Intelligent control system, method and equipment of underwater leveling machine and underwater leveling machine

Abstract

The application is divisional application, and the application number of the original application is 2025118241829. The invention provides an intelligent control system, method and equipment of an underwater leveling machine and the underwater leveling machine, and relates to the field of ocean engineering equipment and automatic control. The intelligent control system comprises a sensing subsystem configured to obtain working state data of the underwater leveling machine; the decision-making subsystem comprises a stage decision-making unit and a fuzzy crowd-sourcing optimization unit; the discretization module is configured to discretize the working process of the underwater leveling machine into a plurality of stages and determine frame parameters of the current stage based on working state data in each stage; the fuzzy crowd-sourcing optimization unit is configured to determine the liquid level variation of each target shipping space based on the frame parameters so as to generate a control instruction of the current stage; and the execution subsystem is configured to control the underwater leveling machine to work underwater based on the multi-stage control instruction. The system can improve the attitude control stability of the underwater leveling machine in a complex operation environment.

Description

[0001] This application is a divisional application. The original application has the application number 2025118241829, the application date is December 5, 2025, and the application title is Intelligent Control System, Method, Equipment and Underwater Leveling Machine. Technical Field

[0002] This application relates to the field of intelligent and automatic control technology for marine engineering equipment, specifically to an intelligent control system, method, equipment, and underwater leveling machine. Background Technology

[0003] In seabed screed leveling operations, the stability of the underwater screed's attitude control directly determines the construction accuracy and equipment safety. However, existing technologies have long faced the problem of unstable attitude control in practical applications, mainly manifested in phenomena such as continuous oscillation, sluggish response, or slow convergence during the adjustment process. The root cause of this problem lies in the insufficient coordination capability of the control system in multi-objective, strongly coupled environments.

[0004] Existing control strategies largely rely on local feedback mechanisms and lack global dynamic planning. For example, when a leveler tilts, the system only triggers the liquid level adjustment of the corresponding compartment based on the instantaneous attitude error, without considering the continuous process of submersion and leveling as a whole for phased optimization. This "single-point response" mode leads to conflicting adjustment actions in different compartments in terms of timing; for example, overlapping water injection and drainage operations cause buoyancy imbalance, resulting in repeated oscillations of the machine. Meanwhile, the difficulty of real-time optimization of high-dimensional decisions exacerbates instability. Levelers are typically equipped with multiple ballast tanks, and their liquid level adjustment involves coupled calculations of multiple parameters such as weight, center of buoyancy, and center of gravity. Traditional algorithms struggle to find the global optimum within a limited time, often falling into local optimization traps. The generated control commands may over-rely on a few compartments or ignore dynamic disturbances, making it difficult for the leveler to stabilize quickly under complex water flow or load changes.

[0005] In summary, existing technologies struggle to achieve rapid and stable control of the leveling machine's attitude due to issues such as a lack of global planning, insufficient real-time optimization capabilities, and disconnect from physical constraints. Summary of the Invention

[0006] Based on this, this application aims to provide an intelligent control system, method, device, and underwater leveling machine, which generates multi-stage control commands through multi-stage global planning and fuzzy swarm optimization to improve the stability of the underwater leveling machine's attitude control.

[0007] In a first aspect, this application provides an intelligent control system for an underwater leveling machine, comprising: a sensing subsystem configured to acquire working status data of the underwater leveling machine; a decision-making subsystem comprising a stage decision-making unit and a fuzzy swarm optimization unit; the decision-making subsystem configured to discretize the working process of the underwater leveling machine into multiple stages, and determine the framework parameters of the current stage based on the working status data within each stage; the fuzzy swarm optimization unit configured to determine the liquid level change of each target compartment based on the framework parameters under the condition of satisfying multiple preset constraints, and generate control instructions for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage; and an execution subsystem configured to receive multi-stage control instructions sent from the decision-making subsystem, and control the underwater leveling machine to perform underwater work based on the multi-stage control instructions.

[0008] According to one embodiment of this application, the stage decision unit is configured to discretize the working process of the underwater leveler into multiple stages based on time intervals or depth intervals.

[0009] According to one embodiment of this application, the stage decision unit is further configured to switch to the next stage in each stage upon triggering conditions such as reaching a fixed duration, an intermediate draft threshold, or an inclination angle threshold.

[0010] According to one embodiment of this application, the fuzzy swarm optimization unit is configured as follows: for each compartment, calculate the antecedent variables for fuzzy inference; wherein, the antecedent variables include correction alignment degree, lever arm efficiency, and liquid level margin, the correction alignment degree is calculated based on the consistency between the attitude gradient and the target correction direction, the lever arm efficiency is calculated based on the lever arm weight of the compartment for roll or pitch, and the liquid level margin is calculated based on the safety margin between the current liquid level and the limit; the antecedent variables are mapped to fuzzy sets through a predefined membership function, and a fuzzy rule base constructed based on domain knowledge is applied for inference to obtain the participation level representing the compartment; based on the participation level, determine whether the compartment is the target compartment.

[0011] According to one embodiment of this application, the fuzzy swarm optimization unit is further configured to use a fuzzy particle swarm optimization algorithm, along with constraints of hyperplane projection and boundary repair processing, to search for the optimal decision variable within a sparse search space composed of the framework parameters of the current stage; the optimal decision variable is the liquid level change of each target compartment.

[0012] According to one embodiment of this application, the multiple preset constraints on the fuzzy swarm optimization unit include at least one of attitude safety envelope constraints, mass conservation constraints, cabin geometry limit constraints, and pump and valve capability constraints.

[0013] According to one embodiment of this application, the decision subsystem is further configured to calculate the opening duration and action sequence of the pump or valve corresponding to each target compartment based on the sign and magnitude of the liquid level change in each target compartment, combined with the target hardware parameters of the underwater leveling machine, and generate corresponding control commands.

[0014] According to one embodiment of this application, the decision subsystem is further configured to monitor working status data when executing control commands. If the working status data indicates that the attitude or draft of the underwater leveler deviates from a preset threshold, a local correction operation is triggered to adjust the liquid level change and reissue the control command.

[0015] According to one embodiment of this application, the decision subsystem further includes a digital twin unit, a job suggestion unit, and a job control unit; the digital twin unit is configured to perform state estimation and three-dimensional terrain modeling based on the working state data to construct a digital twin model of the leveling machine's operation; the job suggestion unit is configured to generate job suggestions based on the working state data and the leveling machine's job tasks; and the job control unit is configured to control the execution subsystem to execute the job tasks based on the job suggestions.

[0016] According to one embodiment of this application, the work task includes at least a material placement task; after obtaining the material placement task and work status data, the work suggestion unit calculates the feasibility of material placement in the area using a material placement scoring function; recommends material thickness and density values ​​using fuzzy rules; predicts potential construction defect areas and high-risk advancement paths using a dynamic window algorithm and a dynamic cost weighting mechanism; and generates a material placement work suggestion including material quantity and advancement direction.

[0017] According to one embodiment of this application, the intelligent control system of the underwater leveling machine further includes an underwater measurement subsystem, which includes: a binocular camera device configured to acquire binocular visual images of the target water area; and an edge computing device configured to acquire point clouds of the target water area. The underwater measurement subsystem is installed on the underwater leveling machine, the binocular camera device is located at the front of the main frame of the underwater leveling machine, and the edge computing device is communicatively connected to the material laying device in the underwater leveling machine.

[0018] According to one embodiment of this application, an edge computing device is configured to calibrate and correct a binocular vision image to obtain the disparity corresponding to each pixel; extract the original grayscale information of the binocular vision image, and then enhance the binocular vision image to generate gradient information; calculate the matching cost based on the disparity of each pixel in the binocular vision image and the corresponding original grayscale information and gradient information, and construct a cost function; calculate the disparity map of the binocular vision image by combining the cost function with an improved SGM algorithm; and back-project the disparity map according to the camera parameters of the binocular camera device to obtain the point cloud of the target water area; wherein, the improved SGM algorithm includes: aggregating the cost function based on four directions; setting a dynamic adjustment strategy to control the penalty for small observation changes and the penalty for large disparity jumps; and performing confidence screening and hole filling on the aggregated disparity map.

[0019] According to one embodiment of this application, the decision subsystem further includes a material laying control unit; the material laying control unit is configured to: when the underwater screed is performing screeding construction, acquire the weight of the stone in the hopper of the underwater screed and the tilt angle of the hopper based on the weighing sensor and inclinometer of the sensing subsystem; determine the stone mass based on the stone mass in the hopper and the tilt angle of the hopper, and calculate the real-time material laying mass flow rate based on the time series data of the stone mass; acquire the current terrain elevation of the subgrade, compare the current terrain elevation of the subgrade with the pre-stored design target elevation, and calculate the terrain residual and the volume to be laid; generate material laying control instructions through an optimized decision algorithm based on the real-time material laying mass flow rate, terrain residual, volume to be laid, and predicted thickness distribution; wherein, the predicted thickness distribution is obtained by the target prediction model based on historical construction data and multimodal sensing information, and the material laying control instructions are used to adjust at least one of the following: the conveyor belt speed of the underwater screed, the opening degree of the material laying valve, or the moving speed of the material laying head.

[0020] According to one embodiment of this application, the optimization decision algorithm is constructed based on the SAC deep reinforcement learning algorithm; the state space of the agent of the SAC deep reinforcement learning algorithm includes a multimodal fusion state tensor composed of stone quality, real-time material flow rate, encoded terrain residual, material head pose, material head speed, and working parameters; the action space of the agent of the deep reinforcement learning framework is continuous and bounded, and the action space corresponds to the control commands of conveyor belt speed, material valve opening, or material head movement speed; the agent of the deep reinforcement learning framework is trained to optimize the material laying strategy by maximizing the comprehensive reward function.

[0021] According to one embodiment of this application, the sensing subsystem further includes multi-source mapping sensors deployed in the operating sea area for collecting multi-source sensing data.

[0022] According to one embodiment of this application, the decision subsystem further includes a smoothness assessment unit, configured to: receive multi-source sensor data; generate a multi-source fusion bed surface model based on the multi-source sensor data; use the multi-source fusion bed surface model as input to a target bed surface prediction model to output an elevation prediction map of the bed surface based on the target bed surface prediction model; and generate a smoothness evaluation index for the operating sea area based on the elevation prediction map. The target bed surface prediction model is obtained by training an initial surface prediction model, which is constructed using a convolutional neural network and a multi-head attention mechanism. The hyperparameters of the initial surface prediction model are optimized using an alpha evolution algorithm, and the loss function of the initial surface prediction model is the Huber function. The value of each grid point in the elevation prediction map represents the deviation of the bed surface at that location from the ideal design plane or target elevation.

[0023] This application also provides an intelligent control method for an underwater leveling machine, comprising: acquiring the working status data of the underwater leveling machine; discretizing the working process of the underwater leveling machine into multiple stages, and determining the framework parameters of the current stage based on the working status data within each stage; determining the liquid level change of each target compartment based on the framework parameters under the condition of satisfying multiple preset constraints, and generating control commands for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage; and controlling the underwater leveling machine to perform underwater work according to the multi-stage control commands.

[0024] Based on the same concept, this application also provides a computer device, which is installed on or connected in communication with an underwater leveling machine. The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the methods described above.

[0025] Based on the same concept, this application also provides an underwater leveling machine, which is equipped with the intelligent control system of the underwater leveling machine described above.

[0026] According to one embodiment of this application, the underwater leveling machine is a walking underwater leveling machine.

[0027] Compared with the prior art, the beneficial effects of this application are: the intelligent control system can decompose the complex continuous control process into a discrete and ordered sequence of stages through a multi-stage control strategy, so that the control system can transform from a passive instantaneous response to an active and predictive global optimization, and make fine decisions by using fuzzy crowd intelligence optimization. Through fuzzy reasoning, the system can intelligently select the set of target compartments that are most effective for current attitude correction, and within this simplified space, collaboratively calculate the optimal liquid level adjustment of each compartment to ensure that the buoyancy distribution is accurately matched to the current needs, which can improve the stability of the attitude control of the underwater leveling machine in complex operating environments. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the architecture of the intelligent control system of the underwater leveling machine provided in the embodiments of this application.

[0029] Figure 2 A schematic diagram of the main frame of the underwater leveling machine provided in this application.

[0030] Figure 3 This is a schematic diagram of the material spreading mechanism of the underwater leveling machine provided in the embodiments of this application.

[0031] Figure 4 This is a schematic diagram of the lower part of the upper material tube provided in an embodiment of this application.

[0032] Figure 5 This is a schematic diagram of the lower material tube provided in an embodiment of this application.

[0033] Figure 6 This is a schematic diagram of the connection between the upper and lower feed tubes provided in an embodiment of this application.

[0034] Figure 7 This is a schematic diagram of the entire workflow of the underwater measurement subsystem provided in the embodiments of this application.

[0035] Figure 8 This is a schematic diagram of the application process of the material spreading control unit provided in the embodiments of this application.

[0036] Figure 9 A schematic diagram of the framework of the target substrate surface elevation prediction model provided in the embodiments of this application.

[0037] Figure 10 This is a schematic diagram illustrating the specific process of the AE algorithm provided in the embodiments of this application. Detailed Implementation

[0038] The present application will now be described in further detail with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the subject matter of the present application to the following embodiments. All technologies implemented based on the content of the present application fall within the scope of protection of the present application.

[0039] Unless otherwise specified, the terms "upper," "lower," "left," "right," "center," "inner," "outer," and "side" used in the description of specific embodiments of this application to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is usually placed during use. These terms are merely for the purpose of facilitating the description of the solution in this application or simplifying the description in specific embodiments, so as to enable those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on this application.

[0040] In the description of the embodiments of this application, technical terms such as "first" and "second" only distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0041] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0042] Unless otherwise specified, the terms "upper," "lower," "left," "right," "center," "inner," "outer," and "side" used in the description of specific embodiments of this application to indicate orientation or positional relationships are based on the orientation or positional relationships shown in the accompanying drawings, or the orientation or positional relationship in which the product / equipment / device is usually placed during use. These terms are merely for the purpose of facilitating the description of the solution in this application or simplifying the description in specific embodiments, so as to enable those skilled in the art to quickly understand the solution, and do not indicate or imply that a particular device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship. Therefore, they should not be construed as limitations on this application.

[0043] In the description of the embodiments of this application, technical terms such as "first" and "second" only distinguish one entity or operation from another, and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary or secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0044] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0045] The intelligent control system for the underwater leveling machine provided in this embodiment aims to achieve intelligent management of the walking underwater leveling machine during buoyancy, attitude control, and leveling operations. Through multi-source sensing, intelligent decision-making, and precise execution, it improves operational efficiency, safety, and adaptability. Please refer to... Figure 1 , Figure 1 This is a schematic diagram of the architecture of the intelligent control system for an underwater leveling machine provided in an embodiment of this application. The intelligent control system 10 of the underwater leveling machine may include: The sensing subsystem 11 acquires the working status data of the underwater leveling machine; the decision subsystem 12 includes a stage decision unit and a fuzzy swarm optimization unit; the decision subsystem 12 is configured to discretize the working process of the underwater leveling machine into multiple stages, and determine the framework parameters of the current stage based on the working status data within each stage; the fuzzy swarm optimization unit is configured to determine the liquid level change of each target compartment based on the framework parameters under the condition of satisfying multiple preset constraints, and generate control commands for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage; the execution subsystem 13 is configured to receive the multi-stage control commands sent from the decision subsystem 12, and control the underwater leveling machine to perform underwater work based on the multi-stage control commands.

[0046] In this embodiment, the underwater leveling machine refers to an engineering device used for underwater foundation leveling operations, capable of leveling, elevation verification, and adjustment of riprap foundations underwater. The underwater leveling machine's structure may include multiple pontoons or compartments, with buoyancy and attitude controlled by adjusting the liquid level within the compartments. Specifically, the underwater leveling machine may be a walking underwater leveling machine, a modular leveling machine with adjustable buoyancy cylinders, a semi-submersible underwater leveling machine, etc.

[0047] The sensing subsystem 11 is used to acquire real-time operating status data of the underwater leveler under various working conditions from all directions and at high frequency. This subsystem is not composed of a single sensor, but is a complex network integrating multiple types of sensors. All these sensor data are transmitted and integrated at high speed and with high reliability through an industry-standard data bus (such as Modbus / TCP) to form a unified, real-time updated operating status database.

[0048] The decision subsystem 12 employs a hierarchical decision-making mechanism from macro-planning to micro-optimization, transforming the complex continuous control problem into a series of solvable discrete optimization problems. This is primarily achieved through two-level collaboration between stage decision units and fuzzy swarm optimization units. In practice, instead of planning all actions from start to finish at once, the continuous operation of the underwater leveler (e.g., diving from the surface to the operating depth and leveling its attitude) is discretized into several sequentially executed stages. These stages can be divided based on time intervals (e.g., 30, 60, 90, or 120 seconds per stage) or depth intervals (e.g., 1, 2, or 5 meters per stage). At the start of each stage, the stage decision unit determines the framework parameters of the current stage based on the latest operating status data provided by the perception subsystem 11. These framework parameters include the sub-objectives of the stage (e.g., reaching a certain intermediate draft or target inclination angle), the stage's state variables, and the stage's constraints.

[0049] Since the leveling machine may have numerous compartments, adjusting all compartments simultaneously is inefficient and could lead to instability. Therefore, the fuzzy swarm optimization unit receives the framework parameters output by the stage decision unit and first performs "fuzzy reasoning" based on domain knowledge to evaluate each compartment. Evaluation indicators may include: the "contribution direction" (correction alignment) of the compartment's liquid level adjustment to correcting the current attitude, the lever arm efficiency generated by its position, and the safety margin of the current liquid level. Through fuzzy logic calculation, the most suitable target compartments for adjustment in the current stage are selected, thereby narrowing the search space of the optimization problem and improving computational efficiency. After determining the target compartments, the unit initiates the Fuzzy Particle Swarm Optimization (FPSO) algorithm. Under the premise of satisfying multiple preset constraints (attitude safety envelope constraints, mass conservation constraints, compartment geometry limit constraints, pump and valve capacity constraints, etc.), a set of optimal liquid level changes is quickly found within the sparse search space formed by the target compartments. This set of solutions is a refined correction of the theoretical decision of the stage decision unit under actual engineering constraints. Ultimately, the unit maps the optimized liquid level change into a specific, executable sequence of pump and valve control commands, such as issuing commands to pumps with specific numbers, specifying their operating duration and flow rate.

[0050] Decision subsystem 12 can employ a rolling or forward dynamic programming (DP) strategy, with the k-th stage... (The set of participating cabins obtained through fuzzy filtering) serves as the decision dimension, and FPSO is used to find the local optimal decision. Specifically, at each stage k, the decision subsystem 12 focuses on the current stage and uses the set of participating cabins obtained through fuzzy filtering. The dimensionality of the decision space is reduced, and the Fuzzy Particle Swarm Optimization (FPSO) algorithm is invoked to quickly solve for a locally optimal decision at the current stage. To compensate for the potential global shortsightedness caused by this local optimization, a "to-go" lower bound heuristic function is introduced in decision subsystem 12 to predict the state from the current stage k. The minimum residual cost required to reach the final goal state from the starting point. Its calculation formula is:

[0051] in, Let this be the lower bound of the remaining cost in stage k. For the target total equivalent water volume (or the corresponding value of the target draft), if If the performance is worse than the known optimal upper bound, then backtracking / reorganization is triggered. The solution is then obtained again (ensuring near-global optimality).

[0052] The execution subsystem 13 is the final execution system for control commands. It receives multi-stage control commands from the decision subsystem 12 and drives the underlying hardware devices to perform actions, including controlling the start, stop, and speed of the variable frequency pump, adjusting the opening of the proportional solenoid valve, and operating the compressed air valve group, thereby precisely realizing the water injection or drainage operations of each target compartment. The entire system operates in a dynamic closed loop. After the execution subsystem 13 performs an action, the state of the underwater leveler changes accordingly. The sensing subsystem 11 immediately captures these changes (such as new draft, attitude angle, and liquid level) and feeds back the updated working status data to the decision subsystem 12. The decision subsystem 12 then determines whether the current stage objective has been achieved. If it has, it switches to the next stage; if there is a deviation, it may trigger a local correction or re-optimization of the current stage. This cycle repeats until the leveler finally reaches the predetermined operational floating state and attitude.

[0053] In some optional embodiments, the sensing subsystem 11 may include a multi-source sensing layer and a fusion and state estimation layer. The multi-source sensing layer is responsible for acquiring various physical parameters of the leveling machine in real time through a series of high-precision sensing devices. These devices include an inertial measurement unit (IMU), a combination of four-corner measuring tower prisms and a shore-based total station, depth or draft sensors, liquid level and pressure sensors for each compartment, valve position sensors, pump power and flow sensors, and a data bus (using the Modbus / TCP protocol) as a data transmission link. These devices together constitute a comprehensive monitoring network, providing raw input for subsequent data processing.

[0054] Specifically, the IMU is used to measure the three-dimensional attitude angles (such as roll angle θ and pitch angle φ) and angular velocity changes of the underwater leveler, reflecting the spatial orientation and rotational dynamics of the equipment. Prisms on the four-corner measuring towers work in conjunction with the shore-based total station to acquire the elevation and spatial coordinates of the four corner points of the leveler through optical measurement technology, providing high-precision position references for calibrating the overall attitude and displacement. Depth or draft sensors focus on monitoring the immersion depth of the entire machine in water, i.e., the draft value d. Liquid level and pressure sensors in each compartment provide real-time feedback on changes in water volume and pressure distribution within the compartment, thus indirectly reflecting the weight distribution and local buoyancy state of the leveler. Valve position sensors are used to detect the opening and closing status and operating position of control valves. Pump power and flow sensors monitor pump operating efficiency, including power consumption and media flow rate, to optimize energy consumption and system response. All these sensor data are integrated and transmitted in real-time via a Modbus / TCP data bus, ensuring seamless synchronization of multi-source information.

[0055] The fusion and state estimation layer is built upon the multi-source sensing layer, primarily utilizing data from key devices such as IMUs, total stations, and level sensors for calculations. The role of this layer is to fuse and process the raw sensor information, estimating the leveler's core state parameters through algorithmic models. For example, combining attitude and angular velocity data from the IMU with spatial coordinate information from the total station can correct and refine the measured values ​​of attitude angles θ and φ, reducing errors. Level sensor data is used to infer water volume changes in each compartment, and combined with draft information from depth sensors, the center of mass and center of buoyancy positions are dynamically calculated using a physical model, and the uncertainties of these estimates are assessed.

[0056] In some optional embodiments, the stage decision unit in the decision subsystem 12 uses a stage decision maker (Dynamic Programming, DP) to discretize the underwater leveling machine's operation into multiple stages, and determines the framework parameters of the current stage based on the working status data within each stage. In each stage, the system switches to the next stage upon reaching a trigger condition of a fixed duration, an intermediate draft threshold, or an inclination angle threshold.

[0057] Here, the framework parameters refer to the set of optimization variables calculated by the decision subsystem 12 based on the working status data at each stage, including the target draft. Target attitude angle The phase division strategy and cabin participation selection rules are used to guide the calculation of the fuzzy swarm optimization unit.

[0058] For example, the stage decision unit can discretize the underwater leveling machine's operation into multiple stages based on time intervals or depth intervals. In each stage, the stage state is represented by a vector consisting of the liquid level heights of all compartments at the current moment, denoted as . ,in, This represents the liquid level in the i-th compartment of the k-th stage. Based on the current state... Based on the analysis, the stage decision-making unit needs to formulate an "action plan" for this stage, i.e., stage decision-making. This decision is also a vector, with each component... This represents the amount of liquid level change that needs to be adjusted in the i-th compartment during this phase. A positive value usually indicates that water is injected into the compartment to increase weight and assist in descent, while a negative value indicates that water is drained from the compartment to reduce weight and promote ascent.

[0059] Within each decision stage k, the optimization algorithm, while searching for the optimal liquid level adjustment scheme, must ensure that its result satisfies a series of hard constraints. These constraints are the bottom line for ensuring the structural safety and operational feasibility of the leveling machine; any candidate scheme must meet these conditions before pursuing optimal performance indicators. The preset constraints are checked in real-time within each stage, including: 1) Liquid level limit: This constraint requires that the liquid level not be too low, causing the compartment to empty, nor too high, causing the compartment to overflow.

[0060] 2) Attitude limits: This constraint requires that the planned level adjustments for this phase be completed. Afterwards, the new attitude angle generated by the underwater leveling machine must be controlled within the safety envelope.

[0061] Water intake / drainage strategy: 1) Internal redistribution: This constraint requires that the sum of the water volumes corresponding to changes in the liquid level of all compartments must be zero, meaning that the water only moves within the compartments and does not exchange with the external environment.

[0062] 2) Clean water intake (for diving): (Given target volume for this stage). This constraint requires that the total volume of water to be taken from the environment (positive value) or the total volume of water to be drained (negative value) in this stage be used to change the total weight of the leveler, thereby enabling it to submerge or surface.

[0063] In order to evaluate different liquid level adjustment schemes in real time Regarding the influence of the underwater leveler's attitude θ and φ, this application uses an attitude approximation model to treat the underwater leveler as a rigid body, whose attitude is determined by the restoring torque generated by the weight of the ballast water in each compartment.

[0064] Approximate small angle below:

[0065]

[0066] in, For the first The horizontal coordinates of the cabin relative to its center of mass The restoring moment stiffness is obtained through bench / field identification.

[0067] The stage cost calculation function uses scalar values ​​to quantify and evaluate a decision. The overall advantages and disadvantages of (i.e., liquid level adjustment scheme).

[0068] When calculating the stage cost (time minimization + engineering penalty), let The pumps are connected in parallel:

[0069]

[0070] The time cost is decomposed into pumping time. Valve action time Two parts, This represents the upper limit of the available equivalent flow rate for this cabin. For the time delay of opening and closing a single compartment, This represents the number of non-zero operating cabins in this phase.

[0071] In addition to time as the main cost, the cost function provided in this application embodiment also introduces an engineering penalty term. This refers to the second-order difference penalty for the change in liquid level. The stage cost is defined as:

[0072] in, This is a penalty coefficient used to suppress drastic repetitions between adjacent stages. When the liquid level adjustment in the current stage differs significantly from that in the previous stage, this penalty value increases, thus guiding the optimization algorithm to tend to choose a scheme that more smoothly connects with the instructions of the previous stage.

[0073] In order to evaluate a decision The advantages and disadvantages of introducing a "time-based cost" function into the stage decision-making unit. Calculate and execute the current decision The estimated total time. This total time is a comprehensive indicator, encompassing not only the main pumping time but also the associated auxiliary operation time. Specifically, it mainly includes the following components: First, the time required for inter-compartment transfer and internal distribution, which is closely related to the absolute value of the liquid level adjustment in each compartment, the rated flow rate of the pumps, the number of pumps that can operate simultaneously, and system efficiency; second, the accumulated time delay of the opening and closing actions of all valves involved in the operation, as each valve requires a brief response time to open or close; in addition, if the control strategy allows for water exchange with the external environment (i.e., internal water transfer in a non-closed system), then the time for draining water from or drawing water from the outside compartment must also be included. Using time as the core cost indicator directly addresses the core engineering requirement of improving operational efficiency and shortening the overall cycle time, ensuring that the optimization process is always time-oriented.

[0074] For example, the stage decision-making unit can take minimizing the total buoyancy time as the primary objective, superimposed with secondary terms such as energy consumption or number of actions. The primary objective draws on the "three-segment time" decomposition in ballast optimization:

[0075] in, For inter-cabin transfer and internal allocation of time (with each cabin) Pump flow rate Number of pumps ,efficiency (Related) Cumulative valve opening and closing time (compared to valve opening and closing delay) With "whether to take action" instruction (Related) External discharge / external water intake time (included if water exchange with the environment is permitted; zero for pure internal redistribution). Optional penalties can include energy consumption / action count / cabin switching penalties. To suppress frequent start-stop cycles and liquid level oscillations (to improve pump and valve lifespan and stability), among which, This is a non-negative weighting coefficient, set by the control system designer based on specific hardware characteristics (such as the tolerance of pumps and valves) and operational requirements. It determines the importance of this penalty term in the total cost function. The core of the penalty term is the sum of squares of the control vector u. In the context of leveling machine control, vector u represents the liquid level change command for all compartments within a given period. Therefore, Essentially, it is the sum of the squares of all control command amplitudes in this stage. The underlying reason for introducing this penalty is that excessive liquid level adjustment means that the pump needs to operate at high intensity and high flow rate in a short period of time. This not only significantly increases the instantaneous energy consumption of the system, but also causes greater mechanical stress on the pump, valves and related pipelines, accelerating equipment aging. It is also a non-negative weighting coefficient used to adjust the strength of the penalty for this term; the core of the penalty term. It is the sum of the squares of the accelerations of the liquid level changes, characterizing the rate of change of the liquid level itself, i.e., the degree of jitter or abrupt change in the control command. For example, a decision may cause the drainage rate of a compartment to increase or decrease sharply in a short period of time, resulting in significant [various effects]. This penalty, introduced to suppress frequent oscillations and drastic fluctuations in control signals, is crucial. In practical control systems, excessively frequent start / stop signals or rapidly changing flow commands can easily trigger oscillations in actuators, leading to decreased system stability. By penalizing high-frequency abrupt changes, the optimization algorithm tends to generate smoother, more coherent control sequences, which is beneficial for the stable operation of pumps and valves and extends their service life.

[0076] In some optional embodiments, before generating control commands for each stage, the fuzzy swarm optimization unit performs participation determination for each cabin based on fuzzy inference to determine the implementation process of the target cabin, which may include: For each compartment, the antecedent variables for fuzzy inference are calculated. These antecedent variables include correction alignment, lever arm effectiveness, and level margin. Correction alignment is calculated based on the consistency between the attitude gradient and the target correction direction. Lever arm effectiveness is calculated based on the lever arm weight of the compartment for roll or pitch. Level margin is calculated based on the safety margin between the current level and the limit. The antecedent variables are mapped to fuzzy sets using a predefined membership function, and a fuzzy rule base built based on domain knowledge is applied for inference to obtain the participation level representing the compartment. The participation level determines whether the compartment is the target compartment.

[0077] Correct alignment The direction of the cabin's contribution to attitude correction can be characterized by the consistency between the attitude gradient and the desired correction towards zero.

[0078]

[0079]

[0080]

[0081] in, Let be the theoretical correction efficiency of the i-th compartment for correcting the current attitude deviation. Let i be the liquid level in the i-th compartment. The density of water, It is the acceleration due to gravity. Let be the horizontal cross-sectional area of ​​the i-th compartment. and Let be the horizontal coordinate of the center of gravity of the i-th compartment in the coordinate system of the leveling machine. and This represents the restoring moment stiffness coefficient of the underwater leveling machine in the lateral (about the x-axis) and longitudinal (about the y-axis) directions.

[0082] Lever arm efficiency This indicates the lever arm weight of the compartment with respect to roll / pitch:

[0083] (Percentiles are 30% and 70%) Liquid level margin This indicates the minimum safety margin from the upper and lower limits:

[0084] Predefined membership functions transform precise numerical values ​​(i.e., antecedent variables) into "fuzzy language" concepts that computers can process. These predefined membership functions can be triangular or trapezoidal functions. For example, the predefined membership functions provided in this application include: : Reverse / Normal / Same direction (Center 0 / 0.5 / 1).

[0085] Small / Medium / Large (using sample quantiles to define a, b, c).

[0086] Low / Medium / High (typical threshold 0.15 / 0.4 / 1.0).

[0087] For correcting alignment This application defines three fuzzy sets in its embodiments: "reverse", "general", and "same direction". For lever arm efficiency... Its fuzzy set is "small", "medium", and "large". For level margin... The fuzzy sets are categorized as "low," "medium," and "high." After fuzzification, calculations are performed using a fuzzy rule base built upon domain knowledge. For example, calculating the fuzzy set for each cabin... After mapping to "participation level" (non-participation / medium / high priority) via member functions, the following is obtained: or weight ,Sure Solve for it.

[0088] For example, the fuzzy rule base provided in this application (example 12 rules, output as participation level) )include: R1 A=Same direction &E=Large &M=High → ; R2 A=Same direction&E=Middle&M=Middle→ ; R3 A=Same direction & M=Low → ; R4 A = General & E = Large & M = High → ; R5 A = General & E = Small → ; R6 A = Reverse → ; R7 Upper limit & A = same direction → ; R8 Rapid Descent Phase & M=High → ; R9 The same cabin has been continuously focused on ≥2 stages → ; R10 symmetrical compartments on the same side are all of equal importance → retain the one with the larger lever arm; R11 Number of participating cabins → Take forward ; R12 valve group interlock conflict → Downshift.

[0089] For example, R1 states that if a compartment has the correct correction direction, a large lever arm efficiency, and a high level safety margin, then it should be a key focus of participation. R6, as the safety baseline, specifies any compartment that might exacerbate attitude deterioration. The reverse () must be excluded ( ).

[0090] The result obtained through fuzzy inference is a fuzzy set, which needs to be defuzzified and converted into a precise numerical value before it can be used for control. In this embodiment, a Mamdani (min-max) aggregation combined with the centroid method is used for output. For each compartment, all activated rules (i.e., rules whose antecedent conditions are satisfied to a certain extent) generate an output fuzzy set. These output fuzzy sets are aggregated through minimum and maximum operations to finally obtain a comprehensive output fuzzy set that represents the total participation level of the compartment.

[0091] It is a number between 0 and 1, quantitatively representing the degree of importance that the compartment should be given at the current stage. To provide a clear, either-or search space for subsequent optimization algorithms, it is necessary to consider continuous... This is converted to a binary decision. In this application, a threshold method can be used. Alternatively, the Top-K method can be used for conversion.

[0092] single cabin stride during phase (i.e., the maximum permissible liquid level change in a compartment within a phase), then based on the participation level of each target compartment. The liquid level change boundary for this stage is calculated dynamically:

[0093]

[0094] Get the participating set Among them, the set From all conditions The set consists of cabin number i, which is a list of all target cabins for the current stage.

[0095] In some optional embodiments, after all target compartments are determined, the fuzzy swarm optimization unit uses the fuzzy particle swarm optimization algorithm, along with constraints such as hyperplane projection and boundary repair processing, to search for the optimal decision variable in a sparse search space composed of the framework parameters of the current stage; the optimal decision variable is the liquid level change of each target compartment.

[0096] For example, after receiving the framework parameters provided by the stage decision-maker, the fuzzy swarm optimization unit is responsible for calculating the optimal liquid level change for each target compartment in the current stage, under the premise of satisfying multiple engineering constraints, and finally generating control commands that can be issued and executed. That is, it uses the fuzzy particle swarm optimization algorithm to optimize the swarm optimization. Calculate the liquid level change for each target compartment. It generates control commands for the current stage based on the liquid level change.

[0097] In the fuzzy particle swarm optimization algorithm, variables The equality constraints (internal redistribution or net water intake) are:

[0098] Under the clean water intake model that allows water exchange with the environment, this sum must equal the target volume change for this stage. Inequality constraints:

[0099] That is, the change in liquid level in each compartment It must be restricted to the feasible range calculated dynamically.

[0100] The optimization objective is a comprehensive objective function. Defined as follows, the objective function aims to minimize the overall cost of this stage:

[0101]

[0102] in, This represents the total pumping time required to complete the level adjustment of all compartments. Its value is related to the absolute value of the adjustment amount of each compartment, the rated flow rate of the pump, and the number of pumps that can be pumped in parallel. This represents the total accumulated time delay of all valve actions. A safety penalty term is used to ensure attitude safety; a significant penalty is introduced to prevent the optimized solution from exceeding the safety envelope. and These are penalty weighting coefficients for the rake and pitch angles. and This is the safe upper limit for attitude angles. The function ensures that penalties are applied only when limits are exceeded. The feasibility penalty term is then implemented through... To penalize the optimized solution that causes the liquid level in the compartment to exceed its physical upper and lower limits, Π(·) is usually a squared penalty function. Its value is 0 when the liquid level is within the limit, and increases with the square of the distance of the exceedance after exceeding the limit.

[0103] To ensure that candidate solutions generated during particle swarm optimization satisfy strict equality constraints, the algorithm introduces a crucial projection repair step. For any candidate solution vector x generated by the FPSO algorithm, it is mathematically projected onto the hyperplane defined by the equality constraints. superior:

[0104] If a component goes out of bounds after projection, then shear to the boundary and "reproject" only on the free variables (this can be iterated several times until it is feasible or determined at this stage). Not feasible → Add 1 compartment and try again.

[0105] The particle swarm optimization process includes: 1. Initialization: Particle position First project, then repair; speed zero or small perturbation. 2. Speed / position update: ; then projection + repair. A "smoothener" for valve action penalty: in calculation Time Set threshold (e.g., 1–2 mm liquid level equivalent) to suppress false start-stop cycles caused by "jitter". Annealing: w: 0.9 to 0.5, iterations 40–60 times; convergence criteria Or reach the limit. Failure recovery: If multiple out-of-bounds errors / infeasibility occur, expand. or reduce Re-filter.

[0106] In some optional embodiments, the specific process of the decision subsystem 12 generating control commands may include: calculating the opening duration and action sequence of the pump or valve corresponding to each target compartment based on the sign and magnitude of the liquid level change in each target compartment, combined with the target hardware parameters of the underwater leveler, and generating corresponding control commands. When the execution subsystem 13 executes the control commands, it monitors the operating status data. If the operating status data indicates that the attitude or draft deviation of the underwater leveler exceeds a preset threshold, a local correction operation is triggered to adjust the liquid level change and reissue the control commands.

[0107] For example, the end-to-end solution process is as follows: 1. Read in 2. Calculation 3. Determine Constructing boundaries 4. Running FPSO: Initialization → (Update → Projection → Repair → Evaluation) × Iteration → Obtain 5. Map to pump / valve timing and execute; 6. Monitor and perform closed-loop correction; Reach stage threshold → Enter 7. If Inferior to current best → Backtrack / Reorganize Recalculate; 8. Achieve →End.

[0108] In this algorithm, equipped with a projection repair mechanism, the FPSO algorithm searches within the dimensionality-reduced space (only for participating compartments) according to its standard procedures. Particle positions are initialized to a uniform random distribution and undergo initial projection repair to ensure feasibility. In each iteration, particles update their velocity and position based on their own historical best position and the group's historical best position. Immediately after the update, projection repair is performed, and the fitness value (i.e., the value of the objective function J(x)) of the new position is calculated. To improve engineering practicality, this application's algorithm incorporates several domain-specific technologies, including: 1. A "smoothing" process, setting a minimum threshold ε for liquid level changes and ignoring adjustments below this threshold to avoid frequent minor pump valve movements (jitter); 2. An annealing strategy, gradually reducing the inertia weight w from 0.9 to 0.5 with each iteration to balance global exploration and local development; 3. Setting an upper limit for the number of iterations or a convergence criterion for fitness changes; 4. A failure recovery mechanism, where if a feasible solution cannot be found multiple times, the algorithm is fed back to the front-end fuzzy filtering stage to expand the participating set. Or lower the screening threshold To broaden the search space.

[0109] In summary, the staged decision-making unit in this application decomposes the high-dimensional overall problem into "intra-stage optimization + inter-stage connection" through staged decomposition, and uses to-go lower bounds and backtracking to ensure approximate global optimization; in fuzzy filtering, it uses... As an antecedent, domain knowledge is written into a computable rule base, producing... Used for dimensional compression and boundary scaling. The fuzzy swarm optimization unit combines time-based costs with attitude, level, or volume constraints on the selected chamber set; it ensures particle feasibility and convergence speed through "hyperplane projection + boundary repair"; finally, it maps to pump / valve cycle execution. It can coordinate and optimize buoyancy and attitude, explicitly incorporating pump / valve delays and parallel constraints, significantly shortening the total buoyancy / descent time while reducing start-stop frequency and overall energy consumption. Employing a phased DP algorithm and fuzzy FPSO sparsity solution, combined with equation projection and boundary repair, it ensures that the lateral and longitudinal tilt are controlled within the safety envelope throughout the entire process, resulting in minimal attitude fluctuations and strong anti-disturbance capabilities. Furthermore, it requires no modification to existing hardware, seamlessly integrating with IMU, level, total station, and digital twin systems, supporting online rolling optimization, adapting to multiple chamber types and water depths, and demonstrating high engineering feasibility.

[0110] In some optional embodiments, the intelligent control system 10 provided in this application can also be used for real-time perception and virtual mapping of the working status of the underwater leveling machine. The intelligent control system 10 can be seamlessly integrated with the digital twin platform to achieve precision, visualization and intelligence in the leveling operation.

[0111] The decision subsystem 12 also includes a digital twin unit, a job suggestion unit, and a job control unit; wherein, the digital twin unit is configured to perform state estimation and three-dimensional terrain modeling based on the working status data to construct a digital twin model of the leveling machine operation; the job suggestion unit is configured to generate job suggestions based on the working status data and the job tasks of the leveling machine; and the job control unit is configured to control the execution subsystem 13 to execute the job tasks based on the job suggestions.

[0112] When the intelligent control system 10 performs virtual mapping, the sensing subsystem 11 also includes an edge gateway, multi-beam sonar, DVL speedometer, etc.; the execution subsystem 13 also includes a thruster and a fabric placement device. Please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram of the main frame of the underwater leveling machine provided in this application. The fabric placement mechanism 7 is mounted on the fabric placement chassis 1 and can move laterally and longitudinally relative to the fabric placement chassis 1; the sonar and multibeam sonar are installed at the fabric placement longitudinal beam 12 of the leveling machine.

[0113] Please Figure 2 Based on the above, refer to Figure 3 , Figure 3 This is a schematic diagram of the material placement mechanism of the underwater leveling machine provided in this application embodiment. The material placement mechanism 7 includes an upper material pipe 72 and a lower material pipe 73: the upper material pipe 72 is inserted above the lower material pipe 73, and a first channel is formed between the upper material pipe 72 and the lower material pipe 73. The lower material pipe 73 includes a bottom pipe structure 731 and a first funnel structure 732 connected to the top of the bottom pipe structure 731, with the larger end of the first funnel structure 732 facing the upper material pipe 72.

[0114] Please Figure 3 Based on the above, refer to Figure 4 , Figure 4 This is a schematic diagram of the lower part of the upper material tube provided in an embodiment of this application.

[0115] The upper tube 72 includes an upper tube structure 721. Several protrusions 722 are arranged circumferentially on the outer wall of the upper tube structure 721. There is a first gap 723 between adjacent protrusions 722. The outer surface of the protrusions 722 is an inclined surface 724 corresponding to the first funnel structure 732. The lower part of the upper tube structure 721 is inserted into the bottom tube structure 731. There is a second gap 725 between the outer wall of the upper tube structure 721 and the inner wall of the bottom tube structure 731, which is connected to the first gap 723.

[0116] Please refer to Figure 5 , Figure 5This is a schematic diagram of the lower material pipe structure provided in an embodiment of this application. The bottom pipe structure 731 and the first funnel structure 732 are welded together, and a first connecting rib 733 is welded between the outer wall of the bottom pipe structure 731 and the outer wall of the first funnel structure 732. A plurality of lower lifting lugs 734 are connected to the top outer wall of the first funnel structure 732, and all the lower lifting lugs 734 are arranged circumferentially along the first funnel structure 732. All the lower lifting lugs 734 are close to the large opening end of the first funnel structure 732.

[0117] Please refer to Figure 6 , Figure 6 This is a schematic diagram of the insertion of the upper and lower feed tubes provided in an embodiment of this application. The upper feed tube 72 also includes a second funnel structure 726 sleeved on the outside of the upper tube structure 721. The second funnel structure 726 faces the first funnel structure 732 and can cover the large end of the first funnel structure 732. The first funnel structure 732 and the second funnel structure 726 are connected by a third gap 741 with a first gap 723. The third gap 741, the first gap 723 and the second gap 725 form a first channel.

[0118] The upper tube structure 721 has a first block 720 located on the outer side of the lower part of the first funnel structure 732. At least one side of the first block 720 can laterally abut against the inner wall of the bottom tube structure 731. This is to increase the connection stability between the upper tube structure 721 and the lower material tube 73.

[0119] The upper tube structure 721 also includes at least two upper tube sections 727 that are detachably connected in sequence along the length of the upper tube structure 721, and the second funnel structure 726 is located on the lowermost upper tube section 727.

[0120] The upper tube structure 721 also includes a third funnel structure 728 connected to the top of the upper tube structure 721, with the large opening of the third funnel structure 728 facing upwards.

[0121] Underwater sensor data acquisition is susceptible to disturbances such as water flow, sediment obstruction, and sound wave scattering, resulting in unstable and inconsistent data. Traditional filtering algorithms such as EKF are prone to fusion bias when dealing with multi-source data with low signal-to-noise ratios and high uncertainty. To address this, the intelligent control system 10 in this embodiment can introduce an edge gateway with a terrain confidence weighting mechanism to dynamically assign weights to data from different sources and of different qualities, thereby improving the robustness and real-time reliability of the twin model.

[0122] The edge gateway uses a fusion algorithm based on terrain confidence weighting to dynamically assign weights to the raw data collected by different sensing devices in the sensing subsystem 11. The working status data acquired by the sensing subsystem 11 includes at least: leveling machine attitude, leveling machine speed, terrain elevation difference in the leveling area, and hydrodynamics.

[0123] Suppose that at time step t, the edge gateway 130 collects n types of observation data sources (data collected by n types of sensing devices 120), including IMU inertial navigation (z 1 ), multibeam sonar (z 2 ), depth gauge (z 3 ), DVL speedometer (z 4 ), etc. The observation model for each type of sensor is defined as follows:

[0124] Where, x t This represents the actual state of the system; h i (x t Let be the observation function of the i-th sensor; Let be the observation noise of type i, which follows a zero-mean Gaussian distribution.

[0125] Construct a terrain confidence map of the current area based on historical multi-period point cloud data. The terrain reliability metric, used to measure the reliability of each coordinate point, is defined as follows:

[0126] in, The length of the sliding window, in For the first At that moment, point The original observed terrain elevation at the location—obtained from point cloud modeling, in meters; This represents the local historical average elevation. For error tolerance, For indicator functions, the range of values ​​is... The closer the value is to 1, the more stable the historical data is and the more reliable the data is.

[0127] For each type of observation data, its observation weight is dynamically adjusted based on the confidence level of its corresponding region. :

[0128] in, For the first Each sensor observes the coverage area; after weight normalization, it satisfies... .

[0129] An improved weighted EKF prediction-update process is used for state estimation: 1. Prediction Phase:

[0130]

[0131] 2. Weighted update phase: Calculate the gain for each observation:

[0132] After considering all observations, the status is updated as follows:

[0133] Covariance update:

[0134] In some optional embodiments, the task includes at least a material placement task. After obtaining the material placement task and raw data, the task recommendation unit calculates the feasibility of material placement in the area using a material placement scoring function; recommends material thickness and density values ​​using fuzzy rules; predicts potential construction defect areas and high-risk advancement paths using a dynamic window algorithm and a dynamic cost weighting mechanism; and generates a material placement task recommendation that includes the material quantity and advancement direction.

[0135] For example, the method by which the work suggestion unit calculates the feasibility of land placement in a region using the land placement scoring function includes: obtaining the height difference between the target flat surface and the current terrain elevation in the digital twin model; obtaining the soil difficulty score of the target flat surface area; and performing a weighted linear combination of the height difference and the difficulty score to obtain the land placement score.

[0136] In some alternative embodiments, the target flat surface generated by digital twin modeling. With current terrain elevation , definition point The residual at is: This represents the height difference between the current point and the ideal flat state; the soil difficulty level score for this area is defined by sonar echo intensity, construction experience, or manual marking.

[0137] A value closer to 1 indicates a more complex soil type (such as soft silt or sand), requiring a higher density of substrate to ensure sedimentation quality. Based on the two environmental factors mentioned above, a substrate placement feasibility scoring function is defined. For weighted linear combinations:

[0138] in, All weights are either manually set or adaptively learned, satisfying... .

[0139] In some optional embodiments, the method by which the job suggestion unit recommends fabric thickness and density values ​​using fuzzy rules may include: mapping fabric scores to fuzzy input variables; setting recommended fabric thickness and recommended fabric density as fuzzy output variables; using Mamdani-type inference to synthesize multiple preset fuzzy rules, and using the centroid method to defuzzify. Optionally, the job suggestion unit 220 performs fuzzy rule recommendation processing on all areas of the target flat surface to obtain a recommended thickness heatmap and a recommended density heatmap for the corresponding target flat surface.

[0140] Will Mapped to fuzzy input variables Define three membership functions: Low: ;middle: ;high: Triangular or trapezoidal membership functions can be used. .

[0141] Fuzzy output variable definition: Recommended fabric thickness The range is such as [2cm, 10cm]; recommended fabric density. The range is such as [200kg / m³, 600kg / m³]. It is also divided into three levels: low, medium, and high; the output is a fuzzy set.

[0142] Examples of fuzzy rules (IF-THEN statements) are shown in Table 1.

[0143] Table 1

[0144] Fuzzy reasoning: Multi-rule synthesis using Mamdani-type reasoning; Defuzzification: Obtaining the final suggestion value using the centroid method.

[0145]

[0146] Final recommended fabric values: Recommended thickness heatmap: Recommended density heatmap: .

[0147] This heat map is displayed through a manual reference system to assist in manual control of the fabric application rate. After the operation is completed, the actual construction results can be fed back to fine-tune the weights. Or update the fuzzy rule base.

[0148] In some optional embodiments, the work suggestion unit predicts potential construction defect areas and risky advancement paths using a dynamic window algorithm and a dynamic cost weighting mechanism by introducing a "terrain risk heat map" constructed from a digital twin model and weighting the influencing factors of the prediction results. Optionally, the influencing factors of the prediction results include: the angle between the target direction and the current path direction, the minimum distance between the path and the nearest obstacle, the magnitude of the advancement speed corresponding to the current path, and the dynamic risk heat layer derived from the twin model.

[0149] To improve the robustness of path planning in underwater emergencies, a "terrain risk heat map" constructed from a digital twin model can be introduced, and a dynamic cost weighting mechanism can be introduced into the dynamic window algorithm (DWA) scoring function.

[0150] The topographic risk heat map is generated by integrating data such as seabed elevation grids reconstructed in real time by digital twins, historical stability (confidence), target flat surface, current velocity and direction, obstacles / restricted areas, and sonar echoes. The decision subsystem 12 first extracts the source data of the aforementioned seabed elevation grids, historical stability (confidence), target flat surface, current velocity and direction, obstacles / restricted areas, and sonar echoes into several "0-1 standardized" risk factors (e.g., elevation difference exceeding limits, slope / undulation, historical fluctuations or insufficient data credibility, soil difficulty, distance from obstacles, hydrodynamic transport risk, etc.), and then performs weighted fusion according to the weights preset by the engineering or obtained through online learning to obtain the comprehensive risk value of each grid. Subsequently, spatial smoothing and anomaly suppression are performed (anisotropic filtering along the flow direction is used under strong current conditions), and a conservative degradation strategy is activated when data frames are lost or anomalies occur. The final output is a heatmap with values ​​ranging from 0 to 1 (typical update frequency of about 1Hz, resolution of 0.2 to 0.5m). The higher the value, the greater the risk of terrain operations, which is used for risk constraints and highlighting in path planning and material placement strategies.

[0151] The DWA scoring function has been modified as follows:

[0152] Among them, heading measures whether the current path direction is towards the target point. The smaller the angle between the target direction and the current path direction, the higher the score. The minimum distance between the path and the nearest obstacle; This represents the speed of propulsion corresponding to the current path. The dynamic risk thermal layer, derived from twin modeling, reflects areas of sudden changes in local terrain, material accumulation, or substandard construction. , , , The weights of the aforementioned influencing factors are given. This mechanism can effectively guide the path away from high-risk areas and improve the stability of the advance. The dynamic risk heat map is a time-series risk layer formed by introducing time-dimension smoothing, short-term forward-looking and stability constraints on the "terrain risk heat map". It is used to reflect the operational risk trend in a short period of time in the future. Its output is a raster value between 0 and 1, with a larger value indicating a higher risk.

[0153] In some optional embodiments, the operation suggestion unit can transmit the suggestion results to the operation control unit for display in the form of heatmaps, path lines, etc. Optionally, the operation control unit can be a shore-based control console or a shipborne terminal.

[0154] Optionally, operators may refer to the suggested results to: adjust the direction of advance, adjust the amount of material laid, change the path, or request a replanning.

[0155] Alternatively, the operation control unit does not automatically drive the execution subsystem 13 to run, but instead adopts a "suggested collaboration" approach, relying on the operator to manually control the execution subsystem 13 to run.

[0156] Optionally, the decision subsystem 12 also includes a feedback optimization module. This module is used to optimize the algorithms of the digital twin unit and the operation suggestion unit by obtaining feedback from the fabric placement operation after completion. The feedback optimization module collects data on the actual fabric coverage, the error between the actual fabric placement trajectory and the predictions of the digital twin model through sensing devices, and adaptively updates the following parameters: scoring function parameters; fuzzy system membership degrees; and terrain modeling algorithms. This continuously optimizes the control strategy, improves prediction accuracy and recommendation rationality, and forms a long-term evolving virtual-real closed-loop control chain.

[0157] In some optional embodiments, the intelligent control system 10 of the underwater screed can also be used to provide fine measurements for underwater screed operations. When the intelligent control system 10 of the underwater screed is used to provide fine measurements for underwater screed operations, it further includes an underwater measurement subsystem, which includes: A binocular camera device is configured to acquire binocular visual images of the target water area; an edge computing device is configured to acquire point clouds of the target water area.

[0158] The underwater measurement subsystem is installed on the underwater leveling machine, the binocular camera equipment is set at the front of the main frame of the underwater leveling machine, and the edge computing equipment is connected to the material laying equipment in the underwater leveling machine.

[0159] Specifically, the binocular camera equipment adapts to strong underwater currents and highly turbid environments through shock absorption and sealing structures. It includes a binocular stereo camera, an auxiliary structured light module, an IMU inertial navigation module, and an embedded data acquisition unit, mounted on the front crossbeam of the leveling machine, and possesses waterproof and pressure-resistant properties. To further meet the requirements of underwater operations, the binocular stereo camera utilizes a low-distortion underwater wide-angle lens and is equipped with a hydraulically pressurized housing. It can be understood that the underwater leveling machine moves underwater in the predetermined direction of the construction operation, and the binocular camera equipment primarily acquires binocular visual images of the target water area in the direction of the underwater leveling machine's movement.

[0160] Edge computing devices can be central processing units (CPUs), microcontrollers (MCUs), hardware chips, etc. Hardware chips can be application-specific integrated circuits (ASICs), programmable logic devices (PLDs), or combinations thereof. The aforementioned PLDs can be complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), generic array logic (GALs), or any combination thereof.

[0161] The edge computing device is configured to calibrate and correct the binocular vision image to obtain the disparity of each pixel; extract the original grayscale information of the binocular vision image, and then enhance the binocular vision image to generate gradient information; calculate the matching cost based on the disparity of each pixel in the binocular vision image and the corresponding original grayscale information and gradient information, and construct a cost function; calculate the disparity map of the binocular vision image by combining the cost function with the improved SGM (Semi-Global Block Matching) algorithm; and back-project the disparity map according to the camera parameters of the binocular camera device to obtain the point cloud of the target water area. The improved SGM algorithm includes: The aggregation cost function is based on four directions; a dynamic adjustment strategy is set to control the penalty for small observation changes and the penalty for large disparity jumps; confidence screening and hole filling are performed on the aggregated disparity map.

[0162] For example, when performing detailed underwater measurements, the underwater measurement subsystem's operational steps include: Binocular visual images of the target water area are acquired using a binocular camera system. The binocular visual images are calibrated and corrected to obtain the disparity for each pixel. The original grayscale information of the binocular visual images is extracted, and then the images are enhanced to generate gradient information. A matching cost is calculated based on the disparity of each pixel in the binocular visual images, along with the corresponding original grayscale and gradient information, thus constructing a cost function. The improved SGM algorithm is used to simultaneously solve the cost function to calculate the disparity map of the binocular visual images. The disparity map is then back-projected based on the camera parameters of the binocular camera system to generate a point cloud of the target water area.

[0163] The calibration and correction can be performed using Zhang Zhengyou's calibration method to obtain intrinsic and extrinsic parameters, followed by stereo correction and binocular alignment to obtain the disparity for each pixel. Pre-processing optimization is necessary because the binocular vision images suffer from varying degrees of defects due to the underwater environment during underwater leveling machine construction, such as low contrast, unclear texture, and high noise. Therefore, enhancement processing of the acquired binocular vision images is required, specifically including: The CLAHE technique is used to enhance the local contrast of the binocular vision image, and bilateral filtering is used to suppress high-frequency noise while preserving the edges. Then, Sobel gradient enhancement is applied to the binocular vision image to obtain gradient information.

[0164] For example, the underwater measurement subsystem can use multi-channel matching and a composite cost function to calculate the matching cost, as shown in the formula:

[0165] in, The cost function to be constructed, , , The weights are the cost function weights. This represents the position of a pixel in a binocular vision image. For disparity values, This is the original grayscale information. For gradient information, Let Census cost function be used. Let gradient cost function, The SAD cost function is calculated on both the original grayscale channel and the gradient channel. It should be noted that the weights of this cost function can be adjusted as needed for different underwater environments.

[0166] SGM aims to aggregate cost functions in multiple directions on an image, balancing computational complexity and accuracy between local disparity calculation and global smoothness constraints. Its core is the construction of the cost volume. Then, the costs are aggregated along multiple paths, and the disparity value with the lowest cost is selected:

[0167] The final total aggregation cost is: .

[0168] in, The position of a pixel in the image, in two-dimensional coordinates. ; This is the disparity value, which is the horizontal offset of the pixel between the matching points in the left and right images; Other candidate disparity values ​​during traversal, used to compare with... Compare the differences in costs; It is along the direction Cumulative path cost; It is the cost of local matching; It is a penalty for small parallax variations (encouraging smoothness); It is a penalty for large parallax jumps (to prevent jumps); This is a term introduced to prevent cost explosion.

[0169] In some alternative embodiments, the improved SGM algorithm includes: First, a cost function based on four directions is used. Traditional SGM algorithms aggregate paths in eight directions; this invention optimizes this by selecting four directions (left / right, up / down, top-left-right-bottom, and top-right-left-bottom) to balance accuracy and real-time performance. Furthermore, weights are increased for paths in areas with large residuals to improve disparity accuracy in scattering regions. Specifically, the aggregated path set includes only four directions: left / right, up / down, top-left-right-bottom, and top-right-left-bottom. Weights are also increased for paths located in areas with large residuals (residuals greater than a threshold). The pixel value, the path cost along that direction. Multiply by the gain factor This enhances the reliability of the match in the scattering region.

[0170] Secondly, a dynamic adjustment strategy is set to control the penalties for small changes in observation and large jumps in parallax. Specifically, this is expressed as follows:

[0171]

[0172] in The gradient magnitude of the image represents the edge intensity; The constant is small, avoiding division by zero; this strategy can reduce smoothing in edge regions and preserve high-frequency information.

[0173] Finally, the aggregated disparity map undergoes confidence filtering and hole filling. Specifically, the confidence score for the disparity value of each pixel in the aggregated disparity map is calculated, and pixels below a confidence threshold are removed. Then, for the hole regions in the filtered disparity map, a front-to-back filling and edge-guided interpolation method based on the left view of the binocular vision image is used. This includes finding the nearest valid disparity value within a preset pixel range for the removed pixels; for example, for the removed pixels (invalid disparity points). Find the nearest valid disparity value within 5 pixels to its left and right. Also, use image edge information to constrain the interpolation direction, preserving the boundaries.

[0174] Based on the above measurement method, the final improved SGM disparity estimation process is as follows: 1. Acquire binocular images 2. Perform CLAHE, bilateral filtering, and Sobel enhancement operations on the image; 3. Construct the cost volume. 4. Aggregate costs along four directional paths to form 5. Obtain the disparity map using the Winner-Takes-All strategy. 6. Perform parallax confidence screening and hole filling; 7. Generate dense point clouds by back-projection using camera parameters.

[0175] It should be noted that Winner-Takes-All (WTA) is a commonly used standard selection strategy after SGM (Small Generalized Selection). Specifically, it employs the WTA strategy to select from... S ( p,d The disparity value d corresponding to the minimum cost is selected as the final disparity estimate, and thus the following is obtained. D ( p ).

[0176] Furthermore, in order to enhance the measurement system's ability to identify abnormal disparity values, the confidence calculation provided in this embodiment of the invention specifically includes calculating a consistency score, a local consistency score, and a back projection score based on the disparity map completed by aggregation, and performing weighted fusion to obtain a confidence score, and then selecting the corresponding pixel points based on the confidence score.

[0177] Specifically, for each pixel parallax The consistency check of the left and right views is adopted, assuming... and For the corresponding points in the left and right disparity maps, the consistency error is: , Set threshold (e.g., 1 pixel), then the consistency score is:

[0178] definition Centered on The neighborhood window, for the mean disparity within the neighborhood with standard deviation Calculate the local consistency score:

[0179] Will After back-projecting to the 3D point cloud, reprojecting back to the left view using system extrinsic parameters yields the result. Calculate pixel residuals:

[0180] The constructed back projection score is:

[0181] The final confidence level is:

[0182] in, , , To integrate the weighting coefficients of the above scores, they can be set according to the screening requirements during actual implementation. Similar to camera extrinsic parameters, system extrinsic parameters are pose extrinsic parameters or transformation matrices describing the transformation from the camera's coordinate system to the world coordinate system.

[0183] Based on the above technical solution, point clouds of the target water area can be obtained. This invention overcomes the limitations of traditional underwater measurement methods that rely on a single sonar or manual operation. By using an improved SGM algorithm to calculate disparity maps and generate point clouds, it significantly enhances the ability to acquire high-precision 3D terrain data in complex underwater environments. Simultaneously, this invention features adaptive disparity confidence assessment and residual back projection mechanisms, effectively suppressing the impact of mismatches and image noise on measurement accuracy, ensuring the stability and reliability of the measurement results.

[0184] In practical applications, binocular camera equipment acquires binocular video streams of the target water area. Based on the movement speed of the underwater leveling machine, the binocular video stream can be split and extracted using a reasonable number of frames to obtain processable binocular visual images.

[0185] Further, please see Figure 7 , Figure 7 This is a schematic diagram of the entire workflow of the underwater measurement subsystem provided in this application embodiment. The edge computing device is also used to obtain the design surface elevation of the target water area, convert the point cloud of the target water area into the terrain elevation, and then calculate the difference between the corresponding design surface elevation and the terrain elevation based on the material placement position of the target water area to obtain the material thickness value. Based on the material thickness, the material quantity is determined, and the material placement equipment is controlled to place the material at the material placement position according to the material quantity.

[0186] The edge computing device calculates the difference between the design elevation and the terrain elevation to obtain the fabric thickness value, specifically including: Calculate the elevation difference and local fluctuation standard deviation between the corresponding design surface elevation and the terrain elevation based on the fabric location; The fabric thickness value is obtained by fusing elevation difference and local fluctuation standard deviation based on a fuzzy rule table.

[0187] For example, converting point clouds into terrain elevation grids. Elevation of the design surface Compare and construct a difference diagram:

[0188] Constructing fabric area :

[0189] For each fabric position According to its To account for local terrain variability, a fuzzy rule table is used to construct recommended fabric levels (light, normal, reinforced) and fabric thickness values. .

[0190] Example of a rule:

[0191] in, For local fluctuation standard deviation, This is the redundancy control coefficient.

[0192] Furthermore, the aforementioned measurement system also includes a user interface. The edge computing device is also used to output the fabric location and corresponding fabric thickness value to the user interface in the form of a two-dimensional heat map or table, allowing the user to confirm the amount of fabric. In one possible implementation, the edge computing device can also output a point cloud map and a difference map of the target water area to the user interface, allowing the user to visualize the measurement results of the target water area through the user interface.

[0193] The underwater measurement subsystem links the measurement results with the actual material placement decision of the underwater leveling machine in a closed loop, which can provide refined measurement results for underwater leveling operations and provide a reliable basis for subsequent construction operations.

[0194] In some optional embodiments, the intelligent control system 10 can also be used to automatically adapt to different water depths, flow rates, or material diameters during material spreading. In this application scenario, the sensing subsystem 11 also includes a weighing sensor and an inclinometer, and the decision subsystem 12 also includes a material spreading control unit.

[0195] The weighing sensor is mounted on the hopper support frame to measure the weight of the stones in the hopper in real time. It is also equipped with an inclinometer (IMU) to monitor the hopper's attitude angle. The weighing sensor, combined with the inclinometer data, enables precise weighing compensation under tilted conditions, ensuring accurate calculation of stone mass even when the device is tilted. A volumetric flow meter is integrated on the conveyor arm to continuously monitor the stone conveying rate, i.e., the volume of stone passing through the distribution head per unit time. The distribution head is also equipped with multimodal environmental sensing sensors: a laser rangefinder (for accurate distance measurement in shallow or emerging water) and a sonar thickness gauge are installed to scan the terrain elevation of the distribution area in real time. Furthermore, pressure sensors are deployed on the underwater leveling machine to sense the leveling force and provide feedback on the leveling effect.

[0196] The material laying control unit is configured as follows: when the underwater screed is performing screeding construction, it acquires the weight of the stone in the hopper and the tilt angle of the hopper based on the weighing sensor and inclinometer of the sensing subsystem 11; determines the stone mass based on the stone weight and tilt angle of the hopper, and calculates the real-time material laying mass flow rate based on the time series data of the stone mass; acquires the current terrain elevation of the subgrade, compares the current terrain elevation of the subgrade with the pre-stored design target elevation, and calculates the terrain residual and the volume to be laid; and generates material laying control commands through an optimized decision algorithm based on the real-time material laying mass flow rate, terrain residual, volume to be laid, and predicted thickness distribution. The predicted thickness distribution is obtained by the target prediction model based on historical construction data and multimodal sensing information, and the material laying control commands are used to adjust at least one of the following: the conveyor belt speed of the underwater screed, the opening of the material laying valve, or the moving speed of the material laying head.

[0197] In some optional embodiments, the sensing subsystem 11 can employ information fusion and redundant configuration to improve reliability in response to potential interference and malfunctions caused by the underwater environment. For example, a weight sensor and a flow meter provide two estimates of material supply, which can be cross-checked; a combination of laser and sonar sensors is used for ranging, with laser used to obtain high-precision local thickness in good water quality, and sonar used for scanning in turbid water. Multi-sensor data fusion not only improves measurement accuracy but also enhances the system's robustness under harsh conditions. Studies have shown that by combining multiple redundant sensor information, the accuracy and stability of measurements can be maintained even when sensors degrade or become inaccurate. Therefore, the overall structure of the construction control system, through a closed-loop hardware "perception-decision-execution" link, achieves real-time fusion and redundancy fault tolerance of multi-sensor data. Through the cross-checking of the weight sensor and flow meter, and the adaptive switching between laser and sonar, the accuracy of construction measurements and the continuity of system operation in harsh underwater environments are ensured, thereby directly improving the anti-interference capability of the intelligent construction process.

[0198] Please refer to Figure 8 , Figure 8This is a schematic diagram illustrating the application flow of the material laying control unit provided in this application embodiment. In the application flow of the construction control system, the sensing subsystem 11 collects data as the sensing layer. The software algorithm of the material laying control unit includes data fusion in the sensing layer and estimation of the laying thickness distribution in the prediction layer, and controls the material laying rhythm and conveying speed as the optimization decision layer. The material laying control unit provides an intelligent closed-loop control flow of "sensing-prediction-decision".

[0199] In the perception layer, raw data is collected through deployed multimodal sensors (including weighing sensors, tilt sensors, volumetric flow meters, and sonar). This layer corresponds to the hardware function of the perception subsystem 11, responsible for acquiring key physical information such as the weight of the stone in the hopper, the attitude of the leveling machine, the material flow rate, and the elevation of the subgrade. Subsequently, these raw data are preliminarily processed and verified through data fusion technology (such as tilt compensation to calculate the actual stone mass), transforming the raw, isolated sensor readings into reliable state information with clear engineering significance (such as real-time stone mass, terrain residual, and expected paving volume), providing accurate and consistent input for subsequent prediction and decision-making. In the prediction layer, the paving control unit uses software algorithms to make forward-looking predictions of future construction effects. This layer is based on the real-time state perceived in the previous stage and combines historical construction data, using time-series prediction models such as LSTM (Long Short-Term Memory) to estimate the thickness distribution of the stone after paving. In the decision-making layer, this application uses the SAC (Soft Actor-Critic) reinforcement learning algorithm based on the maximum entropy framework for optimization decision-making. This layer combines the real-time state provided by the perception layer with the thickness distribution estimate output by the prediction layer to form a complete "state" input; it abstracts the actuators that need to be controlled (such as conveyor belt speed, fabric valve opening, etc.) into "actions"; and it designs a "reward function" that comprehensively considers laying uniformity, material utilization, and action smoothness. The SAC algorithm continuously optimizes the strategy and finally outputs the optimal control action for the current state. These action instructions are sent to the actuators to directly control the fabric laying rhythm and conveying speed, thereby precisely guiding the construction operation.

[0200] In the sensing layer, the mass of the stone in the hopper is calculated in real time using data from the weighing sensor and the inclinometer. If the weighing sensor reading is... And the hopper tilt angle is The calculation module can compensate for the gravity component based on the tilt sensor data and the weighing sensor readings to calculate the mass of the stone when the hopper is tilted. For example, the mass of the stone in the hopper. It can be approximated as:

[0201] in, This represents the gravitational acceleration constant. The above formula utilizes the attitude angle provided by the tilt sensor to compensate for the gravitational component of the weight reading, ensuring accurate mass estimation under different tilt conditions. The material spreading control unit reads the data at high frequency. and Calculate the continuous mass time series By differentiating or subtracting this sequence, we can obtain the amount of stone reduction per unit time, i.e., the real-time material distribution mass flow rate. .

[0202] For example, in time intervals Internal mass reduction Then the fabric flow rate Using data from the volumetric flow meter, the material distribution control unit verifies the estimated mass value and obtains the stone volumetric flow rate; the conversion relationship between the two is as follows: ( (This refers to the bulk density of the stone). Real-time mass and flow rate data provide precise material supply status for subsequent paving rhythm control. Simultaneously, the IMU (Inertial Measurement Unit) provides the acceleration and angular velocity of the material distribution head to infer dynamic factors during the discharge process (such as the impact of ship swaying on instantaneous discharge), thereby compensating for these disturbances in the control. Through this multi-source data, the paving control unit can constantly determine the current remaining material quantity and the subsequent feed rate per second.

[0203] Furthermore, in the sensing layer, sonar / laser thickness sensors at the front end of the fabric head scan the current topographic elevation of the underwater substrate in real time. The material spreading control unit pre-stores the designed target elevation surface. Each time a scan is completed, the paving control unit compares the current elevation of the subgrade with the pre-stored design target elevation and calculates the elevation residual:

[0204] In areas where filler is needed If the value is positive, multiplying it by the corresponding area unit yields the volume to be laid: (For areas that do not require filling or require excavation, then) (Take 0 to avoid overfilling), where, This represents the area of ​​a single area unit. To reduce the impact of noise, it can be... The field is filtered and smoothed, or the local average residual is extracted as a control reference.

[0205] Furthermore, the sonar thickness measurement simultaneously records the orientation of the paving head and the water depth, allowing for accurate residual calculations through coordinate transformation. For example, when the paving head is tilted, the sonar ranging is corrected using the IMU attitude angle to align the measured terrain point cloud with the global coordinates. In environments with high flow velocities, multiple measurements can estimate the impact of water flow on the erosion and displacement of the paved stones, enabling timely corrections. Ultimately, the terrain residuals Matrix and required layup volume at each location This will serve as an important input for subsequent material optimization decisions.

[0206] In the prediction layer, the predicted thickness distribution is obtained by the target prediction model based on historical construction data and multimodal sensing information. The material placement control command is used to adjust at least one of the following: the conveyor belt speed of the underwater leveling machine, the opening of the material placement valve, or the moving speed of the material placement head.

[0207] To make more intelligent decisions regarding the laying strategy, a data-driven predictive model is introduced into the laying control unit to estimate the future laying thickness distribution. The predictive model can utilize historical construction data and multimodal sensor information to predict the thickness distribution after stone laying under the current control strategy, as well as potential weak areas. Considering the temporal correlation and nonlinear characteristics of the laying process, a Long Short-Term Memory (LSTM) neural network is chosen to construct the predictor. LSTM is suitable for processing time-series data; it can take as input a sequence of sensor data over a time window (such as the discharge mass flow rate, the trajectory of the laying head, and terrain changes over the past few seconds) and output the predicted increase in future regional thickness. Distribution. This allows for the simultaneous prediction of stone settlement and distribution trends over a short period of time during the laying process. The LSTM model of the material laying control unit can be trained on the settlement and diffusion behavior of stones underwater, taking into account features such as the amount of material added, flow velocity, and material diameter, and outputting an estimate of the planar thickness distribution at the corresponding time. By continuously comparing and correcting with sonar measurement results, the LSTM model can continuously adapt to the environment and improve prediction accuracy.

[0208] Predictive models require training and validation using extensive historical or simulation data. Training samples can be obtained through offline simulation combined with field calibration: by varying the fabric laying rhythm and environmental parameters in a simulated environment to generate a laying thickness distribution, and then fine-tuning the model using a small amount of on-site sonar mapping data. The trained model will be embedded in the fabric laying control unit, reading the current multimodal state in each decision cycle. Outputs a prediction of the thickness distribution for the next cycle. .

[0209] In the decision-making layer, based on the perceived state and the output of the prediction model, the material spreading control unit needs to decide in real time how to control the material spreading rhythm and conveying speed to optimize the uniformity of the spreading thickness and construction efficiency. This is a dynamic sequential decision problem, which can be abstracted as a Markov Decision Process (MDP). This application employs Deep Reinforcement Learning (DRL) to solve for the optimal control strategy, enabling the material spreading control unit to automatically learn the optimal material spreading control decision. .

[0210] Specifically, the state of the reinforcement learning (RL) agent at each time step is defined as including the currently encoded terrain residual, the already laid thickness distribution, the remaining stone mass in the hopper, the real-time material flow rate, the material head pose, the material head speed, and operating parameters. Actions are defined as adjustments to the material laying rhythm, such as adjusting the conveyor belt speed, starting and stopping the material laying valve, and changing the material head movement speed. The reward function characterizes the laying quality and efficiency, for example, negative thickness variance minus time / energy cost, which encourages faster construction while ensuring uniform thickness. Through repeated interaction with the environment (simulated construction process or actual process), the reinforcement learning agent gradually optimizes its strategy, maximizing the comprehensive reward function to train and optimize the material laying strategy.

[0211] In this embodiment, the optimization decision algorithm is constructed using the SAC (Soft Actor-Critic) deep reinforcement learning algorithm, combined with Priority Experience Replay (PER) and Random Network Distillation (RND) techniques to accelerate training convergence and enhance the robustness of the policy exploration. The action space of the agent in the deep reinforcement learning framework is continuous and bounded, corresponding to control commands for conveyor belt speed, fabric valve opening, or fabric head movement speed. The output of the reinforcement learning policy is the optimized fabric control command. For example, when the thickness in a local area is low, the policy may increase the fabric placement frequency or slow down the movement speed to increase the amount of fabric placed at that location; conversely, if a certain area tends to be too high, the policy will reduce the amount of fabric placed or increase the passing speed. Since the policy performs global optimization based on cumulative effects, the fabric control unit can dynamically adjust the current output speed and movement path to ensure that the thickness is as uniform as possible when completing the construction of the entire area. Furthermore, the RL agent can adaptively adjust the policy according to different water flow and terrain conditions, improving its adaptability to complex environments.

[0212] In some optional embodiments, the training steps of the agent in the deep reinforcement learning framework include: By periodically collecting stone mass, real-time material flow rate, encoded terrain residual, material head pose, material head speed, and working parameters, a multimodal fusion state tensor of the intelligent agent is generated. The policy network outputs Gaussian distributed parameters based on the current multimodal fusion state tensor, and obtains the agent's actions through reparameterization sampling and tanh transformation; where the actions, after linear mapping, correspond to the actual control quantities of conveyor belt speed, fabric valve opening, or fabric head moving speed. The soft action values ​​of state-action pairs are estimated separately using a dual-commenter network, and transition tuples are sampled from the experience replay buffer based on a priority experience replay mechanism. The target soft action value is calculated based on the transition tuple, where the target soft action value is calculated by the target critic network and the target policy network based on the next state in the transition tuple; The parameters of the dual critic network are updated by minimizing the mean squared error loss between the soft action value output by the dual critic network and the target soft action value; the parameters of the policy network are updated by maximizing the weighted sum of the soft action value and the policy entropy; the parameters of the target critic network are tracked by the parameters of the dual critic network through a soft update method.

[0213] For example, Multimodal fusion state tensor at time step It consists of the following observations:

[0214] in This refers to the mass of the stone in the hopper after tilt angle compensation. For weighing reading, For tilt angle, For gravitational acceleration, tilt compensation is used to eliminate the influence of attitude changes on weight, ensuring the stability of mass and flow rate estimates; Mass flow rate; For the elevation residual heatmap generated by sonar or laser After encoder The representation of; Position of the machine body / fabric head; For travel and swing speed; These are operating parameters such as water depth, flow velocity, and material diameter.

[0215] The agent's actions are:

[0216] in, For the conveyor belt speed, This refers to the opening degree of the fabric valve. The lateral movement speed of the fabric head. This represents the scanning or oscillating angular velocity of the fabric head. The last layer of the neural network uses the tanh activation function, whose output range is (-1, 1). By normalizing each action component to the interval (-1, 1), it facilitates direct mapping.

[0217] In the reinforcement learning framework, the environment transition probability Defined from the current state Execute action Then transition to the next state The dynamic characteristics of the material paving control unit are determined by a complex system coupled with multiple physical processes. The dynamic characteristics are mainly determined by the stone settling dynamics, carrier kinematics and dynamics, and hydrodynamic disturbances.

[0218] The simulation includes several key components: Stone settling dynamics, which simulating the stone's trajectory, settling velocity, and diffusion behavior under water flow after discharge from the screed head; and carrier kinematics and dynamics, which describing the kinematic response of the screed body, screed arm, and screed head under controlled actions. This includes the kinematic relationships of the mechanisms and the dynamic characteristics of each actuator (e.g., motor, hydraulic cylinder), used to accurately predict the position and velocity of the screed head after the action. Hydrodynamic disturbances simulate the drifting force of the water flow on the screed carrier, the scouring and transporting effect on the settled stones, and the resistance of the water to the mechanical motion.

[0219] The reward function of the SAC deep reinforcement learning algorithm consists of the terrain residual at the next time step, the underfill penalty, the over-height penalty, the smoothing regularization term of the action amplitude, and the efficiency reward term related to construction time or energy consumption cost. Its specific reward function is as follows:

[0220] in, Thickness uniformity is measured using variance, while underfill and overfill are measured using different methods. Norm constraints are applied, and motion amplitude and rhythm smoothing regularization are superimposed to suppress jerk / surge; The weighting coefficient for the thickness uniformity penalty term. For thickness uniformity, Indicates in At any given moment, the construction area is being monitored by the intelligent agent. Inside, the thickness distribution of the stone paving; This is the weighting coefficient for the penalty for underfilling. The weighting coefficient for the extremely high penalty term. The weighting coefficient for the penalty based on the range of motion. This is the weighting coefficient for the penalty of action change.

[0221] Furthermore, the SAC algorithm, as a maximum entropy deep reinforcement learning algorithm, differs from traditional reinforcement learning in its core objective. It not only requires the agent to maximize cumulative reward but also encourages the policy to maintain a certain degree of randomness, i.e., maximizing the policy entropy. The objective function of the maximum entropy framework can be expressed as:

[0222] in, Indicates the strategy Take the maximum value. Let be the discount factor raised to the power of t, representing the reward at step t. It is a temperature parameter used to weigh the importance of reward against entropy; For strategy In state Entropy. Entropy represents the degree of randomness of a strategy; higher entropy means the strategy has stronger exploration capabilities and robustness, and is less likely to get trapped in local optima when facing the uncertainties of the underwater environment. Based on this objective, this application defines a soft action value function. and soft state value function This is used to guide the updates of the commentator network and the policy network. The soft-action value function in the state... Take action below And have continued to follow the strategy ever since. The expected cumulative soft reward that can be obtained from actions is measured by the soft state value function. The agent begins to follow the policy The expected cumulative soft returns that can be obtained from taking action.

[0223] The soft action value function and the soft state value function are:

[0224]

[0225] in, For state Take action below The reward value directly returned by the environment afterward. The discount factor represents the weight of future rewards in the current value. Indicates about the next state Expectations Indicates the next state Probability distribution of state transitions from the environment Obtained from sampling, For the soft state value function in the next state The value below; Regarding actions Expectations Indicates action From the current strategy It is obtained by sampling from the probability distribution. Used to measure the randomness of a strategy.

[0226] In function approximation, to stabilize training and avoid overly optimistic value estimation, the SAC algorithm in this embodiment employs a dual-critic network design. Specifically, the algorithm maintains two sets of critic networks in parallel. and The target value is calculated by taking its minimum value to suppress the overestimation bias of the Q value, providing a more robust value assessment for decision-making. The target value is calculated using a target network. The calculation formula for the minimization strategy is as follows:

[0227] in, For instant rewards, The expectation for the next action, This indicates taking the minimum value from the outputs of the two target critic networks. Let be the logarithmic probability of the strategy.

[0228] The update objective of each commenter network is to minimize the mean squared error loss between its predicted value and the target value y. :

[0229] in, The expectation is about the empirical data, where the expectation is the transition tuple sampled from the empirical playback buffer D. Calculated above.

[0230] To further improve training stability, the parameters of the target critic network... It is not a direct copy of the parameters from the online network. Instead, it is done through a soft update. (in Approximately 1) Slowly track online network parameters. The data samples required for training come from the experience replay buffer. To accelerate the learning process and enable the agent to learn more effectively from key experiences, the algorithm introduces a priority experience replay mechanism. This mechanism is based on the temporal difference error of each transition tuple. The absolute value is used to set the probability of it being sampled. The larger the error, the higher the probability of the sample being sampled. At the same time, to correct for the bias caused by non-uniform sampling, importance weights are also introduced. Adjustments will be made.

[0231] The SAC algorithm employs a stochastic policy and utilizes reparameterization techniques to efficiently handle the bounded action space. (Policy network) Instead of directly outputting the action, it outputs the parameters of a Gaussian distribution. and .use Reparameterize the action to (-1,1):

[0232]

[0233] in, This is the mean vector output by the policy network. The standard deviation vector output by the policy network. For random noise sampled from a standard normal distribution, Indicates a normal distribution. For the final action to be performed, These are the unnormalized raw action values.

[0234] The optimization objective of the policy network is to maximize the weighted sum of the expected soft Q value and its own entropy, and the corresponding loss function is:

[0235]

[0236] When calculating the logarithmic probability of a strategy, the Jacobian determinant correction introduced by the tanh transformation must be considered. Among these, For the expectations regarding state and noise, Let be the log probability of the strategy. In the state Select action The probability, This refers to the final generated action.

[0237] The optimization objective of the policy network can be decomposed into two trade-offs: maximizing the Q-value and maximizing the entropy. The loss function guides the agent to learn the optimal cloth control policy, including: 1. State input, the policy network receives the state. 2. Action output: Distributed parameters of network output control commands. and 3. Loss calculation, including maximizing Q value and maximizing entropy, is used to encourage the selection of actions that can improve paving quality (uniformity), reduce material waste, and maintain a certain degree of randomness to adapt to uncertainties such as changes in water flow and stone properties.

[0238] In this embodiment, the SAC employs an adaptive mechanism to adjust. This eliminates the need for manual configuration. By setting a target entropy... (Usually taken as a negative number of the action space dimension), and then adjusted by optimizing the loss function. This ensures that the average entropy of the strategy remains near the target entropy, allowing the algorithm to automatically adjust the exploration level according to different working conditions (such as changes in water depth and flow velocity).

[0239]

[0240] in, Regarding the expectation of action, action From the current strategy Sampled from the distribution, The learning rate for the temperature parameter is... For loss function pairs The gradient. The specific adjustment process includes: when the policy entropy is too low (representing insufficient exploration), then increase... This strengthens the importance of the entropy term, encouraging more exploration, while reducing it when the policy entropy is too high (representing excessive randomness). This reduces the importance of the entropy term and decreases random exploration.

[0241] To ensure the safety of the underwater leveling machine, all generated motions must be projected onto a safe set of motions. In this context, the set defines the amplitude constraints of the action. and rate of change constraints ,Right now This can effectively prevent sudden changes in control commands and protect the mechanical structure of the underwater leveling machine.

[0242] in, For amplitude constraints, This is the lower bound of the action vector. The amplitude constraint is the upper bound of the action vector, and it limits the individual action vector. The value range of each component is designed to prevent the agent from outputting commands that exceed the physical capabilities of the underwater leveler, thereby avoiding equipment overload or damage. For example, it ensures that the conveyor belt is not commanded to operate at an unrealized ultra-high speed, or that the valve opening is set to 120%. As a rate of change constraint, For the change in motion, The rate of change constraint is used to suppress the "surge" phenomenon, protect the underwater screed from mechanical shock caused by sudden changes in commands, and make the construction process smoother and improve the laying quality.

[0243] Finally, the algorithm was designed with an end-to-end implementation process. On the data channel, the material distribution control unit... Collect sensor data periodically. (have to Sonar generation and via encoder , will the action After normalization to (-1,1), it is linearly or proportionally mapped to the actuator (such as frequency converter, proportional valve, servo, etc.), and the transfer tuple is stored in the experience pool to form a transfer tuple. ) into the warehouse, of which (e.g., RND enabled). During training, the commentator, policy, and temperature parameters are updated alternately, and soft updates are performed. To improve efficiency, mechanisms such as priority experience replay and random network distillation can be used to enhance exploration (e.g., using PER to sample high TD error segments and using RND to improve early exploration). During online deployment, the trained policy network can be placed on the edge computing device or GPU-NPU module of the decision subsystem 12 to ensure inference latency is less than 50 milliseconds. The material placement control unit forms a multi-layer closed-loop control: the upper-layer SAC decision gives the target material placement rhythm, the lower layer is precisely tracked by the PID controller of the decision subsystem 12, and second-level correction is performed in conjunction with the terrain feedback from sonar, and the correction is refreshed second-level in conjunction with the sonar-terrain loop, and the reliable operation of the underwater leveling machine is jointly ensured by motion safety projection combined with sensor redundancy (e.g., combining weight, flow, sonar and laser mutual verification) and human-machine takeover channel.

[0244] In the construction control system of the underwater screed provided in this application, the weight and tilt angle of the stones in the hopper are sensed in real time, and the precise material flow rate is dynamically calculated by combining time series data, providing accurate instantaneous working condition data for construction. Simultaneously, the material laying control unit compares the real-time collected subgrade elevation with the design target, calculating the terrain residual and the volume to be laid, thereby accurately grasping the actual state and requirements of the construction surface. Based on this, the control subsystem integrates thickness distribution prediction based on historical data and multimodal information, and comprehensively regulates key execution parameters such as conveyor belt speed and material laying valve opening through optimized decision-making algorithms. By combining real-time working condition monitoring, dynamic flow calculation, terrain deviation analysis, and predictive control strategies, closed-loop precise control is achieved throughout the entire process from material delivery to subgrade formation, effectively avoiding uneven laying or elevation deviation problems caused by information lag or human judgment errors in traditional construction, improving the control accuracy of underwater screed construction, and thus improving the construction accuracy and quality of the underwater subgrade.

[0245] In some optional embodiments, after the underwater leveling operation is completed, the intelligent control system 10 can also be used for substrate surface flatness assessment. The intelligent control system 10 can collect multi-source sensing data through multi-source mapping sensors deployed in the operation area, and generate substrate surface flatness assessment results for the operation area based on the flatness assessment unit.

[0246] The smoothness assessment unit is equipped with a trained target bed surface prediction model. The prediction unit is used to generate a multi-source fusion bed surface model based on multi-source sensor data, and uses the multi-source fusion bed surface model as input to the target bed surface prediction model to output an elevation prediction map of the bed surface based on the target bed surface prediction model; and generates a smoothness evaluation index for the operating sea area based on the elevation prediction map.

[0247] The target subgrade surface prediction model is obtained by training the initial surface prediction model. The initial surface prediction model is constructed by a convolutional neural network and a multi-head attention mechanism. The hyperparameters of the initial surface prediction model are optimized by the alpha evolution algorithm, and the loss function of the initial surface prediction model is the Huber function. The value of each grid point in the elevation prediction map represents the deviation of the subgrade surface from the ideal design plane or target elevation.

[0248] In this embodiment of the application, the operating sea area refers to the area where the underwater leveling operation has been completed, and the multi-source sensing data refers to the multi-source sensing data (including three-dimensional morphology and texture data) of the substrate surface obtained from different physical dimensions (such as acoustics and optics) using different types of sensors.

[0249] For example, after the leveling construction is completed, detailed information about the subgrade surface can be obtained by multi-source mapping sensors deployed in the operating sea area, including (1) multibeam echo sounding: multibeam echo sounders can be deployed on the operating vessel to perform high-resolution scanning of the leveled area and obtain data such as the digital elevation model (DEM) and reflection intensity map of the subgrade surface; (2) lidar / structured light scanning: depending on the water depth and water clarity, underwater lidar, structured light projector and camera devices can be used to perform fine three-dimensional scanning of the local area and obtain high-precision point cloud data; (3) underwater camera acquisition: underwater high-definition cameras can be deployed to capture visible light image sequences of the subgrade surface for photometric three-dimensional reconstruction or as a texture auxiliary means.

[0250] By performing operations such as correction, noise filtering, and outlier removal on the acquired multi-source sensor data, and transforming all data into a unified coordinate system, the limitations of a single data source can be overcome (e.g., sonar is susceptible to noise interference, and optical equipment is affected by water clarity). Through coordinate unification, registration, and complementarity, a more complete and accurate multi-source fusion substrate surface model can be generated. Ultimately, a multi-source fusion substrate surface 3D model is generated. This multi-source fusion substrate surface model can be represented as a Digital Elevation Model (DEM) with texture information.

[0251] For example, data acquired by each sensor can be transmitted to a support platform via an underwater communication link or buoy relay. The support platform then generates a multi-source fusion bed surface model based on the multi-source sensor data. The support platform can be a data processing system on a work vessel, a shore-based or fixed platform control center, or a cloud computing service platform.

[0252] By performing sound velocity correction, noise filtering, and outlier removal on multibeam bathymetry data, a seabed topographic point cloud or raster elevation map covering the leveled area is generated. Distortion correction and coordinate transformation (considering the effects of water refraction) are performed on laser / structured light point cloud data, which is then registered and fused with sonar data. Color calibration and distortion correction are performed on underwater images, and sparse point clouds are extracted using photogrammetry algorithms, then combined with laser point clouds for densification. After these preprocessing steps, a multi-source fused seabed surface model and its image texture in a unified coordinate system are obtained. This data will then be used as input to a deep learning model.

[0253] Please refer to Figure 9 , Figure 9 This is a schematic diagram of the framework of the target subgrade surface elevation prediction model provided in this application embodiment. The target subgrade surface prediction model is an intelligent agent for assessing the flatness of the subgrade surface. The initial surface prediction model is an end-to-end model constructed using Convolutional Neural Networks (CNNs) and a multi-head attention mechanism. CNNs excel at extracting local spatial features (such as minute undulations) from images or raster data, while the multi-head attention mechanism can capture the global dependencies between different data sources, weigh the importance of information from each source, and achieve more effective feature fusion. The target subgrade surface prediction model achieves accurate prediction of the subgrade surface elevation through this end-to-end model.

[0254] The model's input is a multi-source fused substrate surface model, specifically represented by three data extraction modules: raster elevation, depth, and texture features. The 3D model generated by the platform is converted into regular raster data, where each grid point contains not only elevation values ​​but also depth information derived from multibeam data and texture features extracted from underwater imagery.

[0255] These input data are fed into a convolutional neural network (CNN) module. The CNN layer shown in the diagram typically contains a series of sequential operations, including: convolution operations that use learnable convolutional kernels to perform sliding window calculations on the input data to capture local spatial features ranging from edges and shapes to more complex terrain structures; the ReLU activation function introduces a non-linear transformation into the network, enabling it to fit complex non-linear relationships; and max pooling operations, which reduce the spatial resolution of the feature maps while preserving salient features, increase the model's translation invariance and receptive field. This series of operations works together to transform the raw input into a high-level, abstract feature map.

[0256] The core of convolutional neural networks lies in the convolution operator and feature fusion. By sliding and multiplying the convolution kernel with the input feature map spatially and accumulating the results, weighted extraction of local patterns is achieved. The output of each convolutional layer undergoes nonlinear activation. (ReLU) mapping introduces sparsity and nonlinear discriminative power. The mathematical expression of the multi-head attention mechanism is: for the input feature sequence... After calculating the key, query, and value vectors, the self-attention output is performed. This mechanism can capture global dependencies. By combining local features extracted through convolution, the model can simultaneously obtain local details and global context, improving the accuracy of terrain prediction.

[0257] The hyperparameters of the initial surface prediction model are optimized using the Alpha Evolution (AE) algorithm, which achieves an efficient balance between global exploration and local exploitation through a core Alpha operator, thereby automatically finding an optimal set of hyperparameters.

[0258] After the CNN extracts spatial features, it enters a multi-head attention mechanism layer. This allows the model to autonomously focus on deeper dependencies between different features. For example, one set of heads might focus on the correlation between elevation and texture features, while another set might focus on the spatial relationships between distant grid points. In this way, the model can more intelligently fuse information from different data sources and understand the broader terrain context, rather than just the local neighborhood, thus making more accurate predictions.

[0259] For example, raster elevation, depth, and texture are fused using a multi-head attention mechanism and then input into the same convolutional network. During the encoding stage, multiple convolutional and pooling layers are stacked sequentially to extract high-level features. The convolutional layers use a certain kernel size (e.g., 3×3) to slide and extract local region features. Each convolutional layer is followed by a ReLU activation function to introduce non-linearity, combined with batch normalization (BN) to accelerate convergence and improve generalization performance. For example, four convolutional layers are used, each with a 3×3 kernel size, and the number of channels increases progressively (e.g., 32, 64, 128, 256), with ReLU and batch normalization applied after each layer. The convolutional computation satisfies: ,in Represents the input image tensor. For convolution kernel weights, It outputs the position index. The kernel's height and width range are defined. Local feature extraction is achieved through convolution operations, while pooling layers downsample the feature map using a 2×2 window, reducing resolution while extracting invariant features. After feature extraction via convolution, a fully connected layer merges spatial information to output a global flatness index for predicting continuous elevation or deviation values.

[0260] The high-level features, weighted and integrated by the multi-head attention mechanism, are flattened and fed into one or more fully connected layers. The fully connected layer acts like a powerful regressor, mapping the distributed features learned from all preceding steps to the final output space. In this task, the output layer is typically a linear unit, whose output value directly corresponds to the predicted elevation value for each grid point.

[0261] The elevation prediction map directly displays the elevation distribution of the leveled subgrade surface. Due to the use of a deep learning model, it can obtain more detailed terrain features than traditional interpolation methods. (Output elevation prediction map) The measured true elevation can be subtracted. (Or subtract the reference elevation surface) to obtain the deviation The deviation diagram visually displays which areas of the subgrade still have protrusions or depressions, and its numerical value reflects the unevenness. This output allows for convenient calculation of flatness evaluation indicators, such as root mean square error, for acceptance assessment.

[0262] In some embodiments, to train an initial surface prediction model, this application defines a loss function to measure the deviation between the predicted elevation and the true elevation. This application uses the Huber function as the loss function for the initial surface prediction model, measuring the difference between the model's predicted value and the true value. The Huber function combines the advantages of Mean Squared Error (MSE) and Mean Absolute Error (MAE), is insensitive to outliers in the training data, and improves the stability of model training.

[0263] The mathematical form of Huber loss is:

[0264] in, The difference between the predicted value and the actual value. The threshold hyperparameter controls the inflection point at which the loss transitions from quadratic to linear. This is achieved by selecting an appropriate threshold hyperparameter. The Huber loss is equivalent to the MSE when the error is small, and equivalent to the absolute error when the error is large, thus improving robustness to outliers. Model training aims to minimize the above loss, and the network parameters are adjusted through backpropagation to make the predicted elevation approximate the true value.

[0265] Please refer to Figure 10 , Figure 10 This is a schematic diagram illustrating the specific process of the AE algorithm provided in an embodiment of this application. In some embodiments, the step of optimizing the hyperparameters of the initial surface prediction model based on the AE algorithm may include: In the hyperparameter space, multiple candidate solutions are generated in a uniformly distributed manner. Each candidate solution represents a hyperparameter configuration vector of an initial surface prediction model. The hyperparameters include the kernel size, learning rate, batch size, and network depth of the initial surface prediction model. The value range of each hyperparameter in the hyperparameter space is determined based on the characteristics of the bed topography data. An evolution matrix is ​​constructed by randomly sampling from the current candidate solutions. Each row vector of the evolution matrix corresponds to a hyperparameter configuration. The evolution matrix is ​​initialized, and an evaluation function is calculated. The evaluation function is the loss function value calculated by the initial surface prediction model on the bed validation set. It is determined whether the current function evaluation count has reached the maximum number of evaluations. If not, the evolution matrix, perturbation matrix, and decay factor are calculated. The calculation path of the basis vector is selected by generating random numbers. When the random number is less than a set threshold, ... Calculate the basis vectors using the first path, otherwise calculate them using the second path. After obtaining the basis vectors based on the selected path, select the first individual with a fitness higher than a preset threshold and the second individual with a fitness lower than a preset threshold from the current population, and calculate the control parameters. Update the search operator using the basis vectors, control parameters, decay factor, and perturbation matrix through the alpha operator. Apply boundary constraints to the updated search operator and use a binary backoff strategy to adjust the solution components that exceed the boundary. Compare the fitness of individuals before and after the update using a greedy selection strategy to decide whether to replace the original individuals. After updating all individuals in the current generation, check if the individual index has reached the population size; if not, continue processing the next individual. After processing all individuals, check again if the function evaluation count has reached the maximum value; if so, output the optimal hyperparameter combination.

[0266] For example, in In the 3D hyperparameter space, the AE algorithm first generates the hyperparameters in a uniform distribution manner. There are several candidate solutions to form a candidate matrix:

[0267] in, Indicates the first One hyperparameter configuration vector, These are the lower and upper bounds of each dimension; This indicates the generation of a length of The vector, each component in Uniform sampling.

[0268] The Alpha operator, by simultaneously fusing three types of information—adaptive global starting point, global random perturbation, and local difference correction—takes into account both global exploration and local development in a single update, thereby efficiently mining and refining candidate solutions. Its mathematical expression is:

[0269] Among them, the evolution matrix By analyzing the candidate matrix A sample obtained by sampling with replacement Matrix. Specifically, in each generation, the algorithm starts from the candidate matrix... Randomly selected from candidate solutions (Repetition allowed) , its first OK That is the first A solution awaiting update; This represents the basis vector, which determines the starting position of the evolution; As a decay factor, it controls the exploration and development of algorithms; Indicates the first A random step size; To control the parameters, control the differential vector (adaptive step size); and That is, from The solution extracted from the middle satisfies .

[0270] For adaptive basis vectors, the evolutionary starting point is initially calculated in two ways:

[0271] in From the candidate matrix Sampling with replacement The square array obtained this time This indicates taking its diagonal; From Sampling without replacement The matrix obtained this time This is a fitness-based weight vector. To make... As historical information accumulates across generations, the AE algorithm introduces evolutionary paths for the two sampling methods.

[0272] The methods for calculating basis vectors through the first path include: ; Methods for calculating basis vectors using the second path include: ; in, As basis vectors, and The learning rate parameter, and These represent the historical basis vectors based on sampling paths with replacement and those based on sampling paths without replacement, respectively; A is a D×D matrix obtained by sampling D times with replacement from the candidate solution matrix X, and B is a matrix obtained by sampling K times without replacement from the candidate solution matrix X. This is a fitness-based weight vector.

[0273] Learning rate , This represents the current number of times the target function has been called. This represents the maximum number of calls allowed.

[0274] For random step size It provides a global search function. The attenuation factor is a non-linear decreasing value, related to the perturbation matrix. Closely related. Its decay process is as follows:

[0275] perturbation matrix The calculation is as follows:

[0276] in and Indicates by The generated random real matrices are used to generate perturbations. Among them, It is a collection of lines and A matrix of columns, express The first in Row vectors.

[0277] To ensure the search remains within the feasible region, AE employs a "binary backoff" approach for variables that exceed the bounds:

[0278] The updated system will use a greedy strategy to select and retain the best performers.

[0279] in, This is the value of the j-th dimension (i.e., a specific hyperparameter, such as the learning rate) of the i-th candidate solution after being updated by the alpha operator. Let j be the upper bound of the j-th hyperparameter in the search space. This is the lower bound of the j-th hyperparameter in the search space. Let be the hyperparameter configuration vector for the i-th candidate solution in the t-th generation of the population. The vector of new candidate solutions is generated after the i-th candidate solution is updated by the alpha operator and boundary constraints are applied.

[0280] Supervised learning was employed during model training, with label data derived from high-precision measured bed elevations (e.g., real elevations obtained from underwater laser point clouds or multibeam bathymetry). The measured elevation of a specific area after leveling construction was used as the ground truth. Multi-source observation data (sonar depth maps, image textures, etc.) of the corresponding region are used as model input. Construct training samples Yes. The loss function should use either Huber or MSE as mentioned above to measure the model output. and The difference is negligible. During training, the model parameters are optimized by minimizing the loss, making the predicted results approximate the true elevation distribution. The optimization algorithm uses the Adam adaptive gradient optimizer, with an initial learning rate set to 0.001, combined with a momentum factor. Adam optimization exhibits good convergence efficiency and robustness to hyperparameters in deep learning. To ensure training stability, embodiments of this application employ a learning rate decay strategy during training, for example, multiplying the learning rate by 0.1 every few epochs to prevent later oscillations.

[0281] Regarding hyperparameter selection and validation, the batch size during training depends on the hardware memory and data size. For raster input, batch sizes such as 32 or 64 can be chosen; if training with small image patches, even larger batch sizes can be set (some studies have used 512 batches to train 9×9 small patches to fully utilize the data). Model training requires dividing the training and validation sets, typically randomly partitioning the data in a ratio of, for example, 8:2, while ensuring data coverage across different regions. To fully utilize limited data and evaluate model robustness, this application's embodiments introduce a K-fold cross-validation scheme. For example... The dataset is divided into five equal parts. One part is used for validation each time, and the remaining four parts are used for training. This process is repeated five times to obtain the average performance. This approach provides a more reliable evaluation of the model compared to a single partition, reducing bias caused by random partitions. A study compared the effects of traditional random partitioning with K-fold validation, finding that the latter improved test accuracy from 84.1% to 88.3% and reduced the spatial error standard deviation by 60%, indicating that cross-validation helps improve the model's generalization performance. Preventing overfitting: Due to the high cost of acquiring underwater terrain data and the limited number of training samples, models are prone to overfitting. To address this, several regularization strategies are employed: First, Dropout layers are used, randomly discarding neuron outputs with a certain probability (e.g., 30%) in fully connected layers or during the decoding stage to break feature dependencies. Dropout reduces over-reliance on certain local features during training, improving the model's adaptability to unseen data. Second, early stopping is applied, monitoring the validation set loss and stopping training when there is no improvement after several epochs to avoid overtraining. Third, data augmentation can be performed. If the original imagery and depth measurement data allow, random rotations, translational perturbations, or noise can be added to the training samples to increase data diversity and thus improve model robustness. Finally, introducing residual connections and regularization terms into the network structure is also helpful. Residual connections allow for training deeper networks without sacrificing stability; regularization weight decay is added to the loss. This approach can suppress excessively large model parameters. By comprehensively utilizing the above strategies, the embodiments of this application can focus more on the performance of validation error while ensuring a reduction in training error, striving to obtain a smoothness evaluation model with strong generalization ability and applicability to different environments.

[0282] Furthermore, to comprehensively evaluate model performance, embodiments of this application select multiple metrics for quantitative evaluation on the validation set, including RMSE, MAE, and These indicators measure the degree of agreement between predicted and actual elevations from different perspectives. Among them, the root mean square error (RMSE) measures the overall error magnitude, reflecting the standard deviation of the predicted value from the actual value. The mean absolute error (MAE) is the average of the absolute values ​​of the errors, and the formula is... The median (MedAE) directly represents the magnitude of the mean deviation. Because the median is highly resistant to outliers, MedAE better reflects the typical error level of most points, ignoring a very small number of outliers. (Coefficient of determination) The measure of how well a model explains actual elevation changes is defined as follows: . The value ranges from 0 to 1. The closer it is to 1, the more terrain variance the model explains, meaning the better the prediction fits the true value.

[0283] In some embodiments, the smoothness evaluation index may include an overall smoothness index and a local smoothness index; wherein, the overall smoothness index includes the root mean square error, used to quantify the overall elevation fluctuation of the operating sea area; the local smoothness index includes the average slope change and local extreme value detection results of the operating sea area, and the local smoothness index is used to characterize the surface smoothness and local unevenness rate of the operating sea area. Step S12 may specifically include: inputting the multi-source fusion substrate surface model into the target substrate surface prediction model, and obtaining the elevation prediction map of the substrate surface by forward inference from the target substrate surface prediction model; based on the elevation prediction map and the reference elevation data, calculating the elevation deviation value of each grid point, and generating an elevation deviation distribution map; wherein, the reference elevation data is the measured substrate elevation or the ideal design plane elevation, and the elevation deviation distribution map includes the convex or concave areas of the substrate surface, and the magnitude of the deviation value in the elevation deviation distribution map characterizes the unevenness of the operating sea area.

[0284] The trained target subgrade surface prediction model can output a corresponding elevation prediction map based on the input multi-source fused subgrade surface model. Each pixel or grid point in the map corresponds to a predicted absolute elevation value. By comparing this prediction map with the ideal design plane or target elevation (i.e., the perfectly flat surface required for construction), the deviation at each point can be obtained, thus forming an elevation deviation distribution map. The variation in the elevation deviation distribution map is the relative deviation value; a positive deviation value indicates a convexity, and a negative deviation value indicates a depression.

[0285] The steps for generating flatness evaluation indicators for the operational sea area based on the elevation deviation distribution map can specifically include: calculating the mean square error or root mean square error of the deviation values ​​at each point, and quantifying the overall elevation fluctuation of the operational sea area based on the mean square error or root mean square error; calculating the slope change of the subgrade surface, determining the local slope through the elevation difference between adjacent grid points, and evaluating the maximum rate of change of the operational sea area.

[0286] For example, after training, the model is deployed for flatness assessment of field data. Preprocessed and fused multi-source data is input into the trained CNN model, and forward inference is performed to obtain the predicted flatness of the subgrade surface. Next, flatness indices are calculated from the model output: the mean square error (MSE) or root mean square error (RMSE) is calculated based on the deviation values ​​at each point to quantify the overall elevation fluctuation; simultaneously, the slope variation of the subgrade surface is calculated, for example, by obtaining the local slope through the elevation difference between adjacent grid points and evaluating its maximum rate of change to capture steep slopes or uneven areas. These indices comprehensively reflect the flatness after leveling. Finally, a flatness assessment report is generated, and the construction quality is judged by comparing it with pre-set acceptance thresholds: if the mean square error and other indices are within the allowable range, the flatness of the subgrade surface is considered acceptable; otherwise, areas exceeding the tolerance are marked for rework and repair.

[0287] In this embodiment, a visualization method can be used to compare and analyze the elevation / deviation map output by the model with the actual measurement. On the one hand, a comparison map of the predicted elevation and the measured elevation (such as a profile comparison, a 3D surface map, etc.) is drawn to visually check whether the model has captured the key undulation features of the subgrade surface. For an ideally flat area, the predicted and measured curves should basically coincide; if there is a large deviation in the prediction at a certain point, it will be shown as a significant elevation or depression on the map. On the other hand, a heat map is used to display the distribution of deviations, and the magnitude of the deviation is represented by color, which can clearly locate the location and degree of residual unevenness. This scatter plot comparison map and deviation heat map can quantitatively and intuitively evaluate the leveling effect: dense point clouds that are close to the ideal line indicate good overall flatness, while outliers and areas with concentrated deviations indicate areas that need to be re-examined. Through the analysis of the above-mentioned multiple visualization methods, we not only verify the accuracy of the model prediction, but also can further guide the rework and optimization of construction based on the deviation map, ultimately achieving a comprehensive and objective evaluation of the leveling quality of the underwater subgrade.

[0288] In the process of generating the subgrade surface smoothness assessment results for the aforementioned operational sea area, a multi-source fusion subgrade surface model was constructed by acquiring multi-source sensor data from the operational sea area, effectively improving the accuracy and robustness of the basic data. Subsequently, the fusion model was processed using an optimized target subgrade surface prediction model. This model combines the feature extraction capabilities of convolutional neural networks and the advantages of multi-head attention mechanisms in capturing long-distance dependencies. Furthermore, the hyperparameters were optimized through an evolutionary algorithm, resulting in a model with stronger adaptability and accuracy in predicting subgrade surface elevation. The use of the Huber loss function further ensured the model's training stability in the presence of outliers. The final generated elevation prediction map intuitively reflects the deviation of each point on the subgrade from the design elevation, providing a direct basis for generating smoothness evaluation indicators. This allows for a comprehensive and objective assessment of the subgrade surface smoothness in the operational sea area, significantly improving the accuracy and reliability of the subgrade surface smoothness assessment.

[0289] It should be understood that when the various modules of the system provided in the above embodiments are working, the division of each functional module in the above description is only used as an example. In actual applications, the above functions can be assigned to different functional modules as needed. That is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0290] The functional modules in the above embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0291] Based on the same concept, embodiments of this application also provide an intelligent control method for an underwater leveling machine, which may include: The system acquires the working status data of the underwater leveling machine; discretizes the working process of the underwater leveling machine into multiple stages, and determines the framework parameters of the current stage based on the working status data within each stage; under the condition of satisfying multiple preset constraints, determines the liquid level change of each target compartment based on the framework parameters, and generates control commands for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage; and controls the underwater leveling machine to perform underwater work according to the multi-stage control commands.

[0292] Optionally, the method may further include: discretizing the working process of the underwater leveler into multiple stages based on time intervals or depth intervals; and in each stage, switching to the next stage when a trigger condition is reached, such as a fixed duration, an intermediate draft threshold, or an inclination angle threshold.

[0293] Optionally, the method may further include: for each compartment, calculating antecedent variables for fuzzy inference; wherein the antecedent variables include correction alignment, lever arm effectiveness, and level margin, the correction alignment is calculated based on the consistency between the attitude gradient and the target correction direction, the lever arm effectiveness is calculated based on the lever arm weight of the compartment for roll or pitch, and the level margin is calculated based on the safety margin between the current level and the limit; mapping the antecedent variables to a fuzzy set through a predefined membership function, and applying a fuzzy rule base built based on domain knowledge for inference to obtain the participation level representing the compartment; and determining whether the compartment is the target compartment based on the participation level.

[0294] Optionally, the method may further include: using a fuzzy particle swarm optimization algorithm, along with constraints such as hyperplane projection and boundary repair processing, to search for the optimal decision variable within a sparse search space composed of the framework parameters of the current stage; the optimal decision variable is the liquid level change of each target compartment.

[0295] Optionally, the method may further include: calculating the opening duration and action sequence of the pump or valve corresponding to each target compartment based on the sign and magnitude of the liquid level change in each target compartment, combined with the target hardware parameters of the underwater leveling machine, and generating corresponding control commands.

[0296] Optionally, the method may further include: monitoring working status data when the execution subsystem executes control commands; if the working status data indicates that the attitude or draft deviation of the underwater leveler exceeds a preset threshold, triggering a local correction operation to adjust the liquid level change and reissue control commands.

[0297] Based on the same concept, embodiments of this application also provide a computer device, which is mounted on or communicatively connected to an underwater leveling machine. This computer device may include a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the methods described above.

[0298] Based on the same concept, this application also provides an underwater leveling machine, which is equipped with the intelligent control system 10 of the underwater leveling machine described above.

[0299] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An intelligent control system for an underwater leveling machine, characterized in that, include: A sensing subsystem configured to acquire the working status data of the underwater leveling machine; The decision-making subsystem includes a stage decision-making unit and a fuzzy swarm optimization unit. The decision subsystem is configured to discretize the working process of the underwater leveling machine into multiple stages, and within each stage, determine the framework parameters of the current stage based on the working status data; The fuzzy swarm optimization unit is configured to determine the liquid level change of each target compartment based on the framework parameters under the condition of satisfying multiple preset constraints, and generate control commands for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage. An execution subsystem is configured to receive multi-stage control commands sent from the decision subsystem and control the underwater leveling machine to perform underwater operations based on the multi-stage control commands. The decision subsystem is configured as follows: A stage decision maker is used to discretize the underwater leveling machine's operation into multiple stages. The stage state of each stage is represented by a vector composed of the liquid level heights of all compartments at the current moment, and the decision of each stage is represented by a vector composed of the liquid level changes of all compartments in this stage. An optimization algorithm minimizes the cost function of each stage to determine the liquid level change decision for each stage. The cost function consists of pumping time cost, valve action time cost, and a penalty term. The penalty term is used to suppress frequent start-stop and liquid level oscillation.

2. The intelligent control system for the underwater leveling machine according to claim 1, characterized in that, The stage decision unit is configured to discretize the working process of the underwater leveling machine into multiple stages based on time intervals or depth intervals.

3. The intelligent control system for the underwater leveling machine according to claim 2, characterized in that, The stage decision unit is also configured to switch to the next stage in each stage when a fixed duration, intermediate draft threshold, or tilt angle threshold is reached.

4. The intelligent control system for the underwater leveling machine according to claim 1, characterized in that, The fuzzy crowd intelligence optimization unit is configured as follows: For each compartment, antecedent variables for fuzzy inference are calculated; wherein, the antecedent variables include correction alignment, lever arm effectiveness, and level margin, wherein the correction alignment is calculated based on the consistency between the attitude gradient and the target correction direction, the lever arm effectiveness is calculated based on the lever arm weight of the compartment for roll or pitch, and the level margin is calculated based on the safety margin between the current level and the limit; the antecedent variables are mapped to fuzzy sets through a predefined membership function, and a fuzzy rule base built based on domain knowledge is applied for inference to obtain the participation level representing the compartment; based on the participation level, it is determined whether the compartment is the target compartment.

5. The intelligent control system for the underwater leveling machine according to claim 4, characterized in that, The fuzzy swarm optimization unit is further configured to use a fuzzy particle swarm optimization algorithm, along with constraints such as hyperplane projection and boundary repair processing, to search for the optimal decision variable within a sparse search space composed of the framework parameters of the current stage; the optimal decision variable is the liquid level change of each target compartment.

6. The intelligent control system for the underwater leveling machine according to claim 1, characterized in that, The decision subsystem is also configured to calculate the opening duration and action sequence of the pump or valve corresponding to each target compartment based on the sign and magnitude of the liquid level change in each target compartment, combined with the target hardware parameters of the underwater leveling machine, and generate corresponding control commands.

7. The intelligent control system for the underwater leveling machine according to claim 1, characterized in that, The decision subsystem is further configured to monitor the working status data when the execution subsystem executes the control command. If the working status data indicates that the attitude or draft deviation of the underwater leveler exceeds a preset threshold, a local correction operation is triggered to adjust the liquid level change and reissue the control command.

8. An intelligent control method for an underwater leveling machine, characterized in that, include: Acquire the operating status data of the underwater leveling machine; The working process of the underwater leveling machine is discretized into multiple stages, and the frame parameters of the current stage are determined based on the working status data in each stage. Under the condition of satisfying multiple preset constraints, the liquid level change of each target compartment is determined based on the framework parameters, so as to generate the control command for the current stage based on the liquid level change; wherein, the target compartment is the compartment that needs to perform liquid level adjustment operation in the current stage. The underwater leveling machine is controlled to perform underwater operations according to multi-stage control commands; The step of determining the liquid level change of each target compartment based on the framework parameters under the condition of satisfying multiple preset constraints, and generating control commands for the current stage based on the liquid level change, includes: A stage decision maker is used to discretize the underwater leveling machine's operation into multiple stages. The stage state of each stage is represented by a vector composed of the liquid level heights of all compartments at the current moment, and the decision of each stage is represented by a vector composed of the liquid level changes of all compartments in this stage. An optimization algorithm minimizes the cost function of each stage to determine the liquid level change decision for each stage. The cost function consists of pumping time cost, valve action time cost, and a penalty term. The penalty term is used to suppress frequent start-stop and liquid level oscillation.

9. A computer device, characterized in that, The computer device is mounted on or communicates with the underwater leveling machine. The computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method as described in claim 8.

10. An underwater leveling machine, characterized in that, The underwater leveling machine is equipped with an intelligent control system as described in any one of claims 1-7.

Images