Method for accurate control of dredging and filling flow rate

Abstract

The application discloses a kind of dredging hydraulic filling flow accurate control method, belong to dredging engineering and fluid conveying automatic control technical field.For existing method is difficult to adapt to solid-liquid two-phase flow dynamic characteristics, disturbance lag and energy consumption is high problem, the method includes: real-time acquisition pipeline pressure, flow, concentration, mud pump speed torque and underwater topography point cloud;With acquisition signal driving by transient flow differential equation and the dynamic hybrid model of network construction by Gated Recurrent Unit online evolution forms digital twin, and utilizes meta-learning strategy adaptive update time-varying parameter;According to topography calculation expected flow, with flow deviation and system energy consumption minimization as target, optimal speed and valve opening sequence are solved using multi-objective evolutionary algorithm;Disturbance observer is constructed to estimate resistance disturbance and is converted into feedforward speed compensation;Feedforward compensation and target speed are superimposed to generate final instruction, and valve opening is synchronously output.The method is used for dredging hydraulic filling flow accurate control, improves flatness and energy efficiency.

Description

Technical Field

[0001] This invention belongs to the field of dredging engineering and automatic control technology of fluid transport, and specifically relates to a method for precise control of dredging and reclamation flow. Background Technology

[0002] During dredging and reclamation construction, the flow control of the pipeline transportation system directly affects the flatness of the reclamation area and the construction efficiency. Existing flow control methods are usually based on steady-state hydraulic models, treating the slurry as a homogeneous single-phase fluid and using classical friction formulas to calculate pressure loss along the flow path. However, the actual flow transported in the dredging pipeline is a solid-liquid two-phase flow, whose density, viscosity, and rheological properties vary significantly with factors such as slurry concentration, particle size, pipe diameter, and transportation distance. Simultaneously, there is a coupling effect between the speed regulation of the slurry pump and the opening adjustment of the slurry discharge pipe outlet valve. Traditional proportional-integral-derivative (PID) controllers struggle to handle this multivariable, highly nonlinear dynamic process, easily leading to large flow overshoot, response lag, or continuous oscillations.

[0003] On the other hand, the topography of reclamation areas is constantly changing, and the required flow rate needs to be adjusted in real time according to the underwater topography. However, existing systems often use fixed flow rate setpoints, which cannot dynamically allocate flow rate according to the actual filling progress of each block in the reclamation area, resulting in overfilling or underfilling in some areas. In the control loop, volumetric concentration sensors are usually installed at the pump outlet or along the pipeline. Their measurement signals contain high-frequency noise, and there is a significant time lag in the propagation of concentration disturbances from the measurement point to the mud pump inlet. When the mud concentration changes abruptly (e.g., high-concentration mud or clear water layer is sucked in), traditional feedback control can only respond after the disturbance has already affected the flow rate, resulting in untimely compensation and easily causing pipeline blockage or flow interruption accidents.

[0004] Furthermore, most existing control methods only focus on minimizing flow deviation as the sole optimization objective, neglecting the energy consumption of the mud pump drive and valve throttling losses, resulting in high overall energy consumption of the conveying system. In practical engineering, construction personnel often rely on experience to manually adjust the mud pump speed and valve opening, which is cumbersome and difficult to guarantee control accuracy. To overcome these problems, those skilled in the art have attempted to use model predictive control, but the lack of a dynamic model that can accurately describe the transient process of solid-liquid two-phase flow, and the difficulty in identifying the time-varying friction coefficient and slurry rheological parameters in the pipeline online, limits the effectiveness of predictive control. At the same time, the high-frequency noise and transmission delay of the concentration signal further increase the difficulty of disturbance suppression. Therefore, there is an urgent need for a precise flow control method that can adapt to the dynamic characteristics of solid-liquid two-phase flow, correct model parameters online, coordinate multi-objective optimization, and compensate for the time delay of concentration disturbances. Summary of the Invention

[0005] This invention provides a method for precise control of dredging and reclamation flow rate, which can adapt to the dynamic characteristics of solid-liquid two-phase flow, correct time-varying friction coefficient and rheological parameters online, achieve joint optimization of flow rate deviation and total system energy consumption, and compensate for the time delay of concentration disturbance transmission, thereby improving the smoothness of reclamation and construction energy efficiency.

[0006] To achieve these objectives and other advantages of the present invention, a method for precise control of dredging and reclamation flow rate is provided, comprising the following steps: S1. Real-time acquisition of pressure, mud flow rate, volume concentration, and mud pump speed and torque signals at multiple sections along the dredging and conveying pipeline, while acquiring underwater topographic point cloud data of the reclamation area; S2. Using pressure, slurry flow rate, volume concentration, pump speed, and torque signals as inputs, the dynamic mixing model of solid-liquid two-phase flow, constructed from the transient flow fundamental differential equation and a gated cyclic unit network, is driven to evolve online, forming a digital twin of the pipeline transportation process. The output of the gated cyclic unit network is the predicted value of pressure and slurry velocity at each cross-section along the pipeline. The transient flow fundamental differential equation uses this predicted value as a state variable, calculates its physical residual, and uses it as a constraint for model evolution. A meta-learning strategy guided by physical information is used to adaptively update the time-varying friction coefficient and slurry rheological parameters in the dynamic mixing model. S3. Calculate the required filling rate for each block of the reclamation area based on the underwater topographic point cloud data, and convert it into the expected transport flow rate; take minimizing the deviation between the expected transport flow rate and the actual flow rate and minimizing the total energy consumption of the transport system as the joint optimization objectives, use the digital twin as the prediction model to perform multi-step flow trajectory prediction in the finite prediction time domain, and use a decomposition-based multi-objective evolutionary algorithm to solve for the Pareto optimal mud pump target speed sequence and mud discharge pipe outlet valve opening sequence; S4. Construct a nonlinear disturbance observer based on the real-time volume concentration signal to estimate the pipeline resistance disturbance caused by concentration change, and convert it into feedforward speed compensation based on the mud pump head-flow characteristics. S5. The feedforward speed compensation amount is superimposed with the first control amount in the target speed sequence of the mud pump to generate the final mud pump speed command, and output to the actuator synchronously with the first control amount in the valve opening sequence; in the next control cycle, return to step S1 and repeat the above process.

[0007] Preferably, the physical information-guided meta-learning strategy in step S2 specifically includes: In the offline phase, a multi-source historical working condition dataset covering different mud types, pipe diameters, and conveying distances is used to train the gated recurrent unit network using a model-independent meta-learning algorithm to obtain the initial meta-network weights. The topology of the gated recurrent unit network is as follows: the input layer receives time window sequences of pressure, mud flow rate, volume concentration, mud pump speed, and torque; the hidden layer uses stacked gated recurrent units; and the output layer simultaneously outputs the pressure prediction vector and slurry velocity prediction vector for each cross-section along the pipeline. During the online update phase, a composite loss function is constructed within the sliding time window. This composite loss function consists of a weighted sum of two terms: the first term is the mean square error between the network-predicted and measured values ​​of mud flow and pressure; the second term is a physical residual regularization term, which is the squared L2 of the residual of the transient flow fundamental differential equation on the predicted values ​​of the gated cyclic unit network. The pressure, velocity, and density state variables in the transient flow fundamental differential equation are directly substituted into the corresponding predicted values ​​output by the network to calculate the partial differential equation residuals at each discrete cross-section, and the sum of the squared residuals of all cross-sections is used as the regularization term. Using the initial network weights as a starting point, the gradient of the composite loss function with respect to the input weight matrix and bias vector in the top-level gated unit of the gated recurrent unit network is calculated, and gradient descent is performed to update the network according to a preset learning rate and a preset number of iterations K, where K is no greater than 5, and the parameters of the remaining network layers are kept frozen. When performing automatic differentiation using a digital twin in step S3, the digital twin used is the complete model after the above online update, which includes the frozen bottom-level network parameters and the updated top-level parameters.

[0008] Preferably, in step S4, before constructing the nonlinear perturbation observer based on the real-time volume concentration signal, a conditioning step for the volume concentration signal is further included, specifically: The volume concentration signal is input into the improved adaptive noise complete set empirical mode decomposition module and decomposed into multiple intrinsic mode functions; Calculate the permutation entropy value of each intrinsic mode function, identify intrinsic mode functions whose permutation entropy value exceeds the preset noise threshold as noise dominant components and remove them, and superimpose the remaining intrinsic mode functions to reconstruct a denoised volume concentration signal; The denoised volume concentration signal is input into the nonlinear perturbation observer, and the gain parameter of the nonlinear perturbation observer is adjusted online according to the current absolute value of the rate of change of the denoised volume concentration signal. The larger the absolute value of the rate of change, the larger the gain value.

[0009] Preferably, when converting the mud pump head-flow characteristics into feedforward speed compensation in step S4, a time-delay lead compensation operation is also included: Using the digital twin that evolves online in step S2, the average slurry velocity v of the cross section where the volume concentration measurement point is located is calculated online based on its real-time updated slurry velocity distribution in the pipeline; or, the convection equation or correlation analysis in the digital twin is used directly to estimate the material transport lag time τ required for the solid phase concentration disturbance to be transmitted from the measurement point to the mud pump inlet along the pipeline. The laying distance L from the volume concentration measurement point along the pipeline axis to the mud pump inlet is obtained, and the lag time of the concentration disturbance is calculated according to τ = L / v. When the direct estimation method is used, τ is obtained online by the digital twin by solving the one-dimensional convection equation or by input-output cross-correlation analysis. Construct a disturbance propagation predictor: The volume concentration signal detected at the current time t is used to obtain the current disturbance estimate through a nonlinear disturbance observer. The estimate is then input into the gated cyclic unit network. The pressure, flow rate, and concentration sequence within the current sliding time window are used as inputs. The end hidden state is taken, and the gated cyclic unit network is used to directly perform τ-step multi-step recursive prediction to obtain the flow deviation that will be caused when the disturbance propagates to the mud pump inlet at time t+τ. Based on the predicted deviation, the feedforward speed compensation is calculated in advance according to the pump head-flow characteristics and output to the speed controller at the current time t. When τ is less than the preset threshold, linear extrapolation is used to approximate the multi-step recursion of the gated loop unit.

[0010] Preferably, when using a decomposition-based multi-objective evolutionary algorithm to solve the problem in step S3, additional real-time enhancement processing is applied, specifically: Pareto nondominated solution sets covering typical dredging conditions are stored in the prior solution set library of the shipboard controller in advance. At least 50% of the individuals in the initial population of the online solution are randomly selected from this prior solution set library, and the remaining individuals are randomly generated in the decision variable space. The joint optimization objective is decomposed into multiple single-objective sub-problems using a reference vector-guided Chebyshev decomposition strategy. During the iterative solution of each sub-problem, the automatic differentiation function of the digital twin is used to calculate the Jacobian matrix of the flow rate and pressure output against the input sequences of the mud pump speed and valve opening. This Jacobian matrix information is used to guide the search direction of the crossover and mutation operators. The digital twin refers to the complete dynamic hybrid model containing all network layer parameters after online updating in step S2, and its automatic differentiation is based on the current parameter values.

[0011] Preferably, the objective function expression of the joint optimization objective is a weighted sum of four terms: The first term is the integral of the square of the deviation between the expected and actual flow rates in the prediction time domain. The second term is the estimated total power consumption of the conveying system, which includes the sum of the input power of the mud pump drive motor and the throttling loss power of the mud discharge pipe outlet valve. The third term is the sum of the squares of the differences between adjacent moments in the mud pump speed command sequence. The fourth term is the sum of the squares of the differences between adjacent moments in the mud discharge pipe outlet valve opening command sequence. The weighting coefficients of the four terms are preset according to the current construction stage of the dredging vessel. When solving the optimization problem, additional constraints on the maximum rate of change of the mud pump speed and the maximum rate of change of the valve opening are applied as hard constraints.

[0012] Preferably, the method for obtaining underwater topographic point cloud data of the reclamation area in step S1 is as follows: The point cloud of the reclamation area above water and shallow water is obtained by scanning with a lidar mounted on a drone in the first cycle, and underwater topographic profile data is obtained by a shipborne multibeam echo sounder in the second cycle, which is shorter than the first cycle. Before calculating the expected transport flow rate in step S3, the lidar point cloud and multibeam echo sounder data are time-stamped and spatially registered. The adaptive Kalman filter algorithm is used to recursively update the elevation value and estimated variance of each grid in the digital elevation model of the reclamation area. Based on the difference between the updated digital elevation model and the design elevation, the required filling rate of each block is recalculated and converted into the expected transport flow rate.

[0013] Preferably, step S1 also involves collecting vibration signals from the mud pump bearing housing and three-phase current signals from the drive motor; the method further includes online assessment and model correction steps for the mud pump's health status. The time-series segments of the vibration and current signals are input into a pre-trained convolutional autoencoder, which is trained using historical vibration and current data under the healthy state of the mud pump and includes an encoder and a decoder. The root mean square error between the input signal and the reconstructed output signal of the convolutional autoencoder is calculated in real time. When the root mean square error exceeds the preset degradation threshold for 10 consecutive monitoring windows, model correction is triggered. The correction method is as follows: Modify the mud pump head-flow characteristic curve H = aQ in the digital twin described in step S2. 2 All coefficients (a, b, c) of + bQ+ c are multiplied by the same preset attenuation factor γ, where 0 < γ < 1; During subsequent normal control, the measured pressure and flow rates collected in real time at the mud pump outlet are used to perform online adaptive calibration of γ using the recursive least squares method, so as to minimize the sum of squared errors between the model output and the measured values; the calibrated γ is used for updating the digital twin in the next control cycle. When γ is below 0.85 after three consecutive calibrations, a mud pump maintenance alarm will be issued.

[0014] Preferably, in step S5, before generating the final mud pump speed command and outputting it to the actuator, a mud pump physical constraint protection step is also performed: Calculate the difference between the final mud pump speed command and the actual output speed command of the previous control cycle. If the absolute value of the difference exceeds the maximum acceleration / deceleration rate allowed by the frequency conversion drive system at the current speed of the mud pump, then truncate the difference according to the maximum acceleration / deceleration rate to obtain the restricted speed command. If the restricted speed command exceeds the upper or lower limit of the rated speed of the mud pump, it will be clamped to the corresponding limit. When cutoff or clamping occurs, the equivalent flow difference is calculated based on the restricted speed difference using the combined flow characteristic curve of the mud pump and the outlet valve. The equivalent flow difference is then converted into a valve opening compensation increment and superimposed on the first control quantity of the valve opening sequence obtained in step S3 to obtain the compensated valve opening command.

[0015] Preferably, during each control cycle of the iterative execution of steps S2 to S5, pipeline safety monitoring and emergency protection steps are also performed: The pressure values ​​measured by the pressure sensors at each cross-section in step S1 are acquired in real time, and the average pressure drop rate of each cross-section within a 2-second time window is calculated. When the pressure drop rate at any section exceeds the preset safety threshold of 0.5 MPa / s, the optimization solution in step S3 and the disturbance observation calculation in step S4 of the current cycle are immediately interrupted. The mud pump speed command is forcibly set to the preset safe low speed, the mud discharge pipe outlet valve opening command is set to the preset minimum safe opening, and an alarm message containing the over-limit section location code is sent to the dredging monitoring system via industrial Ethernet.

[0016] The present invention has at least the following beneficial effects: First, this invention addresses the shortcomings of traditional steady-state models in capturing pressure wave propagation and velocity transients within pipelines by constructing a dynamic mixing model of solid-liquid two-phase flow composed of transient flow fundamental differential equations and a gated cyclic unit network, and by evolving it online to form a digital twin. A meta-learning strategy guided by physical information adaptively updates the time-varying friction coefficient and rheological parameters, enabling the model to track changes in parameters such as mud concentration and particle size. Simultaneously, the expected flow rate of each block is calculated based on the real-time underwater topography of the reclamation area, and a multi-objective evolutionary algorithm is used to simultaneously optimize the mud pump speed and valve opening, using flow deviation and total system energy consumption as joint optimization objectives, avoiding energy waste caused by single-objective optimization. The introduction of a nonlinear disturbance observer allows for feedforward compensation of resistance disturbances caused by concentration changes, improving the response speed and steady-state accuracy of flow control.

[0017] Secondly, this invention addresses the two major challenges of efficiency and physical consistency in online updates of digital twins by proposing a meta-learning strategy guided by physical information. In the offline phase, a model-independent meta-learning algorithm is used to train on multi-source historical conditions. The resulting initial network weights possess the ability to quickly adapt to new conditions, avoiding the need for extensive data retraining for each online update. The composite loss function constructed during online updates includes two terms: data fitting error ensures consistency between the network output and the measured signal; the physical residual regularization term forces the network output to satisfy the fundamental differential equation of transient flow, preventing non-physical predictions that might arise from a purely data-driven model. Online updates only perform a finite-step gradient descent on a small number of parameters in the top-level gating unit, while keeping the parameters of other network layers frozen, significantly reducing the online computational load. Simultaneously, the digital twin used for automatic differentiation and Jacobian matrix calculation in the multi-objective evolutionary algorithm is the complete model after the aforementioned online update. Its automatic differentiation is based on all current parameter values, ensuring the accuracy of sensitivity information.

[0018] Third, this invention decomposes the original volume concentration signal into multiple intrinsic mode functions (EMFs) through an improved adaptive noise-complete ensemble empirical mode decomposition (AMD). It then uses permutation entropy to quantitatively assess the randomness and complexity of each component, accurately identifying and eliminating noise-dominated high-frequency components while preserving the low-frequency trend of the true concentration change. Compared to traditional low-pass filters, this method does not introduce phase lag and can adaptively handle signals with different signal-to-noise ratios. After the reconstructed denoised signal is input into the nonlinear disturbance observer, the observer's gain parameter is adjusted online based on the absolute value of the signal change rate: the gain is increased when the change rate is large to improve disturbance tracking speed, and decreased when the change rate is small to suppress measurement noise amplification. This design enables the disturbance observer to output smooth and accurate disturbance estimates under both stable and drastic concentration fluctuations, thereby improving the reliability of feedforward compensation.

[0019] Fourth, this invention utilizes a digital twin to identify the average slurry velocity *v* under the current operating conditions online, accurately calculating the lag time τ = L / v for concentration disturbances transmitted from the measurement point to the mud pump inlet. The constructed disturbance transmission predictor inputs the estimated concentration disturbance detected at the current moment into the gated recurrent unit network. This network directly performs τ-step multi-step recursive prediction to obtain the flow deviation caused by the disturbance propagating to the mud pump inlet, and outputs the feedforward speed compensation amount at the current moment accordingly. Compared to traditional methods that require resolving partial differential equations, this invention directly reuses the pre-trained gated recurrent unit network for recursive prediction, significantly reducing computational costs. Simultaneously, this recursive process utilizes the solid-liquid two-phase flow time evolution law learned by the network during the offline meta-learning stage, eliminating the need for additional solution of transient flow differential equations. When the lag time τ is less than a preset threshold, a linear extrapolation approximation can be used to further reduce the computational load. This time-delay advance synchronization mechanism ensures that the compensation effect coincides exactly with the arrival time of the disturbance, thus offsetting the inherent lag effect of traditional feedback control, suppressing flow fluctuations caused by sudden changes in concentration, and reducing the risk of pipeline blockage and flow interruption.

[0020] Fifth, this invention employs a UAV-based lidar and a shipborne multibeam echo sounder to collaboratively measure the underwater topography of the reclamation area. The lidar uses a longer scanning cycle to obtain point clouds over a wide area of ​​the surface and shallow water, while the multibeam echo sounder provides local underwater topographic profile data with a shorter cycle. After spatiotemporal registration, the two are fused using an adaptive Kalman filter algorithm. The Kalman filter recursively updates the elevation value and estimated variance of each grid cell in the digital elevation model using the observations from the multibeam data, ensuring a reasonable elevation estimate is maintained even in areas without actual multibeam data, based on the model predictions from the previous cycle. This fusion method leverages both the wide coverage of lidar and the high accuracy and rapid updates of multibeam data, achieving continuous and accurate monitoring of the reclamation area's topography. The filling rate of each block is recalculated based on the difference between the real-time updated digital elevation model and the design elevation, and converted into the desired transport flow rate, providing precise flow rate setpoints for subsequent optimization.

[0021] Sixth, this invention proposes a simplified single-parameter correction scheme for changes in the head-flow characteristic curve caused by mud pump wear. When a convolutional autoencoder based on vibration and current signals detects mud pump performance degradation, all coefficients of the head-flow characteristic curve are multiplied by the same attenuation factor γ. Based on this, using real-time pressure and flow measurement values ​​collected at the mud pump outlet, a recursive least squares method is employed to perform online adaptive calibration of γ, enabling the single attenuation factor to dynamically approximate the overall scaling ratio under the current wear state. This method only requires identification of one parameter, eliminating the need for complex multi-parameter identification or pre-storing multiple models, resulting in extremely low hardware and computational costs. When the continuously calibrated γ falls below 0.85, an automatic maintenance alarm is issued, balancing model accuracy and engineering practicality.

[0022] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0023] Figure 1 This is a schematic flowchart of the method for precise control of dredging and reclamation flow rate according to the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail so that those skilled in the art can implement it based on the description.

[0025] It should be understood that terms such as “having,” “comprising,” and “including” as used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0026] like Figure 1 As shown, the present invention provides a method for precise control of dredging and reclamation flow rate, comprising: Step S1: Real-time acquisition of pressure, slurry flow rate, volumetric concentration, and pump speed and torque signals at multiple cross-sections along the dredging and conveying pipeline, while simultaneously acquiring underwater topographic point cloud data of the reclamation area. The pressure signal reflects the pressure distribution at different locations within the pipeline; the slurry flow rate and volumetric concentration are used to calculate the actual solid mass flow rate being transported; and the pump speed and torque characterize the pump's current operating point and input mechanical power. Underwater topographic point cloud data can be acquired in various ways, such as using a drone equipped with lidar to scan the above-water and shallow water areas of the reclamation area, combined with a shipborne multibeam echo sounder to obtain the deep-water topographic profile, or using a single-beam echo sounder combined with GPS positioning for gridded measurement. In practice, pressure sensors are typically installed at key cross-sections such as the pump outlet, the middle section of the pipeline, and the end of the discharge pipe, with two redundant sensors installed at each cross-section to improve reliability; slurry flow rate can be measured using an electromagnetic flowmeter or an ultrasonic flowmeter, and volumetric concentration can be measured using a gamma-ray concentration meter or a conductivity meter. The sampling frequency of all signals should be higher than twice the highest frequency of the pressure wave inside the pipe, generally set to 10 Hz to 50 Hz, depending on the pipe length and wave velocity. As a reference, when the pipe length is about two kilometers and the wave velocity is about one kilometer per second, the pipe fundamental frequency is about 0.25 Hz, and a sampling frequency of 10 Hz would be sufficient. The acquisition period of underwater topographic point clouds can be decoupled from the control period, for example, the topographic data can be updated every 30 seconds, while the control period is usually set to 1 second.

[0027] Step S2: Using pressure, slurry flow rate, volumetric concentration, pump speed, and torque signals as inputs, the dynamic mixing model of solid-liquid two-phase flow, constructed from the transient flow fundamental differential equation and a gated recurrent unit network, is driven to evolve online, forming a digital twin of the pipeline transportation process. The transient flow fundamental differential equation, namely the continuity equation and momentum equation, describes the propagation of pressure waves within the pipeline and the variation of flow velocity with time and space, accurately characterizing the behavior of solid-liquid two-phase flow under non-constant conditions. The gated recurrent unit network is a lightweight recurrent neural network adept at learning dynamic features from time-series data; its output is the predicted values ​​of pressure and slurry velocity at each cross-section along the pipeline. The combination of the two is as follows: based on the input signal at the current moment, including pressure, flow rate, and concentration data within the historical time window, the gated recurrent unit network outputs the predicted values ​​of pressure and velocity at each cross-section at the next or current moment. These predicted values ​​are then substituted into the transient flow fundamental differential equation to calculate the residuals on both sides of the equation. If the predicted values ​​perfectly conform to physical laws, the residual should be zero; otherwise, the residual is non-zero. The residual is fed back into the network training process as a physical constraint, forcing the network output to simultaneously satisfy data consistency and physical consistency. To adapt to the time-varying characteristics of parameters such as mud concentration, particle size distribution, and pipe wall roughness during dredging, this method adopts a physical information-guided meta-learning strategy to adaptively update the time-varying friction coefficient and slurry rheological parameters in the dynamic hybrid model. In specific implementation, multiple historical datasets under different working conditions can be used offline, such as data covering different mud types, pipe diameters, and conveying distances. The initial weights of the gated recurrent unit network are trained through a model-independent meta-learning algorithm, enabling these initial weights to quickly adapt to new working conditions after a small number of gradient updates. During online operation, a composite loss function is constructed within the sliding time window. This loss function includes two terms: the first term is the mean square error between the network's predicted values ​​and the measured pressure and flow rates, and the second term is the L2 squared of the residual of the transient flow fundamental differential equation. By minimizing the composite loss function and performing gradient descent updates on only a limited number of parameters at the top layer of the network—for example, no more than five steps—time-varying parameters can be updated online, thus maintaining a high degree of consistency between the digital twin and the actual pipeline dynamics. In step S3, when performing automatic differentiation using the digital twin, the digital twin used is the complete model after the aforementioned online update, containing the frozen bottom-level network parameters and the updated top-level parameters.

[0028] Step S3: Calculate the required filling rate for each block of the reclamation area based on underwater topographic point cloud data, and convert it into the expected transport flow rate. Specifically, the reclamation area is divided into several grid blocks. The design elevation of each block is known. The difference between the current elevation and the design elevation is calculated based on the latest acquired point cloud data. Combined with the scheduled completion time of the block in the construction plan, the average filling rate required for that block can be obtained. The filling rates of all blocks are summed by volume weight, and then multiplied by the density or solids content of the mud to obtain the expected volumetric flow rate or mass flow rate required for the entire system. Then, with the joint optimization objectives of minimizing the deviation between the expected transport flow rate and the actual flow rate and minimizing the total energy consumption of the transport system, the digital twin generated in step S2 is used as the prediction model to perform multi-step flow trajectory prediction within a finite prediction time domain, such as predicting the flow rate change in the next 10 to 30 seconds. Since this is a multi-objective optimization problem, traditional weighted methods are insufficient to obtain the Pareto front. Therefore, a decomposition-based multi-objective evolutionary algorithm is employed to solve for the Pareto optimal sequence of target pump speeds and discharge pipe outlet valve openings. The basic principle of this algorithm is to decompose the multi-objective optimization problem into multiple single-objective subproblems, each corresponding to a weight vector. Through evolutionary operations such as crossover, mutation, and selection, all subproblems are optimized simultaneously, ultimately yielding a set of non-dominated solutions. The solution that best suits the current construction preferences is then selected as the output. To improve the real-time performance of the online solution, a set of Pareto non-dominated solutions covering typical dredging conditions can be pre-stored in the shipboard controller and used as part of the initial population during online solving. Simultaneously, the automatic differentiation function of the digital twin is used to calculate the Jacobian matrix, guiding the search direction of the crossover and mutation operators, thereby accelerating convergence. The digital twin refers to the complete dynamic hybrid model containing all network layer parameters after online updating in step S2, and its automatic differentiation is based on all current parameter values.

[0029] Step S4: Construct a nonlinear disturbance observer based on the real-time volumetric concentration signal to estimate the pipeline resistance disturbance caused by concentration changes, and convert it into a feedforward speed compensation amount according to the mud pump head-flow characteristics. Since changes in mud concentration alter the density and viscosity of the slurry, they affect the pipeline friction loss, causing a shift in the delivered flow rate even when the mud pump speed remains constant. The nonlinear disturbance observer uses some dynamic equations from a system model, such as a digital twin, and measured speed and flow signals to calculate the equivalent resistance disturbance value currently acting on the system in real time. To improve observation accuracy, the volumetric concentration signal needs to be conditioned first. An improved adaptive noise complete ensemble empirical mode decomposition is used to decompose the original signal into multiple intrinsic mode functions (EMFs). The permutation entropy value of each mode component is calculated, and components with permutation entropy values ​​exceeding a preset noise threshold, such as 0.6, are identified as noise-dominant and removed. The remaining components are superimposed and reconstructed to obtain the denoised volumetric concentration signal. The denoised concentration signal is input into a nonlinear disturbance observer, and the observer's gain parameter is adjusted online based on the absolute value of the rate of change of the signal: the larger the absolute value of the rate of change, the more drastic the concentration fluctuation, and the gain is increased to improve the disturbance tracking speed; the smaller the rate of change, the gain is decreased to suppress measurement noise amplification. After obtaining the estimated resistance disturbance value, the head-flow characteristic curve of the mud pump is usually a quadratic function H = aQ. 2 + bQ + c, converting the disturbance into a speed increment that needs compensation. When the disturbance is positive, i.e., the concentration increases and the resistance increases, the speed needs to be increased to maintain the same flow rate; conversely, the speed should be decreased. Since the concentration sensor is usually installed at the mud pump outlet or along the pipeline, the concentration disturbance it detects needs to propagate with the slurry flow to the mud pump inlet before affecting the pump's operating point, resulting in a significant time lag. This lag time τ is equal to the pipeline length L from the measurement point to the mud pump inlet divided by the average slurry velocity v under the current operating conditions. To address this, a time lag advance compensation operation is added: using the online evolving digital twin in step S2, based on its real-time updated slurry velocity distribution within the pipeline, the average slurry velocity v at the cross section where the volume concentration measurement point is located is calculated online; or, the convection equations or correlation analysis in the digital twin are directly used to estimate the material transport lag time τ required for the solid phase concentration disturbance to travel from the measurement point along the pipeline to the mud pump inlet.

[0030] A disturbance propagation predictor is constructed: the volumetric concentration signal detected at the current time t is used by a nonlinear disturbance observer to obtain the current disturbance estimate. This estimate is input into the gated cyclic unit network. Using the pressure, flow rate, and concentration sequence within the current sliding time window as initial conditions, the gated cyclic unit network directly performs a multi-step recursive prediction for τ steps to obtain the flow rate deviation caused by the disturbance propagating to the sludge pump inlet at time t+τ. The recursive process of the gated cyclic unit network utilizes the time evolution law of solid-liquid two-phase flow learned during the offline meta-learning stage, eliminating the need to resolve the transient flow differential equation. Based on this predicted deviation, the feedforward speed compensation is pre-calculated according to the sludge pump head-flow characteristics and output to the speed controller at the current time t, ensuring that the compensation effect is synchronized with the arrival time of the disturbance. When τ is less than a preset threshold, linear extrapolation is used to approximate the multi-step recursive prediction of the gated cyclic unit.

[0031] Step S5: Superimpose the feedforward speed compensation amount with the first control quantity in the target speed sequence of the mud pump obtained in Step S3 to generate the final mud pump speed command, and output it synchronously to the actuator along with the first control quantity of the valve opening sequence. The target speed sequence is the first value used for the current moment in the multi-step optimization results, while the feedforward compensation amount is a correction to this value. The sum of the two yields the actual speed command. The first control quantity of the valve opening sequence is output directly. The actuator includes the variable frequency drive of the mud pump and the electric or hydraulic valve at the mud discharge pipe outlet. Before outputting, physical constraint protection must be considered: check whether the rate of change of the speed command relative to the previous cycle exceeds the maximum allowable acceleration / deceleration rate of the variable frequency drive system. As a reference, this maximum acceleration / deceleration rate can be set to no more than 10% of the rated speed per second; if it exceeds this, it is truncated. Check whether the speed command exceeds the upper or lower limit of the mud pump's rated speed, for example, 20% to 100% of the rated speed; if it exceeds this, it is clamped to the corresponding limit. When cutoff or clamping occurs, the restricted speed difference causes a deviation between the actual flow rate and the desired flow rate. In this case, the equivalent flow difference is calculated using the combined flow characteristic curve of the mud pump and the outlet valve, and converted into a compensating increment for the valve opening. This increment is then added to the original valve opening command to compensate for the flow loss as much as possible. Afterwards, in the next control cycle, the process returns to step S1 and repeats, forming a closed-loop real-time control.

[0032] Compared to traditional methods based on steady-state hydraulic models and single-loop PID control, this invention constructs a digital twin by integrating transient flow physical equations with a gated cyclic unit network. This accurately captures the pressure wave propagation and transient velocity characteristics of solid-liquid two-phase flow in pipelines, significantly improving the model's adaptability to time-varying friction and rheological parameters. Simultaneously, it incorporates real-time topographic data from the reclamation area to dynamically adjust the desired flow rate and combines multi-objective optimization to minimize flow deviation and total system energy consumption. Compared to single-objective control, this effectively reduces energy loss while ensuring the smoothness of the reclamation. The combination of a nonlinear disturbance observer and feedforward compensation rapidly suppresses resistance disturbances caused by concentration changes, exhibiting a significantly faster response than methods relying solely on feedback regulation.

[0033] In another technical solution, the specific implementation of the physical information-guided meta-learning strategy in step S2 is further defined. Offline stage: Using a multi-source historical operating condition dataset covering different mud types, pipe diameters, and transport distances, a model-independent meta-learning algorithm is used to train a gated recurrent unit network to obtain initial meta-network weights. The topology of the gated recurrent unit network is as follows: the input layer receives time window sequences of pressure, mud flow rate, volume concentration, mud pump speed, and torque; the hidden layer uses stacked gated recurrent units; and the output layer simultaneously outputs pressure prediction vectors and slurry velocity prediction vectors for each cross-section along the pipeline. The core idea of ​​the model-independent meta-learning algorithm is to train a set of initial network weights that can quickly adapt to new operating conditions after only a few gradient update steps. In specific implementation, the historical dataset should contain complete records of at least dozens of different operating conditions, and the sampling duration for each condition should cover the complete cycle of the pipeline transient process. The number of stacked gated loop units can be selected as two or three layers, with each layer containing 64 or 128 hidden units. The time window length can be determined based on the pipe length and wave velocity, for example, 5 to 10 seconds. As a preferred option, the hidden layer adopts a two-layer stacked structure with 128 units per layer and a time window length of 8 seconds.

[0034] Online Update Phase: A composite loss function is constructed within the sliding time window. This composite loss function consists of a weighted sum of two terms. The first term is the mean square error between the network predictions and measured values ​​of mud flow and pressure, used to ensure consistency between the network output and the actual sensor measurements. The second term is the physical residual regularization term, which is the squared L2 norm of the residuals of the transient flow fundamental differential equation over the predicted values ​​of the gated cyclic unit network. Specifically, the state variables such as pressure, velocity, and density from the transient flow fundamental differential equation are directly substituted into the corresponding predicted values ​​of the network output. The residuals of the partial differential equations at each discrete cross-section are calculated, and then the squared residuals of all cross-sections are summed as the regularization term. The weighting coefficients between the two terms can be adjusted according to the actual operating conditions. For example, the weight of the data term can be set to 1, and the weight of the regularization term can be set to a value between 0.1 and 0.5. As an optimal choice, the weight of the regularization term is 0.2. The length of the sliding time window can be set to twice the length of the time window, for example, 16 seconds, sliding one step per control cycle.

[0035] Parameter update method: Using the initial network weights obtained in the offline stage as a starting point, the gradient of the composite loss function with respect to the input weight matrix and bias vector in the top-level gated unit of the gated recurrent unit network is calculated. Gradient descent updates are then performed according to a preset learning rate and a preset number of iterations K, where K is no greater than 5. The parameters of the remaining network layers remain frozen. Furthermore, when performing automatic differentiation using a digital twin in step S3, the digital twin used is the complete model after the above online update, containing the frozen bottom-level network parameters and the updated top-level parameters. The principle of this design is that the bottom and middle layers of the gated recurrent unit network are responsible for extracting general time-series features, which have good generalization ability under different operating conditions, thus requiring no frequent updates. The top-level gated unit is directly related to the output, and its parameters are most sensitive to changes in operating conditions; updating only the top-level parameters is sufficient to achieve rapid adaptation to new operating conditions. The learning rate can be chosen to be a small value, such as 0.001 or 0.0005, and the number of iterations K can be chosen to be 3 or 4. As an optimal choice, the learning rate is 0.001, and K is 3. Online updates are performed once per control cycle or every few control cycles to keep the digital twin up-to-date.

[0036] Compared to traditional methods of training from scratch or fine-tuning the entire network online, this approach uses initial weights obtained through meta-learning, which can quickly adapt to new operating conditions, significantly reducing the amount of data and computation time required for online updates. Updating only a small number of top-level parameters and with minimal iterations results in extremely low online computational load, meeting the real-time requirements of dredging control. Simultaneously, the physical residual regularization term in the composite loss function forces the network output to satisfy the fundamental differential equation of transient flow, effectively avoiding non-physical predictions that might arise from purely data-driven models and improving the reliability of the digital twin under extreme or untrained conditions.

[0037] In another technical solution, the specific steps of conditioning the volume concentration signal in step S4 are performed before constructing the nonlinear perturbation observer. First, signal decomposition and noise removal are performed: the original volume concentration signal is input into an improved adaptive noise complete set empirical mode decomposition module, which decomposes it into multiple intrinsic mode functions (EMFs). This decomposition method is an improvement on traditional empirical mode decomposition, effectively solving the mode aliasing problem and reducing reconstruction errors. Each EMF obtained from the decomposition represents a component of the signal at different frequencies and time scales, arranged from high frequency to low frequency. In specific implementation, the decomposition stopping criterion can be set to the energy of the remaining signal being less than a certain percentage of the energy of the original signal, such as 5%, or the maximum number of decomposition layers can be set to 8 to 10 layers. Then, the permutation entropy value of each EMF is calculated. Permutation entropy is an indicator that measures the complexity and randomness of a time series; a higher value indicates a more random sequence and more noise components. EMFs with permutation entropy values ​​exceeding a preset noise threshold are identified as noise-dominant components and removed, while EMFs with permutation entropy values ​​below the threshold are considered valid signal components. As a possible reference, the noise threshold can be set between 0.5 and 0.7, with 0.6 being the most preferred. The remaining intrinsic mode functions are then superimposed to reconstruct the denoised volume concentration signal.

[0038] Online adjustment of observer gain: The denoised volumetric concentration signal is input into a nonlinear disturbance observer, and the gain parameter of this nonlinear disturbance observer is adjusted online based on the current absolute value of the rate of change of the denoised volumetric concentration signal. The adjustment rule is that the larger the absolute value of the rate of change, the larger the gain value. The basic form of the nonlinear disturbance observer is to calculate the disturbance estimate using measured speed and flow rate signals based on the system dynamic equations. The gain parameter determines the convergence speed and noise sensitivity of the disturbance estimate: the larger the gain, the faster the observer tracks the disturbance, but the more sensitive it is to measurement noise; the smaller the gain, the smoother the output, but the response is lagging. By dynamically adjusting the gain based on the rate of change of the concentration signal, a small gain can be used to suppress noise when the concentration is stable, and a large gain can be used to quickly track the disturbance when the concentration fluctuates sharply. The specific gain adjustment function can adopt a linear relationship or a piecewise linear relationship, for example, setting the gain as a proportional coefficient proportional to the absolute value of the rate of change multiplied by a reference gain. The absolute value of the rate of change can be obtained by numerically differentiating the denoised concentration signal and smoothing it through a first-order low-pass filter. As an optimal implementation, the base gain is set to 5, the scaling factor is set to 10, the upper limit of gain is set to 50, and the lower limit is set to 1.

[0039] Compared to directly using the original concentration signal or simple low-pass filtering, this scheme combines adaptive noise complete set empirical mode decomposition with permutation entropy to adaptively separate the noise-dominant component from the true concentration change component in the signal without introducing phase delay. The introduction of permutation entropy makes noise component identification independent of a preset frequency range, making it suitable for real-world conditions with dynamically changing signal-to-noise ratios. The observer gain is adjusted online based on the concentration change rate, resulting in good noise suppression under stable conditions and rapid tracking response under drastic fluctuations, thus improving the overall accuracy and robustness of disturbance estimation.

[0040] In another technical solution, step S4 involves a time-delay advance compensation operation when converting the disturbance into a feedforward speed compensation amount based on the slurry pump head-flow characteristics. Specifically, this involves online identification of the slurry velocity: using the digital twin evolved online in step S2, the average slurry velocity v at the cross-section where the volume concentration measurement point is located is identified online based on its real-time updated velocity distribution within the pipeline. In solid-liquid two-phase flow, the concentration disturbance, i.e., the change in the solid phase volume fraction, is carried and transported by the slurry flow, and its propagation speed is approximately equal to the average slurry velocity. This velocity depends on factors such as slurry pump speed, pipeline resistance, slurry density, and viscosity, and varies significantly with operating conditions; therefore, a fixed value cannot be used. By matching the measured flow signal with the velocity distribution predicted by the model, the digital twin can inversely calculate the average slurry velocity under the current conditions. In practice, the flow velocity can be obtained by directly dividing the instantaneous flow rate measured by an electromagnetic or ultrasonic flow meter by the pipe cross-sectional area, followed by low-pass filtering. Alternatively, the automatic differentiation and parameter estimation functions of a digital twin can be used to treat the slurry flow velocity as the parameter to be identified, and the flow velocity value can be updated online by minimizing the error between the measured and predicted flow rates. As a preferred option, flow velocity identification is performed every 5 to 10 control cycles to reduce the computational burden.

[0041] Calculation of lag time: Obtain the laying distance L from the volume concentration measurement point along the pipeline axis to the mud pump inlet, i.e., the actual length of the pipeline. Calculate the lag time τ for the concentration disturbance to travel from the measurement point to the mud pump, τ = L / v, where v is the average slurry velocity under the current operating conditions. In solid-liquid two-phase flow, the resistance disturbance caused by concentration change moves with the slurry at a velocity v; therefore, calculating the lag time based on the slurry velocity is physically reasonable. In practice, the distance L can be obtained from pipeline design drawings or measured via GPS. As a reference, for typical long-distance dredging pipelines, L can reach hundreds to thousands of meters, and the slurry velocity is usually between 1 and 5 meters per second; therefore, the lag time τ is generally in the range of tens to thousands of seconds. When τ is long, it is necessary to combine the predictive capabilities of the digital twin for advance compensation; if τ is less than one control cycle, it can be treated as one cycle.

[0042] Construction and Compensation of the Disturbance Transmission Predictor: The disturbance transmission predictor is constructed as follows: The volumetric concentration signal detected at the current time t is used by a nonlinear disturbance observer to obtain an estimated value of the current resistance disturbance, which manifests as an equivalent increment in the pipe friction coefficient. This estimated value is input into the gated cyclic unit network, using the pressure, flow rate, and concentration sequence within the current sliding time window as initial conditions. The gated cyclic unit network directly performs a multi-step recursive prediction for τ steps to obtain the flow deviation caused by the disturbance propagating to the sludge pump inlet at time t+τ. The recursive process of the gated cyclic unit network utilizes the solid-liquid two-phase flow time evolution law learned during the offline meta-learning stage, eliminating the need to resolve the transient flow differential equation. Based on this predicted deviation, the feedforward speed compensation is pre-calculated according to the sludge pump head-flow characteristics and output to the speed controller at the current time t, ensuring that the compensation effect is synchronized with the arrival time of the disturbance. As a low-cost alternative, when τ is less than a preset threshold, such as two control cycles, linear extrapolation can be directly used to approximate the multi-step recursive prediction of the gated cyclic unit.

[0043] Compared with traditional feedback control or feedforward control that does not consider transmission time delay, this scheme can accurately predict the time when concentration disturbances propagate to the mud pump inlet and the resulting flow deviation by identifying the slurry flow velocity online and constructing a disturbance transmission predictor. It can also output the speed compensation amount in advance, so that the compensation effect arrives synchronously with the disturbance, effectively offsetting the control lag caused by transmission time delay, significantly suppressing the flow fluctuations caused by sudden changes in concentration, and reducing the risk of pipeline blockage and flow interruption accidents.

[0044] In another technical solution, step S3 employs a real-time enhancement process during the solution process using a decomposition-based multi-objective evolutionary algorithm. Specifically, this involves the use of a priori solution set: Pareto non-dominated solutions covering typical dredging conditions are pre-stored in the priori solution set library of the shipboard controller. At least 50% of the individuals in the initial population for online solving are randomly selected from this priori solution set library, with the remaining individuals randomly generated within the decision variable space. Typical dredging conditions can be divided according to parameters such as mud concentration range, pipe length range, and pump characteristics. For each condition, offline solving yields a set of Pareto non-dominated solutions, representing the optimal trade-off between flow deviation and energy consumption under that condition. These solutions are stored in the priori solution set library. During online solving, the closest condition is retrieved from the library based on current condition parameters such as average concentration and pipe length, and individuals are randomly selected from this library as part of the initial population. The proportion of individuals from the priori solution set library in the initial population can be set to 50% to 70%. As an optimal choice, this proportion is 60%. The remaining individuals are randomly generated within the decision variable space to ensure population diversity and avoid getting trapped in local optima. The decision variables include the mud pump speed increment and valve opening increment for each control step in the prediction time domain, and their range is determined based on the equipment's rated values.

[0045] The use of the Jacobian matrix: A Chebyshev decomposition strategy guided by reference vectors decomposes the joint optimization objective into multiple single-objective subproblems. During the iterative solution of each subproblem, the automatic differentiation function of the digital twin is used to calculate the Jacobian matrix of the flow and pressure outputs with respect to the input sequences of pump speed and valve opening. This Jacobian matrix information is then used to guide the search direction of the crossover and mutation operators. The Chebyshev decomposition strategy transforms the multi-objective optimization problem into multiple scalar optimization subproblems, each corresponding to a reference vector. Its objective function is the maximum value of the weighted Chebyshev distances among the objectives. Automatic differentiation is a feature of digital twins, which can automatically calculate the partial derivatives of the output variables with respect to the input variables, i.e., the Jacobian matrix. This matrix reflects how small changes in speed or valve opening affect future flow and pressure; essentially, it is the system's local sensitivity information. Integrating this sensitivity information into the crossover and mutation operators of the evolutionary algorithm makes newly generated individuals more likely to improve towards the Pareto front, rather than blindly searching randomly. In practical implementation, the mutation length and direction can be adjusted based on the sign and magnitude of the Jacobian matrix during the mutation operation. This allows for a smaller mutation length on highly sensitive decision variables and a larger mutation length on less sensitive variables. The digital twin refers to the complete dynamic hybrid model containing all network layer parameters after online updating in step S2, with its automatic differentiation based on all current parameter values. Preferably, each subproblem undergoes a maximum of 50 iterations per generation, and the population size is set to 100.

[0046] For handling hard constraints, any individual that violates the single-step rate of change of rotational speed, the single-step rate of change of valve speed, the upper or lower limits of rotational speed, or the upper or lower limits of valve opening is directly eliminated. For the test vectors generated through crossover mutation, their Chebyshev objective values ​​are calculated, and individuals that are superior to those in the atomic problem are replaced. The reference point is updated every ten generations. The maximum number of iterations is set to fifty, or adjusted according to the actual computation time, ensuring that the total time does not exceed one control cycle. After termination, a compromise solution is selected from the current Pareto front, usually the individual that minimizes the sum of the normalized objectives, and the rotational speed increment and valve increment corresponding to the first control step of this solution are output.

[0047] Compared to multi-objective evolutionary algorithms that directly solve from a random initial population, this scheme provides high-quality initial individuals through a prior solution set library, reducing the number of iterations required for convergence. The Jacobian matrix obtained through automatic differentiation guides the search direction, avoiding blind random mutation and crossover, thus improving the quality of individuals in each generation and further accelerating the solution process. This combination enables the computationally intensive multi-objective evolutionary algorithm to complete optimization within a typical one- to two-second control cycle of a dredging control system, meeting the real-time requirements of engineering projects.

[0048] In another technical solution, the specific expression of the objective function of the joint optimization objective and the hard constraints during optimization are further defined. The objective function expression of the joint optimization objective is a weighted sum of four terms: The first term is the integral of the square of the deviation between the expected and actual flow rates in the prediction time domain, which ensures the accuracy of flow tracking. The second term is an estimate of the total power consumption of the conveying system, which includes the sum of the input power of the mud pump drive motor and the throttling loss power of the sludge discharge pipe outlet valve. The mud pump drive power can be estimated by multiplying the mud pump torque and speed and then dividing by the motor efficiency; the valve throttling loss power is estimated based on the pressure difference before and after the valve and the flow rate through the valve. The third term is the sum of the squares of the differences between adjacent moments in the mud pump speed command sequence, used to penalize frequent and large changes in speed to reduce mechanical shock and extend the life of the frequency converter. The fourth term is the sum of the squares of the differences between adjacent moments in the sludge discharge pipe outlet valve opening command sequence, used to penalize frequent valve operations and avoid excessive wear of the actuator. The weight coefficients of the four terms are preset according to the current construction stage of the dredging vessel. The construction stage can be divided into the start-up stage, the stable conveying stage, and the finishing stage. During the startup phase, the weight of the flow deviation term can be appropriately increased to ensure rapid achievement of the set flow rate. During the stable delivery phase, the weight of the energy consumption term can be increased to save energy. During the shutdown phase, the weight of the speed and valve change rate penalty terms can be increased to ensure smooth shutdown. As a possible reference, the sum of the four weight coefficients is normalized to 1. For the startup phase, the values ​​can be 0.5 for the first term, 0.2 for the second, 0.15 for the third, and 0.15 for the fourth. For the stable phase, the values ​​can be 0.3 for the first term, 0.4 for the second, 0.15 for the third, and 0.15 for the fourth. For the shutdown phase, the values ​​can be 0.2 for the first term, 0.2 for the second, 0.3 for the third, and 0.3 for the fourth. The optimal values ​​should be determined during on-site commissioning based on the specific equipment characteristics and construction requirements.

[0049] When solving the optimization problem, additional constraints on the maximum rate of change of pump speed and valve opening in a single step are applied as hard constraints. These constraints mean that during the optimization process, the difference in speed and valve opening between adjacent control steps in any candidate solution cannot exceed the preset maximum values; otherwise, the solution is considered infeasible and eliminated. The maximum rate of change of pump speed in a single step can be set according to the acceleration and deceleration capabilities of the variable frequency drive system, for example, a change not exceeding 10% of the rated speed per second. The maximum rate of change of valve opening in a single step can be set according to the response speed of the actuator, for example, a change not exceeding 20% ​​of full scale per second. These constraint values ​​should not be violated by the optimization algorithm, thus ensuring that the optimized control sequence is physically feasible in actual execution. As an optimal choice, for a typical large dredging pump, the maximum rate of change of pump speed in a single step is set to 8% of the rated speed, and the maximum rate of change of valve opening in a single step is set to 15% of full scale.

[0050] Compared to single-objective optimization that only considers flow deviation, this scheme introduces an energy consumption term, enabling the optimizer to proactively select more energy-efficient speed-opening combinations while ensuring flow tracking accuracy. This avoids energy waste caused by excessively increasing speed or closing valves to eliminate small deviations. The introduction of penalty terms for speed and valve change rate effectively suppresses drastic fluctuations in control commands and extends the service life of the actuators. Simultaneously, dynamically adjusting weighting coefficients according to the construction stage allows the control strategy to flexibly adapt to the focus of different operational phases. The presence of hard constraints ensures the executability of the optimization results, preventing execution failures or equipment damage due to commands exceeding equipment capabilities.

[0051] In another technical solution, the specific method for acquiring underwater topographic point cloud data of the reclamation area in step S1 and the calculation method for the expected transport flow are further defined. A lidar mounted on a drone acquires point clouds of the surface and shallow water areas of the reclamation area in a first-cycle scan, while an underwater topographic profile data is acquired in a second-cycle scan using a shipborne multibeam echo sounder, where the second-cycle is shorter than the first. The drone lidar can quickly acquire surface elevation data over a large area, but in areas where the water depth exceeds a certain limit, the laser cannot penetrate the water, therefore it is mainly used for surface and shallow water areas. The multibeam echo sounder, mounted on a survey vessel, acquires detailed underwater topographic profiles by emitting fan-shaped acoustic beams. It has high accuracy but low coverage efficiency, hence the shorter scan cycle. The first cycle can be set to 30 to 60 seconds, and the second cycle can be set to 5 to 10 seconds. As an optimal choice, the first cycle is 30 seconds, and the second cycle is 5 seconds. The two data sources are spatially complementary and updated at different frequencies in time.

[0052] Before calculating the expected transport flow in step S3, the lidar point cloud and multibeam echo sounder data are time-stamp aligned and spatially registered. An adaptive Kalman filter algorithm is then used, with the registered multibeam echo sounder data as the observation value, to recursively update the elevation values ​​and estimated variances of each grid cell in the digital elevation model of the reclamation area. Time-stamp alignment refers to aligning the two types of data to a unified time base according to the acquisition time. Spatial registration refers to transforming the lidar point cloud coordinate system and the multibeam echo sounder's hull coordinate system to the same geographic coordinate system, such as UTM coordinates. The adaptive Kalman filter is a recursive filtering algorithm that automatically adjusts the filter gain based on the quality of the observation values. The digital elevation model divides the reclamation area into a regular grid, with each grid cell storing the currently estimated elevation value and the uncertainty variance of that estimate. When new multibeam echo sounder data arrives, it is used as the observation value, and the elevation value and variance of the corresponding grid cell are updated according to the Kalman filter formula. The lidar point cloud data can be used for batch updates after each acquisition, or it can also be used as an observation value input to the filter. Because multibeam data updates more frequently, it can promptly correct digital elevation models, while the wide-area data provided by lidar ensures the integrity of the model in areas without multibeams. As a preferred option, the process noise variance in adaptive Kalman filtering can be dynamically adjusted according to the degree of terrain change.

[0053] The required filling rate for each block is recalculated based on the difference between the updated digital elevation model (DEM) and the design elevation, and then converted into the desired transport flow rate. The specific steps are as follows: First, the reclamation area is divided into M regular grid blocks. Let Ai be the area of ​​the i-th block. The current average elevation of this block is obtained from the DEM, and the design elevation is obtained from the design drawings. The difference between the two yields the remaining filling height; a negative value indicates overfilling and is set to zero. Then, based on the remaining allowable construction time for this block in the construction plan, the block area is multiplied by the remaining filling height and divided by the remaining time to obtain the required average volumetric filling rate for this block. The filling rates of all blocks are summed to obtain the solid volume increase rate for the entire reclamation area. Let the current average volumetric concentration of the mud be Cv, obtained by filtering the volumetric concentration signal. Dividing the solid volume increase rate by Cv yields the required mud volumetric flow rate. Considering the possibility of minor sedimentation or leakage during transport, a correction factor can be applied, typically between 1.0 and 1.2, calibrated based on historical experience, to finally obtain the desired transport flow rate. When the volume concentration is too low, such as below 5 percent or the solid phase volume increase rate is zero, the expected flow rate is forced to be set to the minimum safe flow rate of the system, usually 20% of the rated flow rate, to prevent pipe blockage.

[0054] Compared to relying solely on shipborne depth sounders or manual measurements, this solution achieves continuous, accurate, and efficient monitoring of the reclamation area's topography by integrating large-area, rapid scanning data from UAV lidar with high-precision local profile data from a shipborne multibeam echo sounder. Adaptive Kalman filtering combines the advantages of both data sources, maintaining reasonable elevation estimates even in areas with sparse multibeam data. Dynamically adjusting the desired transport flow rate based on a real-time updated digital elevation model allows the flow rate setpoint to adaptively respond to topographic changes, avoiding overfilling or underfilling issues caused by fixed flow rate setpoints and improving the smoothness of the reclamation area.

[0055] In another technical solution, the online assessment and model correction steps for the mud pump's health status are further defined. This step is based on the vibration and current signals additionally acquired in step S1. Step S1 also acquires vibration signals from the mud pump bearing housing and three-phase current signals from the drive motor. Vibration signals are typically acquired using accelerometers mounted on the bearing housings at both the drive and non-drive ends of the mud pump. The sampling frequency should be high enough to capture the characteristic frequencies of the rolling bearings, for example, 10 kHz. The three-phase current signals can be acquired using current transformers inside the frequency converter or external current clamps, with a sampling frequency of 1 kHz. A convolutional autoencoder is an unsupervised deep learning model consisting of an encoder and a decoder. The encoder compresses the high-dimensional input signal into a low-dimensional latent feature vector, while the decoder attempts to reconstruct the original input signal from this feature vector. This convolutional autoencoder is trained using historical vibration and current data from the mud pump's healthy state, i.e., only data from when the mud pump is in good condition is used. After training, the parameters of the encoder and decoder are fixed.

[0056] The root mean square error (RMSE) between the input signal and the reconstructed output signal of the convolutional autoencoder is calculated in real time. When this RMSE exceeds a preset degradation threshold for 10 consecutive monitoring windows, the pump's performance is considered to have degraded. The principle is that the convolutional autoencoder can only reconstruct signals similar to the training data relatively well. When the pump experiences abnormalities such as wear, imbalance, or bearing failure, the statistical characteristics of the vibration and current signals change, and the autoencoder's reconstruction error increases significantly. Therefore, the reconstruction RMSE can be used as a health indicator. The monitoring window length can be set to 1 second or 2 seconds, with the RMSE calculated once per window. The degradation threshold can be determined by adding a certain number of standard deviations to the historical mean RMSE under healthy conditions, for example, the mean plus 3 standard deviations. As a reference, the degradation threshold is set to 1.5 to 2 times the mean RMSE under healthy conditions, with 1.8 times being the optimal value. Exceeding the threshold for 10 consecutive windows avoids false triggering caused by single noise spikes.

[0057] The correction method is as follows: Modify the mud pump head-flow characteristic curve H = aQ in the digital twin described in step S2. 2All coefficients (a, b, c) of + bQ+ c are multiplied by the same preset attenuation factor γ, where 0 < γ < 1. Then, during subsequent normal control, the measured pressure and flow rates collected in real time at the mud pump outlet are used to perform online adaptive calibration of γ using the recursive least squares method, minimizing the sum of squared errors between the model output and the measured values. The calibrated γ is used for updating the digital twin in the next control cycle. When γ is below 0.85 after three consecutive calibrations, a mud pump maintenance alarm is issued.

[0058] Compared to methods that rely on periodic manual inspections or simply on runtime to estimate wear, this solution utilizes a convolutional autoencoder for unsupervised health assessment of vibration and current signals. This automatically identifies early performance degradation in the mud pump without needing to label fault samples. A continuous window decision mechanism for reconstructing error indices effectively avoids false alarms. Upon detecting degradation, the head-flow characteristic curve coefficients are automatically corrected, ensuring the digital twin matches the actual performance of the mud pump and preventing control deviations caused by model inaccuracies. Simultaneously, the attenuation factor γ is calibrated online using a recursive least squares method, enabling the model to adaptively track gradual changes in wear. This requires identifying only one parameter, resulting in extremely low hardware and computational costs. An automatic alarm is triggered when γ falls below 0.85, balancing model accuracy with engineering practicality.

[0059] In another technical solution, the physical constraint protection step for the mud pump, performed in step S5 before generating the final mud pump speed command and outputting it to the actuator, is further defined. Regarding speed change rate truncation: the difference between the final mud pump speed command and the actual output speed command of the previous control cycle is calculated. If the absolute value of this difference exceeds the maximum acceleration / deceleration rate allowed by the frequency converter drive system at the current mud pump speed, the difference is truncated according to the maximum acceleration / deceleration rate, resulting in a restricted speed command. The final mud pump speed command is obtained by superimposing the feedforward compensation amount in step S5 with the first value of the target speed sequence in step S3. The frequency converter drive system limits the speed change rate to prevent excessive current and mechanical shocks. The maximum acceleration / deceleration rate is usually related to the current speed; for example, a larger absolute value of the speed change rate is allowed in the low-speed region, and a smaller rate of change is allowed in the high-speed region. A maximum acceleration / deceleration rate curve related to speed can be pre-calibrated experimentally. The specific method of truncation is to reduce the excess portion to the maximum allowable value, i.e., the restricted speed command equals the speed command of the previous cycle plus or minus the maximum acceleration / deceleration rate multiplied by the control cycle. For example, if the speed in the previous cycle was 800 rpm, the current final command is 900 rpm, the control cycle is 1 second, and the maximum acceleration / deceleration rate is 50 rpm per second, then the actual speed can only increase by 50 rpm, and the limited speed command is 850 rpm.

[0060] For speed limit clamping: If the speed limit command exceeds the upper or lower limit of the mud pump's rated speed, the speed will be clamped to the corresponding limit. Mud pumps have a minimum operating speed, typically 20% to 30% of the rated speed; operating below this speed may cause vibration or poor lubrication. The maximum speed is 100% to 105% of the rated speed; exceeding this speed may damage the impeller or bearings. Therefore, the speed limit command is compared with the upper and lower limits. If it is greater than the upper limit, it is set to the upper limit; if it is less than the lower limit, it is set to the lower limit. As an optimal choice, the lower speed limit is set to 25% of the rated speed, and the upper speed limit is set to 100% of the rated speed.

[0061] For valve opening compensation: When cutoff or clamping occurs, the equivalent flow difference is calculated using the combined flow characteristic curve of the mud pump and outlet valve based on the restricted speed difference. This equivalent flow difference is converted into a valve opening compensation increment and superimposed on the first control variable of the valve opening sequence obtained in step S3 to obtain the compensated valve opening command. The restricted speed difference refers to the difference between the original final speed command and the actual speed command after restriction and clamping. Based on the mud pump head-flow characteristic curve and pipeline resistance characteristics, the flow change caused by this speed difference, i.e., the equivalent flow difference, can be estimated. Then, based on the valve flow characteristics, i.e., the relationship curve between valve opening and flow, the valve opening change required to compensate for this flow difference is calculated as the compensation increment. Note that the change in valve opening also affects the flow, and there may be nonlinearity and coupling; therefore, the calculation of the compensation increment can use an approximate linearization method. The superimposed valve opening command also needs to be subject to rate of change constraints and upper and lower limit constraints, but this is not explicitly required in this technical solution; similar handling can be used in actual implementation. After compensation, the flow loss caused by the speed limit is partially or completely made up by the valve opening adjustment, thus maintaining the flow tracking accuracy. As a preferred option, if the compensated valve opening command exceeds the allowable range, valve safety is prioritized, and partial compensation is abandoned.

[0062] Compared to directly executing unconstrained speed commands, this solution effectively prevents the variable frequency drive system from tripping due to excessive acceleration or deceleration, and the mud pump from being damaged due to overspeed or underspeed, through physical constraint protection steps. When constraints are unavoidable, the restricted speed difference is converted into valve opening adjustment through a valve opening compensation mechanism, maximizing the tracking accuracy of the desired flow rate and avoiding flow control failure caused by simple cutoff or clamping. This collaborative compensation method fully utilizes the redundant degrees of freedom of both the valve and pump actuators, improving the fault tolerance and flow maintenance capabilities of the control system.

[0063] In another technical solution, the pipeline safety monitoring and emergency protection steps are further defined to be executed in parallel within each control cycle. For real-time calculation of the pressure drop rate: the pressure values ​​measured by the pressure sensors at each cross-section in step S1 are acquired in real time, and the average pressure drop rate of each cross-section within a 2-second time window is calculated. Pressure sensors are typically installed at key cross-sections along the pipeline, such as the mud pump outlet, pipeline midpoint, high point, low point, and near the mud discharge pipe outlet. The pressure values ​​at each cross-section are continuously read according to the sampling frequency, and then, within a sliding time window of 2 seconds, the pressure data are linearly fitted or differentially averaged to obtain the average pressure drop rate within that window. The average pressure drop rate can filter out instantaneous noise spikes and reflect the trend of pressure changes. As a specific implementation method, the least squares method can be used to perform linear regression on the pressure-time data within the window, and the absolute value of the slope is the average drop rate. Alternatively, the pressure difference between the start and end times of the window can be divided by the window length. Considering that the pressure signal may contain noise, the linear regression method is recommended.

[0064] For emergency protection triggering and action: When the pressure drop rate at any section exceeds the preset safety threshold of 0.5 MPa per second, the optimization solution in step S3 and the disturbance observation calculation in step S4 of the current cycle are immediately interrupted. The mud pump speed command is forcibly set to the preset safe low speed, and the opening command of the mud discharge pipe outlet valve is set to the preset minimum safe opening. An alarm message containing the location code of the exceeded section is sent to the dredging monitoring system via industrial Ethernet. The safety threshold of 0.5 MPa per second is a reference value and can be adjusted according to the pipe material, wall thickness, design pressure, etc. For example, it can be set to 0.8 MPa per second for high-strength steel pipes. Rapid pressure drop usually indicates dangerous conditions such as pipe rupture, seal failure, pump inlet flow interruption, or valve accidental opening. Continuing to perform complex optimization solutions and disturbance observation calculations at this time is a waste of time, and emergency measures must be taken immediately. The safe low speed can be set to 20% to 30% of the rated speed of the mud pump, and the minimum safe opening can be set to 5% to 10% of the full valve opening to ensure that a certain flow rate is maintained in the pipeline to prevent blockage, while reducing pressure and flow rate. The alarm message should include the number or location coordinates of the section that exceeded the limit, as well as the current pressure value and rate of pressure drop, to facilitate quick fault location by operators. After emergency protection is activated, the control system can enter manual mode or remain in a safe state awaiting operator intervention. Alternatively, once the pressure stabilizes and is manually confirmed, the normal control process can be restarted.

[0065] Compared to systems without dedicated safety monitoring or relying solely on slow responses from conventional control loops, this solution rapidly identifies hazardous conditions such as pipeline rupture, leakage, or cavitation by monitoring the pressure drop rate at each cross-section in real time. The average rate calculation within a 2-second time window effectively filters out transient interference, preventing false triggers. Once a safety threshold is exceeded, the system immediately interrupts the computationally intensive optimization and observation processes, forcibly switching to a preset safe state; the response speed is significantly faster than the convergence speed of conventional control. Simultaneously, alarm messages provide operators with clear fault location information, facilitating timely response. This emergency protection mechanism significantly improves the safety of dredging and reclamation operations, effectively preventing serious accidents such as pipeline bursts, collapses, and mud pump cavitation.

[0066] The number of devices and processing scale described herein are for the purpose of simplifying the description of the invention. Applications, modifications, and variations of the invention will be readily apparent to those skilled in the art.

[0067] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details.

Claims

1. A method for precise control of dredging and reclamation flow rate, characterized in that, Includes the following steps: S1. Real-time acquisition of pressure, mud flow rate, volume concentration, and mud pump speed and torque signals at multiple sections along the dredging and conveying pipeline, while acquiring underwater topographic point cloud data of the reclamation area; S2. Using pressure, slurry flow rate, volume concentration, pump speed, and torque signals as inputs, the dynamic mixing model of solid-liquid two-phase flow, constructed from the transient flow fundamental differential equation and a gated cyclic unit network, is driven to evolve online, forming a digital twin of the pipeline transportation process. The output of the gated cyclic unit network is the predicted value of pressure and slurry velocity at each cross-section along the pipeline. The transient flow fundamental differential equation uses this predicted value as a state variable, calculates its physical residual, and uses it as a constraint for model evolution. A meta-learning strategy guided by physical information is used to adaptively update the time-varying friction coefficient and slurry rheological parameters in the dynamic mixing model. S3. Calculate the required filling rate for each block of the reclamation area based on the underwater topographic point cloud data, and convert it into the expected transport flow rate; take minimizing the deviation between the expected transport flow rate and the actual flow rate and minimizing the total energy consumption of the transport system as the joint optimization objectives, use the digital twin as the prediction model to perform multi-step flow trajectory prediction in the finite prediction time domain, and use a decomposition-based multi-objective evolutionary algorithm to solve for the Pareto optimal mud pump target speed sequence and mud discharge pipe outlet valve opening sequence; S4. Construct a nonlinear disturbance observer based on the real-time volume concentration signal to estimate the pipeline resistance disturbance caused by concentration change, and convert it into feedforward speed compensation based on the mud pump head-flow characteristics. S5. The feedforward speed compensation amount is superimposed with the first control amount in the target speed sequence of the mud pump to generate the final mud pump speed command, and output to the actuator synchronously with the first control amount in the valve opening sequence; in the next control cycle, return to step S1 and repeat the above process.

2. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, The physical information-guided meta-learning strategy described in step S2 specifically includes: In the offline phase, a multi-source historical working condition dataset covering different mud types, pipe diameters, and conveying distances is used to train the gated recurrent unit network using a model-independent meta-learning algorithm to obtain the initial meta-network weights. The topology of the gated recurrent unit network is as follows: the input layer receives time window sequences of pressure, mud flow rate, volume concentration, mud pump speed, and torque; the hidden layer uses stacked gated recurrent units; and the output layer simultaneously outputs the pressure prediction vector and slurry velocity prediction vector for each cross-section along the pipeline. During the online update phase, a composite loss function is constructed within the sliding time window. This composite loss function consists of a weighted sum of two terms: the first term is the mean square error between the network-predicted and measured values ​​of mud flow and pressure; the second term is a physical residual regularization term, which is the squared L2 of the residual of the transient flow fundamental differential equation on the predicted values ​​of the gated cyclic unit network. The pressure, velocity, and density state variables in the transient flow fundamental differential equation are directly substituted into the corresponding predicted values ​​output by the network to calculate the partial differential equation residuals at each discrete cross-section, and the sum of the squared residuals of all cross-sections is used as the regularization term. Using the initial network weights as a starting point, the gradient of the composite loss function with respect to the input weight matrix and bias vector in the top-level gated unit of the gated recurrent unit network is calculated, and gradient descent is performed to update the network according to a preset learning rate and a preset number of iterations K, where K is no greater than 5, and the parameters of the remaining network layers are kept frozen. When performing automatic differentiation using a digital twin in step S3, the digital twin used is the complete model after the above online update, which includes the frozen bottom-level network parameters and the updated top-level parameters.

3. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, In step S4, before constructing the nonlinear perturbation observer based on the real-time volume concentration signal, a conditioning step for the volume concentration signal is further included, specifically: The volume concentration signal is input into the improved adaptive noise complete set empirical mode decomposition module and decomposed into multiple intrinsic mode functions; Calculate the permutation entropy value of each intrinsic mode function, identify intrinsic mode functions whose permutation entropy value exceeds the preset noise threshold as noise dominant components and remove them, and superimpose the remaining intrinsic mode functions to reconstruct a denoised volume concentration signal; The denoised volume concentration signal is input into the nonlinear perturbation observer, and the gain parameter of the nonlinear perturbation observer is adjusted online according to the current absolute value of the rate of change of the denoised volume concentration signal. The larger the absolute value of the rate of change, the larger the gain value.

4. The method for precise control of dredging and reclamation flow rate as described in claim 3, characterized in that, When converting the mud pump head-flow characteristics into feedforward speed compensation in step S4, a time-delay lead compensation operation is also included: Using the digital twin that evolves online in step S2, the average slurry velocity v of the cross section where the volume concentration measurement point is located is calculated online based on its real-time updated slurry velocity distribution in the pipeline; or, the convection equation or correlation analysis in the digital twin is used directly to estimate the material transport lag time τ required for the solid phase concentration disturbance to be transmitted from the measurement point to the mud pump inlet along the pipeline. Obtain the laying distance L from the volume concentration measurement point along the pipeline axis to the mud pump inlet, and calculate the lag time of the concentration disturbance according to τ= L ​​ / v. When using the direct estimation method, τ is obtained online by the digital twin through solving the one-dimensional convection equation or through input-output cross-correlation analysis; Construct a disturbance propagation predictor: The volumetric concentration signal detected at the current time t is used to obtain the current disturbance estimate through a nonlinear disturbance observer. This estimate is then input into the gated cyclic unit network. Using the pressure, flow rate, and concentration sequence within the current sliding time window as input, and taking its end hidden state, the gated cyclic unit network is used to directly perform τ-step multi-step recursive prediction to obtain the flow rate deviation that will be caused when the disturbance propagates to the mud pump inlet at time t+τ. Based on the predicted deviation, the feedforward speed compensation amount is calculated in advance according to the mud pump head-flow characteristics and output to the speed controller at the current time t. When τ is less than the preset threshold, linear extrapolation is used to approximate the multi-step recursion of the gated loop unit.

5. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, In step S3, when using a decomposition-based multi-objective evolutionary algorithm to solve the problem, additional real-time enhancement processing is applied, specifically: Pareto nondominated solution sets covering typical dredging conditions are stored in the prior solution set library of the shipboard controller in advance. At least 50% of the individuals in the initial population of the online solution are randomly selected from this prior solution set library, and the remaining individuals are randomly generated in the decision variable space. The joint optimization objective is decomposed into multiple single-objective sub-problems using a reference vector-guided Chebyshev decomposition strategy. During the iterative solution of each sub-problem, the automatic differentiation function of the digital twin is used to calculate the Jacobian matrix of the flow rate and pressure output against the input sequences of the mud pump speed and valve opening. This Jacobian matrix information is used to guide the search direction of the crossover and mutation operators. The digital twin refers to the complete dynamic hybrid model containing all network layer parameters after online updating in step S2, and its automatic differentiation is based on the current parameter values.

6. The method for precise control of dredging and reclamation flow rate as described in claim 5, characterized in that, The objective function expression for the joint optimization objective is a weighted sum of four terms: The first term is the integral of the square of the deviation between the expected and actual flow rates in the prediction time domain. The second term is the estimated total power consumption of the conveying system, which includes the sum of the input power of the mud pump drive motor and the throttling loss power of the mud discharge pipe outlet valve. The third term is the sum of the squares of the differences between adjacent moments in the mud pump speed command sequence. The fourth term is the sum of the squares of the differences between adjacent moments in the mud discharge pipe outlet valve opening command sequence. The weighting coefficients of the four terms are preset according to the current construction stage of the dredging vessel. When solving the optimization problem, additional constraints on the maximum rate of change of the mud pump speed and the maximum rate of change of the valve opening are applied as hard constraints.

7. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, The method for obtaining underwater topographic point cloud data of the reclamation area in step S1 is as follows: The point cloud of the reclamation area above water and shallow water is obtained by scanning with a lidar mounted on a drone in the first cycle, and underwater topographic profile data is obtained by a shipborne multibeam echo sounder in the second cycle, which is shorter than the first cycle. Before calculating the expected transport flow rate in step S3, the lidar point cloud and multibeam echo sounder data are time-stamped and spatially registered. The adaptive Kalman filter algorithm is used to recursively update the elevation value and estimated variance of each grid in the digital elevation model of the reclamation area. Based on the difference between the updated digital elevation model and the design elevation, the required filling rate of each block is recalculated and converted into the expected transport flow rate.

8. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, Step S1 also involves collecting vibration signals from the mud pump bearing housing and three-phase current signals from the drive motor; the method further includes online assessment and model correction steps for the mud pump's health status. The time-series segments of the vibration and current signals are input into a pre-trained convolutional autoencoder, which is trained using historical vibration and current data under the healthy state of the mud pump and includes an encoder and a decoder. The root mean square error between the input signal and the reconstructed output signal of the convolutional autoencoder is calculated in real time. When the root mean square error exceeds the preset degradation threshold for 10 consecutive monitoring windows, model correction is triggered. The correction method is as follows: Modify the mud pump head-flow characteristic curve H = aQ in the digital twin described in step S2. 2 All coefficients (a, b, c) of + bQ + c are multiplied by the same preset attenuation factor γ, where 0 < γ < 1; During subsequent normal control, the measured pressure and flow rates collected in real time at the mud pump outlet are used to perform online adaptive calibration of γ using the recursive least squares method, so as to minimize the sum of squared errors between the model output and the measured values; the calibrated γ is used for updating the digital twin in the next control cycle. When γ is below 0.85 after three consecutive calibrations, a mud pump maintenance alarm will be issued.

9. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, In step S5, before generating the final mud pump speed command and outputting it to the actuator, a mud pump physical constraint protection step is also performed: Calculate the difference between the final mud pump speed command and the actual output speed command of the previous control cycle. If the absolute value of the difference exceeds the maximum acceleration / deceleration rate allowed by the frequency conversion drive system at the current speed of the mud pump, then truncate the difference according to the maximum acceleration / deceleration rate to obtain the restricted speed command. If the restricted speed command exceeds the upper or lower limit of the rated speed of the mud pump, it will be clamped to the corresponding limit. When cutoff or clamping occurs, the equivalent flow difference is calculated based on the restricted speed difference using the combined flow characteristic curve of the mud pump and the outlet valve. The equivalent flow difference is then converted into a valve opening compensation increment and superimposed on the first control quantity of the valve opening sequence obtained in step S3 to obtain the compensated valve opening command.

10. The method for precise control of dredging and reclamation flow rate as described in claim 1, characterized in that, During each control cycle of the iterations from steps S2 to S5, pipeline safety monitoring and emergency protection steps are also performed: The pressure values ​​measured by the pressure sensors at each cross-section in step S1 are acquired in real time, and the average pressure drop rate of each cross-section within a 2-second time window is calculated. When the pressure drop rate at any section exceeds the preset safety threshold of 0.5 MPa / s, the optimization solution in step S3 and the disturbance observation calculation in step S4 of the current cycle are immediately interrupted. The mud pump speed command is forcibly set to the preset safe low speed, the mud discharge pipe outlet valve opening command is set to the preset minimum safe opening, and an alarm message containing the over-limit section location code is sent to the dredging monitoring system via industrial Ethernet.

Images