Marine vehicle large time-delay actuator input sequence optimization method

Abstract

This invention aims to optimize the input sequence of actuators with significant time delays in marine vehicles (such as variable buoyancy hydraulic pumps and valves, and slow-response actuators like rudders). Under complex sea conditions and model uncertainties, this method achieves rapid and stable attainment of the target operational state with minimal action cost, while simultaneously satisfying safety and lifespan constraints. The method disclosed in this invention uses a parameter identification model as the prediction kernel, incorporating terminal task deviation, the number of execution actions and non-zero duration, and the safety volume boundary into a multi-objective cost function. It employs an improved LM algorithm with adaptive damping for sequence-level optimization and rolling updates, thereby overcoming the shortcomings of conventional closed-loop compensation and pure gradient methods in terms of slow convergence, oscillation, and difficulty in balancing safety and wear on time-delayed objects. This significantly improves convergence speed, control accuracy, and reliability, while reducing energy consumption and mechanical wear.

Description

Technical Field

[0001] This invention relates to a method for optimizing the input sequence of large time-delay actuators for marine vehicles, specifically in the field of marine intelligent equipment and control technology, for practical marine sampling tasks. Background Technology

[0002] In the past five years, several domestic inventions have focused on buoyancy system control and trajectory / heading control for actuators of marine vehicles (including AUVs, USVs, and gliders) in highly nonlinear and strongly coupled environments. For example, an oil-filled underwater glider achieves neutral hovering near any operating depth by constructing a temperature, salinity, and density profile model and adjusting the oil bladder's drainage volume using a hydraulic pump. This approach relies on feedback regulation based on density field estimation (CN107544526A). Another example is the use of a "depth-velocity dual closed loop + nonlinear extended state observer (ESO)" to counteract environmental disturbances for rapid AUV diving and surfacing. However, the core of this approach remains continuous regulation within the loop rather than feedforward input sequence optimization (CN111547212A). At the device level, buoyancy adjustment mechanisms for fuel bladders / pressure tanks have also formed a series of configurations and control processes, emphasizing the improvement of diving and buoyancy efficiency and energy consumption performance through pump-valve / attitude coordination (CN110803270B, CN1923614A, CN114194365B). For residual buoyancy compensation and pitch angle errors during deep-diving navigation, patents have also provided compensation laws based on residual terms caused by changes in density, temperature, salinity, and pressure (CN107776859A). The above technical approaches are engineering feasible for "point-to-depth" tasks, but they share a common characteristic: they do not characterize the significant time delay of actuators as an "optimizable object," and they lack unified cost modeling and input sequence-level optimization for "number of actions, duration, and hard / soft constraints on safety volume."

[0003] On the other hand, mature time-domain control and compensation methods exist for ships and unmanned vessels to address the problem of "large time delay / large inertia," but these methods are mostly geared towards propellers / servo motors rather than slow-response actuators such as hydraulic bladders. For example, Generalized Predictive Control (GPC) is used for dynamic positioning to cope with large time delays, rolling optimization, and multi-step prediction (CN101963784A); Smith predictors + Active Disturbance Rejection Control (ADRC) are introduced into heading angle control to improve the positioning accuracy and control smoothness of objects with time delays (CN112558481B); for dynamic positioning systems with "time delay + saturation," there are cascaded system designs of time delay compensators + actuator saturation compensators (CN110045726A); in the field of path tracking, there are also approaches such as event triggering + input quantization to reduce the execution frequency, and nonlinear control considering the influence of input dead zone (CN111045432B). On the AUV side, Model Predictive Control (MPC) is used for constrained path / velocity control and roll optimization, but it focuses on trajectory error and control constraints, failing to abstract the slow-response / large-time-delay hydraulic volume adjustment into a "discrete input sequence" for sparsification and time-penalty optimization (CN112947068A). Academic reviews also point out that surface / underwater unmanned systems generally face challenges such as time delays, actuator constraints, and environmental uncertainties, but most studies remain at the level of compensation and steady-state analysis, lacking a task-oriented optimal design framework for the "input structure of time-delayed actuators (such as pulse width / interval / number of pulses)." Summary of the Invention

[0004] The technical problem this invention aims to solve is that existing technologies either focus on density field modeling and feedback regulation (oil bladder / pump valve closed loop) or on time delay compensation / predictive control (propeller / servo motor). Overall, they have not formed an input sequence-level optimization paradigm for "large time delay actuators (such as hydraulic oil bladders / variable buoyancy systems) of marine vehicles". That is, under a unified criterion, the terminal task error, number of actions, non-zero duration and safety volume (hard / soft) constraints are quantified and compromised, and the sequence optimization is carried out in combination with the environmental coupling (temperature-salt-pressure) model.

[0005] To address the aforementioned technical problems, the present invention discloses a method for optimizing the input sequence of a large time-delay actuator for marine vehicles. The method is characterized by: using a parameter identification model based on environment-dynamic coupling, shifting and predicting the input according to the actuator time delay; constructing a unified cost function that includes task error, action count penalty, non-zero duration penalty, energy consumption / wear surrogate terms, and hard / soft constraint terms for safety boundaries; iteratively solving the finite-time-domain input sequence using an improved LM algorithm with adaptive damping factor updates; converting the obtained sequence into executable pulse / duty / opening commands via an executable mapper, and satisfying amplitude and other conditions through structured projection.

[0006] This invention targets actuators in marine vehicles that exhibit significant response hysteresis (such as variable buoyancy hydraulic pumps / valve, steering gear / thrusters in low-speed, high-inertia conditions, and propulsion power-variable devices), optimizing the input sequence within a finite prediction time domain. Preferably, the operational output is denoted as... (Can be longitudinal displacement, attitude, or other task states), the discrete sampling period is The number of time-domain steps is planned. The actuator time delay ( ).

[0007] Preferably, the parameter identification model based on environment-dynamic coupling is:

[0008]

[0009]

[0010] In the formula, For state, This represents the input after time-delay propagation. The parameters obtained from online / offline identification include temperature-salt-pressure, added mass, damping, shell compressibility, propulsion efficiency, etc. For disturbance, For discrete-time systems, state transition functions / dynamic update operators are used to describe the state transition functions / dynamic update operators in the parameters. and disturbance Under its influence, the system state changes from Lag input Driven by evolution to The function may include a coupling term between environmental quantities (such as temperature, salinity, pressure, and flow rate) and kinetic parameters. This is an observation / output mapping function used to map the state... With parameters Mapped to a measurable or controllable output quantity; For the first The output measurements or mission outputs of the step (e.g., depth, longitudinal displacement, heading / attitude angle, velocity, track error or a combination thereof).

[0011] Preferably, an input serialization and executable mapper (from ideal to pump / valve / servo motor pulse) is used. To eliminate the disconnect between the "algorithm solution and hardware executable stream," an input mapper is introduced. Ideal input The mapping is used to generate pulse / duty cycle / opening trajectory that can be sent out. The same mapping is used for both prediction simulation and actual sending out, ensuring consistency between executable prediction and executable sending out.

[0012] Preferably, the performance characteristics include: upper / lower limits of amplitude. ,in, Input the minimum allowable amplitude or lower limit for the actuator (which can correspond to the minimum opening, minimum duty cycle, minimum thrust / torque, or the value corresponding to the "off" state). Input the maximum allowable amplitude or upper limit for the actuator (which can correspond to the maximum opening, maximum duty cycle, and maximum thrust / torque). ,in Input the upper limit of the rate of change / the threshold of the slope (reflecting the actuator's climbing / descending speed capability); minimum start / stop time. ,in The minimum duration that must be maintained in a single on or off state (to avoid wear and vibration caused by frequent switching); minimum interval ,in It is the minimum time interval between two consecutive non-zero pulses (or two activation actions). Discretization conversion factors or scaling factors for minimum interval constraints (used to convert continuous time intervals into discrete steps / sampling intervals); quantization and dead zone: introducing a threshold. With quantization step size This compresses small disturbances into zero motion, ensuring sparse motion and avoiding high-frequency jitter.

[0013] Variable buoyancy examples (pumps / valves):

[0014] in, For the first The equivalent liquid volume / equivalent volume state of the buoyancy adjustment system within the step (used to characterize the amount of buoyancy adjustment). and A mapping function for input to pump / valve on / off state, duty cycle, or opening degree (transforming ideal input) (Converted into a sendable pump / valve control trajectory). and This refers to the rated flow rate of the pump and valve. and This refers to the efficiency factor or flow rate effective coefficient. This is the leakage coefficient / leakage attenuation coefficient, used to characterize the attenuation effect caused by leakage or backflow as the system's volume increases. Safety hard constraints: ,in This is the maximum permissible liquid volume limit or the system safety limit (exceeding this limit may lead to excessive structural stress, buoyancy saturation, or safety risks).

[0015] Preferably, when constructing the unified cost function, a single criterion function is constructed to uniformly weigh key indicators (the terminal / trajectory composite form can be selected according to the task), as shown in the following formula:

[0016] For arbitrary precision (tracking precision term), where, This is the terminal error weighting coefficient. To predict the time-domain terminal step (the first step) The output corresponding to step ) For the desired output / target value, For process error weighting coefficients, To predict / optimize the time domain length (number of walks); This represents the number of actions (non-zero action count items), where... Penalty weight for the number of actions, This is a scaling factor or normalization factor for the count term (used for dimension matching; in engineering implementation, it can be taken as...). ), The function is an indicator function (1 if the condition is true, 0 otherwise), and it is implemented using a smooth approximation (e.g.) log-sum or reweighting ), This is the dead zone threshold (actions below this threshold are considered zero). For non-zero duration terms, among which, Penalty weight based on duration; For wear / energy consumption proxy items, among which, Total variation (TV) penalty weight (to suppress high-frequency changes and wear). Input amplitude penalty weight (reflecting energy consumption or action intensity); For soft constraint terms, use quadratic penalties or barrier functions (such as ReLU squared penalties or logbarriers), where... For safety reasons, penalty weight, This is the safety cost function / soft constraint penalty function, used to measure the degree to which the predicted state and input violate safety constraints (such as boundary margin, collision distance, CBF condition, physical upper bound constraint). The greater the violation, the larger the penalty value. Adaptive near / far field weighting: when Time reduction , To reduce constraints on motion sparsity and accelerate approximation; when Time increases , To suppress over-control and chattering, among which, The far-field error threshold. The near-field error threshold is given, and it satisfies the following conditions: .

[0017] Preferably, the improved LM algorithm for iteratively solving finite-time-domain input sequences includes an improved LM sequence-level least squares solver (adaptive damping + time-delay sensitivity): Will Written as residual least squares Construct residual stacking:

[0018] In the formula, To output the tracking residual stack (including the weighted residuals of process error and terminal error, for example, by...) and constitute); Number of actions (pseudo) The smoothed approximate residual of the term (used to approximate counting non-zero actions); The residuals of the non-zero duration term (smoothed approximation indicated by non-zero duration) (Weighted composition) The residuals of the total variation (TV) term (from the difference) (Constructs and can smooth absolute values). Input amplitude Item residual (by (The smooth approximation is constructed). For the residuals of the soft constraint term (by (or its square root form).

[0019] Using an improved LM iteration:

[0020] In the formula, For the first The input sequence vector for the next iteration (by...) Stacking); Let be the Jacobian matrix of the residuals with respect to the input sequence, i.e. ; For the first The LM damping coefficient of the next iteration; For the identity matrix (and) (same dimensions) In order to be in The residual vector calculated at the location; Damping adaptive:

[0021] In the formula, The ratio of the actual decrease to the decrease predicted by the quadratic model (reliability index). For the first The cost function value for the next iteration; For the first The update step size obtained in the next iteration; The damping coefficient; In order to be in The residual vector at the location; like : Approaching the Newtonian pace, increasing speed, among which... The current damping coefficient, Let be the damping scaling factor and satisfy ; like : Then transition to a conservative gradient step and reach steady state.

[0022] Jacobian acquisition: analytical sensitivity / numerical difference / automatic differentiation; explicit partial derivatives are provided for time-delay chain structures to maintain block sparsity and accelerate the solution; Robustness: Action counting, TV, Using Huber / pseudo Reweighting reduces the impact of sharp non-smooth terms on convergence; if necessary, Armijo backtracking can be used as a fallback.

[0023] Preferred constraints and feasibility: a combination of hard / soft approaches + structured projection, including: Perform structured projection after each iteration : Amplitude / slope limiting: point-by-point cropping and TV smoothing; Minimum start / stop / minimum interval: Merge or delete adjacent small pulses to ensure minimum start / stop time and minimum interval; Safety hard constraints (such as) , , When a prediction is violated, subsequent pulses are contracted and the corresponding soft constraint weights are increased.

[0024] Preferably, rolling optimization and event triggering: online adaptation to environment and model drift, that is, the method supports rolling optimization with fixed period or event triggering.

[0025] Preferably, the scrolling optimization includes: Rolling horizon: every Focusing solely on optimizing the frontier M-step reduces real-time computing power. ; Triggering condition: When In the event of drastic environmental changes or a decrease in safety margin, immediate replanning should be implemented, including... Based on the model The predicted output (or the predicted value of the critical task output) obtained under the current input sequence. The event trigger threshold / maximum allowed prediction deviation threshold; Parameters updated online: We employ robust RLS / EKF / IV-LS small-step updates to avoid “mistaken model – miscontrol”.

[0026] This invention, based on an environment-coupled global parameter identification model and an improved Levenberg-Marquardt (LM) algorithm incorporating an adaptive damping factor update mechanism, achieves coordinated satisfaction of precise control, energy consumption and wear suppression, and safety constraints. Under complex sea conditions and model uncertainties, it achieves rapid and stable attainment of the target operational state with minimal action cost, while simultaneously satisfying safety and lifespan constraints. The method disclosed in this invention uses the parameter identification model as the prediction kernel, unifying terminal task deviation, number of actions executed, non-zero duration, and safety volume boundary into a multi-objective cost function. It employs an improved LM algorithm with adaptive damping for sequence-level optimization and rolling updates, thereby overcoming the shortcomings of conventional closed-loop compensation and pure gradient methods in terms of slow convergence, oscillation, and difficulty in balancing safety and wear on time-delayed objects. This significantly improves convergence speed, control accuracy, and reliability, while reducing energy consumption and mechanical wear.

[0027] Compared with existing technical solutions, the present invention has the following beneficial effects: 1) Faster convergence and more accurate steady-state. By employing an improved LM with adaptive damping and a time-delay sensitive Jacobian, under conditions of strong nonlinearity, significant time delay, and non-smoothness penalty, the iteration exhibits an adaptive switching characteristic of "Newton-gradient", avoiding oscillation and step rejection, and achieving rapid positioning and small steady-state error. 2) Sparser movements and shorter non-zero durations result in a simultaneous decrease in wear and energy consumption. By explicitly incorporating action count penalties and non-zero duration penalties into a single cost function, and combining this with wear / energy consumption proxying of total variation (TV) and L1, the output input sequence has fewer start-stop cycles and shorter durations, effectively suppressing high-frequency jitter and ineffective fine-tuning. In the comparison of the implementation examples, the number of actions and cumulative power-on duration are significantly reduced, and the corresponding energy consumption and mechanical wear decrease simultaneously. 3) Higher safety redundancy and lower risk of exceeding boundaries Soft constraints (ReLU) 2 The parallel combination of the obstacle function and hard limit constraint, along with structured projection (amplitude / slope / minimum start and stop time / minimum interval), ensures that the sequence cannot exceed the limits at both the numerical and engineering levels; when approaching the critical point, it automatically enters a high-damping, small-step, and safe mode, which significantly reduces the probability of exceeding the limits. 4) "Executable prediction = Executable deployment", reducing repeated parameter adjustments on actual ships / drills. The ideal input is mapped to pulse / duty / opening by an executable mapper, and isomorphic models are used on both the prediction and transmission sides to eliminate the inconsistency between the algorithm solution and the hardware bitstream, reduce the failure of "calculated but not transmitted" in the field, shorten the parameter tuning cycle and improve the first success rate. 5) More robust to environment and model drift By introducing event-triggered rolling optimization and online small-step parameter updates, the sequence can be reprogrammed in real time and maintain feasibility when temperature, salinity, pressure field, wave and current disturbances, or structural parameters drift. It can still operate stably even when there is model mismatch, data noise, and large hysteresis uncertainty.

[0028] In summary, compared with existing schemes that rely on closed-loop compensation or single-objective optimization, this invention takes sequences as the object, unified cost as the core, improved LM as the engine, mapping and projection as feasibility assurance, and rolling optimization as the adaptation mechanism, forming systematic advantages in terms of convergence speed, action sparsity, energy consumption and wear, safety redundancy, and robustness. Attached Figure Description

[0029] Figure 1 This is a flowchart of an input sequence optimization method for a large time-delay actuator of a marine vehicle according to the present invention.

[0030] Figure 2 This diagram illustrates a comparison of fuel quantity control sequences under three different methods. Note: The vertical axis represents the input. The horizontal axis represents time. The 650 s switching point of GD and the 2457 s switching point of LM are marked.

[0031] Figure 3 This illustrates the input sequence response results of a traditional method. Note: This includes three subplots or three curves. , , ), the figure caption should be written as follows .

[0032] Figure 4 This illustrates the response results of the input sequence to the GD algorithm. Note: Same style as above; figure captions indicate... .

[0033] Figure 5 The diagram illustrates the response results of the input sequence to the improved LM algorithm. Note: Same style as above; figure captions indicate... . Detailed Implementation

[0034] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0035] The present invention aims to propose an input sequence optimization method for actuators with significant response hysteresis (such as variable buoyancy hydraulic pump valves, rudders, and other slow-response actuators) in marine vehicles.

[0036] 1. Test subjects and operational objectives Object: Prototype of an aircraft (internal oil bladder + hydraulic pump / valve).

[0037] Operational objective: In a specific area of ​​the South China Sea, to enable the vehicle to descend from its initial depth and stabilize at a certain point. To meet depth accuracy and safety constraints, while reducing energy consumption and mechanical wear, among other things, This indicates the target depth / desired stable depth (the depth setting value that controls the target).

[0038] Safety boundary: Inner oil bladder volume The remaining electrical and attitude limits are subject to hard limiting based on the equipment nameplate values.

[0039] Stability criterion:

[0040]

[0041] in, Indicates the time of the spacecraft The actual depth (usually oriented downwards as positive or according to the depth coordinates agreed upon by the system); This represents the depth steady-state error tolerance threshold (depth accuracy tolerance bandwidth). This represents the rate of change of depth / vertical velocity (descent or ascent speed). This represents the allowable vertical velocity threshold (a velocity threshold in stability criteria, used to exclude cases where the vehicle continues to move at high speed even after crossing the target depth). Definition This is the minimum time when both of the above equations are satisfied simultaneously for the first time.

[0042] 2. Global parameter identification model and environmental field A global parameter identification model is adopted:

[0043]

[0044]

[0045] in: The second derivative / vertical acceleration representing depth; The first derivative of depth / vertical velocity; This represents the equivalent static term or the net buoyancy bias term caused by density stratification that varies with depth (which can be understood as the gravity-buoyancy imbalance term caused by the density difference between the reference depth and the current depth). This represents the combined disturbance term (including equivalent external forces / uncertainties caused by waves, internal waves, turbulence, unmodeled hydrodynamics, non-ideal pumps and valves, etc.). This represents the density of seawater at the depth where the vehicle is located (or the equivalent seawater density used for buoyancy calculations). This indicates the reference volume of the internal oil bladder / buoyancy adjustment system (e.g., the initial volume or the volume corresponding to neutral buoyancy). This represents the volume of the oil bladder at the current moment (a controllable buoyancy state quantity). Indicates target depth Seawater density at that location; Indicates the displacement volume of the aircraft (equivalent displaced water volume). The amount of correction to the drainage volume (e.g., volume change due to shell compressibility or effective drainage volume change due to structural deformation). Indicates a reference density (e.g., sea surface or nominal seawater density, used to define a baseline buoyancy / weight balance). This represents the seawater density profile as a function of depth (determined by salinity, temperature, and hydrostatic pressure).

[0046]

[0047]

[0048] in, Represents the depth variable.

[0049] 3. Input variables, actuator mapping, and time delay Optimization variable: Hydraulic oil quantity input sequence ,in, Indicates the prediction of the time-domain terminal step (the first step) The hydraulic oil quantity control input (or the ideal actuator drive quantity corresponding to the step) of the step.

[0050] Time delay modeling: Actuator time delay (in This represents the number of discrete steps / delay step size corresponding to the time delay, typically... In this embodiment, the value is taken based on the device identification result. The prediction uses either a shift chain or Smith's estimate for equivalent processing. This indicates the discrete sampling period / control update period.

[0051] Executable mapping (pump / valve pulse / duty cycle):

[0052] in, Indicates the first The volume state of the inner oil bladder at discrete moments (controllable buoyancy state quantity). This represents the pump-side efficiency factor or effective flow coefficient. Indicates the pump's rated flow rate; Indicates that the input will be entered. An executable mapping function that maps to pump-side switch state / duty cycle / opening degree; Indicates the first Step-by-step hydraulic oil quantity control input (ideal input or mapper input); This represents the valve-side efficiency factor or effective flow coefficient. Indicates the valve's rated flow rate; Indicates that the input will be entered. An executable mapping function that maps to valve-side switching state / duty cycle / opening degree; This represents the leakage coefficient / backflow decay coefficient (characterizing the intensity of volume decay over time caused by leakage, seepage, or backflow due to volume conditions).

[0053] Subject to hard constraints on amplitude / slope / minimum start / stop time / minimum interval and quantization step size Dead zone threshold Constraints; Full-process enforcement ,in This indicates the safe upper limit of the internal oil bladder volume (the upper limit of the maximum allowable filling volume).

[0054] 4. Use of multi-objective criterion functions and weights Employing a multi-objective criterion function:

[0055] Includes four categories: terminal depth error, number of actions, duration of non-zero events, and soft safety constraints. Represents the total task duration divided by the terminal time (which can be a continuous time terminal and usually satisfies...). ); Indicates the terminal time The actual depth; This represents the terminal depth error weighting coefficient; This represents the penalty weighting coefficient based on the number of actions performed; This represents the scaling / normalization coefficient of the count term (used for dimension matching; suitable for engineering implementation). ); This represents an indicator function (1 if the condition is true, 0 otherwise), and can be approximated by a smooth function for numerical optimization; This represents the penalty weight coefficient for non-zero duration; Indicates the sampling period / distance from the walk; This represents the penalty weight coefficient for soft safety constraints; This represents the number of time-domain steps (steps away) for planning / prediction.

[0056] 5. Solution Method and Numerical Settings (Improved LM) Residual construction: The four types of indices are mapped to residual blocks, and Sigmoid / IRLS are used where necessary. The smooth approximation facilitates LM solving.

[0057] Iteration formula:

[0058] in, Indicates the first The input sequence vector of the next iteration (i.e.) In the (values ​​obtained in the next LM iteration); The Jacobian matrix of the residuals with respect to the input sequence (i.e.) ); Indicates the first The LM damping coefficient for the next iteration (used for adaptive switching between Gauss-Newton steps and gradient steps). Represents the identity matrix (and) (same dimensions) Indicates the input sequence The residual vector calculated at the point is obtained by stacking residual blocks such as terminal error, action count approximation, duration approximation and safety penalty.

[0059] Damping adaptive (with gain factor) (For reference only) (Approaching the speed of Newtonian steps); (Conservative approach to gradient step), where, This indicates a low confidence threshold (used to determine that the current step size is ineffective and that damping should be increased and updates made more conservatively). This represents a high confidence threshold (used to determine that the current step size is effective and damping should be reduced to accelerate convergence), and it typically satisfies... .

[0060] Termination conditions: or If the number of iterations is below the threshold, or the maximum number of iterations is reached.

[0061] 6. Comparison of the "input sequence-response" of the three methods (including illustrations) 6.1 Figure 2 Comparison of oil quantity control sequences under three methods The input instructions for the traditional method, GD algorithm, and improved LM algorithm are drawn from top to bottom respectively. Changes over time.

[0062] Traditional method: Initially, quickly adjust the volume of the inner oil bladder to close to the volume required for a fixed depth. Then stop immediately, and the curve shows a "one-step approach followed by a constant curve". The advantage is its simplicity; the disadvantage is the lack of subsequent fine-tuning, making it difficult to resist disturbances.

[0063] GD Algorithm: The input is initially corrected discretely and gradually stabilizes; the significant switching time from 0 to 1 is approximately... It has a certain degree of self-adaptation to errors, but is easily affected by step size sensitivity, and suffers from overshoot and oscillation.

[0064] Improved LM: Initially, the input amplitude is larger to quickly narrow the error, then it converges rapidly and suppresses oscillations; the switching time from 0 to 1 is approximately... (Then a small, smooth adjustment is made).

[0065] 6.2 Figure 3 : Input sequence response results of traditional methods Demonstrating the depth trajectory under the traditional "one-step" strategy Internal oil sac volume With input sequence The depth tends to stabilize over a long period of time. Since the oil quantity is fixed at the initial moment, there is a lack of subsequent adjustments to adapt. The disturbance changes significantly prolonged the arrival time. The entire process did not exceed [a certain timeframe]. However, it lacks sufficient control flexibility.

[0066] 6.3 Figure 4 Response results of the input sequence of the GD algorithm The diagram shows the depth, volume, and input curves for GD under the same objective and constraints. GD can gradually reduce depth error, but under strong nonlinearity and time-delay coupling, it is prone to overshoot and oscillation; the final settling time... This is later than traditional methods. The root cause is that the step size is difficult to balance convergence speed and steady-state smoothness, and no structural inhibition is applied to the "number of actions / duration".

[0067] 6.4 Figure 5: Improved LM algorithm input sequence response results Based on the multi-objective criterion and adaptive damping LM proposed in this invention, depth, volume, and input are also demonstrated. Under the initial large input drive, the vehicle rapidly approaches. Subsequently, adaptive damping is used to reduce the step size, avoiding oscillations and achieving a smooth landing. The settling time is only... It is significantly superior to traditional methods and GD. Throughout the entire process... Always lower Sufficient safety redundancy; the input sequence exhibits the characteristics of "sparse action, short non-zero duration, and small total variation".

[0068] Under the same model and constraints, the improved LM method is superior in terms of arrival speed, stability, motion economy and safety. Its advantage comes from the fast convergence and steady-state smoothness brought about by the suppression of "number of motions / duration / safety" in the unified cost and the damping adaptation.

[0069] 7. Numerical efficiency comparison (offline) Under equivalent hardware and software conditions, the offline runtime of GD is approximately 0.19472 s, while that of the improved LM is approximately 0.17407 s. The improved LM, due to its near-Newtonian orientation and adaptive damping, achieves equal or better solution quality with fewer steps and more favorable linear algebra condition numbers, resulting in higher numerical efficiency.

[0070] 8. Operating Procedures 1) Initialization: Let , and , , , ; estimate initial Section and .

[0071] 2) Prediction and mapping: using Will the current Convert to executable pulses / duty cycles and perform executable predictions within the model.

[0072] 3) Residuals and Jacobi: Assembly Calculation or approximation (The time-delay structure gives preference to explicit partial derivatives).

[0073] 4) LM Update: Solve according to the aforementioned formula. ,calculate And adaptive adjustment .

[0074] 5) Structured projection: Apply amplitude / slope / minimum start / stop / minimum interval / hard limit; enter safe mode if necessary.

[0075] 6) Issuance and Rollback: Issuing several steps of instructions (MPC style) to the front end, then returning to step 2) until the stability criterion is met or the task ends.

Claims

1. A method for optimizing the input sequence of a large time-delay actuator for a marine vehicle, characterized in that: Based on the parameter identification model of environment-dynamic coupling, the input is shifted and predicted according to the actuator time delay; a unified cost function is constructed that includes task error, action number penalty, non-zero duration penalty, energy consumption / wear surrogate term and safety boundary hard / soft constraint term; an improved LM algorithm with adaptive damping factor update is used to iteratively solve the finite time domain input sequence; the obtained sequence is converted into a printable pulse / duty / opening command through an executable mapper, and the amplitude and other conditions are satisfied through structured projection.

2. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 1, characterized in that, Record the job output as The discrete sampling period is The number of time-domain steps is planned. The actuator time delay .

3. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 1, characterized in that, The parameter identification model based on environment-dynamic coupling is as follows: In the formula, State; This represents the input after time delay propagation; Parameters obtained from online / offline identification; For disturbance; For the system state transition function / discrete-time dynamics update operator, used to describe the parameter... With disturbance Under the influence of the action, the current state Delayed input Driven to the next moment state The evolutionary pattern; For observation / output mapping functions, used to map states... With parameters Calculate measurable output; To output the measured quantity or the performance output that is of interest for control.

4. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 3, characterized in that, Introducing an input mapper Ideal input The mapping is used to generate pulse / duty cycle / opening trajectory that can be sent out. The same mapping is used for both prediction simulation and actual sending out, ensuring consistency between executable prediction and executable sending out.

5. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 4, characterized in that, Performance characteristics include: upper and lower limits of amplitude. ,in, Input the minimum allowable amplitude to the actuator; Input the maximum allowable amplitude to the actuator; , Input the upper limit of the rate of change / the threshold of the slope; minimum start and stop time. , The minimum duration that must be maintained in a single "on" or "off" state; minimum interval. ,in, It is the minimum time interval between two consecutive non-zero pulses. The scaling factor or time discretization conversion factor is used to characterize the minimum margin constraint; quantization and dead zone: introducing a threshold. With quantization step size This compresses small disturbances into zero motion, ensuring sparse motion and avoiding high-frequency jitter.

6. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 1, characterized in that, When constructing the unified cost function, a single criterion function is constructed to uniformly weigh the key indicators, as shown in the following equation: For arbitrary precision, where, This refers to the terminal error weighting coefficient; To predict the output corresponding to the time-domain terminal step; The desired output / target value; This refers to the process error weighting coefficient; To predict / optimize the time domain length; Let be the number of actions, where Weight the penalty for the action count; This is the scaling factor or normalization factor for the count term; The indicator function is implemented using a smooth approximation. ; This is the dead zone threshold; The duration is non-zero, where, To determine the duration of the penalty weight, The sampling period is the distance from the walk; For wear / energy consumption agents, where, The total variation penalty weight; Input amplitude penalty weight; For soft constraints, a quadratic penalty or barrier function is used, where, Weighting of penalties for safety; This is the safety cost function / soft constraint penalty function, used to measure the degree to which the predicted state and input violate safety constraints; the greater the violation, the larger the penalty value. Adaptive near / far field weighting: when Time reduction , To reduce constraints on motion sparsity and accelerate approximation; when Time increases , To suppress over-control and chattering, among which, The far-field error threshold. The near-field error threshold is given, and it satisfies the following conditions: .

7. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 1, characterized in that, The improved LM algorithm iteratively solves for finite-time-domain input sequences, including: Will Written as residual least squares Construct residual stacking: In the formula, To output the tracking residual; The smoothed approximate residuals are the number of actions. The residual is penalized for non-zero duration; The total variation residuals; Input amplitude Punish residuals; For safety soft constraint residuals; Using an improved LM iteration: In the formula, For the first The input sequence vector for the next iteration; The Jacobian matrix of the residuals with respect to the input sequence ; For the first The LM damping coefficient of the next iteration; It is the identity matrix; In order to be in The residual vector calculated at the location; Damping adaptive: In the formula, This is the ratio of the actual decrease to the model-predicted decrease. For the first The cost function value for the next iteration; For the first The update step size is calculated in the next iteration; The damping coefficient; It is the residual vector; like : Approaching Newtonian steps, increasing speed. in, The current damping coefficient, This is the damping scaling factor; like : Switch to conservative gradient step, and reach steady state; Jacobian acquisition: analytical sensitivity / numerical difference / automatic differentiation; explicit partial derivatives are provided for time-delay chain structures to maintain block sparsity and accelerate the solution; Robustness: Action counting, TV, Using Huber / pseudo Reweighting reduces the impact of sharp non-smooth terms on convergence; if necessary, Armijo backtracking can be used as a fallback.

8. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 7, characterized in that, Perform structured projection after each iteration : Amplitude / slope limiting: point-by-point cropping and TV smoothing; Minimum start / stop / minimum interval: Merge or delete adjacent small pulses to ensure minimum start / stop time and minimum interval; Safety hard constraints: When a prediction is violated, subsequent pulses are contracted and the corresponding soft constraint weights are increased.

9. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 1, characterized in that, The method supports rolling optimizations based on fixed periods or event triggers.

10. The method for optimizing the input sequence of a large time-delay actuator for a marine vehicle as described in claim 9, characterized in that, The rolling optimization includes: Rolling horizon: every Focusing solely on optimizing the frontier M-step reduces real-time computing power. ; Triggering condition: When In the event of drastic environmental changes or a decrease in safety margin, immediate replanning should be implemented, including... For model-based The predicted value of the output obtained under the current input sequence; The trigger threshold / maximum allowed prediction deviation threshold; Parameters updated online: We employ robust RLS / EKF / IV-LS small-step updates to avoid "mistaken model - miscontrol".

Images