Intelligent control system, method and device of underwater screed and underwater screed

Abstract

The application provides an intelligent control system, method and device of an underwater screed and the underwater screed, and relates to the field of ocean engineering equipment and automatic control. The intelligent control system comprises: a perception subsystem configured to obtain working state data of the underwater screed; a decision subsystem comprising a stage decision unit and a fuzzy group intelligence optimization unit; the decision subsystem is configured to discretize the working process of the underwater screed into multiple stages, and determine the framework parameters of the current stage based on the working state data in each stage; the fuzzy group intelligence optimization unit is configured to determine the liquid level change amount of each target cabin based on the framework parameters to generate the control instruction of the current stage; and an execution subsystem configured to receive the multi-stage control instruction sent from the decision subsystem, and control the underwater screed to work underwater based on the multi-stage control instruction. The system can improve the stability of the posture control of the underwater screed in a complex working environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligentization and automatic control of ocean engineering equipment, and particularly relates to an intelligent control system and method for an underwater screed, equipment and underwater screed. BACKGROUND

[0002] In seabed base leveling operations, the stability of the posture control of the underwater screed directly determines the construction accuracy and equipment safety. However, the existing technology has long been facing the problem of unstable posture control in actual application, mainly manifested in that the screed is prone to sustained oscillation, response lag or slow convergence during adjustment. The root cause of this problem lies in the insufficient coordination ability of the control system in the multi-target and strong coupling environment.

[0003] The existing control strategy mainly relies on local feedback mechanism and lacks global dynamic planning. For example, when the screed is tilted, the system only triggers the liquid level adjustment of the corresponding cabin according to the instantaneous attitude error, without optimizing the continuous process of diving and leveling as a whole. This "single-point response" mode leads to the conflict between the adjustment actions of different cabins in time sequence, such as overlapping of water injection and drainage operations, causing imbalance of buoyancy distribution, and further causing repeated fluctuations of the machine body. At the same time, the real-time optimization of high-dimensional decision-making exacerbates the instability. The screed is usually equipped with multiple ballast tanks, and the liquid level adjustment involves multi-parameter coupling calculation of weight, floating center and stable center. The traditional algorithm is difficult to solve the global optimal solution in a limited time, often falling into a local optimization trap, and the generated control instructions may excessively rely on a small number of cabins or ignore dynamic disturbances, so that the screed cannot be quickly stabilized in complex water flow or load change.

[0004] In summary, due to the problems of lack of global planning, insufficient real-time optimization ability and disconnection of physical constraints, the existing technology is difficult to realize the rapid and stable control of the posture of the screed. SUMMARY

[0005] Therefore, the present application aims to provide an intelligent control system and method for an underwater screed, equipment and underwater screed, to generate multi-stage control instructions through multi-stage global planning and fuzzy swarm optimization, so as to improve the stability of the posture control of the underwater screed.

[0006] In a first aspect, the present application provides an intelligent control system of an underwater screed machine, comprising: a perception subsystem configured to obtain working state data of the underwater screed machine; a decision subsystem comprising a stage decision unit and a fuzzy swarm intelligence optimization unit; the decision subsystem is configured to discretize the working process of the underwater screed machine into multiple stages, and determine the framework parameters of the current stage based on the working state data in each stage; the fuzzy swarm intelligence optimization unit is configured to determine the liquid level change amount of each target tank based on the framework parameters under the condition of meeting multiple preset constraints, to generate the control instruction of the current stage based on the liquid level change amount; wherein the target tank is the tank that needs to perform liquid level adjustment operation in the current stage; an execution subsystem configured to receive the multi-stage control instruction sent from the decision subsystem, and control the underwater screed machine to work underwater based on the multi-stage control instruction.

[0007] According to an embodiment of the present application, the stage decision unit is configured to discretize the working process of the underwater screed machine into multiple stages based on time intervals or depth intervals.

[0008] According to an embodiment of the present application, the stage decision unit is further configured to switch to the next stage in each stage when the trigger conditions of reaching a fixed time length, an intermediate draft threshold or an inclination threshold are met.

[0009] According to an embodiment of the present application, the fuzzy swarm intelligence optimization unit is configured to: for each tank, calculate the antecedent variables of fuzzy reasoning; wherein the antecedent variables include correction alignment, force arm efficiency and liquid level margin, the correction alignment is calculated based on the consistency of the attitude gradient and the target correction direction, the force arm efficiency is calculated based on the force arm weight of the tank to the roll or pitch, and the liquid level margin is calculated based on the safety margin of the current liquid level and the limit value; map the antecedent variables to fuzzy sets through a predefined membership function, and apply a fuzzy rule base constructed based on domain knowledge for reasoning to obtain a participation level representing the tank; determine whether the tank is a target tank based on the participation level.

[0010] According to an embodiment of the present application, the fuzzy swarm intelligence optimization unit is further configured to search for optimal decision variables in a sparse search space composed of the framework parameters of the current stage by using a fuzzy particle swarm optimization algorithm and constraint conditions of hyperplane projection and boundary repair processing; the optimal decision variables are the liquid level change amounts of each target tank.

[0011] According to an embodiment of the present application, the multiple preset constraints of the fuzzy swarm intelligence optimization unit include at least one of attitude safety envelope constraint, mass conservation constraint, tank geometry limit constraint, and pump valve capacity constraint.

[0012] According to an embodiment of the present application, the decision subsystem is further configured to calculate the opening duration and action timing of the pump or valve corresponding to each target compartment according to the positive and negative and size of the liquid level change amount of each target compartment, in combination with the target hardware parameters of the underwater screed machine, and generate corresponding control instructions.

[0013] According to an embodiment of the present application, the decision subsystem is further configured to monitor the working state data in the case that the execution subsystem executes the control instructions, and if it is detected that the working state data represents that the attitude or draft of the underwater screed machine deviates beyond a preset threshold, a local correction operation is triggered to adjust the liquid level change amount and reissue the control instructions.

[0014] According to an embodiment of the present application, the decision subsystem further comprises 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 according to the working state data, and construct a digital twin model of the screed machine operation; the job suggestion unit is configured to generate operation suggestions according to the working state data and the operation task of the screed machine; and the job control unit is configured to control the execution subsystem to perform the operation task according to the operation suggestions.

[0015] According to an embodiment of the present application, the operation task at least includes a material distribution task; after obtaining the material distribution task and the working state data, the job suggestion unit calculates the regional material distribution feasibility by using a material distribution scoring function, recommends the material thickness and density values by using fuzzy rules, predicts potential construction defect areas and high-risk advancing paths by using a dynamic window algorithm and a dynamic cost weighting mechanism, and generates a material distribution operation suggestion including the material amount and the advancing direction.

[0016] According to an embodiment of the present application, the intelligent control system of the underwater screed machine further comprises an underwater measurement subsystem, which comprises: a binocular camera device configured to obtain binocular vision images of a target water area; and an edge computing device configured to obtain point clouds of the target water area; wherein the underwater measurement subsystem is installed on the underwater screed machine, the binocular camera device is arranged at the front part of the main frame of the underwater screed machine, and the edge computing device is in communication connection with the material distribution device in the underwater screed machine.

[0017] According to one embodiment of the present application, the edge computing device is configured to calibrate and correct the binocular vision image to obtain the parallax corresponding to each pixel point; extract the original gray information of the binocular vision image, and then enhance the binocular vision image to generate gradient information; calculate the matching cost according to the parallax of each pixel point in the binocular vision image and the corresponding original gray information and gradient information, and construct a cost function; calculate the parallax map of the binocular vision image by simultaneously solving the cost function through the improved SGM algorithm; and perform back projection on the parallax 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: a cost function based on four-direction aggregation; setting a dynamic adjustment strategy to control the punishment of small parallax jump and the punishment of large parallax jump; and performing confidence screening and hole filling on the aggregated parallax map.

[0018] According to one embodiment of the present application, the decision subsystem further includes a paving control unit; the paving control unit is configured to: in the case that the underwater screed machine is performing leveling construction, obtain the stone weight in the hopper of the underwater screed machine and the hopper inclination angle based on the weighing sensor and the inclinometer of the perception subsystem; determine the stone quality based on the stone weight in the hopper and the hopper inclination angle, and calculate the real-time paving mass flow rate based on the time series data of the stone quality; and obtain the current terrain elevation of the base bed, compare the current terrain elevation of the base bed with the pre-stored design target elevation, calculate the terrain residual error and the to-be-paved volume; generate a paving control instruction through an optimization decision algorithm based on the real-time paving mass flow rate, the terrain residual error, the to-be-paved volume, and the predicted thickness distribution; wherein the predicted thickness distribution is obtained by a target prediction model based on historical construction data and multi-modal sensing information, and the paving control instruction is used to adjust at least one of the conveying belt speed, the paving valve opening degree, or the paving head moving speed of the underwater screed machine.

[0019] According to one embodiment of the present 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 multi-modal fusion state tensor composed of the stone quality, the real-time paving mass flow rate, the encoded terrain residual error, the paving head pose, the paving head speed, and the working condition 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 instructions of the conveying belt speed, the paving valve opening degree, or the paving head moving speed; and the agent of the deep reinforcement learning framework trains the optimization paving strategy by maximizing the comprehensive reward function.

[0020] According to one embodiment of the present application, the perception subsystem further includes a multi-source surveying sensor arranged in the working sea area, which is used to collect multi-source sensing data.

[0021] According to one embodiment of the present application, the decision subsystem further comprises a flatness evaluation unit configured to: receive multi-source sensing data, generate a multi-source fusion base surface model based on the multi-source sensing data; input the multi-source fusion base surface model as an input of a target base surface prediction model to output an elevation prediction map of the base surface based on the target base surface prediction model; and generate a flatness evaluation index of the work sea area based on the elevation prediction map. The target base surface prediction model is obtained by training an 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 an alpha evolutionary algorithm, and the loss function of the initial surface prediction model is a Huber function. The value of each grid point in the elevation prediction map represents the deviation of the base surface at this point from the ideal design plane or target elevation.

[0022] The present application also provides an intelligent control method of an underwater screed machine, comprising: obtaining working state data of the underwater screed machine; discretizing a working process of the underwater screed machine into multiple stages, and determining a framework parameter of a current stage based on the working state data in each stage; determining a liquid level change amount of each target tank based on the framework parameter under the condition of satisfying multiple preset constraints, to generate a control instruction of the current stage based on the liquid level change amount; wherein the target tank is a tank that needs to perform a liquid level adjustment operation in the current stage; and controlling the underwater screed machine to perform underwater work according to the multi-stage control instruction.

[0023] Based on the same application concept, the present application also provides a computer device, which is arranged on or in communication connection with the underwater screed machine, and comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method described in the above content when executing the computer program.

[0024] Based on the same application concept, the present application also provides an underwater screed machine, which is provided with the intelligent control system of the underwater screed machine according to the above embodiments.

[0025] According to one embodiment of the present application, the underwater screed machine is a walking type underwater screed machine.

[0026] Compared with the prior art, the beneficial effects of the present application are that the intelligent control system can decompose the complex continuous control process into a discrete and ordered phase sequence through a multi-stage control strategy, change the control system from passive instantaneous response to active and predictable global optimization, and make fine decisions through fuzzy group wisdom optimization, intelligently screen out the target cabin set most effective for current posture correction through fuzzy reasoning, and in this simplified space, cooperatively calculate the optimal liquid level adjustment amount of each cabin to ensure accurate matching of the buoyancy distribution to the current demand, thereby improving the stability of the posture control of the underwater screed machine in complex operation environment. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 The schematic diagram of the intelligent control system of the underwater screed machine provided by the embodiment of the present application is shown.

[0028] Figure 2 The schematic diagram of the main frame of the underwater screed machine provided by the present application is shown.

[0029] Figure 3 The schematic diagram of the material distribution mechanism of the underwater screed machine provided by the embodiment of the present application is shown.

[0030] Figure 4 The schematic diagram of the structure of the lower part of the upper material pipe provided by the embodiment of the present application is shown.

[0031] Figure 5 The schematic diagram of the structure of the lower material pipe provided by the embodiment of the present application is shown.

[0032] Figure 6 The schematic diagram of the plug-in of the upper material pipe and the lower material pipe provided by the embodiment of the present application is shown.

[0033] Figure 7 The schematic diagram of the whole working process of the underwater measurement subsystem provided by the embodiment of the present application is shown.

[0034] Figure 8 The schematic diagram of the application process of the material paving control unit provided by the embodiment of the present application is shown.

[0035] Figure 9 The schematic diagram of the framework of the target base surface elevation prediction model provided by the embodiment of the present application is shown.

[0036] Figure 10 The schematic diagram of the specific process of the AE algorithm provided by the embodiment of the present application is shown. DETAILED DESCRIPTION

[0037] The present application will be further described in detail below in combination with test examples and specific embodiments. However, this should not be understood as limiting the scope of the above-mentioned subject matter of the present application to the following embodiments, and any technology realized based on the content of the present application falls within the scope of protection of the present application.

[0038] In the description of specific embodiments of the present application, the orientation or positional relationship indicated by the terms "up", "down", "left", "right", "center", "inner", "outer", "side", etc. appearing in the description of specific embodiments of the present application are based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship when the product / device / apparatus is normally used. These orientation or positional relationship terms are only for the convenience of describing the present application or simplifying the description in specific embodiments, to facilitate the quick understanding of the scheme by the technicians, and are not intended to indicate or imply that a specific device / component / element must have a specific orientation, or be constructed and operated in a specific positional relationship, and therefore cannot be understood as a limitation of the present application.

[0039] In the description of specific embodiments of the present application, the technical terms "first", "second", etc. only distinguish one entity or operation from another entity or operation, and cannot be understood as indicating or implying relative importance or implicitly indicating the number, specific order or primary and secondary relationship of the indicated technical features. In the description of specific embodiments of the present application, the meaning of "multiple" is two and more than two, unless otherwise explicitly and specifically limited.

[0040] Reference herein to "embodiments" means that the specific features, structures, or characteristics described in connection with the embodiments can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily a separate or alternative embodiment to other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0041] The intelligent control system of the underwater screed machine provided in the embodiments of the present application aims to realize intelligent management of the walking type underwater screed machine in the process of floating, attitude control and screeding operation, and to improve the operation efficiency, safety and adaptability through multi-source perception, intelligent decision-making and precise execution. Please refer to Figure 1 , Figure 1 The architecture schematic diagram of the intelligent control system of the underwater screed machine provided in the embodiments of the present application. The intelligent control system 10 of the underwater screed machine can include:

[0042] The perception subsystem 11 is configured to acquire working state data of the underwater screed machine; the decision subsystem 12 includes a stage decision unit and a fuzzy swarm intelligence optimization unit; the decision subsystem 12 is configured to discretize the working process of the underwater screed machine into a plurality of stages, and determine the framework parameters of the current stage based on the working state data in each stage; the fuzzy swarm intelligence optimization unit is configured to determine the liquid level change amount of each target compartment based on the framework parameters under the condition of meeting a plurality of preset constraints, to generate the control instruction of the current stage based on the liquid level change amount; wherein the target compartment is a compartment that needs to perform a liquid level adjustment operation in the current stage; the execution subsystem 13 is configured to receive the multi-stage control instruction sent from the decision subsystem 12, and control the underwater screed machine to perform underwater work based on the multi-stage control instruction.

[0043] In the embodiments of the present application, the underwater screed machine refers to an engineering equipment used for underwater foundation leveling operation, which can perform leveling, elevation review and leveling operation on a riprap bed on the seabed. The structure of the underwater screed machine can include a plurality of pontoons or compartments, and the buoyancy and attitude are controlled by adjusting the liquid level in the compartment. The underwater screed machine can be a walking underwater screed machine, a modular screed machine with adjustable buoyancy cylinder, a semi-submersible underwater screed machine, etc.

[0044] The perception subsystem 11 is used to acquire real-time working state data of the underwater screed machine in various working conditions in all directions and at high frequency. The subsystem is not composed of a single sensor, but a complex network integrating multiple types of sensors. All these transmission data are transmitted and integrated at high speed and reliably through an industrial standard data bus (such as Modbus / TCP) to form a unified and real-time updated working state database.

[0045] The decision subsystem 12 adopts a hierarchical decision mechanism from macro planning to micro optimization, which converts a complex continuous control problem into a series of solvable discrete optimization problems. It is mainly realized through the cooperation of the stage decision unit and the fuzzy swarm intelligence optimization unit. In the specific working process, the continuous working process of the underwater screed machine (such as diving from the water surface to the working depth and leveling the attitude) is discretized into a plurality of sequentially executed stages. The basis for dividing the stages can be time interval (such as every 30 seconds, 60 seconds, 90 seconds, 120 seconds for a stage) or depth interval (such as every 1 meter, 2 meters, 5 meters for a stage). At the beginning of each stage, the stage decision unit will determine the framework parameters of the current stage based on the latest working state data provided by the perception subsystem 11. These framework parameters include the sub-targets of the current stage (such as reaching a certain intermediate draft depth or target inclination), the state variables of the stage, and the constraint conditions of the stage.

[0046] Since the leveller can have numerous tanks, and it is inefficient and can cause instability to adjust all tanks at the same time, the fuzzy group intelligence optimization unit receives the framework parameters output by the stage decision unit, first performs "fuzzy reasoning" based on domain knowledge, and evaluates each tank. The evaluation indicators can include: the "contribution direction" of the tank level adjustment to correcting the current attitude (correction alignment), the force arm effect generated by its position, and the safety margin of the current level. Through fuzzy logic calculation, the most suitable target tank for adjustment in the current stage is selected, thereby narrowing the search space of the optimization problem and improving the calculation efficiency. After determining the target tank, the unit starts the fuzzy particle swarm optimization algorithm (FPSO). Under the premise of meeting multiple preset constraint conditions (attitude safety envelope constraint, mass conservation constraint, tank geometry limit constraint, pump valve capacity constraint, etc.), a set of optimal level change amounts is quickly found in the sparse search space formed by the target tank. This set of solutions is a fine-tuned correction of the theoretical decision of the stage decision unit under actual engineering constraints. Finally, the unit maps the optimized level change amount to specific and executable pump valve control instruction sequences, such as issuing instructions to a specific numbered pump to specify its opening duration and flow rate.

[0047] The decision subsystem 12 can adopt a rolling or forward DP strategy, and the decision space of the kth stage is The decision subsystem 12 uses FPSO to find a local optimal decision in the decision space of the kth stage Specifically, for each stage k, the decision subsystem 12 focuses on the current stage, uses the set of participating tanks obtained by fuzzy screening to reduce the dimension of the decision space, and calls the fuzzy particle swarm optimization (FPSO) algorithm to quickly solve a local optimal decision for the current stage. To make up for the global short-sightedness that may be caused by this local optimization, a "to-go" lower bound heuristic function is introduced in the decision subsystem 12 to estimate the minimum remaining cost required to reach the final target state from the state of the kth stage. The calculation formula is:

[0048]

[0049] wherein, is the lower bound of the remaining cost of the kth stage, is the target total equivalent water amount (or the target draft corresponding value), and if is worse than the known optimal upper bound, backtracking / restructuring is triggered to re-solve to ensure approximate global optimization.

[0050] The execution subsystem 13 is the final execution system of the control instructions, receives the multi-stage control instructions from the decision subsystem 12, drives the underlying hardware devices to act, including controlling the start and stop and rotating speed of the frequency conversion pump, adjusting the opening of the proportional electromagnetic valve, operating the compressed air valve group, and thus accurately realizing the water filling or draining operation of each target cabin. The whole system runs in a dynamic closed loop, after the execution subsystem 13 executes the action, the state of the underwater screed machine changes, the sensing subsystem 11 immediately captures these changes (such as the new draft depth, the attitude angle, the liquid level value), and feeds back the updated working state data to the decision subsystem 12. The decision subsystem 12 judges whether the current stage target is reached accordingly. If it is reached, it switches to the next stage; if there is a deviation, it may trigger local correction or re-optimization in the current stage. This cycle continues until the screed machine finally reaches the predetermined working float state and attitude.

[0051] In some optional embodiments, the sensing subsystem 11 can include a multi-source sensing layer and a fusion and state estimation layer, the multi-source sensing layer is responsible for collecting various physical parameters of the screed machine in real time through a series of high-precision sensing devices, including an inertial measurement unit (IMU), a combination of a four-corner measuring tower prism and a shore-based total station, a depth or draft sensor, a liquid level and pressure sensor of each cabin, a valve state sensor, a pump group power and flow sensor, and a data bus (using Modbus / TCP protocol) as a data transmission link. These devices together form a comprehensive monitoring network to provide raw input for subsequent data processing.

[0052] Specifically, the IMU is used to measure the three-dimensional attitude angle (such as the roll angle θ and the pitch angle φ) and the angular velocity change of the underwater screed machine, reflecting the spatial orientation and rotational dynamics of the device. The prism on the four-corner measuring tower works with the shore-based total station to obtain the elevation and spatial coordinates of the four corner points of the screed machine through optical measurement technology, providing a high-precision position reference for calibrating the overall attitude and displacement. The depth or draft sensor focuses on monitoring the immersion depth of the whole machine in the water, i.e. the draft value d. The liquid level and pressure sensor of each cabin feeds back the change of the water quantity and the pressure distribution in the cabin in real time, thereby indirectly reflecting the weight distribution and local buoyancy state of the screed machine. The valve state sensor is used to detect the opening and closing state and action position of the control valve. The pump group power and flow sensor monitors the running efficiency of the pump, including power consumption and medium flow, to optimize energy consumption and system response. All these sensing data are integrated and transmitted in real time through the Modbus / TCP data bus, ensuring seamless synchronization of multi-source information.

[0053] The fusion and state estimation layer is built on the basis of the multi-source perception layer, and mainly uses the data of key devices such as IMU, total station and liquid level sensor for calculation. The role of the fusion and state estimation layer is to fuse and process the original sensing information, and estimate the core state parameters of the leveling machine through the algorithm model. For example, the attitude and angular velocity data provided by the IMU combined with the spatial coordinate information of the total station can correct and refine the measured values of the attitude angles θ and φ, and reduce the error; the liquid level sensor data is used to infer the water volume change of each cabin, combined with the draft information of the depth sensor, the mass center position and the floating center position are dynamically calculated through the physical model, and the uncertainty of these estimated values is evaluated.

[0054] In some optional embodiments, the stage decision unit in the decision subsystem 12 uses a dynamic programming (DP) to discretize the working process of the underwater leveling machine into multiple stages, and determines the framework parameters of the current stage based on the working state data in each stage. In each stage, when the trigger conditions of reaching a fixed time length, an intermediate draft threshold or an inclination threshold are reached, the next stage is switched to.

[0055] Among them, the framework parameters refer to the set of optimization variables calculated by the decision subsystem 12 based on the working state data in each stage, including the target draft , the target attitude angle , the stage division strategy, and the cabin participation screening rule, which are used to guide the calculation of the fuzzy swarm intelligence optimization unit.

[0056] For example, the stage decision unit can discretize the working process of the underwater leveling machine into multiple stages based on time intervals or depth intervals ; in each stage, the stage state is represented by a vector composed of the liquid level heights of all cabins at the current time, denoted as , where represents the liquid level in the i th cabin in the k th stage. Based on the analysis of the current state , the stage decision unit needs to develop an “action plan” for this stage, i.e. the stage decision , which is also a vector, each component of which represents the liquid level change amount of the i th cabin that needs to be adjusted in this stage. A positive value usually indicates filling water into the cabin to increase weight and assist in diving, and a negative value indicates draining water from the cabin to reduce weight and promote floating.

[0057] In each decision stage k, the optimization algorithm must ensure that the results meet a series of hard constraints when searching for the optimal liquid level adjustment scheme. These constraints are the bottom line to ensure the safety of the leveling machine structure and the feasibility of operation, and any candidate scheme must meet these conditions before pursuing the optimal performance indicator. The preset constraint conditions are checked immediately within the stage, including:

[0058] 1) Level limits: This constraint requires that the level cannot be too low to empty the tank, nor too high to overflow the tank.

[0059] 2) Attitude limits: This constraint requires that the new attitude angle generated by the underwater screed after performing the level adjustment of this phase plan must be controlled within the safety envelope.

[0060] Water intake / discharge strategy:

[0061] 1) Internal redistribution: This constraint requires that the sum of the water volumes corresponding to the level changes of all tanks must be zero, i.e. water is only internally transferred between tanks, not exchanged with the environment.

[0062] 2) Net water intake (for diving): (Given phase target volume). This constraint requires that the total water volume (positive value) or discharge volume (negative value) from the environment for this phase plan, in order to change the total weight of the screed, to achieve diving or surfacing.

[0063] To evaluate in real-time the impact of different level adjustment schemes on the underwater screed attitude θ and φ, the present application considers the underwater screed as a rigid body whose attitude is determined by the restoring moments generated by the weights of the water in each tank, through an attitude approximation model.

[0064] Under small angle approximation:

[0065]

[0066]

[0067] wherein, is the horizontal coordinate of the tank #i relative to the center of mass, is the restoring moment stiffness identified from the bench / field. The phase cost computation function quantifies the overall merit of a certain decision

[0068] (i.e. level adjustment scheme) based on a scalar numerical value. When computing the phase cost (time minimization + engineering penalty), let

[0069] ; pumps are in parallel:

[0070]

[0071] ​​

[0072] wherein the time cost is decomposed into pumping time and valve action time two parts, is the upper limit of the equivalent flow rate available for the tank, is the one-time opening and closing delay of the tank, is the number of non-zero operation tanks in this stage.

[0073] In addition to the main cost of time, the cost function provided in the embodiments of the present application also introduces an engineering penalty term , that is, the second-order difference penalty of the liquid level change amount. The stage cost is defined as:

[0074]

[0075] wherein, is the penalty coefficient, used to suppress the drastic repetition of adjacent stages. When the liquid level adjustment amount of the current stage is very different from the last stage, the penalty value will increase, thereby guiding the optimization algorithm to tend to select a smoother solution that connects with the last stage instruction.

[0076] In order to judge the pros and cons of a decision , the "time type cost" function is introduced in the stage decision unit, which calculates the total time expected to be spent in executing the current decision . This total time is a comprehensive index, which not only includes the main pumping time, but also carefully includes the auxiliary operation time related thereto. Specifically, it mainly includes the following parts: first, the time required for inter-tank transfer and internal distribution, which is closely related to the absolute value of the tank liquid level adjustment amount, the rated flow of the pump, the number of pumps that can work at the same time, and the system efficiency; second, the cumulative delay of the opening and closing action of all valves involved in the operation, because the opening or closing of each valve requires a short response time; in addition, if the control strategy allows water exchange with the external environment (i.e. internal water transfer of a non-closed system), the time of draining water to the outside of the tank or taking water from the outside of the tank also needs to be counted. Using time as the core cost index directly reflects the core engineering demand of improving work efficiency and shortening the overall cycle, so that the optimization process is always time-oriented.

[0077] For example, the stage decision unit can take the minimum total lifting time as the main target, with the secondary items of energy consumption or action times. The main target draws on the "three time" decomposition in ballast optimization:

[0078]

[0079] wherein, is the inter-tank transfer and internal distribution time (related to the absolute value of the tank liquid level adjustment amount, the rated flow of the pump, the number of pumps that can work at the same time, and the system efficiency) , pump flow , number of pumps , efficiency related); : cumulative valve opening / closing time (related to valve opening / closing time delay and "whether to act" indication related); : external discharge / withdrawal time (if water exchange with the environment is allowed; if it is a pure internal redistribution, this item is zero). Optional penalties can be added to energy consumption / action number / cabin switching penalty to suppress frequent start-stop and liquid level oscillation (convenient for pump valve life and stability), wherein is a non-negative weight coefficient, which is set by the control system designer according to the specific hardware characteristics (such as pump valve tolerance) and operation requirements, and determines the importance of this penalty in the total cost function; the core of the penalty term is the sum of the squares of the control vector u, in the context of the leveling machine control, the vector u represents the liquid level change amount instruction of all cabins in a stage. Therefore, is essentially the sum of the squares of all control instruction amplitudes in this stage. The deep reason for introducing this penalty is that too large liquid level adjustment amount means that the pump needs to work in high intensity and large flow in a short time, which not only significantly increases the instantaneous energy consumption of the system, but also causes greater mechanical stress to the pump, valve and related pipeline, and accelerates the equipment aging. is also a non-negative weight coefficient, used to adjust the intensity of this penalty; the core of the penalty term is the sum of the squares of the liquid level change acceleration, representing the change speed of the liquid level change itself, i.e. the degree of jitter or mutation of the control instruction. For example, a decision causes the drainage rate of a cabin to increase or decrease sharply in a short time, resulting in a larger value. The introduction of this penalty can suppress the frequent oscillation and violent fluctuation of the control signal. In the actual control system, too frequent start-stop signal or sharp change of flow instruction is easy to cause oscillation of the actuator, leading to the decline of system stability. By punishing high-frequency mutations, the optimization algorithm will tend to generate smooth and continuous control sequences, which is beneficial to the stable operation of the pump valve and prolongs its service life.

[0080] In some optional embodiments, before generating the control instruction of each stage, the fuzzy swarm optimization unit determines the participation of each cabin based on fuzzy reasoning, and the implementation process of the target cabin can include:

[0081] For each tank, calculate the antecedent variables of fuzzy inference; wherein, the antecedent variables include correction alignment degree, force arm effectiveness and liquid level margin, the correction alignment degree is calculated based on the consistency of the attitude gradient and the target correction direction, the force arm effectiveness is calculated based on the force arm weight of the tank to the roll or pitch, and the liquid level margin is calculated based on the safety margin of the current liquid level and the limit value; 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 tank; and it is determined whether the tank is the target tank based on the participation level.

[0082] Correction alignment degree Reflecting the contribution direction of the tank to the attitude correction, the consistency of the attitude gradient and the "desired correction to 0" can be used to describe:

[0083]

[0084]

[0085]

[0086]

[0087] wherein, is the theoretical correction effectiveness of the i-th tank for correcting the current attitude deviation, is the liquid level of the i-th tank, is the density of water, is the acceleration of gravity, is the horizontal cross-sectional area of the i-th tank, and is the horizontal coordinate of the center of gravity of the i-th tank in the screed body coordinate system, and are the restoring moment stiffness coefficients of the underwater screed in the roll (around the x-axis) and pitch (around the y-axis) directions.

[0088] Force arm effectiveness represents the force arm weight of the tank to the roll / pitch:

[0089]

[0090] (interquartile is 30% and 70%)

[0091] Liquid level margin represents the minimum safety margin from the upper and lower limits:

[0092]

[0093] A predefined membership function converts the precise numerical input (i.e. antecedent variable) into a "fuzzy language" concept that the computer can process. The predefined membership function can be a triangular or trapezoidal function. For example, the predefined membership functions provided in the present application include:

[0094] : Reverse / General / Co-directional (center 0 / 0.5 / 1).

[0095] : Small / Medium / Large (a, b, c are set by sample quantiles).

[0096] : Low / Medium / High (typical thresholds 0.15 / 0.4 / 1.0).

[0097] For the correction alignment degree , the present application defines three fuzzy sets: "reverse", "general", and "co-directional". For the force arm effectiveness , the fuzzy sets are "small", "medium", and "large". For the liquid level margin , the fuzzy sets are "low", "medium", and "high". After completing the fuzzification, the fuzzy rule base constructed based on domain knowledge is used for calculation. For example, the calculation of each cabin , the member function is mapped to the "participation level" (not involved / medium / important) to obtain or weight , to determine , which is solved.

[0098] For example, the fuzzy rule base provided in the present application (12 examples, output is the participation level ) includes:

[0099] R1 A=co-directional & E=large & M=high → ;

[0100] R2 A=co-directional & E=medium & M=medium → ;

[0101] R3 A=co-directional & M=low → ;

[0102] R4 A=general & E=large & M=high → ;

[0103] R5 A=general & E=small → ;

[0104] R6 A=reverse → ;

[0105] R7 upper limit & A=co-directional → ;

[0106] R8 fast dive phase & M=high → ;

[0107] R9 same tank has consecutive focus ≥2 phase → ;

[0108] R10 same side symmetric tanks both focus → keep the one with larger retention arm;

[0109] R11 number of participating tanks → take the one with the largest focus value ;

[0110] R12 valve group interlock conflict → downshift.

[0111] For example, R1 states that if a tank has the correct correction direction, large retention arm effectiveness, and high liquid level safety margin, it should be the focus of attention ( ). R6, as a safety bottom line, stipulates that any tank that may exacerbate the attitude deterioration ( reverse direction) must be excluded ( ).

[0112] The result of fuzzy reasoning is a fuzzy set, which needs to be de-fuzzified into an exact numerical value for control. In the embodiments of the present application, the Mamdani (min-max) aggregation combined with the centroid method is used to output . For each tank, all activated rules (i.e., rules with a certain degree of satisfaction of the antecedent conditions) will generate an output fuzzy set. These output fuzzy sets are aggregated by min and max operations to finally obtain an integrated output fuzzy set representing the total participation level of the tank.

[0113] is a number between 0 and 1, which quantitatively represents the degree to which the tank should be valued in the current phase. In order to provide a clear, non-binary search space for the subsequent optimization algorithm, it is necessary to convert the continuous into a binary decision. In the present application, the threshold method or the Top-K method can be used for conversion.

[0114] Let the single-tank step size of the phase (i.e., the maximum liquid level change allowed for a tank in a phase), then according to the participation level of each target tank , the liquid level change boundary of the current phase is dynamically calculated for it:

[0115]

[0116]

[0117] Get the set of participants . Where the set is composed of all cabin numbers i that satisfy the condition , i.e. the set is a list of all target cabins for the current stage.

[0118] In some optional embodiments, after determining all target cabins, the fuzzy swarm intelligence optimization unit employs a fuzzy particle swarm optimization algorithm, and the constraint conditions of hyperplane projection and boundary repair processing, to search for optimal decision variables in the sparsified search space composed of the frame parameters of the current stage; the optimal decision variables are the liquid level change amounts of each target cabin.

[0119] For example, the fuzzy swarm intelligence optimization unit, after receiving the frame parameters provided by the stage decision maker, is responsible for calculating the optimal liquid level change amount of each target cabin in the current stage under the premise of satisfying multiple engineering constraints, and finally generating control instructions that can be issued for execution, i.e. finding the liquid level change amount of each target cabin on the set by the fuzzy particle swarm optimization algorithm, and generating the control instructions of the current stage based on the liquid level change amount.

[0120] In the fuzzy particle swarm optimization algorithm, the variable , the equality constraint (internal redistribution or net water withdrawal) is:

[0121]

[0122] In the net water withdrawal mode that allows water exchange with the environment, the sum must be equal to the target volume change amount of the current stage, the inequality constraint is:

[0123]

[0124] That is, the liquid level change amount of each cabin must be limited within its dynamically calculated feasible interval.

[0125] The optimization objective is defined by a comprehensive objective function , aiming to minimize the comprehensive cost of the current stage, and the objective function is:

[0126]

[0127] Where, ​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.

[0128] 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:

[0129]

[0130] 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.

[0131] The particle swarm optimization process includes:

[0132] 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.

[0133] In some optional embodiments, the specific process of the decision subsystem 12 generating control instructions can include: according to the positive and negative and size of the liquid level change amount of each target compartment, combining the target hardware parameters of the underwater screed, calculating the opening duration and action timing of the pump or valve corresponding to each target compartment, and generating corresponding control instructions. In the case where the execution subsystem 13 executes the control instructions, the working state data is monitored, and if it is detected that the working state data represents that the attitude or draft of the underwater screed deviates more than a preset threshold, a local correction operation is triggered to adjust the liquid level change amount and reissue the control instructions.

[0134] For example, the end-to-end solving process is: 1, read in ; 2, calculate 3, determine , construct boundary ; 4, run FPSO: initialization→(update→projection→repair→evaluation)×iteration→get ; 5, map to pump / valve timing and execute; 6, monitor and closed-loop correction; reach stage threshold→enter ; 7, if Worse than the current best→backtracking / reorganization Recalculate; 8, reach →end.

[0135] Wherein, after being equipped with a projection repair mechanism, the FPSO algorithm searches in the reduced dimension space (only for participating compartments) according to its standard steps. The particle position is initialized as a uniform random distribution, and is subjected to initial projection repair to ensure feasibility. In each iteration, the particle updates the speed and position according to its own historical optimal position and the group historical optimal position. After updating, projection repair is immediately performed, and the fitness value (i.e. the value of the objective function J(x)) of the new position is calculated. In order to improve engineering practicability, multiple field knowledge is introduced in the algorithm, including 1, "smoother" processing, setting a minimum threshold ε for the liquid level change amount, ignoring adjustments below the threshold, to avoid frequent small actions (jitter) of the pump valve; 2, adopting annealing strategy, gradually reducing the inertia weight w from 0.9 to 0.5 with the number of iterations, to balance global exploration and local development; 3, setting an upper limit for the number of iterations or a convergence criterion for fitness change; 4, containing a failure recovery mechanism, if a feasible solution cannot be obtained for multiple times, feedback to the fuzzy screening stage, and expand the participating set or lower the screening threshold To relax the search space.

[0136] Overall, the stage decision unit in the embodiments of the present application decomposes the high-dimensional whole-process problem into "intra-stage optimization + inter-stage connection" through staging, and guarantees approximate global optimization with to-go lower bound and backtracking; in fuzzy screening, the To the former, write the domain knowledge as a computable rule base, yielding For dimension compression and boundary scaling. The fuzzy swarm intelligence optimization unit combines time type cost and attitude, liquid level or volume constraints on the selected cabin set; "hyperplane projection + boundary repair" is used to ensure the feasibility and convergence speed of particles; finally it is mapped to the pump / valve beat to execute. The lifting and attitude can be optimized simultaneously, and the pump / valve time delay and parallel constraints are explicitly taken into account, which can significantly shorten the total time of lifting / descent, reduce the number of start / stop and comprehensive energy consumption. The DP algorithm and fuzzy FPSO sparse solution are used in stages, combined with equation projection and boundary repair, to ensure that the roll and pitch are controlled within the safety envelope throughout the process, with small attitude fluctuations and strong anti-disturbance ability. And without modifying the existing hardware, it can be seamlessly integrated with IMU, liquid level, total station and digital twin system, supporting online rolling optimization, adapting to multi-cabin and multi-water depth conditions, and having high engineering landing performance.

[0137] In some optional embodiments, the intelligent control system 10 provided by the embodiments of the application can also be used for real-time sensing and virtual mapping of the working state of the underwater screed, and the intelligent control system 10 can be seamlessly integrated with a digital twin platform to realize the precision, visualization and intelligence of the screeding operation.

[0138] The decision subsystem 12 further 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 according to the working state data, and to construct a digital twin model of the screed operation; the job suggestion unit is configured to generate a job suggestion according to the working state data and the job task of the screed; and the job control unit is configured to control the execution subsystem 13 to execute the job task according to the job suggestion.

[0139] In the case of executing virtual mapping, the sensing subsystem 11 of the intelligent control system 10 further includes an edge gateway, a multi-beam sonar, a DVL speed meter, etc.; and the execution subsystem 13 further includes a pusher and a material distribution device. Please refer to Figure 2 , Figure 2 The schematic diagram of the underwater screed main frame provided by the application is shown in

[0140] Please refer to Figure 2 on the basis of Figure 3 , Figure 3The schematic view of the distributing mechanism of the underwater screed provided by the embodiment of the present application. The distributing mechanism 7 comprises an upper pipe 72 and a lower pipe 73: the upper pipe 72 is inserted above the lower pipe 73, and the upper pipe 72 and the lower pipe 73 have a first channel therebetween. The lower pipe 73 comprises a bottom pipe structure 731 and a first funnel structure 732 connected to the top of the bottom pipe structure 731, and the large opening end of the first funnel structure 732 is arranged towards the upper pipe 72.

[0141] Please refer to Figure 3 the basis of Figure 4 , Figure 4 The structural schematic view of the lower part of the upper pipe provided by the embodiment of the present application.

[0142] The upper pipe 72 comprises an upper pipe structure 721, and a plurality of protrusions 722 are arranged on the outer wall of the upper pipe structure 721 in a circumferential direction, and a first gap 723 is arranged between adjacent protrusions 722. The outer side surface of the protrusion 722 is an inclined surface 724 arranged corresponding to the first funnel structure 732. The lower part of the upper pipe structure 721 is inserted into the bottom pipe structure 731, and the second gap 725 which is communicated with the first gap 723 is arranged between the outer wall of the upper pipe structure 721 and the inner wall of the bottom pipe structure 731.

[0143] Please refer to Figure 5 , Figure 5 The structural schematic view of the lower pipe provided by the embodiment of the present application. The bottom pipe structure 731 and the first funnel structure 732 are welded, and the first connecting rib plate 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 lugs 734 are connected to the top outer wall of the first funnel structure 732, and all the lower lugs 734 are arranged in a circumferential direction of the first funnel structure 732. All the lower lugs 734 are close to the large opening end of the first funnel structure 732.

[0144] Please refer to Figure 6 , Figure 6 The insertion schematic view of the upper pipe and the lower pipe provided by the embodiment of the present application. The upper pipe 72 further comprises a second funnel structure 726 which is sleeved on the outer side of the upper pipe structure 721, and the second funnel structure 726 is arranged towards the first funnel structure 732 and can cover the large opening end of the first funnel structure 732. The first funnel structure 732 and the second funnel structure 726 have the third gap 741 which is communicated with the first gap 723 therebetween, and the third gap 741, the first gap 723 and the second gap 725 form the first channel.

[0145] The outer side of the part of the upper pipe structure 721 which is located below the first funnel structure 732 is provided with the first block 720, and at least one side of the first block 720 can laterally abut against the inner wall of the bottom pipe structure 731. To increase the connection stability between the upper pipe structure 721 and the lower pipe 73.

[0146] The upper pipe structure 721 further comprises at least two upper pipe sections 727 which are detachably connected in sequence along the length direction of the upper pipe structure 721, and the second funnel structure 726 is located on the lowermost upper pipe section 727.

[0147] The upper pipe structure 721 further comprises a third funnel structure 728 connected to the top of the upper pipe structure 721, and the large end of the third funnel structure 728 is arranged upward.

[0148] Because the underwater sensor is easy to be affected by water flow disturbance, sediment shielding, sound wave scattering and the like when collecting data, the perception data has instability and inconsistency. The traditional filtering algorithm such as EKF is easy to cause fusion deviation when facing low signal-to-noise ratio and large uncertainty multi-source data. Therefore, the intelligent control system 10 in the embodiment of the present application can introduce the edge gateway of the fusion algorithm of the terrain confidence weight mechanism, and dynamically weight the data of different sources and different qualities, so as to improve the data robustness and real-time credibility of the twin model.

[0149] The edge gateway uses the fusion algorithm of the terrain confidence weight mechanism to dynamically weight the original data collected by different sensing devices in the perception subsystem 11. The working state data obtained by the perception subsystem 11 at least includes: the attitude of the grader, the speed of the grader, the terrain elevation difference of the grading area, the hydrodynamic force and the like.

[0150] 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 ), multi-beam sonar (z 2 ), depth gauge (z 3 ), DVL speedometer (z 4 ) and the like. The observation model of each type of sensor is defined as:

[0151]

[0152] Wherein, x t is the true state of the system; h i (x t ) is the observation function of the i th sensor; is the i th observation noise, which satisfies the zero-mean Gaussian distribution.

[0153] Based on historical multi-period point cloud data, a terrain confidence map of the current area is constructed , which is used to measure the terrain reliability of each coordinate point, and is defined as follows:

[0154]

[0155] Wherein, is the sliding window length, and in Z (t) is the original observed terrain elevation value at point at time t, modeled by point cloud, unit: m; Z (t) is the original observed terrain elevation value at point Z (t) is the original observed terrain elevation value at point Z (t) is the original observed terrain elevation value at point Z (t) is the original observed terrain elevation value at point , the closer to 1 indicates that the point is historically stable and the data is reliable.

[0156] For each type of observation data, its observation weight is dynamically adjusted according to the confidence of its corresponding region :

[0157]

[0158] wherein, Z (t) is the original observed terrain elevation value at point is the coverage area of the th sensor; after weight normalization, it satisfies .

[0159] The improved weighted EKF prediction-update process is used for state estimation:

[0160] 1. Prediction stage:

[0161]

[0162]

[0163] 2. Weighted update stage: calculate the gain for each observation:

[0164]

[0165] After integrating all observations, the state update is:

[0166]

[0167] Covariance update:

[0168]

[0169] In some optional embodiments, the work task at least includes a laying task. After the work suggestion unit obtains the laying task and the original data, it calculates the region laying feasibility by using a laying score function; recommends the laying thickness and density values by using fuzzy rules; predicts the potential construction defect region and high-risk advancing path by using a dynamic window algorithm and a dynamic cost weighting mechanism; and generates a laying work suggestion including the laying amount and the advancing direction.

[0170] For example, the job suggestion unit calculates the feasibility of the region material in the following manner using the material score function: obtaining the height difference between the target flat surface of the digital twin model and the current terrain elevation; obtaining the difficulty score of the soil quality of the target flat surface region; and performing weighted linear combination on the height difference and the difficulty score to obtain the material score.

[0171] In some optional embodiments, the target flat surface generated by the digital twin modeling is compared with the current terrain elevation , and the residual at the point is defined as: , which represents the height difference of the current point from the ideal flat state; the soil quality difficulty score of the region is defined by the sonar echo intensity, construction experience, or manual annotation:

[0172]

[0173] The closer to 1 indicates that the soil quality is more complex (such as soft silt or sand layer), and the material density needs to be increased to ensure the deposition quality. By combining the above two environmental factors, the material feasibility score function is defined as a weighted linear combination:

[0174]

[0175] wherein are weight coefficients set manually or learned adaptively, and satisfy .

[0176] In some optional embodiments, the job suggestion unit can recommend the material thickness and density values in the following manner using fuzzy rules: mapping the material score to a fuzzy input variable; setting the recommended material thickness and the recommended material density as fuzzy output variables; using Mamdani-type reasoning to perform multi-rule synthesis on a plurality of preset fuzzy rules, and using the centroid method to defuzzify. Optionally, the job suggestion unit 220 performs fuzzy rule recommendation processing on all regions of the target flat surface to obtain a recommended thickness heat map and a recommended density heat map corresponding to the target flat surface.

[0177] Mapping to a fuzzy input variable , three membership functions are defined: low: ; medium: ; and high: . Triangular or trapezoidal membership functions may be used.

[0178] The fuzzy output variable is defined as: the recommended material thickness , with a range such as [2 cm, 10 cm]; and the recommended material density , range as [200kg / m³, 600kg / m³]. Also divided into: low, medium, high three levels; output is a fuzzy set.

[0179] Fuzzy rule examples (IF-THEN statements) are shown in Table 1.

[0180] Table 1

[0181]

[0182] Fuzzy inference: multi-rule synthesis using Mamdani type inference; defuzzification: use the centroid method to get the final recommended value:

[0183]

[0184]

[0185] The final output recommended cloth value is: recommended thickness heat map: ; recommended density heat map: .

[0186] This heat map is displayed by the artificial reference system, which is used to assist manual control of cloth amount. After the work is completed, the actual construction results can be fed back to fine-tune the weights or update the fuzzy rule base.

[0187] In some optional embodiments, the work recommendation unit predicts the potential construction defect area and risk advancing path in the following way using a dynamic window algorithm and a dynamic cost weighting mechanism: a "terrain risk heat map" constructed by introducing a digital twin model is used to set the weights of the influencing factors of the prediction results. Optionally, the influencing factors of the prediction results include: the included angle between the target direction and the current path direction, the minimum distance between the path and the nearest obstacle, the size of the current path corresponding to the advancing speed, and the dynamic risk heat layer derived from the twin modeling.

[0188] To improve the robustness of path planning in underwater sudden environments, a "terrain risk heat map" constructed by introducing a digital twin model can be used to introduce a dynamic cost weighting mechanism in the scoring function of the dynamic window algorithm (DWA).

[0189] The terrain risk heat map is generated by the seabed elevation grid reconstructed by the digital twin in real time, historical stability (confidence), target flat surface, flow velocity and direction, obstacles / entry prohibited areas, and sonar echoes, etc. The decision subsystem 12 first refines the source data of the seabed elevation grid, historical stability (confidence), target flat surface, flow velocity and direction, obstacles / entry prohibited areas, and sonar echoes, etc. into a plurality of risk factors (for example: height difference exceeding limit, slope / undulation, historical fluctuation or insufficient data reliability, soil difficulty, distance from obstacle, water power transport risk, etc.) standardized to 0-1, then performs weighted fusion according to the weights obtained by engineering preset or online learning to obtain the comprehensive risk value of each grid. Subsequently, spatial smoothing and abnormality suppression (anisotropic filtering along the flow direction under strong flow conditions) are performed, and a conservative degradation strategy is enabled when data is lost or abnormal. The final output is a heat map with a value of 0-1 (typical update frequency of about 1 Hz, resolution of 0.2-0.5 m), and the higher the value, the greater the risk of terrain operation, which is used for risk constraint and highlight prompt of path planning and material distribution strategy.

[0190] The DWA score function is modified as:

[0191]

[0192] Wherein, heading is a measure of 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; The size of the current path corresponding to the propulsion speed. The dynamic risk heat layer derived from the twin modeling reflects the local terrain mutation, material accumulation or non-compliance area of construction. 、 、 、 The weights of the above-mentioned influence factors. This mechanism can effectively guide the path to move away from high-risk areas and improve the stability of propulsion. The dynamic risk heat layer is a time-series risk layer formed by introducing time dimension smoothing, short-term foresight and stability constraint on the basis of the "terrain risk heat map", which is used to reflect the operation risk trend in the future a short period. Its output is a grid value between 0 and 1, and the larger the value, the higher the risk.

[0193] In some optional embodiments, the operation suggestion unit can transmit the suggestion result in the form of a heat map, a path line, etc. to the operation control unit for display. Optionally, the operation control unit can be a shore-based console or a shipborne terminal.

[0194] Optionally, the operator refers to the suggestion result to adjust the propulsion direction, adjust the material amount, change the path, or request re-planning.

[0195] Optionally, the job control unit does not automatically drive the execution subsystem 13 to run, but adopts "suggested cooperation" to rely on the manual control of the operator to control the execution subsystem 13 to run.

[0196] Optionally, the decision subsystem 12 further comprises a feedback optimization module. The feedback optimization module is used to optimize the algorithms of the digital twin unit and the job suggestion unit after the completion of the material job. The feedback optimization module collects the error between the actual material covering condition, the advancing track and the prediction of the digital twin model through a sensing device, and performs adaptive update on: the scoring function parameter; the fuzzy system membership degree; and the terrain modeling algorithm, to continuously optimize the control strategy, improve the prediction accuracy and the recommendation rationality, and form a long-term evolving virtual-real closed-loop control chain.

[0197] In some optional embodiments, the intelligent control system 10 of the underwater screed machine can also be used to provide fine measurement for underwater screeding work. When the intelligent control system 10 of the underwater screed machine is applied to provide fine measurement for underwater screeding work, the intelligent control system 10 of the underwater screed machine further comprises an underwater measurement subsystem, and the underwater measurement subsystem comprises:

[0198] A binocular camera device configured to obtain binocular vision images of the target water area; and an edge computing device configured to obtain point clouds of the target water area.

[0199] Optionally, the underwater measurement subsystem is installed on the underwater screed machine, the binocular camera device is arranged at the front part of the main frame of the underwater screed machine, and the edge computing device is in communication connection with the material device in the underwater screed machine.

[0200] Specifically, the binocular camera device is adapted to the underwater strong flow and high turbidity environment through a damping and sealing structure, and comprises a binocular stereo camera, an auxiliary structured light module, an IMU inertial navigation module and an embedded collector, is installed on the front beam of the screed machine, and has waterproof and pressure-resistant performance. In order to meet the requirements of underwater work, further, the binocular stereo camera selects a low-distortion underwater wide-angle lens and has a water pressure shell. It can be understood that the underwater screed machine moves underwater according to the predetermined direction of the construction work, and the binocular camera device mainly obtains binocular vision images of the target water area in the advancing direction of the underwater screed machine.

[0201] The edge computing device can be a central processing unit (CPU), a microcontroller unit (MCU), a hardware chip, or the like. The hardware chip can be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The PLD can be a complex programmable logic device (CPLD), a field-programmable gate array (FPGA), a generic array logic (GAL), or any combination thereof.

[0202] The edge computing device is configured to calibrate and correct the binocular vision image to obtain a parallax corresponding to each pixel point, extract original gray information of the binocular vision image, and then enhance the binocular vision image to generate gradient information, calculate a matching cost according to the parallax of each pixel point in the binocular vision image and the corresponding original gray information and gradient information, and construct a cost function, solve the cost function by using an improved SGM (Semi-Global Block Matching) algorithm to calculate a parallax map of the binocular vision image, and perform back projection on the parallax map according to camera parameters of the binocular camera device to obtain a point cloud of the target water area.

[0203] The improved SGM algorithm includes:

[0204] The improved SGM algorithm includes:

[0205] For example, in the case of underwater fine measurement, the working execution steps of the underwater measurement subsystem include:

[0206] The binocular vision image of the target water area is obtained by using the binocular camera device. The binocular vision image is calibrated and corrected to obtain a parallax corresponding to each pixel point. The original gray information of the binocular vision image is extracted, and then the binocular vision image is enhanced to generate gradient information. A matching cost is calculated according to the parallax of each pixel point in the binocular vision image and the corresponding original gray information and gradient information, and a cost function is constructed. The cost function is solved by using an improved SGM algorithm to calculate a parallax map of the binocular vision image. The parallax map is back projected according to the camera parameters of the binocular camera device to generate a point cloud of the target water area.

[0207] Wherein, the calibration and correction can adopt Zhang Zhengyou calibration method to obtain internal and external parameters, perform stereo correction and binocular alignment, and obtain the disparity corresponding to each pixel point. The optimization before processing is based on the influence of the underwater flattening machine on the underwater environment during the construction process. The binocular vision image has different degrees of defects, such as low contrast, unobvious texture, and large noise. Therefore, the binocular vision image obtained needs to be enhanced, which specifically includes:

[0208] CLAHE technology is adopted to improve the local contrast of the binocular vision image, bilateral filtering is used to suppress high-frequency noise while maintaining the edge, and then the binocular vision image is subjected to Sobel gradient enhancement to obtain gradient information.

[0209] For example, the underwater measurement subsystem can match multiple channels, calculate the matching cost by using a composite cost function, and the formula is:

[0210]

[0211] Wherein, is the constructed cost function, , , is the cost function weight, is the position of a pixel point in the binocular vision image, is the disparity value, is the original gray information, is the gradient information, is the Census cost function, is the Gradient cost function, is the SAD cost function, which is calculated in the original gray channel and the gradient channel respectively. It should be noted that for different underwater environments, the cost function weights can be set as needed.

[0212] SGM aims to aggregate the cost function in multiple directions on the image, balancing the calculation complexity and accuracy between local disparity calculation and global smooth constraint. Its core is to construct a cost volume Then, the cost is aggregated along multiple paths and the disparity value with the minimum cost is selected:

[0213]

[0214] The final total aggregation cost is: .

[0215] Wherein, is the position of a pixel point in the image, two-dimensional coordinate ; is the disparity value, that is, the horizontal offset between the matching points of the left image and the right image; for other candidate disparity values at the current pixel to compare the cost difference; is the accumulated path cost along the direction is the accumulated path cost along the direction is the local matching cost; is the penalty for small disparity variation (encouraging smoothness); is the penalty for large disparity jump (preventing jump); is the normalization term introduced to prevent cost explosion.

[0216] In some optional embodiments, the improved SGM algorithm comprises:

[0217] First, the aggregation path function is based on four directions. The traditional SGM algorithm has an aggregation path of 8 directions. The present application optimizes the selection of 4 directions (left, right, up, down, left-up-right-down, right-up-left-down) to balance accuracy and real-time performance; and increases the weight of the path in the large residual area to improve the accuracy of the disparity in the scattering area. Specifically, the aggregation path set only includes left, right, up, down, left-up-right-down, and right-up-left-down four directions, and for the pixel points located in the area with large residual (residual greater than threshold , the path cost along the direction is multiplied by the gain coefficient to enhance the matching credibility in the scattering area.

[0218] Second, a dynamic adjustment strategy is set to control the penalty for small disparity variation and the penalty for large disparity jump. Specifically, it is expressed as:

[0219]

[0220]

[0221] wherein is the image gradient module, indicating the edge strength; is a small constant to avoid division by zero; this strategy can reduce smoothing in the edge area and preserve high-frequency information.

[0222] Finally, the aggregated disparity map is subjected to confidence screening and hole filling. Specifically, the confidence of each pixel point in the aggregated disparity map is calculated, and the pixel points below the confidence threshold are screened out; then, the hole area in the screened disparity map is filled in front and back using the left view in the binocular vision image and the edge-guided interpolation method, including finding the nearest legal disparity value within the preset pixel range for the screened pixel points (invalid disparity points) , such as finding the nearest legal disparity value within 5 pixels to the left and right of the screened pixel points. In addition, the image edge information is used to constrain the interpolation direction to preserve the boundary.

[0223] Based on the above measurement method, the final improved SGM disparity estimation process is as follows:

[0224] 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.

[0225] 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 ).

[0226] 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.

[0227] 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:

[0228] ,

[0229] Set threshold (e.g., 1 pixel), then the consistency score is:

[0230]

[0231] definition Centered on The neighborhood window, for the mean disparity within the neighborhood with standard deviation Calculate the local consistency score:

[0232]

[0233] Will After being back-projected to the three-dimensional point cloud, the system extrinsic parameters are used to re-project back to the left view to obtain , and the pixel residual error is calculated:

[0234]

[0235] The back-projection score is constructed as:

[0236]

[0237] The final confidence is:

[0238]

[0239] wherein, , , are the weight coefficients of the fusion of the above scores, which can be set according to the screening requirements in the actual implementation process. The system extrinsic parameters and the camera extrinsic parameters are similar, and both are the pose extrinsic parameters or transformation matrixes describing the coordinate system of the camera device to the world coordinate system.

[0240] Based on the above technical solutions, the point cloud of the target water area can be obtained. The present application breaks through the limitation of the traditional underwater measurement method relying on a single sonar or manual operation, calculates the disparity map through the improved SGM algorithm, generates the point cloud, and significantly improves the ability to obtain high-precision three-dimensional terrain in a complex underwater environment. At the same time, the present application has a self-adaptive disparity confidence evaluation and residual back-projection mechanism, which effectively suppresses the influence of false matching and image noise on the measurement accuracy, and guarantees the stability and reliability of the measurement results.

[0241] In actual application process, the binocular camera device collects binocular video stream of the target water area. According to the motion speed of the underwater screed, the binocular video stream can be split and extracted by reasonable frame number to obtain the binocular vision image that can be processed.

[0242] Further, please refer to Figure 7 , Figure 7 for the full working process schematic diagram of the underwater measurement sub-system provided by the present application. The edge computing device is also used to obtain the design surface elevation of the target water area, and convert the point cloud of the target water area into the terrain elevation. Then, based on the paving position of the target water area, the difference between the corresponding design surface elevation and terrain elevation is calculated to obtain the paving thickness value. According to the paving thickness, the paving amount is confirmed, and the paving device is controlled to discharge according to the paving amount to the paving position.

[0243] Among them, the edge computing device calculates the difference between the corresponding design surface elevation and terrain elevation to obtain the paving thickness value, which specifically includes:

[0244] According to the paving position, the elevation difference between the corresponding design surface elevation and terrain elevation and the local fluctuation standard deviation are calculated.

[0245] The cloth thickness value is obtained by fusing the elevation difference and the local fluctuation standard deviation based on the fuzzy rule table.

[0246] The point cloud is converted into a terrain elevation grid The design surface elevation The difference map is constructed by comparison:

[0247]

[0248] The cloth area is constructed :

[0249]

[0250] For each cloth position , according to its local terrain fluctuation, the cloth recommendation level (light, normal, and enhanced) and the cloth thickness value are constructed by using the fuzzy rule table .

[0251] Rule example:

[0252]

[0253] wherein, is the local fluctuation standard deviation, is the redundancy control coefficient.

[0254] Further, the above measurement system further includes a user interface, and the edge computing device is further configured to output the cloth position and the corresponding cloth thickness value to the user interface in the form of a two-dimensional heat map or a table, so that the user can confirm the cloth amount. In one possible implementation, the edge computing device can also output the point cloud map and the difference map of the target water area to the user interface, so that the user can visually view the measurement results of the target water area through the user interface.

[0255] The underwater measurement subsystem closes the loop with the actual cloth distribution decision of the underwater grading machine, can provide fine measurement results for underwater grading operation, and can provide reliable basis for subsequent operation construction.

[0256] In some optional embodiments, the intelligent control system 10 can also be used for automatically adapting to different water depths, flow rates, or material diameters for paving construction. In this application scenario, the perception subsystem 11 further includes a weighing sensor and an inclinometer, and the decision subsystem 12 further includes a paving control unit.

[0257] The weighing sensor is installed on the hopper support frame to measure the weight of the stone in the hopper in real time. The weighing sensor is also equipped with an inclination meter (IMU) to monitor the inclination angle of the hopper. The weighing sensor combined with the inclination data can achieve accurate weighing compensation in the inclined state, ensuring accurate calculation of the stone mass when the device is inclined. The volume flow meter is integrated on the conveying 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 multi-modal environmental perception sensors: on the one hand, a laser range finder (for accurate ranging in shallow water or out of water) and a sonar thickness measuring device are installed to scan the terrain elevation of the distribution area in real time. In addition, the underwater screed machine is also equipped with pressure sensors to sense the screed force and feedback the screed effect.

[0258] The paving control unit is configured to: when the underwater screed machine is performing paving construction, obtain the weight of the stone in the hopper and the inclination angle of the hopper of the underwater screed machine based on the weighing sensor and the inclination meter of the perception subsystem 11; determine the mass of the stone based on the weight of the stone in the hopper and the inclination angle of the hopper, and calculate the real-time paving mass flow rate based on the time series data of the mass of the stone; obtain the current terrain elevation of the base, compare the current terrain elevation of the base with the pre-stored design target elevation, and calculate the terrain residual and the volume to be paved; based on the real-time paving mass flow rate, the terrain residual, the volume to be paved, and the predicted thickness distribution, generate paving control instructions through an optimization decision algorithm; wherein the predicted thickness distribution is obtained by a target prediction model based on historical construction data and multi-modal sensing information, and the paving control instructions are used to adjust at least one of the conveying belt speed, the distribution valve opening degree, or the movement speed of the distribution head of the underwater screed machine.

[0259] In some optional embodiments, to address the interference and failure of sensors in underwater environment, the perception subsystem 11 can use information fusion and redundancy configuration to improve reliability. For example, the weight sensor and the flow meter provide two estimates of material supply, which can be cross-checked; the laser and sonar sensor combination is used for ranging, with laser for high-precision local thickness in good water quality, and sonar for scanning in turbid water. Multi-sensor data fusion not only improves measurement accuracy, but also enhances the robustness of the system in harsh conditions. Research shows that by combining multiple redundant sensing information, the accuracy and stability of measurement can be maintained when the sensors are degraded or misaligned. Therefore, the overall structure of the construction control system realizes real-time fusion and redundancy fault tolerance of multi-sensor data through the "perception-decision-execution" closed loop link in hardware, ensures the accuracy of construction measurement and the continuity of system operation in harsh underwater environment through cross-checking of weight sensor and flow meter, adaptive switching of laser and sonar, thereby directly improving the anti-interference ability of intelligent construction process.

[0260] Please refer to Figure 8 , Figure 8The application flowchart of the paving control unit provided by the embodiments of the present application is shown in the figure. In the application flowchart of the construction control system, the perception subsystem 11 performs data collection as the perception layer, the software algorithm of the paving control unit includes data fusion in the perception layer and estimation of the paving thickness distribution in the prediction layer, and controls the paving rhythm conveying speed as the optimization decision layer. The paving control unit provides an intelligent closed-loop control flowchart of "perception-prediction-decision".

[0261] In the perception layer, raw data acquisition is performed by deploying multi-modal sensors including weighing sensors, inclinometers, volumetric flow meters, and sonars, etc. This layer corresponds to the hardware function of the perception subsystem 11, which is responsible for real-time acquisition of key physical information such as the weight of stone in the hopper, the attitude of the screed, the material flow, and the elevation of the base terrain. Subsequently, through data fusion technology (such as incline compensation to calculate the actual stone mass), these raw data are preliminarily processed and verified, and the raw, isolated sensor readings are converted into reliable state information with clear engineering significance (such as real-time stone mass, terrain residual error, 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. Based on the real-time state perceived in the previous stage and combined with historical construction data, the thickness distribution after stone paving is estimated using time series prediction models such as LSTM (Long Short-Term Memory). In the decision layer, the SAC (Soft Actor-Critic) reinforcement learning algorithm based on the maximum entropy framework is used for optimization decision-making in the present application. This layer combines the real-time state provided by the perception layer and the thickness distribution estimate output by the prediction layer to form a complete "state" input; abstracts the actuators to be controlled (such as the conveying belt speed and the paving valve opening) as "actions"; and designs a "reward function" that comprehensively considers paving uniformity, material utilization rate, and action smoothness. The SAC algorithm continuously optimizes the strategy and finally outputs the optimal control action under the current state. These action instructions are sent to the actuators to directly control the paving rhythm and conveying speed, thereby accurately guiding the construction operation.

[0262] In the perception layer, the mass of stone in the hopper is calculated in real time through the data of the weighing sensor and the inclinometer. If the weighing sensor reading is and the hopper inclination angle is , the calculation module can compensate the weighing sensor reading based on the inclinometer data to calculate the stone mass in the hopper under the inclined state. For example, the stone mass in the hopper can be approximated as:

[0263]

[0264] wherein, is the constant of gravitational acceleration. The above formula uses the attitude angles provided by the tilt sensor to compensate the weight reading for the gravity component, ensuring accurate estimation of the mass under different tilt conditions. The paving control unit reads the weight at a high frequency and , and calculates the continuous mass time series . Taking the derivative or difference of this series, the mass reduction per unit time, i.e. the real-time paving mass flow rate , can be obtained.

[0265] For example, if the mass reduction is in the time interval , then the paving flow rate is . Together with the volumetric flow meter data, the paving control unit can verify the mass estimation on one hand, and obtain the stone volume flow rate on the other hand, with the conversion relationship being where is the stone bulk density. The real-time mass and flow rate data provide accurate material supply status for subsequent paving rhythm control. Meanwhile, the IMU inertial unit can provide the acceleration and angular velocity of the paving head, which are used to infer the dynamic factors in the paving process (such as the influence of ship body sway on instantaneous discharging), so as to compensate for these disturbances in control. Through these multi-source data, the paving control unit can determine the current remaining amount of material and the amount of material to be fed every second at all times.

[0266] Further, in the perception layer, the current terrain elevation of the underwater foundation bed is scanned in real time by the sonar / laser thickness measurement sensor at the front end of the paving head . The paving control unit has pre-stored the designed target elevation surface , and whenever a scan is completed, the paving control unit compares the current terrain elevation of the foundation bed with the pre-stored designed target elevation, and calculates the elevation residual:

[0267]

[0268] In the area where filling is needed, is positive, and its multiplication with the corresponding area unit can obtain the volume to be paved: (for areas where no filling or excavation is needed, then is taken as 0 to avoid overfilling), where represents the area of an area unit. In order to reduce the influence of noise, the field can be filtered and smoothed, or the local average residual is extracted as a control reference.

[0269] In addition, the posture of the material head and the water depth are recorded synchronously when the thickness is measured by the sonar, and the accuracy of the residual calculation can be ensured through coordinate transformation. For example, when the material head is tilted, the sonar distance measurement is corrected by the IMU posture angle, so that the measured terrain point cloud is aligned with the global coordinates. In an environment with a large flow rate, multiple measurements can estimate the erosion and displacement of the laid stone by the water flow, and timely corrections . Finally, the terrain residual matrix and the required laying volume at each position will be an important input for subsequent laying optimization decisions.

[0270] In the prediction layer, the predicted thickness distribution is obtained by the target prediction model based on historical construction data and multi-modal sensing information, and the material control instruction is used to adjust at least one of the conveying belt speed, the material valve opening degree or the material head moving speed of the underwater screed.

[0271] In order to make more intelligent decisions on material laying strategy, a data-driven prediction model is introduced in the material laying control unit to estimate the future thickness distribution. The prediction model can use historical construction data and multi-modal sensing information to predict the thickness distribution after stone laying under the current control strategy, as well as possible weak areas. Considering the time correlation and nonlinear characteristics of the laying process, a long short-term memory neural network (LSTM) is selected to build the predictor. LSTM is suitable for processing time series data, and can input a time window of sensor sequences (such as the discharge mass flow rate, material head motion trajectory, terrain change amount, etc. in the past few seconds) and output the predicted future area thickness increase distribution. In this way, the stone settlement and distribution trend in the next short period of time can be predicted while the material is being laid. The LSTM model of the material laying control unit can be trained for the settlement and diffusion behavior of the stone under water, with input including input amount, flow rate, stone diameter, etc. and output being the planar thickness distribution estimate at the corresponding time. Through continuous comparison and correction with sonar measurement results, the LSTM model can continuously adapt to the environment and improve the prediction accuracy.

[0272] The prediction model needs to be trained and verified with a large amount of historical data or simulation data. Training samples can be obtained based on offline simulation and actual measurement correction: change the material laying rhythm and environmental parameters in the simulation environment, generate the laying thickness distribution, and then fine-tune the model combined with a small amount of field sonar survey data. The trained model will be embedded in the material laying control unit, which will read the current multi-modal state at each decision-making period and output the prediction of the next period thickness distribution .

[0273] In the decision layer, according to the state obtained by perception and the output of the prediction model, the paving control unit needs to decide in real time how to control the paving rhythm and conveying speed to optimize the uniformity of the paving thickness and the construction efficiency. This belongs to a dynamic sequential decision problem, which can be abstracted as a Markov Decision Process (MDP). In this application, a Deep Reinforcement Learning (DRL) method is used to solve the optimal control strategy, so that the paving control unit can automatically learn the optimal paving control decision .

[0274] Specifically, the state of the Reinforcement Learning (RL) agent at each time step is defined to include the current encoded terrain residual, the paved thickness distribution, the remaining stone mass in the hopper, the real-time paving mass flow rate, the paving head pose, the paving head speed, and the working condition parameters, etc. The action is defined as the adjustment of the paving rhythm, such as adjusting the conveying belt speed, starting and stopping the paving valve, changing the paving head moving speed, etc. The reward function describes the paving quality and efficiency, for example, the negative thickness non-uniformity variance minus the time / energy cost, that is, to encourage faster completion of construction while ensuring thickness uniformity. Through repeated interaction with the environment (simulated construction process or actual process), the reinforcement learning agent gradually optimizes the strategy to maximize the comprehensive reward function to train and optimize the paving strategy.

[0275] In the embodiment of the application, the optimization decision algorithm is constructed by using the SAC (Soft Actor-Critic) deep reinforcement learning algorithm, and combined with the priority experience replay (PER) and random network distillation (RND) techniques to accelerate the training convergence and enhance the robust exploration of the strategy. The action space of the agent in the deep reinforcement learning framework is continuous and bounded, and the action space corresponds to the control instructions of the conveying belt speed, the paving valve opening or the paving head moving speed, and the output of the reinforcement learning strategy is the optimized paving control instruction. For example, when the local area thickness is low, the strategy may increase the paving frequency or slow down the moving speed to increase the amount of delivery at that place; on the contrary, if a certain place tends to be too high, the strategy reduces the delivery or speeds up the passing speed. Since the strategy is optimized globally from the cumulative effect, the paving control unit can dynamically adjust the current delivery speed and moving path to ensure that the thickness is as uniform as possible when the entire area is constructed. In addition, the RL agent can also adaptively adjust the strategy according to different water flow and terrain conditions, which can improve the adaptability to complex environments.

[0276] In some optional embodiments, the training steps of the agent of the deep reinforcement learning framework include:

[0277] The multi-modal fusion state tensor of the agent is generated by periodically collecting stone mass, real-time material mass flow rate, encoded terrain residual, material head pose, material head speed, and working condition parameters;

[0278] The action of the agent is output by the policy network according to the current multi-modal fusion state tensor, and is obtained through reparameterization sampling and tanh transformation; wherein the action corresponds to the actual control quantity of the conveying belt speed, the material valve opening or the material head moving speed after linear mapping;

[0279] The soft action value of the state-action pair is estimated based on the double critic network, and the transition tuple is sampled from the experience replay buffer based on the priority experience replay mechanism;

[0280] The target soft action value is calculated based on the transition tuple, wherein 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;

[0281] The parameters of the double critic network are updated by minimizing the mean square error loss between the soft action value output by the double 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; wherein the parameters of the target critic network track the parameters of the double critic network through soft update.

[0282] For example, The multi-modal fusion state tensor at the moment is composed of the following observations:

[0283] wherein is the stone mass in the hopper after inclination compensation, is the weighing reading, is the inclination, is the acceleration of gravity, and inclination compensation is used to eliminate the influence of attitude change on weighing to ensure the stability of mass and flow rate estimation; is the mass flow rate; is the elevation residual heat map generated by sonar or laser after encoding by the encoder ; is the body / material head pose; is the travel and swing speed; is the water depth / flow rate / material diameter and other working condition parameters.

[0284] The action of the agent is:

[0285]

[0286] wherein, is the conveying 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.

[0287] 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.

[0288] 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.

[0289] 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:

[0290]

[0291] 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.

[0292] 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:

[0293]

[0294] 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 rewards that can be obtained from taking action.

[0295] The soft action value function and the soft state value function are:

[0296]

[0297]

[0298] 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 is sampled from, the value of the soft state value function at the next state ; is the expected reward for the action ; is sampled from the probability distribution of the current policy , is used to measure the randomness of the policy.

[0299] In function approximation, in order to stabilize training and avoid overly optimistic value estimates, the SAC algorithm in the embodiments of the present application adopts the design of double critic networks. Specifically, the algorithm maintains two sets of critic networks and in parallel, and takes the minimum value to calculate the target value, so as to suppress the overestimation bias of the Q value, and provide a more robust value evaluation for decision making. The target value is calculated by the target network and the minimum policy, and the calculation formula is:

[0300]

[0301] wherein, is the immediate reward, is the expected of the next action, is the minimum value of the outputs of the two target critic networks, is the log probability of the policy.

[0302] The update target of each critic network is to minimize the mean square error loss between its predicted value and the target value y:

[0303]

[0304] wherein, is the expected value of the experience data, and the expectation is calculated on the transition tuple sampled from the experience replay buffer D.

[0305] In order to further improve the stability of training, the parameters of the target critic network are not directly copied from the parameters of the online network , but slowly track the online network parameters through a soft update method (wherein is close to 1). The data samples required for training come from the experience replay buffer . In order to speed up the learning process and let the agent learn more effectively from key experiences, the algorithm introduces a priority experience replay mechanism. According to the temporal difference error The probability of being sampled is set by the absolute value of the error , the larger the error, the higher the probability of being sampled. At the same time, in order to correct the bias caused by non-uniform sampling, an importance weight is introduced to adjust the loss function.

[0306] The SAC algorithm adopts a random strategy and uses a reparameterization technique to efficiently handle bounded action spaces. The policy network does not directly output actions, but outputs the parameters of a Gaussian distribution and . The action is reparameterized to (-1, 1) using :

[0307]

[0308]

[0309] where is the mean vector output by the policy network, is the standard deviation vector output by the policy network, is a random noise sampled from a standard normal distribution, denotes a normal distribution, is the final executed action, is the original action value without normalization.

[0310] The optimization goal 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:

[0311]

[0312]

[0313] When calculating the log probability of the policy, the Jacobian determinant correction introduced by the tanh transformation needs to be considered. Where is the expectation with respect to the state and noise, is the log probability of the policy, is the probability of selecting action in state , is the final generated action.

[0314] The optimization goal of the policy network can be decomposed into the trade-off of two parts, including maximizing the Q value and maximizing the entropy, and the loss function guides the agent to learn the optimal cloth control strategy, including: 1. State input, the policy network receives the state , 2. Action output, the network outputs the distribution parameters of the control instruction and ,3, loss calculation, including maximizing Q value and maximizing entropy, to encourage the selection of actions that can improve paving quality (uniformity), reduce material waste, and maintain a certain randomness to adapt to uncertainties such as changes in water flow and changes in stone characteristics.

[0315] In the embodiments of the present application, SAC uses an adaptive mechanism to adjust without manual setting. By setting a target entropy (typically taking the negative of the action space dimension), and then adjusting by optimizing the loss function, the average entropy of the strategy is maintained near the target entropy, so that the algorithm can automatically adjust the exploration level according to different working conditions (such as changes in water depth and flow rate):

[0316]

[0317]

[0318] where, is the expectation about the action, and the action is sampled from the distribution of the current policy , is the learning rate of the temperature parameter, is the gradient of the loss function with respect to . The specific adjustment process includes: when the policy entropy is too low (representing insufficient exploration), increase , thereby strengthening the importance of the entropy term and encouraging more exploration, and when the policy entropy is too high (representing excessive randomness), decrease , thereby reducing the importance of the entropy term and reducing random exploration.

[0319] To ensure the safety of the underwater screed, all generated actions need to be projected into a safe action set which defines the amplitude constraint and the rate constraint of the action, i.e. , which can effectively prevent the mutation of control instructions and protect the mechanical structure of the underwater screed:

[0320] where, is the amplitude constraint, is the lower bound of the action vector, is the upper bound of the action vector, the amplitude constraint limits the value range of each component in the single action vector , which is used to prevent the agent from outputting instructions that exceed the physical capabilities of the underwater screed, thereby avoiding equipment overload or damage. For example, it ensures that the conveyor belt will not be commanded to operate at an unrealistic super-high speed, or that the valve opening will not be set to 120%. a rate constraint, an action change, a maximum allowed instantaneous change, the rate constraint is used to suppress the "surge" phenomenon, protect the underwater screed from mechanical impact caused by sudden changes in instructions, and make the construction process smoother and improve the paving quality.

[0321] Finally, the algorithm designs an end-to-end landing process. On the data channel, the paving control unit collects sensor data at a period of (derived ), the sonar generates and is encoded by the encoder , the action is normalized to (-1, 1) and then linearly or proportionally mapped to the actuator (such as frequency converter, proportional valve, servo, etc.), and the transition tuple is stored in the experience pool to form the transition tuple ) into the warehouse, where (if RND is enabled). During training, the critic, policy and temperature parameters are updated alternately, and is updated softly. To improve efficiency, mechanisms such as priority experience replay and random network distillation can be used to enhance exploration (such as using PER sampling high TD error fragments and using RND to improve early exploration). When deployed online, the trained policy network can be placed on the edge computing device or GPU-NPU module of the decision subsystem 12 to ensure that the inference delay is less than 50 milliseconds. The paving control unit forms a multi-layer closed-loop control: the upper layer SAC decision gives the target paving rhythm, the lower layer is tracked by the PID controller of the decision subsystem 12, and the terrain feedback of the sonar is combined for second-level correction, and the sonar-terrain ring is refreshed by second-level correction, and the action safety projection is combined with the sensor redundancy (such as combining weight, flow, sonar and laser mutual verification) to form a human-machine takeover channel to jointly ensure the reliable operation of the underwater screed.

[0322] In the construction control system of the underwater screed machine provided in the application, the accurate material mass flow rate is dynamically calculated by real-time sensing of the stone weight and the inclination angle in the hopper in combination with time sequence data, thereby providing accurate instantaneous working condition basis for construction. Meanwhile, the paving control unit compares the real-time collected bed surface elevation with the design target, calculates the terrain residual and the volume to be paved, and thereby accurately masters the actual state and demand of the construction surface. The control subsystem, on the basis, integrates the thickness distribution predicted based on historical data and multi-modal information, and comprehensively regulates and controls the key execution parameters such as the conveying belt speed and the material valve opening degree through the optimization decision algorithm. Through the combination of real-time working condition monitoring, dynamic flow calculation, terrain deviation analysis and predictive control strategy, the whole-process closed-loop accurate control from material conveying to bed forming is realized, the problems of uneven paving or elevation deviation caused by information lag or human judgment error in traditional construction are effectively avoided, the control precision of underwater screeding construction is improved, and thereby the construction precision and construction quality of the underwater bed are improved.

[0323] In some optional embodiments, after the underwater screeding operation is completed, the intelligent control system 10 can also be used for bed surface flatness evaluation. The intelligent control system 10 can collect multi-source sensing data through the multi-source mapping sensors arranged in the operation sea area, and generate a bed surface flatness evaluation result of the operation sea area based on a flatness evaluation unit.

[0324] The flatness evaluation unit is configured with a trained target bed surface prediction model, the prediction unit is configured to generate a multi-source fusion bed surface model based on the multi-source sensing data, use the multi-source fusion bed surface model as the input of the target bed surface prediction model, and output an elevation prediction map of the bed surface based on the target bed surface prediction model; and generate a flatness evaluation index of the operation sea area based on the elevation prediction map.

[0325] The target bed surface prediction model is obtained by training an 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 an alpha evolutionary algorithm, and the loss function of the initial surface prediction model is a Huber function; the value of each grid point in the elevation prediction map represents the deviation amount of the bed surface at this point from an ideal design plane or a target elevation.

[0326] In the embodiments of the application, the operation sea area refers to a region in which the underwater screeding operation has been completed, and the multi-source sensing data refers to multi-source sensing data (including three-dimensional topography and texture data) of the bed surface obtained from different physical dimensions (such as acoustics, optics) by using different types of sensors.

[0327] Exemplarily, after the completion of the flattening construction, detailed information of the bed surface can be obtained by the multi-source surveying sensors arranged in the working sea area, including (1) multi-beam sounding: a multi-beam echo sounder can be arranged on the working ship to perform high-resolution scanning on the flattening area to obtain data such as digital elevation model (DEM) and reflection intensity map of the bed surface; (2) laser radar / structure light scanning: according to the water depth and water clarity, underwater laser radar, structure light projector and camera device can be used to perform fine three-dimensional scanning on the local area to obtain high-precision point cloud data; (3) underwater camera shooting: underwater high-definition cameras can be deployed to shoot visible light image sequences of the bed surface, which are used for light measurement three-dimensional reconstruction or as a texture auxiliary means.

[0328] By correcting, filtering noise, and removing abnormal points of the obtained multi-source sensing data, all data can be converted to a unified coordinate system, which can overcome the limitations of a single data source (such as sonar being susceptible to noise interference and optical equipment being affected by water clarity), and through coordinate unification, registration and complementation, a multi-source fusion bed surface model with higher integrity and accuracy can be generated. Finally, a multi-source fusion bed surface three-dimensional model is generated. The multi-source fusion bed surface model can be a digital elevation model (DEM) with texture information.

[0329] Exemplarily, the data obtained by each sensor can be transmitted to the support platform through an underwater communication link or a buoy relay, and the multi-source fusion bed surface model is generated by the support platform based on the multi-source sensing data. The support platform can be a data processing system on the working ship, a shore-based or fixed platform control center, or a cloud computing service platform.

[0330] By performing sound velocity correction, noise filtering and outlier removal on the multi-beam sounding data, seabed terrain point cloud or raster elevation map covering the flattening area is generated. By performing distortion correction and coordinate conversion (considering the influence of water medium refraction) on the laser / structure light point cloud data, and registering and fusing with the sonar data. By performing color calibration and distortion correction on the underwater image, sparse point cloud is extracted by using photogrammetry algorithm, and then combined with the laser point cloud for densification processing. After the preprocessing step, the multi-source fusion bed surface model and its image texture in the unified coordinate system are obtained, and the above data will be used as the input of the deep learning model.

[0331] Please refer to Figure 9 , Figure 9A schematic diagram of the framework of the target base surface elevation prediction model provided by the embodiments of the present application. The target base surface prediction model is an intelligent body for base surface roughness evaluation. The initial surface prediction model is an end-to-end model constructed by a convolutional neural network (CNN) and a multi-head attention mechanism. CNN is good at extracting local spatial features (such as small concave and convex fluctuations) in images or raster data, while the multi-head attention mechanism can capture global dependencies between different data sources, weigh the importance of each source information, and achieve more effective feature fusion. The target base surface prediction model realizes accurate prediction of the base surface elevation through the end-to-end model.

[0332] The input of the model is a multi-source fusion base surface model, which is specifically embodied as three data extraction modules of raster elevation, depth, and texture features. The three-dimensional model generated by the support platform is converted into regular raster data, in which each grid point contains not only the elevation value, but also the depth information derived from the multi-beam data and the texture features extracted from the underwater image.

[0333] These input data are sent to the convolutional neural network module. The CNN layer shown in the figure usually includes a series of continuous operations, including: convolution operation uses a learnable convolution kernel to perform sliding window calculation on the input data to capture local spatial features from edges, shapes to more complex terrain structures; the ReLU activation function introduces a nonlinear transformation for the network, which can fit complex nonlinear relationships; the maximum pooling operation reduces the spatial resolution of the feature map, while preserving significant features, increasing the translational invariance and receptive field of the model. This series of operations work together to transform the original input into high-level, abstract feature maps.

[0334] Among them, the core of the convolutional neural network is the convolution operator and feature fusion. By sliding the convolution kernel in space and multiplying and accumulating the input feature map, the weighted extraction of local patterns is realized. The output of each convolution layer is mapped through a nonlinear activation (ReLU) to introduce sparsity and nonlinear discrimination ability. 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 can capture global dependencies. This mechanism combined with the local features extracted by convolution helps the model to obtain local details and global context at the same time, improving the accuracy of terrain prediction.

[0335] The hyperparameters of the initial surface prediction model are optimized using an Alpha Evolution (AE) algorithm, which automatically finds the best combination of hyperparameters by balancing global exploration and local exploitation through a core Alpha operator.

[0336] After the CNN extracts spatial features, it enters a multi-head attention mechanism layer, which allows the model to autonomously focus on deeper dependencies between different features. For example, one group of heads may focus on the correlation between elevation features and texture features, while another group of heads may focus on the spatial relationship between distant grid points. In this way, the model can more intelligently integrate information from different data sources and understand the context of large-scale terrain, rather than just the local neighborhood, making more accurate predictions.

[0337] For example, grid elevation, depth, and texture are fused through a multi-head attention mechanism and then input into the same convolutional neural network. In the encoding stage, multiple convolutional layers and pooling layers are stacked in sequence to extract high-level features. The convolutional layers use a certain size of convolutional kernel (such as 3x3) to slide and extract local area features, and a ReLU activation function is added after each convolutional layer to introduce nonlinearity. Batch normalization (BN) is used to accelerate convergence and improve generalization performance. For example, 4 layers of convolution are used, with a convolution kernel size of 3x3 and a channel number increasing layer by layer (such as 32, 64, 128, 256), and ReLU and batch normalization are used after each layer. The convolution calculation satisfies: where represents the input image tensor, is the convolution kernel weight, is the output position index, is the kernel height and width range. Local feature extraction is achieved through convolution operations, and the pooling layer then down-samples the feature map with a window such as 2x2 to reduce the resolution while extracting invariant features. After convolutional feature extraction, a fully connected layer is used to merge spatial information and output a global flatness index for predicting continuous elevation or bias values.

[0338] After the multi-head attention mechanism weights and integrates the high-level features, they are flattened and fed into one or more fully connected layers. The fully connected layer acts as a powerful regressor that maps the distributed features learned by all previous stages to the final output space. In this task, the output layer is usually a linear unit, and the output values directly correspond to the predicted elevation values for each grid point.

[0339] The elevation prediction map directly gives the elevation distribution of the flattened subgrade surface. Since a deep learning model is used, more detailed terrain details can be obtained compared to traditional interpolation methods. The output elevation prediction map can be subtracted from the measured true elevation obtained deviation from (or reduction to) a reference elevation surface The deviation map intuitively shows which areas of the roadbed still have protrusions or depressions, and the numerical value reflects the unevenness. Through this output, subsequent calculations of flatness evaluation indicators, such as root mean square error, can be easily performed for acceptance evaluation.

[0340] In some embodiments, to train the initial surface prediction model, a loss function is defined to measure the deviation of the predicted elevation from the true elevation. In the embodiments of the present application, the Huber function is used as the loss function of the initial surface prediction model to measure the gap between the predicted value and the true value. The Huber function can integrate the advantages of mean squared error (MSE) and mean absolute error (MAE), and is not sensitive to outliers in the training data, thereby improving the stability of model training.

[0341] The mathematical form of the Huber loss is:

[0342]

[0343] wherein, is the difference between the predicted value and the true value, is a threshold hyperparameter that controls the inflection point of the loss from quadratic to linear. By selecting an appropriate , the Huber loss is equivalent to MSE when the error is small, and equivalent to absolute error when the error is too large, thereby improving the robustness to outliers. The model training aims to minimize the above loss, and adjusts the network parameters through backpropagation to make the predicted elevation approach the true value.

[0344] Please refer to Figure 10 , Figure 10 is the specific flowchart of the AE algorithm provided by the embodiments of the present application. In some embodiments, the step of optimizing the hyperparameters of the initial surface prediction model based on the AE algorithm can include:

[0345] A plurality of candidate solutions are generated in a uniform distribution manner in a hyperparameter space, each candidate solution representing an initial surface prediction model hyperparameter configuration vector; wherein the hyperparameters include the initial surface prediction model convolution kernel size, learning rate, batch size and network depth, and the value range of each hyperparameter in the hyperparameter space is determined according to the characteristics of the base terrain data; an evolution matrix is constructed by randomly sampling from the current candidate solution, each row vector of the evolution matrix corresponding to a hyperparameter configuration; the evolution matrix is initialized and the evaluation function is calculated, the value of the evaluation function being the loss function value calculated by the initial surface prediction model on the base verification set; it is judged whether the current function evaluation times reaches the maximum evaluation times, if not, the evolution matrix, the perturbation matrix and the decay factor are calculated; the calculation path of the basis vector is selected by generating a random number, when the random number is less than the set threshold, the basis vector is calculated through the first path, otherwise the basis vector is calculated through the second path; after obtaining the basis vector according to the selected path, the first individual with fitness higher than the preset threshold and the second individual with fitness lower than the preset threshold are selected from the current population, and the control parameter is calculated; the basis vector, the control parameter, the decay factor and the perturbation matrix are used to update the search operator through the alpha operator; the updated search operator is subjected to boundary constraint processing, and the binary backoff strategy is used to adjust the solution components exceeding the boundary; the fitness of the individuals before and after updating is compared through the greedy selection strategy to decide whether to replace the original individual; after the update of all individuals in the current generation is completed, it is judged whether the individual index reaches the population size, if not, the next individual is processed; after all individuals are processed, it is judged again whether the function evaluation times reaches the maximum value, if so, the optimal hyperparameter combination is output.

[0346] For example, in a dimensional hyperparameter space, the AE algorithm first generates candidate solutions in a uniform distribution manner to form a candidate matrix:

[0347]

[0348] wherein, represents the th hyperparameter configuration vector, is the lower bound and upper bound of each dimension; represents generating a vector with a length of , each component is uniformly sampled on .

[0349] The Alpha operator fuses three types of information, including adaptive global starting point, global random disturbance and local differential correction, in one update, and considers global exploration and local development, so as to efficiently mine and refine candidate solutions. Its mathematical expression is:

[0350]

[0351] 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 .

[0352] For adaptive basis vectors, the evolutionary starting point is initially calculated in two ways:

[0353]

[0354] 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.

[0355] The methods for calculating basis vectors through the first path include:

[0356] ;

[0357] Methods for calculating basis vectors using the second path include:

[0358] ;

[0359] in, As basis vectors, and The learning rate parameter, and are the history basis vector based on the with-replacement sampling path and the history basis vector based on the without-replacement sampling path, respectively; A is a D x D matrix obtained by with-replacement sampling D times from the candidate solution matrix X, B is a matrix obtained by without-replacement sampling K times from the candidate solution matrix X, is the fitness-based weight vector.

[0360] where the learning rate , is the current number of target function calls, is the maximum allowed number of calls.

[0361] For the random step size , a global search function is provided. The decay factor is a nonlinearly decreasing value closely related to the perturbation matrix . Its decay process is as follows:

[0362]

[0363] The perturbation matrix is calculated as:

[0364]

[0365] where and are random real matrices generated by . Among them, is a matrix containing rows and columns, represents the i-th row vector in . To ensure that the search is within the feasible region, AE uses the "bisection backtracking" idea for variables that exceed the limit:

[0366]

[0367] After updating, the greedy strategy is used to select the optimal one:

[0368]

[0369] where,

[0370] is the value of the j-th dimension (i.e., a certain specific hyperparameter such as the learning rate) of the i-th candidate solution after updating by the alpha operator. is the upper bound of the j-th hyperparameter in the search space, is the lower bound of the j-th hyperparameter in the search space. is the hyperparameter configuration vector of the i-th candidate solution in the t-th generation population. ​The new candidate solution vector generated after the alpha operator is updated and the boundary constraint is applied for the ith candidate solution.

[0371] During the model training, supervised learning is used for training, and the label data comes from high-precision measured bed elevations (for example, true elevations obtained from underwater laser point clouds or multi-beam sounding). The measured elevations of a certain area after flattening construction are taken as true values , and the multi-source observation data (sonar depth map, image texture, etc.) of the corresponding area are taken as model inputs , to construct training samples . The loss function selects the Huber or MSE described above to measure the difference between the model output and . In the training process, the model parameters are optimized by minimizing the loss to make the prediction result approximate the true elevation distribution. The optimization algorithm uses the Adam adaptive gradient optimizer, which can be set to an initial learning rate of 0.001, and combines a momentum factor . Adam optimization shows good convergence efficiency and robustness to hyperparameters in deep learning. To ensure training stability, the learning rate decay strategy is used in the training process, for example, multiplying the learning rate by 0.1 every several rounds to prevent late oscillation.

[0372] Regarding the selection and verification of hyperparameters, the batch size during training depends on the hardware memory and the size of the data. For grid input, a batch size such as 32 or 64 can be selected; if small image patches are used for training, even larger batches can be set (some studies use a batch of 512 to train 9x9 small patches to fully utilize the data). Model training needs to divide the training set and the validation set, and generally the data is randomly divided in a ratio of, for example, 8:2, while ensuring that the data of different areas are covered. In order to fully utilize the limited data and evaluate the robustness of the model, the K-fold cross-validation scheme is introduced in the embodiments of the present application. For example The dataset is equally divided into 5 parts, and 1 part is taken out each time for verification, and the remaining 4 parts are trained, and 5 rounds are repeated to obtain the average performance. Compared with single division, this scheme can more reliably evaluate the model and reduce the deviation caused by accidental division. A study compared the effects of traditional random division and K-fold verification, and found that the test accuracy increased from 84.1% to 88.3%, and the spatial error standard deviation decreased by 60%, indicating that cross-validation helps to improve the generalization performance of the model. Prevent overfitting: Due to the high cost of underwater terrain data acquisition, the training sample is limited, and the model is prone to overfitting. Therefore, various regularization strategies are used: First, use the Dropout layer, randomly discard neuron outputs with a certain probability (such as 30%) in the fully connected layer or decoding stage to break the feature dependence. Dropout reduces the over-reliance on certain local features during training, improving the model's adaptability to unseen data. Second, apply early stopping, monitor the validation set loss, and stop training when there is no improvement for several rounds to avoid overtraining. Third, perform data augmentation. If the original image and depth data allow, random rotation, translation perturbation or noise addition can be applied to the training samples to increase data diversity and improve model robustness. Finally, introducing residual connection and regularization term in network structure is also helpful. Residual connection allows training of deeper networks without loss of stability; regularization weight decay adds a term to the loss to suppress model parameters that are too large. By comprehensively using the above strategies, the embodiments of the present application can ensure the decline of training error while paying more attention to the performance of validation error, striving to obtain a flatness evaluation model with strong generalization ability and suitable for different environments.

[0373] Further, in order to comprehensively evaluate the performance of the model, the embodiments of the present application select a variety of indicators for quantitative evaluation on the validation set, including RMSE, MAE and These indicators measure the degree of conformity of predicted elevations to true elevations from different angles. Among them, the root mean square error (RMSE) measures the overall error magnitude, reflecting the standard deviation of the predicted value from the true value. The mean absolute error (MAE) is the average of the absolute value of the error, and the formula is . MAE directly represents the average deviation. Since the median is strong against outliers, MedAE can better reflect the typical error level of most points, ignoring a small number of abnormal deviations. The determination coefficient ( ) measures the degree of explanation of the model to the true elevation change, defined as . The value range is 0, 1, and the closer to 1 indicates that the model explains more terrain variance, that is, the predicted results fit the true values better.

[0374] ​In some embodiments, the flatness evaluation index can include a global flatness index and a local flatness index; wherein the global flatness index includes a root mean square error, which is used to quantify the overall elevation fluctuation of the work sea area; the local flatness index includes the average slope change of the work sea area and the local extreme value detection result, and the local flatness index is used to represent the surface smoothness and local unevenness rate of the work sea area. Step S12 can specifically include: inputting the multi-source fused base surface model into the target base surface prediction model, and obtaining the elevation prediction map of the base surface by the target base surface prediction model through forward reasoning; based on the elevation prediction map and the reference elevation data, calculating the elevation deviation value of each grid point to generate an elevation deviation distribution map; wherein the reference elevation data is the measured base elevation or the ideal design plane elevation, and the elevation deviation distribution map includes the convex or concave area of the base surface, and the deviation value in the elevation deviation distribution map represents the unevenness degree of the work sea area.

[0375] The trained target base surface prediction model can output a corresponding elevation prediction map according to the input multi-source fused base surface model, and each pixel or grid point in the map corresponds to a predicted absolute elevation value. Comparing this prediction map with the ideal design plane or target elevation (i.e. the perfect flat surface required by construction) can obtain the deviation amount of each point, thereby forming an elevation deviation distribution map. The variation in the elevation deviation distribution map is a relative deviation value, and the deviation amount is positive for convexity and negative for concavity.

[0376] The step of generating the flatness evaluation index of the work sea area based on the elevation deviation distribution map can specifically include: calculating the mean square error or the root mean square error by counting the deviation values of each point, to quantify the overall elevation fluctuation of the work sea area based on the mean square error or the root mean square error; calculating the slope change of the base surface, determining the local slope by the elevation difference of adjacent grid points, and evaluating the maximum change rate of the work sea area.

[0377] For example, after training, the model is deployed for flatness evaluation of field data. The pre-processed fused multi-source data is input into the trained CNN model, and forward reasoning is performed to obtain the flatness prediction result of the base surface. Next, the model output result is calculated for flatness index: the mean square error (MSE) or the root mean square error (RMSE) is obtained by counting the deviation values of each point, to quantify the overall elevation fluctuation; at the same time, the slope change of the base surface is calculated, for example, the local slope is calculated by the elevation difference of adjacent grid points, and the maximum change rate is evaluated to capture the steepness or uneven area. These indexes comprehensively reflect the flatness after leveling. Finally, the flatness evaluation result is generated to generate a report, and the pre-set acceptance threshold is compared to judge the construction quality: if the mean square error and other indexes are within the allowable range, the base surface flatness is considered to be qualified, otherwise the out-of-tolerance area is marked to guide the rework and modification.

[0378] In the embodiments of the present application, a visual method can be used to compare and analyze the elevation map / deviation map output by the model with the real measurement. On the one hand, a control chart of the predicted elevation and the measured elevation (such as a profile line comparison, a three-dimensional surface chart, etc.) is drawn to intuitively check whether the model has captured the key undulating characteristics of the roadbed surface. For an ideally flat area, the predicted and measured curves should basically coincide; if the prediction deviation is large at a certain place, it will be shown as a significant high or low on the chart. On the other hand, a heat map is used to display the deviation distribution, and the deviation size is represented by color, which can clearly locate the position and degree of residual unevenness. This scatter plot comparison chart and deviation heat map can quantitatively and intuitively evaluate the flattening effect: dense point cloud and close to the ideal line mean good overall flatness, and outliers and deviation concentration areas indicate parts that need to be reviewed in detail. Through the analysis of the above-mentioned various visualization methods, we not only verify the accuracy of the model prediction, but also can further guide the rework and optimization of construction according to the deviation map, and finally achieve a comprehensive and objective evaluation of the quality of the underwater roadbed flattening.

[0379] In the implementation process of generating the roadbed surface flatness evaluation result of the work sea area described above, the multi-source fusion roadbed surface model is constructed by acquiring multi-source sensing data of the work sea area, which effectively improves the accuracy and robustness of the basic data. Subsequently, the fusion model is processed by using the optimized target roadbed surface prediction model. The model combines the feature extraction capability of the convolutional neural network and the advantage of capturing long-distance dependency relationship of the multi-head attention mechanism, and optimizes the hyperparameters through the evolutionary algorithm, so that the model has stronger adaptability and precision when predicting the roadbed surface elevation. The use of Huber loss function further guarantees the training stability of the model in the presence of abnormal values. The finally generated elevation prediction map directly reflects the deviation of each point of the roadbed from the design elevation, providing a direct basis for the generation of flatness evaluation indicators, so that a comprehensive and objective evaluation result of the roadbed flatness of the work sea area can be obtained, which can significantly improve the accuracy and reliability of the roadbed flatness evaluation.

[0380] It should be understood that, in the above embodiments, the various modules of the system provided in the embodiments are only exemplified by the division of the functional modules in the above description, and in actual application, the above functions can be completed by different functional modules according to needs, i.e., the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0381] The function modules in the above embodiments can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit, and the integrated unit can be realized in the form of hardware or in the form of a software function unit. In addition, the specific names of the function units and modules are only for the purpose of mutual distinction, and do not limit the protection scope of the embodiments of the present application.

[0382] Based on the same application concept, the embodiments of the present application also provide an intelligent control method of an underwater screed machine, which can include:

[0383] obtaining working state data of the underwater screed machine; discretizing a working process of the underwater screed machine into multiple stages, and determining a frame parameter of a current stage based on the working state data in each stage; under the condition of satisfying multiple preset constraints, determining a liquid level change amount of each target tank based on the frame parameter, to generate a control instruction of the current stage based on the liquid level change amount; wherein the target tank is a tank that needs to perform a liquid level adjustment operation in the current stage; and controlling the underwater screed machine to perform underwater work according to the multi-stage control instruction.

[0384] Optionally, the method can further include: discretizing the working process of the underwater screed machine into multiple stages based on a time interval or a depth interval; and in each stage, switching to the next stage when a trigger condition of reaching a fixed time length, an intermediate draft threshold or an inclination threshold is reached.

[0385] Optionally, the method can further include: for each tank, calculating an antecedent variable of fuzzy reasoning; wherein the antecedent variable includes a correction alignment degree, a force arm efficiency and a liquid level margin, the correction alignment degree is calculated based on the consistency of the attitude gradient and the target correction direction, the force arm efficiency is calculated based on the force arm weight of the tank to the roll or the pitch, and the liquid level margin is calculated based on the safety margin of the current liquid level and the limit value; mapping the antecedent variable to a fuzzy set through a predefined membership function, and applying a fuzzy rule base constructed based on domain knowledge to obtain a participation level representing the tank; and determining whether the tank is a target tank based on the participation level.

[0386] Optionally, the method can further include: using a fuzzy particle swarm optimization algorithm, and constraint conditions of hyperplane projection and boundary repair processing, to search for an optimal decision variable in a sparse search space composed of the frame parameter of the current stage; and the optimal decision variable is the liquid level change amount of each target tank.

[0387] Optionally, the method can further include: according to the positive and negative and size of the liquid level change amount of each target tank, and in combination with target hardware parameters of the underwater screed machine, calculating the opening time length and action timing of the pump or valve corresponding to each target tank, and generating a corresponding control instruction.

[0388] Optionally, the method can further include: monitoring the working state data in the case that the execution subsystem executes the control instruction, and triggering a local correction operation to adjust the liquid level change amount and reissue the control instruction if it is detected that the working state data represents that the attitude or draft of the underwater screed deviates from the preset threshold.

[0389] Based on the same application concept, the embodiments of the present application also provide a computer device, which is arranged on the underwater screed or in communication connection with the underwater screed. The computer device can include a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method as described above when executing the computer program.

[0390] Based on the same application concept, the embodiments of the present application also provide an underwater screed, which is provided with the intelligent control system 10 of the underwater screed as described above.

[0391] The above-described embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. An intelligent control system for an underwater screed, the system comprising: The method comprises: a perception subsystem configured to acquire working state data of the underwater leveling machine; a decision subsystem comprising a stage decision unit and a fuzzy swarm intelligence optimization unit; the decision subsystem is configured to discretize the working process of the underwater leveling machine into multiple stages, and in each stage, determine the framework parameters of the current stage based on the working state data; the fuzzy swarm intelligence optimization unit is configured to determine the liquid level change amount of each target tank based on the framework parameters under the condition of meeting multiple preset constraints, and generate control instructions for the current stage based on the liquid level change amount; wherein the target tank is a tank that needs to perform liquid level adjustment operation in the current stage; an execution subsystem configured to receive multi-stage control instructions sent from the decision subsystem, and control the underwater leveling machine to work underwater based on the multi-stage control instructions; wherein the fuzzy swarm intelligence optimization unit is configured to: for each tank, calculate the antecedent variables of fuzzy reasoning; wherein the antecedent variables include correction alignment, force arm efficiency and liquid level margin, the correction alignment is calculated based on the consistency of attitude gradient and target correction direction, the force arm efficiency is calculated based on the force arm weight of the tank to roll or pitch, and the liquid level margin is calculated based on the safety margin of the current liquid level and the limit value; 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 reasoning to obtain a participation level representing the tank; based on the participation level, it is determined whether the tank is the target tank.

2. The intelligent control system of an underwater screed according to claim 1, wherein, 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 of an underwater screed according to claim 2, wherein, The stage decision unit is also configured to switch to the next stage when the trigger conditions of reaching a fixed time length, an intermediate draft threshold or an inclination threshold are met in each stage.

4. The intelligent control system of an underwater screed according to claim 1, wherein, The fuzzy swarm intelligence optimization unit is also configured to use a fuzzy particle swarm optimization algorithm, and constraint conditions of hyperplane projection and boundary repair processing, to search for optimal decision variables in a sparse search space composed of the framework parameters of the current stage; the optimal decision variables are the liquid level change amounts of each target tank.

5. The intelligent control system for an underwater screed according to claim 1, wherein, The multiple preset constraints of the fuzzy swarm intelligence optimization unit include at least one of attitude safety envelope constraint, mass conservation constraint, tank geometry limit constraint, and pump and valve capacity constraint.

6. The intelligent control system for an underwater screed according to claim 1, wherein, The decision subsystem is also configured to calculate the opening time and action timing of the pump or valve corresponding to each target tank based on the positive and negative and size of the liquid level change amount of each target tank, combined with the target hardware parameters of the underwater leveling machine, and generate corresponding control instructions.

7. The intelligent control system of an underwater screed according to claim 1, wherein, The decision subsystem is also configured to monitor the working state data in the case that the execution subsystem executes the control instructions, and if it is detected that the working state data represents that the attitude or draft of the underwater leveling machine deviates more than a preset threshold, a local correction operation is triggered to adjust the liquid level change amount and reissue the control instructions.

8. The intelligent control system of an underwater screed according to claim 1, wherein, The decision subsystem further comprises a digital twin unit, a work suggestion unit and a work control unit; the digital twin unit is configured to perform state estimation and three-dimensional terrain modeling according to the work state data, and construct a digital twin model of the work of the screed machine; the work suggestion unit is configured to generate a work suggestion according to the work state data and the work task of the screed machine; The work control unit is configured to control the execution subsystem to perform the work task according to the work suggestion.

9. The intelligent control system of an underwater screed according to claim 8, wherein, The work task at least comprises a material distribution task; After the work suggestion unit acquires the material distribution task and the work state data, the work suggestion unit calculates the region material distribution feasibility by using a material distribution score function, and recommends the material thickness and density values by using fuzzy rules; The work suggestion unit predicts the potential construction defect region and the high-risk advancing path by using a dynamic window algorithm and a dynamic cost weighting mechanism, and generates a material distribution work suggestion comprising the material amount and the advancing direction.

10. The intelligent control system for an underwater screed according to claim 1, wherein, The intelligent control system of the underwater screed machine further comprises an underwater measurement subsystem, the underwater measurement subsystem comprising: a binocular camera device configured to acquire binocular vision images of a target water area; an edge computing device configured to acquire point clouds of the target water area; The underwater measurement subsystem is installed on the underwater screed machine, the binocular camera device is arranged at the front part of the main frame of the underwater screed machine, and the edge computing device is in communication connection with the material distribution device in the underwater screed machine.

11. The intelligent control system of an underwater screed according to claim 10, wherein, The edge computing device is configured to calibrate and correct the binocular vision images to obtain the parallax corresponding to each pixel point; The edge computing device is configured to extract the original gray information of the binocular vision images, and then enhance the binocular vision images to generate gradient information; The edge computing device is configured to calculate the matching cost according to the parallax of each pixel point in the binocular vision images and the corresponding original gray information and gradient information, construct a cost function, and calculate the parallax map of the binocular vision images by using the improved SGM algorithm. The edge computing device is configured to perform back projection on the parallax map according to the camera parameters of the binocular camera device to obtain the point clouds of the target water area. The improved SGM algorithm comprises: aggregating the cost function based on four directions, setting a dynamic adjustment strategy to control the penalty of small parallax variation and the penalty of large parallax jump, and performing confidence screening and hole filling on the aggregated parallax map.

12. The intelligent control system for an underwater screed according to claim 1, wherein, The decision subsystem further comprises a material distribution control unit; the material distribution control unit is configured to: acquire the weight of the stone in the hopper and the inclination angle of the hopper of the underwater screed machine based on the load cell and the inclinometer of the perception subsystem when the underwater screed machine is performing the leveling construction; determine the stone quality based on the weight of the stone in the hopper and the inclination angle of the hopper, calculate the real-time material mass flow rate based on the time series data of the stone quality, and acquire the current terrain elevation of the bed, compare the current terrain elevation of the bed with the pre-stored design target elevation, and calculate the terrain residual and the volume to be paved; Based on the real-time material mass flow rate, the terrain residual error, the to-be-paved volume, and a predicted thickness distribution, a material control instruction is generated by an optimization decision algorithm; wherein the predicted thickness distribution is obtained by a target prediction model based on historical construction data and multi-modal sensing information, and the material control instruction is used to adjust at least one of a conveying belt speed, a material valve opening degree, or a material head moving speed of the underwater screed.

13. The intelligent control system for an underwater screed according to claim 12, wherein, The optimization decision algorithm is constructed based on a SAC deep reinforcement learning algorithm; a state space of an agent of the SAC deep reinforcement learning algorithm includes a multi-modal fusion state tensor composed of the stone mass, the real-time material mass flow rate, the encoded terrain residual error, a material head pose, a material head speed, and a working condition parameter; an action space of the agent of the deep reinforcement learning framework is continuous and bounded, and the action space corresponds to control instructions of the conveying belt speed, the material valve opening degree, or the material head moving speed; the agent of the deep reinforcement learning framework trains an optimized paving strategy by maximizing a comprehensive reward function.

14. The intelligent control system for an underwater screed according to claim 1, wherein, The perception subsystem further includes multi-source surveying sensors deployed in the working sea area, for collecting multi-source sensing data.

15. The intelligent control system for an underwater screed according to claim 14, wherein, The decision subsystem further includes a flatness evaluation unit, in which a trained target base surface prediction model is configured, for generating a multi-source fusion base surface model based on the multi-source sensing data, taking the multi-source fusion base surface model as an input of the target base surface prediction model, to output an elevation deviation distribution map of the base surface based on the target base surface prediction model, and generate a flatness evaluation index of the working sea area based on the elevation deviation distribution map; wherein the target base surface prediction model is obtained by training an initial surface prediction model constructed by a convolutional neural network and a multi-head attention mechanism, the hyperparameters of the initial surface prediction model are optimized by an alpha evolution algorithm, and the loss function of the initial surface prediction model is a Huber function; the value of each grid point in the elevation deviation distribution map represents the deviation amount of the base surface at this point relative to the ideal design plane or the target elevation.

16. A method of intelligent control of an underwater screed, characterized by The method comprises: acquiring working state data of the underwater screed; discretizing the working process of the underwater screed into multiple stages, and determining a framework parameter of a current stage based on the working state data in each stage; under the condition of satisfying multiple preset constraints, determining a liquid level change amount of each target compartment based on the framework parameter, to generate a control instruction of the current stage based on the liquid level change amount; wherein the target compartment is a compartment that needs to perform a liquid level adjustment operation in the current stage; controlling the underwater screed to perform underwater work according to the multi-stage control instruction; wherein, under the condition of satisfying multiple preset constraints, determining a liquid level change amount of each target compartment based on the framework parameter, to generate a control instruction of the current stage based on the liquid level change amount, further comprises: For each cabin, calculate the antecedent variables of fuzzy reasoning; wherein, the antecedent variables include correction alignment, force arm efficiency and liquid level margin, the correction alignment is calculated based on the consistency of the attitude gradient and the target correction direction, the force arm efficiency is calculated based on the force arm weight of the cabin to the roll or pitch, and the liquid level margin is calculated based on the safety margin of the current liquid level and the limit value; 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 reasoning to obtain a participation level representing the cabin; and it is determined whether the cabin is the target cabin based on the participation level.

17. A computer device, comprising: The computer device is arranged on or in communication connection with the underwater screed machine, and the computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method according to claim 16 when executing the computer program.

18. An underwater screed, characterized by The underwater screed machine is provided with the intelligent control system of the underwater screed machine according to any one of claims 1-15.

19. The underwater screed of claim 18, wherein, The underwater screed machine is a walking type underwater screed machine.

Images