
[0027]
[0028]
[0029] in, and These are the positive definite linear feedback gain matrix and the positive definite robust switching gain matrix, respectively.
[0030] Furthermore, the thruster health perception weighting mechanism sets the thruster's energy weight as the reciprocal function of its health; the objective function of the quadratic programming optimization problem is:
[0031]
[0032]
[0033] The constraints are as follows:
[0034]
[0035]
[0036]
[0037] in, For thrust energy penalty weight matrix; This is the basic energy weight matrix under the nominal state of the system; The weight matrix is the penalty matrix for slack variables; These are slack variables; The estimated value of the thruster performance matrix is derived from the PINN residuals. Physical thrust saturation constraint for the thruster; For thrust rate of change constraint;
[0038] The quadratic programming optimization problem is solved using the effective set method or the original-dual interior point method.
[0039] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation described above.
[0040] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation described above.
[0041] The present invention also provides a computer program product, including a computer program / instruction which, when executed by a processor, implements the steps of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation described above.
[0042] The beneficial effects of this invention are as follows:
[0043] (1) The method of this invention introduces an adaptive gradient balancing mechanism. In the transient state of a fault, the algorithm automatically increases the gradient contribution of the physical constraint weights and uses the Newton-Euler equation to "forcefully correct" the network output, achieving millisecond-level fast convergence. In the steady state, the data loss weights are automatically increased, and the model is fine-tuned using sensor data. This dynamic balancing mechanism enables the system to maintain optimal recognition performance in both the initial stage of strong noise interference and the final stage of high-precision tracking.
[0044] (2) The method of this invention constructs a variable gain control strategy based on residual confidence. The controller not only uses the output of PINN for feedforward cancellation, but also uses the real-time convergence loss of PINN to dynamically adjust the sliding mode gain. When the network identification is accurate, the gain automatically decreases, and the output is as smooth as silk; when the network is mainly learning (the identification error is large), the gain automatically increases to ensure safety. This deep linkage of "perception-control" eliminates unnecessary chattering at the mathematical level and significantly extends the physical life of the actuator.
[0045] (3) The method of this invention introduces a health-aware dynamic allocation mechanism, which explicitly maps the health status of the thrusters to the weights of the energy cost function. Once a decrease in efficiency of a thruster is detected (such as slight entanglement), the algorithm will immediately increase its usage cost exponentially, actively reducing the load on that thruster at the optimization level and smoothly transferring its working pressure to other healthy thrusters. This "soft protection" mechanism avoids the cascading deterioration of faults and greatly improves the robot's continuous operation capability in deep-sea isolated environments. Attached Figure Description
[0046] Figure 1 This is an overall flowchart of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation of the present invention;
[0047] Figure 2 This is a flowchart of the PINN physical information residual identification part of the method of the present invention;
[0048] Figure 3 This is a flowchart of the non-singular terminal sliding mode (NTSMC) controller part of the method of the present invention;
[0049] Figure 4 This is a flowchart of the dynamic control allocation (DCA) optimization part of the method of the present invention;
[0050] Figure 5 This is a flowchart of the graded degradation fault tolerance strategy part of the method of the present invention. Detailed Implementation
[0051] The present invention will be further described below with reference to the accompanying drawings. The embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.
[0052] This invention discloses a fault-tolerant control system for underwater robots based on physical information residual networks and dynamic control allocation. The system employs a four-level closed-loop architecture: state observation, residual identification, robust control, and optimized allocation. The state observation module monitors the motion state using sensor data; the residual identification module uses a PINN network to fit model uncertainties and fault residuals in real time; the robust control module calculates the required virtual total torque; and the optimized allocation module solves a constrained quadratic programming problem to allocate the actual thrust.
[0053] A fault-tolerant control method for underwater robots based on physical information residual networks and dynamic control allocation, combining... Figures 2 to 5 As shown, it includes the following steps:
[0054] Step S1: System initialization and nominal dynamic modeling to establish a mathematical baseline model for the underwater robot. After system startup, based on rigid body kinematics and fluid dynamics theory, a nominal dynamic equation including an additional mass matrix, Coriolis force matrix, and nonlinear damping terms is constructed. On this basis, a diagonal matrix representing the actuator's health (characterizing actuator performance) and a lumped disturbance term (characterizing external flow field disturbances and model parameter drift) are introduced to establish a generalized dynamic model describing fault conditions. This model clarifies the physical mapping relationship between control input and state response, providing the necessary physical equation constraint foundation for subsequent residual identification using neural networks.
[0055] Step S2: Online identification of dynamic residuals and faults based on PINN, utilizing the Physical Information Neural Network (PINN) to achieve "soft perception" of faults. A lightweight fully connected neural network is constructed to fit the dynamic residuals caused by thruster failure and environmental disturbances in real time. A composite loss function incorporating "physical constraint terms" is innovatively defined. In particular, an adaptive gradient balancing strategy is adopted to dynamically adjust the weight ratio of data loss and physical loss during backpropagation, ensuring that the gradient update rates of both remain on the same order of magnitude, preventing one term from dominating the training process; this forces the network output to not only fit sensor data but also satisfy the Newton-Euler dynamic equations. This mechanism allows PINN to converge quickly and accurately estimate the current lumped fault magnitude with only a small number of online samples, without requiring a large amount of offline training data, thus solving the problem of slow convergence in traditional adaptive methods.
[0056] Step S3: Generate the Non-Singular Terminal Sliding Mode (NTSMC) composite control law and calculate the virtual total control force required to resist faults. A non-singular terminal sliding mode control strategy is adopted, designing a sliding surface containing fractional powers of the error to ensure that the system state error converges to zero within a finite time and avoids singularity issues. The core operation is to use the fault residual identified by PINN in Step S2 as a feedforward compensation term, directly subtracting it from the total control law. Through this feedforward cancellation, the controller no longer requires a very large robust switching gain to maintain stability, thus significantly suppressing the inherent high-frequency chattering phenomenon of sliding mode control and avoiding secondary mechanical damage to the remaining healthy thrusters. When PINN converges well, the switching gain is automatically reduced to weaken chattering; when the identification error is large, the robust gain is enhanced to ensure stability. This deep "sensor-control" coupling mechanism maximizes the smoothness of the control output while ensuring finite-time convergence.
[0057] To address the challenges of high sensor noise and high reliability of physical constraints in the initial stage of a sudden failure, while requiring precise data fitting in the steady-state phase, the system dynamically adjusts the weights of physical constraints by real-time monitoring of the gradient magnitudes of the data loss term and the physical loss term. This mechanism grants the physical equations greater "guiding power" during the transient phase of a fault, forcing network weights to decrease rapidly along the dynamic manifold, achieving millisecond-level convergence. In the steady-state phase, it automatically balances the weights and uses data to fine-tune model parameter deviations. This completely solves the problem of traditional fixed-weight PINNs easily diverging or overfitting in noisy environments.
[0058] Step S4: Considering the dynamic control allocation optimization for actuator saturation, the control allocation is modeled as a constrained quadratic programming (QP) optimization problem, with the goal of minimizing energy consumption while meeting the virtual total torque requirement. In the optimization objective function, a health-aware weighting mechanism is introduced, setting the thruster's energy weight to its health level. The inverse function of the thruster. Thus, the lower the thruster's health, the higher the cost of allocating thrust, thereby guiding the system to prioritize the use of healthy thrusters. Physical saturation constraints (maximum / minimum thrust) and rate of change constraints for the thrusters are explicitly introduced into the optimization model. Compared to the relay-style 'hard cutoff' adopted by traditional methods after a thruster's complete failure, the 'soft switching' mechanism of this invention achieves stepless continuous control reconfiguration at the mathematical level, effectively avoiding severe attitude twitching of the fuselage during fault-tolerant switching, and realizing a protection upgrade from 'passive power-off' to 'active life extension';
[0059] The health matrix identified by PINN Energy cost function of explicitly incorporating quadratic programming (QP) Unlike traditional methods that treat faults solely as hard constraints, this invention constructs a dynamic weight matrix that is inversely correlated with health status. This means that when a thruster shows early signs of failure ( When the load on a thruster decreases (i.e., its operating cost is automatically increased by the algorithm), prompting the optimizer to proactively reduce the load on that thruster before it reaches physical saturation, smoothly transferring its workload to other healthy thrusters. This "soft protection" mechanism represents a technological leap from "passive fault tolerance" to "active life extension."
[0060] Step S5: Execute a graded degradation strategy based on fault severity. The system monitors and controls the slack variables in real time during the allocation optimization process to quantify whether the current fault exceeds the system's physical redundancy. If the slack variables exceed the safety threshold (meaning that the remaining thrust is no longer sufficient to maintain full-degree-of-freedom control), the system immediately triggers a degradation strategy: actively relinquishing control of non-critical degrees of freedom such as lateral movement and roll, locking the current state, and concentrating limited energy on ensuring core survival functions such as depth maintenance and heading control, ensuring the robot can safely hover or surface in an emergency, and preventing system divergence caused by forcibly correcting uncontrollable axes.
[0061] In some implementations, step S1 involves establishing a nominal and fault dynamics model of the underwater robot; specifically, after system startup, a nominal model is constructed based on rigid body dynamics and fluid dynamics theories.
[0062]
[0063] in, The inertia matrix includes the added mass; The Coriolis force matrix; The nominal damping matrix; This is the restoring force term; This represents the desired control force / torque vector.
[0064] Considering model mismatch (such as mass changes caused by deposits) and thruster failures that may occur in actual operation, a failure factor matrix is introduced:
[0065]
[0066] in, For the first Health status of each thruster;
[0067] The dynamic equations of the actual system are corrected as follows:
[0068]
[0069] in, For the thruster configuration matrix; For the input commands of each thruster; This is to address external lumped interference. The core task of the method in this invention is to... and Solve the unknown Make the actual state Approaching, tracking, and converging to the desired state .
[0070] In some embodiments, in step S2, a lightweight fully connected neural network is constructed. To approximate the sum of unknown fault residuals and disturbances:
[0071]
[0072] It is important to emphasize that, in the complex marine environment, propeller physical failures... External flow field interference and model uncertainty They often exhibit strong coupling characteristics, making it difficult to decouple and measure them using conventional sensors.
[0073] This invention overcomes the limitations of traditional analytical methods that attempt to independently separate various disturbances, and unifies the above three highly nonlinear and time-varying uncertainties into a lumped residual model. .
[0074] In the specific implementation of the control system, the system does not need to (and cannot) solve for analytical terms separately. Instead, it uses the aforementioned Physical Information Residual Network (PINN) to directly approximate the sum of the lumped residuals online in real time by utilizing the overall system dynamic response deviation fed back by airborne sensors, and outputs the residual estimate. (i.e., network output) This data is then directly fed into the subsequent non-singular terminal sliding mode controller for global compensation. This 'lump estimation, unified compensation' strategy greatly enhances the system's robustness under extremely harsh sea conditions.
[0075] Define the physical constraint loss function Instead of relying solely on prediction errors, it directly incorporates dynamic residuals as part of the loss:
[0076]
[0077] in, This is the output vector of the Physical Information Residual Network (PINN). This is the set of all training parameters in the neural network (including the weight matrix and bias vector of each hidden layer).
[0078] By minimizing this loss, the neural network N not only fits the data but also satisfies the Newton-Euler equations. This physical constraint allows PINN to converge quickly with small samples or even single-step iterations, providing real-time, accurate estimates of the system's "sub-healthy state." .
[0079] In some embodiments, in step S3, the robot's three-dimensional position coordinates and the attitude angles of roll, pitch, yaw and desired attitude angle in the Earth coordinate system are defined as follows:
[0080] ,
[0081] ,
[0082] but , ;
[0083] Based on the above position coordinates, the position tracking error is calculated as follows:
[0084]
[0085] Constructing non-singular terminal sliding surfaces :
[0086]
[0087] in, ; The advantage of non-singular terminal sliding mode controllers (NTSMCs) is that they can converge in a finite time and have no singularities.
[0088] Design control law It includes equivalent control terms and robust switching terms, and introduces residual compensation for the PINN output:
[0089]
[0090]
[0091]
[0092] By introducing feedforward compensation The controller does not require a large switching gain. This can offset the effects of the fault, thereby significantly reducing the jitter of the control signal and protecting the remaining healthy thrusters.
[0093] After calculating the desired total control force, the system further enters the dynamic control allocation stage. This is the key to achieving fault tolerance, namely, "how to complete the action using the remaining healthy limbs".
[0094] In some embodiments, in step S4, to achieve adaptive fault tolerance and active life extension for sub-healthy thrusters, a soft switching mechanism based on health awareness is introduced into the objective function. The system receives the thruster health matrix identified in real time by PINN. Based on this, a dynamic penalty weight matrix is constructed. :
[0095]
[0096] In the formula, This represents the base energy weight matrix under the system's nominal state. Therefore, it can be seen that when a thruster experiences early wear or entanglement, affecting its health... As it gradually decreases, its dynamic penalty weight in the control allocation... It will rise continuously and smoothly in a reciprocal relationship.
[0097] Subsequently, the dynamic weight matrix Substituting this into the quadratic programming cost function of Dynamic Control Allocation (DCA):
[0098]
[0099] Driven by the aforementioned dynamic weights, the optimization solver minimizes the global cost. During the process, the thrust commands allocated to the sub-health booster will be automatically and continuously reduced. And smoothly reconstruct and transfer the missing control torque to other healthy thrusters.
[0100] Its constraints include:
[0101] Dynamic constraints (introducing slack variables):
[0102]
[0103] Physical thrust saturation constraint of the thruster (i.e., the maximum forward and reverse thrust limit that the motor can generate):
[0104]
[0105] Thrust rate of change constraint (limits the step amplitude of commands between adjacent control cycles to prevent current overload and protect motor hardware):
[0106]
[0107] in, This is the estimated value of the thruster performance matrix derived from the PINN residuals.
[0108] Furthermore, the effective set method or the primal-dual interior-point method are used to solve this QP problem online. When a certain thruster fails ( When this happens, the optimizer will automatically reduce the output weight of that thruster to zero and reallocate the thrust vector synthesis of the other thrusters. This enables seamless "soft switching".
[0109] Example 1
[0110] This embodiment first constructs a generalized dynamic model of an underwater robot (ROV) under fault conditions. In specific implementation, the robot's state vector is defined based on the principles of rigid body dynamics and fluid dynamics. Considering that the thruster may experience a decrease in thrust efficiency (such as entanglement) or complete failure in actual operation, a health diagonal matrix is introduced.
[0111] The revised dynamic equations are expressed as follows:
[0112]
[0113]
[0114]
[0115] in, The inertia matrix includes the added mass; The Coriolis force matrix; The nominal damping matrix; This is the restoring force term; The problem is a thruster malfunction. Due to unknown external environmental disturbances; For the thruster configuration matrix; , For the first Health status of each thruster; For the input commands of each thruster; This is to mitigate external lumped interference. The core objective of this system is to... and Achieve stable attitude control under conditions where everything is unknown.
[0116] Reference Figure 2 The flowchart shown below illustrates the PINN physical information residual identification process, which constructs a fully connected neural network. Its input is the current state. The output is an estimate of the total system disturbance (including fault residuals and environmental disturbances). The system does not rely on external sensors to detect faults, but instead "learns" faults through a neural network. To address the problems of traditional networks requiring massive amounts of data and having slow convergence, this invention includes a loss function with physical constraints in the k-th training iteration. :
[0117]
[0118]
[0119]
[0120] in, Indicates the first The physical mechanism constraint loss under the next iteration. This term directly embeds the prior dynamic model of the underwater robot (Newton-Euler equations) into the training space of the neural network as a strong constraint regularization term. Indicates the first In each iteration, the deviation loss between the neural network's output value and the system's "observation value" represents the degree to which the algorithm fits the real-time sensor data, ensuring that the neural network's identification results do not deviate from the actual observed robot motion state. To control the nominal thrust torque calculated by the distribution module; , and These are the velocity, acceleration, and pose data measured in real time by the sensors, respectively. An update law based on the gradient norm-adaptive penalty coefficient is adopted:
[0121]
[0122] That is, by calculating the maximum gradient of the data loss with respect to the network parameters and dividing it by the average gradient of the physical loss, the weights of the physical constraints are dynamically updated in each k-th iteration. This ensures that the physical mechanism term can provide a sufficiently strong correction gradient in the event of a sudden failure.
[0123] By minimizing physical loss during backpropagation, the weight updates of the neural network are forced to satisfy the Newton-Euler dynamics. This enables PINN to converge rapidly within a few control cycles (milliseconds) after a fault occurs, outputting accurate residual estimates.
[0124] Reference Figure 3 The flowchart of the Non-Singular Terminal Sliding Mode (NTSMC) controller shown illustrates how, based on accurate fault residual estimates, the system designs the NTSMC to calculate the virtual total control force required to resist faults.
[0125] To ensure that the error converges within a finite time and avoid singularities, based on the position tracking error... Constructing non-singular terminal sliding surfaces :
[0126]
[0127] Design a composite control law The control law is composed of equivalent control terms. Robust switching item and feedforward compensation terms from PINN composition:
[0128]
[0129] In the formula, It is a positive definite linear feedback gain matrix, used to determine the speed at which the system state approaches the sliding surface; It is a positive definite robust switching gain matrix, used to overcome system uncertainties and ensure system stability.
[0130] Because of the introduction Actively mitigates fault effects, robust gain It can be set to a very small value, thereby effectively suppressing the high-frequency chattering of traditional sliding mode control and protecting the mechanical structure.
[0131] Reference Figure 4 The flowchart shown illustrates the Dynamic Control Allocation (DCA) optimization process. The system addresses the problem of "how to use the remaining thrusters to generate total torque"; the control allocation is modeled as a quadratic programming (QP) optimization problem with slack variables. The objective function is... Designed to balance control error and energy consumption:
[0132]
[0133] The constraints explicitly include the physical limitations of the thruster:
[0134]
[0135]
[0136]
[0137] The optimization solver searches for the optimal solution online. When a thruster fails, the system automatically redistributes the thrust weights, achieving "soft reconfiguration".
[0138] Finally, combining Figure 5 The diagram shown illustrates a tiered degradation fault tolerance strategy, serving as the system's safety baseline.
[0139] In practice, slack variables are monitored in real time during the optimization process. .
[0140] Decision logic: If This indicates that the fault is within a controllable redundancy range, maintaining full-degree-of-freedom control.
[0141] Degradation logic: If This indicates that the failure has exceeded physical limits (e.g., complete destruction of the lateral thrust). At this point, a degradation strategy is triggered, forcibly locking non-critical degrees of freedom (e.g., lateral movement). and roll ):
[0142]
[0143] The system concentrates its limited power on ensuring depth control (z) and heading control. The system automatically triggers the "degree of freedom degradation" logic, actively abandoning non-critical degrees of freedom (such as lateral movement and roll), locking the control target of the relevant axis to the current value, concentrating energy to ensure depth maintenance and heading control, ensuring that the robot can safely surface or return to shore, preventing divergence caused by forcibly correcting uncontrollable degrees of freedom, and ensuring that the robot can safely surface and be recovered.
[0144] In particular, in some preferred embodiments of the present invention, a computer device is also provided, including a memory and a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation described in any of the above embodiments.
[0145] In some other preferred embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program / instruction is stored, wherein when the computer program is executed by a processor, the steps of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation described in any of the above embodiments are implemented.
[0146] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above embodiments of the underwater robot fault-tolerant control method based on physical information residual network and dynamic control allocation, which will not be repeated here.
[0147] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.
[0148] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0149] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
Claims
1. A fault-tolerant control method for underwater robots based on physical information residual networks and dynamic control allocation, characterized in that, Includes the following steps: Based on the dynamic equations, a generalized dynamic model of the underwater robot under fault conditions is constructed by introducing the thruster health diagonal matrix and the external lumped disturbance term. Construct a physical information neural network; the physical information neural network adopts a fully connected neural network, and its input is the current state. The output is an estimate of the total system disturbance. ; The total system disturbance is the sum of the fault residual and the disturbance, including thruster physical faults, external lumped disturbances, and model uncertainties; in, For the thruster configuration matrix; , For the first The health status of each thruster The value can be 1 to n; It is a unit matrix; For the input commands of each thruster; External lumped interference; This is due to the uncertainty of the model; Based on the design of a composite control law using a non-singular terminal sliding surface, the virtual total control force required to resist faults is calculated. The control law includes an equivalent control term and a robust switching term, and the total system disturbance output by the physical information neural network is introduced. The control allocation is transformed into a constrained quadratic programming optimization problem. A thruster health-aware weighting mechanism and slack variables are introduced into the objective function. The thrust of each thruster is then redistributed through an optimization algorithm. The thruster health-aware weighting mechanism sets the energy weight of each thruster as the reciprocal of its health. The objective function of the quadratic programming optimization problem is: The constraints are as follows: in, For thrust energy penalty weight matrix; This is the basic energy weight matrix under the system's nominal state; The weight matrix is the penalty matrix for slack variables; These are slack variables; The estimated value of the thruster performance matrix is derived from the residuals of the physical information neural network. Physical thrust saturation constraint for the thruster; For thrust rate of change constraint; For control laws; When the underwater robot's propulsion system malfunctions, a graded degradation strategy is determined based on the slack variables in the optimization objective function. The graded degradation strategy is as follows: if the L2 norm of the slack variable is less than or equal to a set threshold, full degree of freedom control is maintained; if the L2 norm of the slack variable is greater than the set threshold, non-critical degrees of freedom are abandoned, depth and heading degree of freedom control is maintained, and the thrust of each thruster is redistributed.
2. The fault-tolerant control method for underwater robots according to claim 1, characterized in that, The dynamic equations of the generalized dynamic model of the underwater robot under fault conditions are as follows: in, The inertia matrix includes the added mass; The Coriolis force matrix; The nominal damping matrix; This is the restoring force term.
3. The fault-tolerant control method for underwater robots according to claim 2, characterized in that, The physical information neural network incorporates the Newton-Euler equations and embeds them into the prior dynamics model of the underwater robot. It utilizes a composite loss function of physical equation constraints and data equation constraints to dynamically and adaptively adjust the weight coefficients between the two terms. By minimizing the composite loss function during backpropagation, the parameters of the neural network are forced to be updated adaptively. The loss function of the physical information neural network is: in, , and These are the velocity, acceleration, and pose data measured in real time by the sensors, respectively. This is the output vector of the physical information residual network; An adaptive penalty coefficient; For the set of all trainable parameters of the physical information residual network The gradient operator.
4. The fault-tolerant control method for underwater robots according to claim 3, characterized in that, A non-singular terminal sliding mode control strategy is adopted, and a non-singular terminal sliding mode surface containing the fractional power of the error is designed. ; in, For tracking error; , , ; ; .
5. The fault-tolerant control method for underwater robots according to claim 4, characterized in that, The control law for: in, and These are the positive definite linear feedback gain matrix and the positive definite robust switching gain matrix, respectively.
6. The fault-tolerant control method for underwater robots according to claim 1, characterized in that, The quadratic programming optimization problem is solved using the effective set method or the original-dual interior point method.
7. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 6.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.
9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.