An intelligent monitoring and control system for the stable operation of unmanned ships

Abstract

This invention discloses an intelligent monitoring and control system for the stable operation of unmanned vessels, comprising a heterogeneous observer group, a dynamic weight allocation module, and an active shift controller. The heterogeneous observer group includes a finite-time extended state observer, an interval type-2 fuzzy adaptive observer, and a data-driven trend predictor, which are used to estimate high-frequency disturbances and abrupt signals, estimate long-term uncertainty drift and model mismatch errors, and output future operating condition prediction results, respectively. The dynamic weight allocation module calculates the estimation errors and their rates of change of each observer in real time based on the Lyapunov energy function, dynamically determines the weight coefficients, and outputs a weighted fusion state estimate. Through an interpolation function, the control law is smoothly transitioned from a topology adapted to the current operating condition to a topology adapted to the predicted operating condition. This invention changes the control mode from lag compensation to proactive pre-adaptation, improving the stable operation capability and reliability of unmanned vessels in complex marine environments.

Description

Technical Field

[0001] This invention relates to the field of unmanned vessel monitoring technology, specifically to an intelligent monitoring and control system for the stable operation of unmanned vessels. Background Technology

[0002] When unmanned vessels operate in complex marine environments, they are affected by various uncertain disturbances such as wind, waves, and currents. Their motion states exhibit strong nonlinearity, large inertia, and strong coupling characteristics. At the same time, during long-duration autonomous navigation missions, the hull dynamic parameters may slowly drift due to factors such as load changes, biological attachment, and sea state evolution. The actuators may also experience performance degradation. These factors together lead to the dual challenges of model uncertainty and external disturbances in the stable control of unmanned vessels.

[0003] Existing unmanned vessel stability monitoring and control systems mainly adopt a control architecture based on disturbance observers. They estimate the total system disturbance by expanding state observers or disturbance observers and perform feedforward compensation in the control loop. Some schemes adopt a parallel structure of multi-modal observers, select the observer output with the smallest estimation error through a selector, or switch control laws to switch between different operating conditions. At the hardware level, existing systems generally adopt an Internet of Things architecture to collect multi-source sensor data and realize remote monitoring through a cloud platform.

[0004] However, existing technologies still have the following shortcomings: First, the operating mode of the observer and controller is "post-compensation," meaning that the observer only converges and outputs an estimated value after a disturbance occurs and affects the system state, and the controller compensates based on this. This lag makes it difficult to guarantee control quality when encountering sudden changes in sea state or sudden failures. Second, multi-observer fusion schemes usually adopt fixed switching logic or hard switching based on the current error magnitude. The output fusion between observers lacks stability theory support, and the switching process is prone to introducing chattering. Third, existing systems lack effective identification and fault-tolerance mechanisms for atypical anomalies such as actuator performance degradation and sensor common mode failures. The observer fusion architecture does not form an information closed loop with actuator load distribution, making it difficult to distinguish between control distribution changes and external disturbances. Fourth, the system lacks proactive self-checking capabilities. Model parameter drift is difficult to calibrate in a timely manner during long-term operation, and maintenance is only passively triggered after a failure occurs or performance deteriorates significantly. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent monitoring and control system for the stable operation of unmanned vessels, so as to solve the problems mentioned in the background art.

[0006] An intelligent monitoring and control system for the stable operation of unmanned vessels includes a heterogeneous observer group comprising at least three types of observers based on different mathematical principles: the first type is a finite-time extended state observer configured to estimate high-frequency disturbances and abrupt changes on a millisecond-level time scale; the second type is an interval type-II fuzzy adaptive observer configured to estimate long-term uncertainty drift and model mismatch error of the system; and the third type is a data-driven trend predictor based on a time-series neural network or Gaussian process regression, which outputs predictions of future operating conditions based on historical navigation parameters, environmental load data, and actuator status. The dynamic weight allocation module has its input connected to the output of each observer in the heterogeneous observer group. This module calculates the estimation error and its rate of change of each observer in real time based on the Lyapunov energy function, dynamically determines the weight coefficient of each observer's output in the final state estimation, and outputs the weighted and fused state estimation value as the system observation result. An active shift controller, whose input is connected to the dynamic weight allocation module and the trend predictor, when the operating condition prediction result output by the trend predictor indicates that the system will enter a new operating condition, smoothly changes the structure or parameters of the control law through an interpolation function before the actual state of the system changes across a threshold, so that the control law transitions from a topology adapted to the current operating condition to a topology adapted to the predicted operating condition.

[0007] Furthermore, when calculating the weight coefficients, the dynamic weight allocation module increases the weight coefficient of any observer to a preset upper limit value within a single control cycle when the rate of change of the estimation error of any observer exceeds a preset threshold, while reducing the weight coefficients of the other observers. The threshold adjustment mechanism based on the rate of change of the estimation error and the continuous adjustment mechanism based on the Lyapunov energy function operate in parallel. When there is a difference between the weight coefficients output by the two, the one that makes the observer's weight coefficient larger is taken as the final weight to prioritize the response to abrupt signals.

[0008] Furthermore, it also includes an observer health status monitoring module, which receives the internal state variables and estimated residual sequences of the finite-time extended state observer, the interval type II fuzzy adaptive observer, and the data-driven trend predictor, respectively. Based on the sliding window statistics of the residual mean, variance, and residual autocorrelation of each observer, it independently generates a health score for each observer. When calculating the weight coefficients, the dynamic weight allocation module introduces the health scores of each observer as product factors into the weight allocation function. When the health score of any observer is lower than a preset threshold, the weight coefficient of that observer is forcibly set to zero, and the system enters a reduced-order observation mode. It also includes an observer consistency constraint module, which calculates in real time the difference between the output of the finite-time extended state observer and the output of the interval type II fuzzy adaptive observer, as well as the deviation between the two and the predicted value output by the trend predictor. When the statistical characteristics of the difference exceed the normal range and the health scores of each observer are higher than a preset threshold, it is determined that there is a common mode fault in the system. At this time, the observer consistency constraint module outputs a weight reduction instruction to the dynamic weight allocation module, forcibly reducing the weight coefficients of all observers and increasing the proportion of multi-source sensor fusion reference values ​​in the final control. When the observer health status monitoring module and the observer consistency constraint module issue instructions at the same time, the weight reduction instruction of the observer health status monitoring module takes priority, first isolating the faulty observer, and then performing consistency verification on the remaining observers.

[0009] Furthermore, the active shift controller includes at least two sets of control laws with different topologies. One set of control laws is configured as a proportional-differential structure with heading maintenance as the main function, and the other set of control laws is configured as a feedforward compensation structure with surge resistance as the main function. When the operating condition prediction result indicates that the system will enter a high sea state or a high wave navigation state, the active shift controller will smoothly transition the current control law to the surge resistance feedforward compensation structure in advance. Before being input to the ship's actuators, the control commands output by the active shift controller pass sequentially through the actuator load allocation module and the nonlinear command shaper. The actuator load allocation module receives the total control demand output by the active shift controller and, in conjunction with the future operating condition predictions output by the trend predictor, dynamically allocates the load among multiple actuators using a model predictive control framework. The allocation objective incorporates the cumulative wear, heat dissipation status, and expected margins of each actuator under future operating conditions into the cost function. This module simultaneously feeds back the allocated load commands to the dynamic weight allocation module, enabling dynamic weight allocation. When evaluating observer error, the module distinguishes between dynamic changes caused by load redistribution and dynamic changes caused by external disturbances or model mismatch. The nonlinear command shaper receives the allocated command output by the actuator load distribution module, calculates the feasible range of command changes in real time based on the current physical constraints of each actuator, including the maximum angular velocity of the servo motor and the propeller speed change rate limit, performs constraint shaping on the command increment that exceeds the range, and outputs the shaped command to the actuator. At the same time, the shaper feeds back the command limit caused by the constraint to the dynamic weight distribution module as a correction item for observer error evaluation.

[0010] Furthermore, the future operating condition prediction results of the trend predictor include at least operating condition type labels and corresponding confidence levels. The operating condition type labels are obtained by offline clustering of historical navigation data. The offline clustering is based on marine environmental parameters, ship load parameters, and actuator health parameters. The data-driven trend predictor internally constructs a working condition transition probability matrix, which is generated based on the statistical analysis of the transition frequency between different working conditions in historical navigation data. When outputting the future working condition prediction result, the trend predictor simultaneously outputs the transition probability from the current working condition to each possible target working condition. The active shift controller calculates the target weight of the pre-switching control law based on the transition probability and the preset risk preference coefficient, and gradually completes the transition of the control law in multiple control cycles using an exponential smoothing method. The trend predictor includes a probability prediction layer based on Monte Carlo dropout or a Bayesian neural network. This layer outputs the most likely target operating condition, the probability distribution of the time window in which the operating condition occurs, and the probability of at least one alternative operating condition. When the confidence level of the main predicted operating condition corresponding to the transition probability is higher than a preset threshold, the active shift controller gradually completes the control law transition using an exponential smoothing method. When the confidence level of the main predicted operating condition is lower than the preset threshold and the difference between the probability of the alternative operating condition and the probability of the main predicted operating condition is less than a preset difference, one or more alternative control law branches are activated with a gain lower than that of the main channel. The alternative control law branches correspond to the alternative operating conditions and are in a hot standby state. When the main predicted operating condition does not occur within the time window and the alternative operating condition occurs first, the system completes the switch from the main control law branch to the alternative control law branch within one control cycle. When the confidence level of the main predicted operating condition is lower than the preset threshold and the probability of the alternative operating condition is significantly lower than the probability of the main predicted operating condition, the active shift controller pauses the pre-switching action and maintains the current control law.

[0011] Furthermore, the finite-time extended state observer, the interval type II fuzzy adaptive observer, and the data-driven trend predictor operate in parallel, each with the same sampling time base frequency, and the calculation cycle of the dynamic weight allocation module is synchronized with this sampling time base frequency.

[0012] Furthermore, it also includes a multi-source sensor data consistency verification module. This module accesses the data streams from the ship's inertial navigation system, global navigation satellite system, radar, and visual perception system. It performs multi-source data fusion through an extended Kalman filter to generate redundant ship motion state reference values. These reference values ​​are compared with the weighted fusion state estimate output by the dynamic weight allocation module. When the deviation between the two exceeds a preset safety threshold, the system determines that the current weighted fusion state estimate is invalid and forcibly switches to a backup control loop based on the multi-source sensor fusion reference value until the heterogeneous observer group completes self-calibration and re-converges. It also includes a self-calibration trigger, which monitors the cumulative deviation between the ship's actual navigation trajectory and the desired trajectory, the frequency of actuator actions, and the parameter drift of the interval type II fuzzy adaptive observer. When the above monitoring indicators are all below the threshold within a set time window, the system is determined to be in a fully excited and steady-state operation state, triggering a background self-calibration process. The system then uses the currently collected input and output data to re-identify the baseline parameters of the ship dynamics model and updates the initial fuzzy rule base of the interval type II fuzzy adaptive observer and the offline clustering center of the trend predictor accordingly.

[0013] Furthermore, it also includes an online excitation and response evaluation module. When the system is in a steady state and the self-calibration trigger is not triggered, and the system is not in the response evaluation period after the probe signal injection, this module injects a probe signal with controlled amplitude and limited spectrum into the control command, with the constraint of not interfering with the main control objective. The probe signal is generated based on a pseudo-random binary sequence or a sinusoidal sweep frequency signal. The module also records the system's response to the probe signal, identifies and calculates the current closed-loop frequency characteristics of the system through online system identification, and compares them with the reference frequency characteristics. When the comparison deviation exceeds a preset threshold, it actively triggers the online correction of the observer parameters or controller parameters. During the probe signal injection period, the self-calibration trigger suspends the monitoring and judgment of the actuator action frequency and trajectory cumulative deviation, and resumes monitoring after the response evaluation period ends.

[0014] Furthermore, it also includes a shipborne real-time processing unit, which operates the heterogeneous observer group, dynamic weight allocation module and active shift controller. The control cycle is set to the millisecond level, and it is connected to the cloud training platform through the Internet of Things communication module. The model parameters of the trend predictor are updated offline by the cloud training platform and then sent to the shipborne real-time processing unit.

[0015] Furthermore, when the confidence level of the trend predictor output is lower than a preset threshold, the dynamic weight allocation module pauses the pre-switching action of the active shift controller and redistributes the weight coefficients to the finite-time extended state observer and the interval type II fuzzy adaptive observer. The parameter matrix of the Lyapunov energy function is adjusted online according to the current system operating point, so that when the system state deviates more from the equilibrium point, the weight coefficient of the finite-time extended state observer assigned by the dynamic weight allocation module increases monotonically.

[0016] By adopting the above technical solution, a complete predictive adaptive control closed loop is realized, which transforms the operation mode of unmanned ships in complex marine environments from the traditional "passive compensation after disturbance" to "active pre-adjustment before the evolution of operating conditions". The system uses three types of observers with different time scales to run in parallel, capturing high-frequency mutations at the millisecond level, parameter drifts at the long time scale, and the trend of operating condition evolution in the next few seconds to tens of seconds. Then, the three are integrated by a dynamic weight allocation module based on the Lyapunov energy function to form a system state estimate that has both theoretical stability support and can respond quickly to mutations.

[0017] Compared with the prior art, the beneficial effects of the present invention are: the intelligent monitoring and control system for stable operation of unmanned vessels, In the heterogeneous observer group, the finite-time extended state observer handles millisecond-level high-frequency abrupt changes, the interval type II fuzzy adaptive observer handles long-term slow-varying drift, and the trend predictor provides forward-looking information on future operating conditions. The three types of observers complement each other on the time scale. Combined with the continuous adjustment mechanism based on the Lyapunov energy function in the dynamic weight allocation module, the weight coefficients change continuously with the system state and have stability theory support, which solves the problem of chattering easily introduced by hard switching in traditional multi-observer schemes. At the same time, the threshold adjustment mechanism based on the rate of change of estimation error and the continuous adjustment mechanism run in parallel, and the larger of the two is taken as the final weight, so that the system can respond quickly in a single control cycle when faced with abrupt signals without sacrificing the smoothness in steady state.

[0018] At the control timing level, the active shift controller, based on the operating condition prediction results output by the trend predictor, smoothly changes the control law topology through an interpolation function before the actual system state changes beyond the threshold. This mechanism transforms the control action from a lagging mode of "compensating after the disturbance occurs" to a proactive mode of "pre-adjusting before the operating condition evolves". When encountering scenarios such as sudden strong winds or surge changes, the pre-switching of the control law is completed synchronously with the actual operating condition changes, avoiding the response delay and overshoot caused by the traditional switching controller waiting for the state to cross the threshold.

[0019] At the fault tolerance level, the observer health status monitoring module independently generates the health score of each observer through a sliding window statistical analysis of residual mean, variance, and autocorrelation, and introduces it as a product factor into the weight allocation function. When the performance of an observer degrades, its weight automatically decays, and in severe cases, it is forced to be set to zero. The system can still operate at a reduced order. The observer consistency constraint module calculates the difference between the outputs of the three types of observers. When the health scores of each observer are all at a high level but the outputs are inconsistent, it can identify common mode faults, forcibly reduce the weight of the observers, and increase the proportion of multi-source sensor fusion reference values. The priority setting of the two modules ensures that the faulty observer is isolated first, and then the consistency verification is performed on the remaining observers, forming a multi-level fault tolerance protection.

[0020] At the information closed-loop level between actuators and observers, the actuator load allocation module feeds back the allocated load command to the dynamic weight allocation module, enabling the system to distinguish between dynamic changes caused by load redistribution and dynamic changes caused by external disturbances or model mismatch. The nonlinear command shaper also feeds back the command limiting amount generated by physical constraints to the dynamic weight allocation module, avoiding the actuator saturation being misjudged as an external disturbance. These two feedback loops solve the problem of independent control allocation and state observation in multi-actuator systems, improving the observer's accuracy in understanding the true state of the system.

[0021] In terms of redundancy and safety, the multi-source sensor data consistency verification module fuses inertial navigation, satellite navigation, radar and visual perception data through an extended Kalman filter to generate a redundant motion state reference value independent of the observer group. When the weighted fused state estimate deviates from this reference value by more than a safety threshold, the system is forced to switch to the backup control loop until the heterogeneous observer group completes self-calibration and reconverges. This mechanism provides a safety net in the event of a complete failure of the observer group.

[0022] At the long-term operation and maintenance level, the self-calibration trigger monitors the cumulative deviation of the trajectory, the frequency of actuator actions, and the drift of the observer parameters. It only triggers background self-calibration when the system is fully excited and in steady-state operation, avoiding the problem of distortion of identification results caused by conventional timed calibration under insufficient excitation conditions. The online excitation and response evaluation module injects a probe signal with controlled amplitude when the system is in steady state and self-calibration is not triggered. By identifying and calculating the closed-loop frequency characteristics of the online system and comparing them with the benchmark, it can proactively detect parameter deviations and trigger corrections before the system performance has significantly degraded. The two modules complement each other, one using natural excitation for calibration and the other actively injecting excitation for verification, giving the system a self-maintenance capability that combines passive and active approaches. Attached Figure Description

[0023] Figure 1 This is a block diagram of the overall system structure of the present invention; Figure 2 This is a block diagram of the internal structure of the heterogeneous observer group and dynamic weight allocation module of the present invention; Figure 3 This is a block diagram of the processing chain of the active shift controller and actuator of the present invention; Figure 4 This is a block diagram of the trend predictor and probability prediction layer structure of the present invention; Figure 5 This is a flowchart illustrating the observer health monitoring and consistency constraint logic of the present invention. Figure 6 This is a flowchart of the multi-source sensor verification and backup control process of the present invention; Figure 7This is a flowchart illustrating the collaboration between the self-calibrating trigger and the online excitation module of the present invention. Figure 8 This is the overall control flowchart of the system of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see Figure 1-3 This invention provides a technical solution: an intelligent monitoring and control system for the stable operation of unmanned ships. At the physical level, this system includes an onboard real-time processing unit. This unit operates a heterogeneous observer group, a dynamic weight allocation module, and an active shift controller. The control cycle is set to milliseconds. The onboard real-time processing unit is connected to a cloud-based training platform via an IoT communication module. The model parameters of the trend predictor are updated offline by the cloud-based training platform and then sent to the onboard real-time processing unit. The system accesses data streams from the ship's inertial navigation system, global navigation satellite system, radar, and visual perception system. It also accesses feedback signals from actuators such as servos and propellers, as well as environmental load monitoring equipment such as accelerometers, gyroscopes, and strain sensors installed throughout the hull.

[0026] The heterogeneous observer group operates in parallel, with the three types of observers sharing the same sampling time fundamental frequency. The calculation cycle of the dynamic weight allocation module is synchronized with this sampling time fundamental frequency. The first type, the finite-time-dilation state observer, receives signals from ship motion state sensors and estimates high-frequency disturbances and abrupt changes on a millisecond-level timescale, outputting high-frequency disturbance estimates. The second type, the interval type-II fuzzy adaptive observer, receives input and output data from the ship dynamics model and uses interval type-II fuzzy rules to process long-term uncertainty drift and model mismatch errors, outputting slow-varying error estimates. The third type, the data-driven trend predictor, is based on a time-series neural network or Gaussian process regression. Based on navigation parameters, environmental load data, and actuator status over a past period, it outputs predictions of future operating conditions. These predictions include operating condition type labels and corresponding confidence levels. The operating condition type labels are obtained through offline clustering of historical navigation data, based on marine environmental parameters, ship load parameters, and actuator health parameters.

[0027] The dynamic weight allocation module receives the outputs of three types of observers, calculates the estimation error and its rate of change for each observer in real time based on the Lyapunov energy function, dynamically determines the weight coefficient of each observer's output in the final state estimation, and outputs the weighted and fused state estimate as the system observation result to the active shift controller. When calculating the weight coefficients, the dynamic weight allocation module simultaneously operates two adjustment mechanisms: a continuous adjustment mechanism based on the Lyapunov energy function and a threshold adjustment mechanism based on the rate of change of estimation error. When the rate of change of estimation error of any observer exceeds a preset threshold, the weight coefficient of that observer is increased to a preset upper limit within a single control cycle, while the weight coefficients of the other observers are decreased. These two mechanisms operate in parallel. When there is a difference in the weight coefficients output by the two mechanisms, the one that results in the larger weight coefficient for that observer is taken as the final weight, prioritizing the response to sudden changes. The parameter matrix of the Lyapunov energy function is adjusted online according to the current system operating point, ensuring that as the system state deviates more from the equilibrium point, the weight coefficient of the finite-time extended state observer monotonically increases.

[0028] The observer health status monitoring module receives the internal state variables and estimated residual sequences of the three types of observers respectively. Based on the sliding window statistics of the residual mean, variance and residual autocorrelation of each observer, it independently generates the health score of each observer. When calculating the weight coefficient, the dynamic weight allocation module introduces the health score of each observer as a product factor into the weight allocation function. When the health score of any observer is lower than the preset threshold, the weight coefficient of that observer is forced to be set to zero, and the system enters the reduced-order observation mode.

[0029] The observer consistency constraint module calculates in real time the difference between the outputs of the finite-time extended state observer and the interval type II fuzzy adaptive observer, as well as the deviation between the two and the predicted value output by the trend predictor. When the statistical characteristics of the difference exceed the normal range and the health scores of each observer are higher than the preset threshold, it is determined that there is a common mode fault in the system. At this time, the observer consistency constraint module outputs a weight reduction instruction to the dynamic weight allocation module, forcibly reducing the weight coefficients of all observers and increasing the proportion of multi-source sensor fusion reference values ​​in the final control. When the observer health status monitoring module and the observer consistency constraint module issue instructions at the same time, the weight reduction instruction of the observer health status monitoring module takes priority. The faulty observer is isolated first, and then the consistency check is performed on the remaining observers.

[0030] The active shift controller receives the weighted fusion state estimate from the dynamic weight allocation module and the operating condition prediction from the trend predictor. Internally, the active shift controller contains at least two sets of control laws with different topologies. One set is configured as a proportional-derivative structure primarily for heading maintenance, while the other is configured as a feedforward compensation structure primarily for surge protection. When the operating condition prediction from the trend predictor indicates that the system will enter a new operating condition, the active shift controller smoothly changes the structure or parameters of the control law through an interpolation function before the actual system state crosses a threshold change, transitioning the control law from a topology adapted to the current operating condition to one adapted to the predicted operating condition. When the operating condition prediction indicates that the system will enter a high sea state or high-wave navigation state, the active shift controller smoothly transitions the current control law to the surge-resistant feedforward compensation structure in advance.

[0031] The data-driven trend predictor internally constructs a condition transition probability matrix, which is generated based on the statistical analysis of the transition frequency between different conditions in historical navigation data. When outputting the future condition prediction results, the trend predictor simultaneously outputs the transition probability from the current condition to each possible target condition. The active shift controller calculates the target weight of the pre-switching control law based on the transition probability and the preset risk preference coefficient, and gradually completes the transition of the control law in multiple control cycles using an exponential smoothing method. The trend predictor also includes a probability prediction layer based on Monte Carlo dropout or Bayesian neural network. This layer outputs the most likely target condition, the probability distribution of the time window in which the condition occurs, and the probability of at least one alternative condition. When the confidence of the main predicted condition corresponding to the transition probability is higher than a preset threshold, the active shift controller gradually completes the transition of the control law using an exponential smoothing method. When the confidence level of the primary predicted operating condition is lower than a preset threshold and the difference between the probability of the alternative operating condition and the probability of the primary predicted operating condition is less than a preset difference, one or more alternative control law branches are activated simultaneously with a gain lower than that of the primary channel. The alternative control law branches correspond to the alternative operating conditions and are in a hot standby state. When the primary predicted operating condition does not occur within the time window but the alternative operating condition occurs first, the system completes the switching from the primary control law branch to the alternative control law branch within one control cycle. When the confidence level of the primary predicted operating condition is lower than a preset threshold and the probability of the alternative operating condition is significantly lower than that of the primary predicted operating condition, the active shift controller suspends the pre-switching action and maintains the current control law. When the confidence level of the trend predictor output is lower than a preset threshold, the dynamic weight allocation module suspends the pre-switching action of the active shift controller and redistributes the weight coefficients to the finite-time extended state observer and the interval type II fuzzy adaptive observer.

[0032] Before being input to the ship's actuators, the control commands output by the active shift controller pass sequentially through the actuator load distribution module and the nonlinear command shaper. The actuator load distribution module receives the total control demand from the active shift controller and, combined with future operating condition predictions from the trend predictor, dynamically distributes the load among multiple actuators using a model predictive control framework. The distribution objective incorporates the cumulative wear, heat dissipation status, and expected margins of each actuator under future operating conditions into the cost function. This module also feeds back the distributed load commands to the dynamic weight distribution module, enabling it to distinguish between dynamic changes caused by load redistribution and those caused by external disturbances or model mismatch when evaluating observer errors. The nonlinear command shaper receives the distributed commands from the actuator load distribution module and, based on the current physical constraints of each actuator, including the maximum angular velocity of the steering gear and the propeller speed change limit, calculates the feasible range of command variation in real time. Command increments exceeding this range are constrained and shaped, and the shaped commands are output to the actuators. Simultaneously, this shaper feeds back the command limit caused by constraints to the dynamic weight distribution module as a correction term for observer error evaluation.

[0033] The multi-source sensor data consistency verification module accesses the data streams from the ship's inertial navigation system, global navigation satellite system, radar, and visual perception system. It performs multi-source data fusion through an extended Kalman filter to generate redundant ship motion state reference values. These reference values ​​are compared with the weighted fusion state estimate output by the dynamic weight allocation module. When the deviation between the two exceeds a preset safety threshold, the system determines that the current weighted fusion state estimate is invalid and forcibly switches to a backup control loop based on the multi-source sensor fusion reference value until the heterogeneous observer group completes self-calibration and re-converges.

[0034] The self-calibration trigger monitors the cumulative deviation between the ship's actual navigation trajectory and the desired trajectory, the frequency of actuator actions, and the parameter drift of the interval type II fuzzy adaptive observer. When all of the above monitoring indicators are below the threshold within the set time window, the system is determined to be in a state of fully excitation and steady-state operation, triggering the background self-calibration process. The system then uses the currently collected input and output data to re-identify the baseline parameters of the ship's dynamics model and updates the initial fuzzy rule base of the interval type II fuzzy adaptive observer and the offline clustering center of the trend predictor accordingly.

[0035] When the system is in steady state, the self-calibration trigger is not triggered, and the system is not in the response evaluation period after the probe signal injection, the online excitation and response evaluation module injects a probe signal with controlled amplitude and limited spectrum into the control command, constrained by not interfering with the main control objective. This probe signal is generated based on a pseudo-random binary sequence or a sinusoidal sweep frequency signal. The module simultaneously records the system's response to the probe signal, calculates the current system closed-loop frequency characteristics through online system identification, and compares them with the reference frequency characteristics. When the comparison deviation exceeds a preset threshold, it actively triggers online correction of the observer parameters or controller parameters. During the probe signal injection period, the self-calibration trigger suspends monitoring and judgment of the actuator's action frequency and cumulative trajectory deviation, resuming monitoring after the response evaluation period ends.

[0036] Example 1 An unmanned survey vessel is conducting seabed topographic mapping in open waters. The sea state is Class II with relatively small swells, and the vessel is cruising along a pre-set survey line at a cruising speed. At this time, the system is in a relatively steady operating state.

[0037] The finite-time extended state observer receives heading, roll, and pitch angular velocity signals from the shipborne inertial navigation system at millisecond intervals, and outputs high-frequency disturbance estimates. Due to the stable sea state, the rate of change of the observer's estimation error is at a low level. The interval-type II fuzzy adaptive observer receives rudder angle commands and actual heading feedback, and outputs model mismatch error estimates based on the fuzzy rule base. Its parameter drift changes slowly within the window set by the self-calibration trigger. The trend predictor outputs the operating condition prediction results for the next 30 seconds based on the navigation parameters and environmental load data of the past five minutes. Due to the stable sea state, the prediction results show that the system will continue to maintain the current operating condition. Among the operating condition transition probabilities, the probability of "maintaining the current operating condition" exceeds 95%, and the probability of alternative operating conditions is less than 3%.

[0038] The dynamic weight allocation module calculates the weight coefficients of the three types of observers based on the Lyapunov energy function. Since the estimation error rate of the finite-time extended state observer is low, the mean and variance of the residuals of the interval type II fuzzy adaptive observer are within the normal range, and the output confidence of the trend predictor is high, the dynamic weight allocation module sets the weight coefficient of the interval type II fuzzy adaptive observer to around 0.6, and the weight coefficients of the finite-time extended state observer and the trend predictor to 0.2 and 0.2, respectively. The observer health status monitoring module scores the health of the three types of observers above 0.95, and no order reduction observation is triggered. The observer consistency constraint module calculates the difference between the outputs of the three types of observers. The statistical characteristics of the difference are within the normal range, and no common mode fault is determined.

[0039] After receiving the output from the trend predictor, the active shift controller determines that the operating condition remains unchanged, so it does not perform a control law switch and maintains the proportional-derivative structure that prioritizes heading maintenance. The total control requirement output by the active shift controller is the rudder angle command required to maintain the heading. The actuator load distribution module receives this total control requirement. Since the ship is configured with a single rudder and a single propeller, the load distribution module directly transmits the command to the nonlinear command shaper. The nonlinear command shaper shapes the command according to the maximum angular velocity constraint of the rudder. Since the rate of change of the rudder angle does not exceed the constraint, the shaped command is directly output to the rudder actuator.

[0040] The multi-source sensor data consistency verification module compares the reference value fused from the inertial navigation system and the global navigation satellite system with the weighted fusion state estimate output by the dynamic weight allocation module. If the deviation between the two is within the preset safety threshold, the system maintains the operation of the main control loop. The self-calibration trigger detects that the trajectory cumulative deviation, actuator action frequency, and observer parameter drift are all below the threshold, and the system has been running continuously for more than the set time window. It then triggers the background self-calibration process, re-identifies the reference parameters of the ship dynamics model using the currently collected input and output data, and updates the initial fuzzy rule base of the interval type II fuzzy adaptive observer and the offline clustering center of the trend predictor accordingly. The online excitation and response evaluation module detects that the self-calibration trigger has been triggered, suspends the injection of detection signals, and resumes it after the self-calibration process is completed.

[0041] Example 2 On its return journey after completing its survey mission, the same unmanned survey vessel encountered a sudden strong wind. The sea state rose from level two to level four within thirty seconds, the swells intensified significantly, and the ship's rolling amplitude increased.

[0042] Based on real-time access to anemometer data and ship roll acceleration data, the trend predictor outputs the operating condition prediction results for the next ten seconds through a Bayesian neural network probabilistic prediction layer. The main predicted operating condition is "navigating in high sea state with high waves" with a confidence level of 78%. At the same time, it outputs the probability of the alternative operating condition "navigating in high sea state with low waves" at 15%. The time window probability distribution shows that the operating condition change will occur within five to eight seconds. The operating condition transition probability matrix shows that the probability of transitioning from the current operating condition to the high sea state operating condition exceeds 70%.

[0043] After receiving the high-confidence output from the trend predictor, the dynamic weight allocation module begins to adjust the weight coefficients. The rate of change of the estimation error of the finite-time extended state observer begins to rise, but has not yet exceeded the preset threshold. The threshold adjustment mechanism based on the rate of change of the estimation error has not yet intervened. The continuous adjustment mechanism based on the Lyapunov energy function gradually increases the weight coefficient of the finite-time extended state observer from 0.2 to 0.4 according to the degree of deviation of the system state from the equilibrium point. At the same time, the weight coefficient of the interval type II fuzzy adaptive observer is reduced from 0.6 to 0.4. The weight coefficient of the trend predictor remains at 0.2. The observer health status monitoring module's health scores for the three types of observers remain high, and no order reduction observation is triggered.

[0044] The active shift controller, based on the fact that the confidence level of the main predicted operating condition output by the trend predictor is higher than a preset threshold, uses an exponential smoothing method to gradually transition the control law from a proportional-derivative structure mainly for heading maintenance to a feedforward compensation structure mainly for surge resistance within six control cycles. During the transition, the controller calculates the target weight of the pre-switching control law by combining the transfer probability and risk preference coefficient, and adjusts it gradually in steps of 20% per cycle. At the same time, since the confidence level of the main predicted operating condition is higher than the threshold and the difference between the probability of the alternative operating condition and the probability of the main predicted operating condition is 63%, which is greater than the preset difference, the system does not activate the alternative control law branch.

[0045] The overall control requirement output by the active shift controller is a combined command for rudder angle and propeller speed required for surge protection. Upon receiving this overall control requirement, the actuator load allocation module, combined with future operating condition predictions from the trend predictor, dynamically allocates the load between the rudder and propeller using a model predictive control framework. The cost function incorporates weights for accumulated wear and heat dissipation of the actuators. The model predictive control framework optimizes the expected rudder angle and propeller speed changes over the next ten seconds, ultimately outputting a load allocation scheme where the rudder load accounts for 70% and the propeller load accounts for 30%. This allocation result is fed back to the dynamic weight allocation module, enabling it to distinguish between dynamic changes caused by load redistribution and those caused by external disturbances during subsequent observer error assessment. The nonlinear command shaper receives the allocated command and calculates the feasible variation range in real time based on the rudder's maximum angular velocity and the propeller speed change rate limit. Command increments exceeding this range are constrained and shaped before being output to the actuators, while the command limit is fed back to the dynamic weight allocation module.

[0046] The multi-source sensor data consistency verification module continuously compares the fused reference value with the weighted fused state estimate, and the deviation between the two remains within a safe threshold. The self-calibration trigger detects that the accumulated trajectory deviation is increasing, and the actuator frequency is rising, determining that the system is no longer in a steady-state operation, and suspends the self-calibration trigger. The online excitation and response evaluation module detects that the system is not in a steady state and suspends the injection of detection signals.

[0047] When the strong wind actually arrived ten seconds later, the system had already completed the pre-switching of the control law, the ship's roll amplitude was controlled within a safe range, and the heading deviation did not exceed the preset threshold.

[0048] Example 3 During the mission, the starboard rudder of the unmanned survey vessel experienced a progressive failure, with the rudder response speed dropping to 60% of the normal value, but it had not yet completely failed. This failure was not directly detected by conventional sensors in the early stages.

[0049] During continuous operation, the mean and variance of the estimated residual sequence of the interval type II fuzzy adaptive observer begin to rise slowly, and the drift of the internal parameters of the observer gradually increases. The observer health status monitoring module performs sliding window statistics on the residual mean, variance, and residual autocorrelation of each observer. The calculated health score of the interval type II fuzzy adaptive observer decreases from 0.95 to 0.72, but is still higher than the preset threshold of 0.7, so it does not trigger a forced zeroing. The health score of the finite-time expansion state observer remains at 0.93, and the health score of the trend predictor remains at 0.88.

[0050] The observer consistency constraint module calculates the difference between the outputs of the three types of observers in real time. Due to the decrease in servo response speed, the deviation between the high-frequency disturbance component estimated by the finite-time extended state observer and the actual actuator output increases, causing the statistical characteristics of the difference between the observer output and the interval type II fuzzy adaptive observer output to exceed the normal range. At the same time, the health scores of each observer are higher than the preset threshold. Based on this, the observer consistency constraint module determines that there is a common mode fault in the system, outputs a weight reduction instruction to the dynamic weight allocation module, forcibly reducing the weight coefficients of all observers to 50% of their original values, and increasing the proportion of multi-source sensor fusion reference values ​​in the final control from 10% to 40%.

[0051] The dynamic weight allocation module receives instructions from both the observer health status monitoring module and the observer consistency constraint module. According to the priority setting, the observer health status monitoring module does not issue a downgrade observation instruction. Therefore, the system first executes the downgrade instruction from the observer consistency constraint module. The proportion of the observer output in the weighted fusion state estimate decreases, and the proportion of the multi-source sensor fusion reference value increases. The system then enters the fault-tolerant operation mode.

[0052] The weighted fusion state estimate received by the active shift controller, having incorporated more sensor fusion reference values, still maintains an effective estimate of the ship's actual state. The confidence level of the operating condition prediction output by the trend predictor decreases due to the decline in the health of the observer, but it is still above the preset threshold. The active shift controller continues to maintain the normal control law pre-switching function.

[0053] When allocating control commands, the actuator load distribution module, due to the decreased response speed of the servo motor, incorporates a constraint on the expected margin of the starboard servo motor in the cost function, distributing more load to the port servo motor and propeller. Simultaneously, the allocated load commands are fed back to the dynamic weight distribution module. The nonlinear command shaper shapes the commands based on the actual physical constraints of the servo motor, and the command limiting is fed back to the dynamic weight distribution module, enabling the system to distinguish between actuator performance degradation and changes in external disturbances.

[0054] The self-calibration trigger detected that the actuator's operating frequency was lower than normal, and the parameter drift of the interval type II fuzzy adaptive observer was close to the threshold. It determined that the system did not meet the conditions for sufficient excitation and steady-state operation, and the self-calibration process was not triggered for the time being. The online excitation and response evaluation module detected that the system was not in a steady state and suspended the injection of probe signals.

[0055] The shipborne real-time processing unit uploads the observer health score, the judgment result of the observer consistency constraint module, and the load allocation record of the actuator load allocation module to the cloud training platform through the Internet of Things communication module. After offline analysis, the cloud training platform determines that the starboard rudder has performance degradation and sends a maintenance warning to the shore-based control center. At the same time, it updates the offline clustering center of the trend predictor and includes the decrease in rudder response speed as a new working condition type in the clustering space.

Claims

1. An unmanned ship stable operation intelligent monitoring control system, characterized in that, include: The heterogeneous observer group comprises at least three types of observers based on different mathematical principles: the first type is a finite-time extended state observer, configured to estimate high-frequency disturbances and abrupt signals on a millisecond-level timescale; the second type is an interval type-II fuzzy adaptive observer, configured to estimate the system's long-term uncertainty drift and model mismatch error; and the third type is a data-driven trend predictor, based on a time-series neural network or Gaussian process regression, which outputs predictions of future operating conditions based on historical navigation parameters, environmental load data, and actuator states. The dynamic weight allocation module has its input connected to the output of each observer in the heterogeneous observer group. This module calculates the estimation error and its rate of change of each observer in real time based on the Lyapunov energy function, dynamically determines the weight coefficient of each observer's output in the final state estimation, and outputs the weighted and fused state estimation value as the system observation result. An active shift controller, whose input is connected to the dynamic weight allocation module and the trend predictor, when the operating condition prediction result output by the trend predictor indicates that the system will enter a new operating condition, smoothly changes the structure or parameters of the control law through an interpolation function before the actual state of the system changes across a threshold, so that the control law transitions from a topology adapted to the current operating condition to a topology adapted to the predicted operating condition.

2. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, When calculating the weight coefficients, the dynamic weight allocation module increases the weight coefficient of any observer to a preset upper limit within a single control cycle when the rate of change of the estimation error of any observer exceeds a preset threshold, while simultaneously reducing the weight coefficients of the remaining observers. The threshold adjustment mechanism based on the rate of change of estimation error and the continuous adjustment mechanism based on the Lyapunov energy function operate in parallel. When there is a difference between the weight coefficients output by the two, the one that makes the observer's weight coefficient larger is taken as the final weight to prioritize the response to abrupt signals.

3. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, It also includes an observer health status monitoring module, which receives the internal state variables and estimated residual sequences of the finite-time extended state observer, the interval type II fuzzy adaptive observer, and the data-driven trend predictor, respectively. Based on the sliding window statistics of the residual mean, variance, and residual autocorrelation of each observer, it independently generates a health score for each observer. When calculating the weight coefficients, the dynamic weight allocation module introduces the health scores of each observer as product factors into the weight allocation function. When the health score of any observer is lower than a preset threshold, the weight coefficient of that observer is forced to zero, and the system enters a reduced-order observation mode. It also includes an observer consistency constraint module, which calculates in real time the difference between the output of the finite-time extended state observer and the output of the interval type II fuzzy adaptive observer, as well as the deviation between the two and the predicted value output by the trend predictor. When the statistical characteristics of the difference exceed the normal range and the health scores of each observer are higher than a preset threshold, it is determined that there is a common mode fault in the system. At this time, the observer consistency constraint module outputs a weight reduction instruction to the dynamic weight allocation module, forcibly reducing the weight coefficients of all observers and increasing the proportion of multi-source sensor fusion reference values ​​in the final control. When the observer health status monitoring module and the observer consistency constraint module issue instructions at the same time, the weight reduction instruction of the observer health status monitoring module takes priority, first isolating the faulty observer, and then performing consistency verification on the remaining observers.

4. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, The active shift controller includes at least two sets of control laws with different topologies. One set of control laws is configured as a proportional-differential structure with heading maintenance as the main function, and the other set of control laws is configured as a feedforward compensation structure with surge resistance as the main function. When the operating condition prediction result indicates that the system will enter a high sea state or a high wave navigation state, the active shift controller will smoothly transition the current control law to the surge resistance feedforward compensation structure in advance. Before being input to the ship's actuators, the control commands output by the active shift controller pass sequentially through the actuator load allocation module and the nonlinear command shaper. The actuator load allocation module receives the total control demand output by the active shift controller and, in conjunction with the future operating condition predictions output by the trend predictor, dynamically allocates the load among multiple actuators using a model predictive control framework. The allocation objective incorporates the cumulative wear, heat dissipation status, and expected margins of each actuator under future operating conditions into the cost function. This module simultaneously feeds back the allocated load commands to the dynamic weight allocation module, enabling dynamic weight allocation. When evaluating observer error, the module distinguishes between dynamic changes caused by load redistribution and dynamic changes caused by external disturbances or model mismatch. The nonlinear command shaper receives the allocated command output by the actuator load distribution module, calculates the feasible range of command changes in real time based on the current physical constraints of each actuator, including the maximum angular velocity of the servo motor and the propeller speed change rate limit, performs constraint shaping on the command increment that exceeds the range, and outputs the shaped command to the actuator. At the same time, the shaper feeds back the command limit caused by the constraint to the dynamic weight distribution module as a correction item for observer error evaluation.

5. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, The future operating condition prediction results of the trend predictor include at least an operating condition type label and a corresponding confidence level. The operating condition type label is obtained by offline clustering of historical navigation data. The offline clustering is based on marine environmental parameters, ship load parameters, and actuator health parameters. The data-driven trend predictor internally constructs a working condition transition probability matrix, which is generated based on the statistical analysis of the transition frequency between different working conditions in historical navigation data. When outputting the prediction results of future operating conditions, the trend predictor simultaneously outputs the transition probability of the current operating condition to each possible target operating condition. The active shift controller calculates the target weight of the pre-switching control law based on the transition probability and the preset risk preference coefficient, and gradually completes the transition of the control law in multiple control cycles using an exponential smoothing method. The trend predictor includes a probability prediction layer based on Monte Carlo dropout or a Bayesian neural network. This layer outputs the most likely target operating condition, the probability distribution of the time window in which the operating condition occurs, and the probability of at least one alternative operating condition. When the confidence level of the main predicted operating condition corresponding to the transition probability is higher than a preset threshold, the active shift controller gradually completes the control law transition using an exponential smoothing method. When the confidence level of the main predicted operating condition is lower than the preset threshold and the difference between the probability of the alternative operating condition and the probability of the main predicted operating condition is less than a preset difference, one or more alternative control law branches are activated with a gain lower than that of the main channel. The alternative control law branches correspond to the alternative operating conditions and are in a hot standby state. When the main predicted operating condition does not occur within the time window and the alternative operating condition occurs first, the system completes the switch from the main control law branch to the alternative control law branch within one control cycle. When the confidence level of the main predicted operating condition is lower than the preset threshold and the probability of the alternative operating condition is significantly lower than the probability of the main predicted operating condition, the active shift controller pauses the pre-switching action and maintains the current control law.

6. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, The finite-time extended state observer, the interval type II fuzzy adaptive observer, and the data-driven trend predictor operate in parallel, each with the same sampling time base frequency, and the calculation cycle of the dynamic weight allocation module is synchronized with this sampling time base frequency.

7. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, It also includes a multi-source sensor data consistency verification module. This module accesses the data streams from the ship's inertial navigation system, global navigation satellite system, radar, and visual perception system. It performs multi-source data fusion through an extended Kalman filter to generate redundant ship motion state reference values. These reference values ​​are compared with the weighted fusion state estimate output by the dynamic weight allocation module. When the deviation between the two exceeds a preset safety threshold, the system determines that the current weighted fusion state estimate is invalid and forcibly switches to a backup control loop based on the multi-source sensor fusion reference value until the heterogeneous observer group completes self-calibration and re-converges. It also includes a self-calibration trigger, which monitors the cumulative deviation between the ship's actual navigation trajectory and the desired trajectory, the frequency of actuator actions, and the parameter drift of the interval type II fuzzy adaptive observer. When the above monitoring indicators are all below the threshold within a set time window, the system is determined to be in a fully excited and steady-state operation state, triggering a background self-calibration process. The system then uses the currently collected input and output data to re-identify the baseline parameters of the ship dynamics model and updates the initial fuzzy rule base of the interval type II fuzzy adaptive observer and the offline clustering center of the trend predictor accordingly.

8. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, It also includes an online excitation and response evaluation module. When the system is in a steady state and the self-calibration trigger is not triggered, and the system is not in the response evaluation period after the probe signal injection, this module injects a probe signal with controlled amplitude and limited spectrum into the control command, with the constraint of not interfering with the main control objective. The probe signal is generated based on a pseudo-random binary sequence or a sinusoidal sweep frequency signal. The module also records the system's response to the probe signal, identifies and calculates the current closed-loop frequency characteristics of the system through online system identification, and compares them with the reference frequency characteristics. When the comparison deviation exceeds a preset threshold, it actively triggers the online correction of the observer parameters or controller parameters. During the probe signal injection period, the self-calibration trigger suspends the monitoring and judgment of the actuator action frequency and trajectory cumulative deviation, and resumes monitoring after the response evaluation period ends.

9. The unmanned ship stable operation intelligent monitoring control system according to claim 1, characterized in that, It also includes a shipborne real-time processing unit, which runs the heterogeneous observer group, dynamic weight allocation module and active shift controller. The control cycle is set to the millisecond level, and it is connected to the cloud training platform through the Internet of Things communication module. The model parameters of the trend predictor are updated offline by the cloud training platform and then sent to the shipborne real-time processing unit.

10. The intelligent monitoring and control system for stable operation of unmanned vessels according to claim 1, characterized in that, When the confidence level of the trend predictor output is lower than a preset threshold, the dynamic weight allocation module pauses the pre-switching action of the active shift controller and redistributes the weight coefficients to the finite-time extended state observer and the interval type II fuzzy adaptive observer. The parameter matrix of the Lyapunov energy function is adjusted online according to the current system operating point, so that when the system state deviates more from the equilibrium point, the weight coefficient of the finite-time extended state observer assigned by the dynamic weight allocation module increases monotonically.

Images