Intelligent control method and system for a trailing suction dredger

Abstract

The application discloses a kind of intelligent control method and system of drag suction dredger, comprising the following steps: S1-the establishment of rabble lip data model: by training DDPG algorithm framework, rabble lip data model is established;S2-determination of theoretical best working condition point: rabble lip data model is established by the step S1, and mud pump flow-mud pump lift (pipeline loss) curve is obtained by fitting, and mud pump flow-mud pump lift (pipeline loss) curve is obtained by mud pump flow-mud pump lift (pipeline loss) curve, and the highest point of mud pump flow-mud pump yield curve is theoretical best working condition point;S3-control of dredging tool: control the working condition point of dredging tool to move to theoretical best working condition point.Compared with prior art, the present application can realize the unmanned intelligent control of rabble lip in dredging process, improve the instantaneous yield of drag suction dredger, and improve the efficiency of dredging operation.

Description

Technical Field

[0001] This invention belongs to the field of dredging engineering technology, specifically relating to an intelligent control method and system for trailing suction hopper dredgers. Background Technology

[0002] With the development of automation and intelligent technologies, dredging operations are gradually moving from intelligent operation to unmanned operation. Compared to the level of intelligence in the dredging process, the intelligent control of the dredging vessel's navigation is relatively weak. Existing vessels mostly use manual operation to control the up and down movement of the dredging lip, resulting in low dredging efficiency. Furthermore, due to the diverse soil conditions in the working range of dredging vessels, working on low-density soil can lead to a large suction vacuum on the dredging lip, easily causing cavitation and damage. On high-density soil, the load on the dredging lip is too high, which may also damage it, resulting in low dredging output. Therefore, how to intelligently control the dredging lip to ensure it operates at the appropriate working point is a problem that needs to be solved. Summary of the Invention

[0003] To address the issues of existing dredgers' inability to achieve intelligent control of the dredger's rake lip, leading to rake lip damage and low dredging output, the following solution is provided:

[0004] A method for intelligent control of a trailing suction hopper dredger includes the following steps:

[0005] Establishment of the S1-Rake Lip Data Model:

[0006] A data model for rake lips is established by training the DDPG algorithm architecture;

[0007] Determination of the theoretical optimal operating point for S2:

[0008] The mud pump flow rate-mud pump head (pipeline loss) curve is obtained by fitting the rake lip data model established in step S1, and the mud pump flow rate-mud production curve is obtained by using the mud pump flow rate-mud pump head (pipeline loss) curve. The highest point of the mud pump flow rate-mud production curve is the theoretical optimal operating point.

[0009] S3 - Control of dredging equipment:

[0010] The mud pump on the trailing suction dredger is controlled with the mud pump flow rate corresponding to the theoretical optimal operating point as the target flow rate, and the ground angle of the rake lip is controlled with the mud density corresponding to the theoretical optimal operating point as the target, so that the operating point of the dredging equipment moves towards the theoretical optimal operating point.

[0011] Furthermore, step 1 includes the following steps:

[0012] S1.1 - Acquisition of Ship Data:

[0013] Data from the ship's sensor system was collected, including data on dredging depth, pump speed, mud pump flow rate, rake head hood angle to the ground, and instantaneous undisturbed soil production, to serve as a historical dataset.

[0014] Building the S1.2-DDPG algorithm architecture:

[0015] Historical datasets are used as the environment dataset for the DDPG algorithm;

[0016] The dredging depth, pump speed, and mud pump flow rate in the environmental dataset are used as the state S in the DDPG algorithm;

[0017] The ground-to-ground angle of the rake head's active cover in the environmental dataset is used as action A in the DDPG algorithm;

[0018] The instantaneous undisturbed soil yield is used as the reward R in the DDPG algorithm;

[0019] S1.3 - Model Training:

[0020] The DDPG algorithm trains a rake lip data model by interacting with environmental datasets.

[0021] Furthermore, step S2 includes the following steps:

[0022] S2.1 - Definition of operating point:

[0023] When the dredging depth of the dredger's rake head and the pumping speed of the mud pump are constant, the intersection of the mud pump head curve and the pipeline loss curve at this point is defined as the operating point.

[0024] S2.2 - Determination of working conditions for different soil densities:

[0025] By using the rake lip data model, it traverses different mud densities to obtain several mud pump head curves and pipeline loss curves corresponding to different mud densities. Several operating points are determined based on the intersection of several mud pump head curves and pipeline loss curves.

[0026] S2.3 Determination of the flow-density curve:

[0027] By fitting all operating points, the mud pump flow rate-mud density curve is obtained when the pump speed and dredging depth are constant.

[0028] Conversion of S2.4-mud pump flow rate-mud production curve:

[0029] The mud pump flow rate-mud density curve is converted into a mud pump flow rate-mud production curve by assuming that mud production is equal to mud pump flow rate multiplied by mud density.

[0030] S2.5 - Determination of target flow rate and target density:

[0031] The highest point of the mud pump flow rate-mud production curve is the optimal operating point. The mud pump flow rate corresponding to the optimal operating point is the target mud pump flow rate, and the mud density corresponding to the optimal operating point is the target density.

[0032] Furthermore, step S3 includes the following steps:

[0033] S3.1 - Control the mud pump flow rate to achieve the target flow rate:

[0034] By controlling the speed of the mud pump, the mud pump flow rate can be shifted towards the target flow rate.

[0035] S3.2 - Control the intake density of the rake lip to achieve the target density:

[0036] The Q-learning algorithm is used to control the vacuum degree of the rake lip intake, so that the intake density of the rake lip moves towards the target density corresponding to the optimal working point.

[0037] Furthermore, step S3.2 also includes the following steps:

[0038] S3.2.1 — Vacuum range setting:

[0039] The vacuum range of the rake lip is set according to the soil conditions and the characteristics of the mud pump on the actual vessel, and the minimum vacuum value is set to vacuum. low The highest vacuum value is vacuum. high The current hollowness of the rake lip is vacuum;

[0040] Partitioning of the state space s in the S3.2.2-Q-learning algorithm:

[0041] The state space is divided into [s1, s2, s3] based on the vacuum value;

[0042] If vacuum low ≤vacuum≤vacuum high If , then s∈s1;

[0043] If vacuum ≥ vacuum high If , then s∈s2;

[0044] If vacuum ≤ vacuum low If , then s∈s3;

[0045] Attribute settings for action space 'a' in the S3.2.3-Q-learning algorithm:

[0046] The motion space a = [rake lip rises, rake lip remains still, rake lip descends];

[0047] If s∈s1, then a∈rlip is stationary;

[0048] If s∈s2, then a∈rlip rises;

[0049] If s∈s3, then a∈rlip descent;

[0050] Attribute settings for the reward value r in the S3.2.4-Q-learning algorithm:

[0051] The reward value for the Q-learning algorithm is r = -vacuum;

[0052] S3.2.5-Q-learning algorithm expected update strategy settings:

[0053] The expected value q of the Q-learning algorithm:

[0054] Furthermore, the construction of the DDPG algorithm architecture in step S1.2 also includes the following steps:

[0055] S1.2.1 - Construction of the lip-raking motion space:

[0056] The motion space of the rake lip is constructed by using the mechanical and hydraulic control principles of the rake lip.

[0057] S1.2.2 - Building the Neural Network:

[0058] Construct the current Action network based on the policy Poily, where the policy function in the current Action network has a parameter of ω1;

[0059] Construct a policy-based Poily target action network, where the policy function of the target action network has a parameter of ω2;

[0060] Construct the current Critic network based on value, where the parameter of the value function of the current Critic network is θ1;

[0061] Construct a value-based target Critic network, where the parameter of the value function of the target Critic network is θ2;

[0062] Build an experience pool to store the interaction data of the current Action network.

[0063] Furthermore, the S1.3 model training includes the following steps:

[0064] S1.3.1 - Initialization of the Neural Network:

[0065] Initialize the current Action network and the current Critic network, and copy the initialization parameters to the target Action network and the target Critic network to complete the initialization of the neural network;

[0066] S1.3.2 - Current Action Network Environment Interaction

[0067] The current Action network is input with the dredging depth, pump speed, and mud pump flow rate from the environmental dataset as state S. The current Action network determines action A in the rake lip action space through a policy function with ω1 as a parameter.

[0068] Extract the digging depth, pump speed, and mud pump flow rate corresponding to action A from the environmental dataset as the state S' for the next time step, and use the instantaneous undisturbed soil yield as the reward R for the DDPG algorithm.

[0069] Using state S, action A, reward R, and the next state S' as samples, the samples are stored in the experience pool;

[0070] S1.3.3 - Current output of the Critic network

[0071] The current Critic network takes state S and action A as input, and outputs the current expected value Q1 corresponding to state S and action A through a value function with θ1 as parameter;

[0072] S1.3.4 - Output of the target Action network:

[0073] The target Action network takes the next time step state S' as input and outputs the next time step action A' through a policy function with ω2 as a parameter;

[0074] S1.3.5 - The operation of the target Critic network:

[0075] The target Critic network takes the next time step state S' and the next time step action A' as inputs, and outputs the corresponding target expected value Q2 through a value function with θ2 as a parameter;

[0076] S1.3.6 - Current Critic network update:

[0077] The current expected value Q1 is propagated backward to the target expected value Q2 using gradient descent, and a new θ1' is generated along the direction of the maximum gradient descent from the current expected value Q1 to the target expected value Q2. The new θ1' is then used to replace θ1, thus updating the parameters of the current Critic network value function.

[0078] S1.3.7 - Current Action network update:

[0079] Update the policy parameter ω1 in the current Action network along the direction of the maximum gradient ascent of the current expected value Q1 in state S;

[0080] S1.3.8 - Updates to the Target Action Network and Target Critic Network

[0081] Set an interval time t. Every interval time t, use the policy parameters ω1' of the current Action network and the value parameters θ1' of the current Critic network to update and iterate the parameters ω2 and θ2 of the target Action network and the target Critic network through a soft update method.

[0082] An intelligent control system for a trailing suction hopper dredger includes:

[0083] Data acquisition module: used to collect the working parameters of ship dredging equipment and store the collected working parameters as historical datasets;

[0084] Model building module: Used to train the DDPG algorithm architecture based on the historical dataset of the data acquisition module, and to obtain the working data model of the dredging equipment;

[0085] Data processing module: Used to fit and determine the optimal operating point based on the working data model obtained from the model building module;

[0086] Intelligent control module: Used to control the rake lip and mud pump to move towards the optimal operating point based on the optimal operating point obtained by the data processing module.

[0087] Furthermore, the data acquisition module includes, but is not limited to, a digging depth sensor and a ground angle sensor installed on the rake head, and a pump speed sensor, a mud pump flow sensor, and a mud density sensor installed in the mud pump.

[0088] The data acquisition module collects data on digging depth, pump speed, mud pump flow rate, the angle of the rake head's movable cover to the ground, and mud density, and stores the collected data as a historical dataset.

[0089] Furthermore, the DDPG algorithm architecture in the model building module includes:

[0090] Experience pool unit: used to store feedback information on the operating status of dredging equipment after the execution of action commands;

[0091] Environmental data unit: Used to store the data collected by the data acquisition module, output the state and reward corresponding to different actions, and then verify and correct the neural network;

[0092] Current Action Network Unit: Used to interact with the environment dataset, generating the corresponding current action based on the current state as input;

[0093] Current Critic network unit: used to derive the expected value corresponding to the current action based on the current state and the current action;

[0094] Target Action Network Unit: Selects the action for the next time step based on the state of the next time step, and interacts with the Target Critic Network Unit;

[0095] Target Critic network unit: Calculates the evaluation value of the next time step state and the next time step optimal action based on the next time step state.

[0096] Compared with existing technologies, the intelligent control method for the drag suction hopper dredger's drag lip can achieve unmanned intelligent control of the drag lip during the dredging process, increase the instantaneous output of the drag suction hopper dredger, and improve the efficiency of dredging operations. Attached Figure Description

[0097] Figure 1 This is the mud pump flow rate-mud pump head (pipeline loss) curve in the intelligent control method for trailing suction hopper dredgers of this invention.

[0098] Figure 2 This is the mud pump flow rate-mud production curve in the intelligent control method for trailing suction hopper dredgers of the present invention.

[0099] Figure 3 This is the vacuum density time-history curve in the intelligent control method for trailing suction hopper dredgers of the present invention.

[0100] Figure 4 This is a flowchart illustrating the training process of the rake lip data model in the intelligent control method for trailing suction hopper dredgers of the present invention.

[0101] Figure 5 This is a schematic diagram of the overall structural modules of the intelligent control system for the trailing suction hopper dredger of the present invention.

[0102] Figure 6 This is a schematic diagram of the data acquisition module in the intelligent control system of the trailing suction hopper dredger of the present invention.

[0103] Figure 7 This is a schematic diagram of the model building module in the intelligent control system of the trailing suction hopper dredger of the present invention.

[0104] In the picture:

[0105] 1-Data acquisition module, 2-Model building module, 3-Data processing module, 4-Intelligent control module;

[0106] 11-Historical dataset, 12-Dig depth sensor, 13-Ground angle sensor of the rake head movable cover, 14-Pump speed sensor;

[0107] 15-Mud pump flow sensor; 16-Mud density sensor;

[0108] 21-Working data model, 22-Experience pool unit, 23-Environmental data unit, 24-Current Action network unit;

[0109] 25 - Current Critic network unit, 26 - Target Action network unit, 27 - Target Critic network unit;

[0110] 31-Optimal operating point, 41-rake lip, 42-mud pump. Detailed Implementation

[0111] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and specific embodiments. It should be understood that the described embodiments are only for illustrating the present invention and are not intended to limit the scope of the present invention.

[0112] Example 1: An intelligent control method for a trailing suction hopper dredger, comprising the following steps:

[0113] Establishment of the S1-Rake Lip Data Model:

[0114] A data model for rake lips is established by training the DDPG algorithm architecture;

[0115] Determination of the theoretical optimal operating point for S2:

[0116] The mud pump flow rate-mud pump head (pipeline loss) curve is obtained by fitting the rake lip data model established in step S1, and the mud pump flow rate-mud production curve is obtained by using the mud pump flow rate-mud pump head (pipeline loss) curve. The highest point of the mud pump flow rate-mud production curve is the theoretical optimal operating point.

[0117] S3 - Control of dredging equipment:

[0118] The mud pump on the trailing suction dredger is controlled with the mud pump flow rate corresponding to the theoretical optimal operating point as the target flow rate, and the ground angle of the rake lip is controlled with the mud density corresponding to the theoretical optimal operating point as the target, so that the operating point of the dredging equipment moves towards the theoretical optimal operating point.

[0119] Furthermore, step 1 includes the following steps:

[0120] S1.1 - Acquisition of Ship Data:

[0121] Data from the ship's sensor system was collected, including data on dredging depth, pump speed, mud pump flow rate, rake head hood angle to the ground, and instantaneous undisturbed soil production, to serve as a historical dataset.

[0122] Building the S1.2-DDPG algorithm architecture:

[0123] Historical datasets are used as the environment dataset for the DDPG algorithm;

[0124] The dredging depth, pump speed, and mud pump flow rate in the environmental dataset are used as the state S in the DDPG algorithm;

[0125] The ground-to-ground angle of the rake head's active cover in the environmental dataset is used as action A in the DDPG algorithm;

[0126] The instantaneous undisturbed soil yield is used as the reward R in the DDPG algorithm;

[0127] S1.3 - Model Training:

[0128] The DDPG algorithm trains a rake lip data model by interacting with environmental datasets.

[0129] Furthermore, step S2 includes the following steps:

[0130] S2.1 - Definition of operating point:

[0131] When the dredging depth of the dredger's rake head and the pumping speed of the mud pump are constant, the intersection of the mud pump head curve and the pipeline loss curve at this point is defined as the operating point.

[0132] S2.2 - Determination of working conditions for different soil densities:

[0133] By using the rake lip data model to traverse different mud densities, the following results were obtained: Figure 1 The diagram shows several pump head curves and pipeline loss curves corresponding to different mud densities. Several operating points are determined based on the intersection of these pump head curves and pipeline loss curves.

[0134] S2.3 Determination of the flow-density curve:

[0135] By fitting all operating points, the mud pump flow rate-mud density curve is obtained when the pump speed and dredging depth are constant.

[0136] Conversion of S2.4-mud pump flow rate-mud production curve:

[0137] The mud production rate is equal to the mud pump flow rate multiplied by the mud density. The mud pump flow rate-mud density curve is then converted as follows: Figure 2 The mud pump flow rate-mud production curve shown below;

[0138] S2.5 - Determination of target flow rate and target density:

[0139] The highest point of the mud pump flow rate-mud production curve is the optimal operating point. The mud pump flow rate corresponding to the optimal operating point is the target mud pump flow rate, and the mud density corresponding to the optimal operating point is the target density.

[0140] Furthermore, step S3 includes the following steps:

[0141] S3.1 - Control the mud pump flow rate to achieve the target flow rate:

[0142] By controlling the speed of the mud pump, the mud pump flow rate can be shifted towards the target flow rate. It should be noted that:

[0143] Since the flow rate of the mud pump is equal to the mud flow velocity multiplied by the pipe cross-sectional area, and the pipe diameter on the ship is a fixed value, and the mud flow velocity is positively correlated with the speed of the mud pump, the mud pump flow rate can be shifted towards the target flow rate by controlling the speed of the mud pump.

[0144] S3.2 - Control the intake density of the rake lip to achieve the target density:

[0145] The Q-learning algorithm is used to control the vacuum degree of the rake lip intake, so that the intake density of the rake lip moves towards the target density corresponding to the optimal working point;

[0146] It should be noted that using the vacuum level of the rake lip as the basis for control is feasible:

[0147] The goal of intelligent control of the rake lips is to ensure that the suction density of the rake lips equals the target density. Because the mud density meter is installed near the mud pump, it lags behind the suction density of the rake lips. Figure 3 The curve showing the relationship between mud density and vacuum degree indicates a strong correlation between mud density and the vacuum degree sucked in by the rake head. The higher the vacuum, the higher the density sucked in by the rake head, and the vacuum response is rapid. Therefore, it is feasible to use the vacuum degree sucked in by the rake lip as the basis for control.

[0148] Furthermore, step S3.2 also includes the following steps:

[0149] S3.2.1 — Vacuum range setting:

[0150] The vacuum range of the rake lip is set according to the soil conditions and the characteristics of the mud pump on the actual vessel, and the minimum vacuum value is set to vacuum. low The highest vacuum value is vacuum. high The current hollowness of the rake lip is vacuum;

[0151] Partitioning of the state space s in the S3.2.2-Q-learning algorithm:

[0152] The state space is divided into [s1, s2, s3] based on the vacuum value;

[0153] If vacuum low ≤vacuum≤vacuum high If , then s∈s1;

[0154] If vacuum ≥ vacuum high If , then s∈s2;

[0155] If uacuum ≤ vacum low If , then s∈s3;

[0156] Attribute settings for action space 'a' in the S3.2.3-Q-learning algorithm:

[0157] The motion space a = [rake lip rises, rake lip remains still, rake lip descends];

[0158] If s∈s1, then a∈rlip is stationary;

[0159] If s∈s2, then a∈rlip rises;

[0160] If s∈s3, then a∈rlip descent;

[0161] Attribute settings for the reward value r in the S3.2.4-Q-learning algorithm:

[0162] The reward value for the Q-learning algorithm is r = -vacuum;

[0163] S3.2.5-Q-learning algorithm expected update strategy settings:

[0164] The expected value q of the Q-learning algorithm: In the formula, q is the expected value of the Q-learning algorithm, r is the reward value of the Q-learning algorithm, s is the state of the Q-learning algorithm, s' is the state value at the next time step, a is the action corresponding to s, and a' is the action at the next time step corresponding to s'. The control process of the Q-learning algorithm is already disclosed in the prior art and will not be described in detail here.

[0165] Furthermore, the construction of the DDPG algorithm architecture in step S1.2 also includes the following steps:

[0166] S1.2.1 - Construction of the lip-raking motion space:

[0167] The motion space of the rake lip is constructed by using the mechanical and hydraulic control principles of the rake lip.

[0168] S1.2.2 - Building the Neural Network:

[0169] Construct the current Action network based on the policy Poily, where the policy function in the current Action network has a parameter of ω1;

[0170] Construct a policy-based Poily target action network, where the policy function of the target action network has a parameter of ω2;

[0171] Construct the current Critic network based on value, where the parameter of the value function of the current Critic network is θ1;

[0172] Construct a value-based target Critic network, where the parameter of the value function of the target Critic network is θ2;

[0173] Build an experience pool to store the interaction data of the current Action network;

[0174] Furthermore, such as Figure 4 As shown, the S1.3 model training includes the following steps:

[0175] S1.3.1 - Initialization of the Neural Network:

[0176] Initialize the current Action network and the current Critic network, and copy the initialization parameters to the target Action network and the target Critic network to complete the initialization of the neural network;

[0177] S1.3.2 - Current Action Network Environment Interaction:

[0178] The current Action network is input with the dredging depth, pump speed, and mud pump flow rate from the environmental dataset as state S. The current Action network determines action A in the rake lip action space through a policy function with ω1 as a parameter.

[0179] Extract the digging depth, pump speed, and mud pump flow rate corresponding to action A from the environmental dataset as the state S' for the next time step, and use the instantaneous undisturbed soil yield as the reward R for the DDPG algorithm.

[0180] In the above formula: Vsoil represents the instantaneous undisturbed soil production, ρ represents the instantaneous mud density during the dredging process, ρw represents the seawater density, ρs represents the dry soil density, and F represents the mud flow rate.

[0181] Using state S, action A, reward R, and the next state S' as samples, the samples are stored in the experience pool;

[0182] S1.3.3 - Current output of the Critic network:

[0183] The current Critic network takes state S and action A as input, and outputs the current expected value Q1 corresponding to state S and action A through a value function with θ1 as parameter;

[0184] S1.3.4 - Output of the target Action network:

[0185] The target Action network takes the next time step state S' as input and outputs the next time step action A' through a policy function with ω2 as a parameter;

[0186] S1.3.5 - The operation of the target Critic network:

[0187] The target Critic network takes the next time step state S' and the next time step action A' as inputs, and outputs the corresponding target expected value Q2 through a value function with θ2 as a parameter;

[0188] S1.3.6 - Current Critic network update:

[0189] The current expected value Q1 is propagated backward to the target expected value Q2 using gradient descent, and a new θ1' is generated along the direction of the maximum gradient descent from the current expected value Q1 to the target expected value Q2. The new θ1' is then used to replace θ1, thus updating the parameters of the current Critic network value function.

[0190] S1.3.7 - Current Action network update:

[0191] Update the policy parameter ω1 in the current Action network along the direction of the maximum gradient ascent of the current expected value Q1 in state S;

[0192] S1.3.8 - Updates to the Target Action Network and Target Critic Network

[0193] Set an interval time t. Every interval time t, use the policy parameters ω1' of the current Action network and the value parameters θ1' of the current Critic network to update and iterate the parameters ω2 and θ2 of the target Action network and the target Critic network through a soft update method.

[0194] The soft update refers to introducing a learning rate τ, taking a weighted average of the old target network parameters and the new corresponding network parameters, and then assigning the values ​​to the target Action network and the target Critic network.

[0195] Example 2: An intelligent control system for a trailing suction hopper dredger, such as Figure 5 As shown, it includes:

[0196] Data acquisition module 1: used to collect the working parameters of ship dredging equipment and store the collected working parameters as historical dataset 11;

[0197] Model building module 2: Used to train the DDPG algorithm architecture based on the historical dataset 11 of data acquisition module 1, and obtain the working data model 21 of dredging equipment;

[0198] Data processing module 3: Used to fit and determine the optimal operating point 31 based on the working data model 21 obtained from model building module 2;

[0199] Intelligent control module 4: Used to control the rake lip 41 and mud pump 42 to move towards the optimal working point based on the optimal working point 31 obtained by data processing module 3.

[0200] Furthermore, in combination Figure 5 and Figure 6 The data acquisition module 1 includes, but is not limited to, a digging depth sensor 12 and a ground angle sensor 13 installed on the rake head, a pump speed sensor 14, a mud pump flow sensor 15 and a mud density sensor 16 installed in the mud pump.

[0201] The data acquisition module 1 collects data on digging depth, pump speed, mud pump flow rate, the ground angle of the rake head movable cover, and mud density, and stores the collected data as a historical dataset 11.

[0202] Furthermore, the DDPG algorithm architecture in the model building module 2 includes:

[0203] Experience pool unit 22: Used to store the operating status feedback information of the dredging equipment after the execution of the action command;

[0204] Environmental data unit 23: Used to store the data collected by the data acquisition module, output the state and reward corresponding to different actions, and then verify and correct the neural network;

[0205] Current Action Network Unit 24: Used to interact with the environment dataset and generate the corresponding current action with the current state as input;

[0206] Current Critic network unit 25: used to derive the expected value corresponding to the current action based on the current state and the current action;

[0207] Target Action Network Unit 26: Selects the action for the next moment based on the state of the next moment and interacts with Target Critic Network Unit 27;

[0208] Target Critic network unit 27: Calculates the evaluation values ​​of the next time step state and the next time step optimal action based on the next time step state.

[0209] Example 3: Application Example

[0210] Engineering examples based on the method of Embodiment 1 above:

[0211] For soils with a dry density of 1.70-1.80 t / m³ 3 The density of seawater is 1.05 t / m³. 3In the dredging operation area, the main dredging equipment of the vessel was controlled according to the intelligent control method of the trailing suction hopper dredger provided in Example 1. First, environmental data was collected through sensors on the vessel to establish an environmental dataset. Then, the DDPG algorithm was trained based on the environmental dataset to obtain the working model of the rake lip. The optimal working point was obtained by fitting the working model of the rake lip. Finally, the Q-learning algorithm was used to control the suction vacuum degree of the rake lip to 1.0≤vacuum≤0.8, so that the dredging equipment worked at the optimal working point. After data analysis, the dredging output increased by 5.53%.

[0212] The specific embodiments described above provide a further detailed explanation of the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific implementation examples of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent control of a trailing suction hopper dredger, characterized in that, Includes the following steps: Establishment of the S1-Rake Lip Data Model: A data model for rake lips is established by training the DDPG algorithm architecture; Determination of the theoretical optimal operating point for S2: The mud pump flow rate-mud density curve is obtained by fitting the rake lip data model established in step S1, and the mud pump flow rate-mud production curve is obtained by the mud pump flow rate-mud density curve. The highest point of the mud pump flow rate-mud production curve is the theoretical optimal operating point. S2.1 - Definition of operating point: When the dredging depth of the dredger's rake head and the pumping speed of the mud pump are constant, the intersection of the mud pump head curve and the pipeline loss curve at this point is defined as the operating point. S2.2 - Determination of working conditions for different soil densities: By using the rake lip data model, it traverses different mud densities to obtain several mud pump head curves and pipeline loss curves corresponding to different mud densities. Several operating points are determined based on the intersection of several mud pump head curves and pipeline loss curves. S2.3 Determination of the flow-density curve: By fitting all operating points, the mud pump flow rate-mud density curve is obtained when the pump speed and dredging depth are constant. Conversion of S2.4-mud pump flow rate-mud production curve: The mud pump flow rate-mud density curve is converted into a mud pump flow rate-mud production curve by assuming that mud production is equal to mud pump flow rate multiplied by mud density. S2.5 - Determination of target flow rate and target density: The highest point of the mud pump flow rate-mud production curve is the optimal operating point. The mud pump flow rate corresponding to the optimal operating point is the target mud pump flow rate, and the mud density corresponding to the optimal operating point is the target density. S3 - Control of dredging equipment: The mud pump on the trailing suction dredger is controlled with the mud pump flow rate corresponding to the theoretical optimal operating point as the target flow rate, and the ground angle of the rake lip is controlled with the mud density corresponding to the theoretical optimal operating point as the target, so that the operating point of the dredging equipment moves towards the theoretical optimal operating point.

2. The intelligent control method for a trailing suction hopper dredger according to claim 1, characterized in that, Step S1 includes the following steps: S1.1 - Acquisition of Ship Data: Data from the ship's sensor system was collected, including data on dredging depth, pump speed, mud pump flow rate, rake head hood angle to the ground, and instantaneous undisturbed soil production, to serve as a historical dataset. Building the S1.2-DDPG algorithm architecture: Historical datasets are used as the environment dataset for the DDPG algorithm; The dredging depth, pump speed, and mud pump flow rate in the environmental dataset are used as the state S in the DDPG algorithm; The ground-to-ground angle of the rake head's active cover in the environmental dataset is used as action A in the DDPG algorithm; The instantaneous undisturbed soil yield is used as the reward R in the DDPG algorithm; S1.3 - Model Training: The DDPG algorithm trains a rake lip data model by interacting with environmental datasets.

3. The intelligent control method for a trailing suction hopper dredger according to claim 1, characterized in that, Step S3 includes the following steps: S3.1 - Control the mud pump flow rate to achieve the target flow rate: By controlling the speed of the mud pump, the mud pump flow rate can be shifted towards the target flow rate. S3.2 - Control the intake density of the rake lip to achieve the target density: The Q-learning algorithm is used to control the vacuum degree of the rake lip intake, so that the intake density of the rake lip moves towards the target density corresponding to the optimal working point.

4. The intelligent control method for a trailing suction hopper dredger according to claim 3, characterized in that, Step S3.2 further includes the following steps: S3.2.1 — Vacuum range setting: The vacuum range of the rake lip is set according to the soil conditions and the characteristics of the mud pump on the actual vessel, and the minimum vacuum value is set to... The highest vacuum value is The current hollowness of the rake lip is ; Partitioning of the state space s in the S3.2.2-Q-learning algorithm: State space is divided according to vacuum value ; like ,but ; like ,but ; like ,but ; Attribute settings for action space 'a' in the S3.2.3-Q-learning algorithm: Action space ; like ,but ; like ,but ; like ;but ; Attribute settings for the reward value r in the S3.2.4-Q-learning algorithm: Q-learning algorithm reward value ; S3.2.5-Q-learning algorithm expected update strategy settings: The expected value q of the Q-learning algorithm: , where s' is the state value at the next moment, and a' is the action at the next moment corresponding to s'.

5. The intelligent control method for a trailing suction hopper dredger according to claim 2, characterized in that, The construction of the DDPG algorithm architecture in step S1.2 also includes the following steps: S1.2.1 - Construction of the lip-raking motion space: The motion space of the rake lip is constructed using the mechanical and hydraulic control principles of the rake lip. S1.2.2 - Building the Neural Network: Construct the current Action network based on the policy Poily, where the policy function in the current Action network has a parameter of ω1; Construct a policy-based Poily target action network, where the policy function of the target action network has a parameter of ω2; Construct the current Critic network based on value, where the parameter of the value function of the current Critic network is θ1; Construct a value-based target Critic network, where the parameter of the value function of the target Critic network is θ2; Build an experience pool to store the interaction data of the current Action network.

6. The intelligent control method for a trailing suction hopper dredger according to claim 5, characterized in that, The S1.3 model training includes the following steps: S1.3.1 - Initialization of the Neural Network: Initialize the current Action network and the current Critic network, and copy the initialization parameters to the target Action network and the target Critic network to complete the initialization of the neural network; S1.3.2 - Current Action Network Environment Interaction The current Action network is input with the dredging depth, pump speed, and mud pump flow rate from the environmental dataset as state S. The current Action network determines action A in the rake lip action space through a policy function with ω1 as a parameter. Extract the digging depth, pump speed, and mud pump flow rate corresponding to action A from the environmental dataset as the state S' for the next time step, and use the instantaneous undisturbed soil yield as the reward R for the DDPG algorithm. ;in, Vsoil represents the instantaneous undisturbed soil yield, ρ represents the instantaneous mud density during the dredging process, ρw represents the seawater density, ρs represents the dry soil density, and F represents the mud flow rate. Using state S, action A, reward R, and the next state S' as samples, the samples are stored in the experience pool; S1.3.3 - Current output of the Critic network The current Critic network takes state S and action A as input, and outputs the current expected value Q1 corresponding to state S and action A through a value function with θ1 as parameter; S1.3.4 - Output of the target Action network: The target Action network takes the next time step state S' as input and outputs the next time step action A' through a policy function with ω2 as a parameter; S1.3.5 - The operation of the target Critic network: The target Critic network takes the next time step state S' and the next time step action A' as inputs, and outputs the corresponding target expected value Q2 through a value function with θ2 as a parameter; S1.3.6 - Current Critic network update: The current expected value Q1 is propagated backward to the target expected value Q2 using gradient descent, and a new θ1' is generated along the direction of the maximum gradient descent from the current expected value Q1 to the target expected value Q2. The new θ1' is then used to replace θ1, thus updating the parameters of the current Critic network value function. S1.3.7 - Current Action network update: Update the policy parameter ω1 in the current Action network along the direction of the maximum gradient ascent of the current expected value Q1 in state S; S1.3.8 - Updates to the Target Action Network and Target Critic Network Set an interval time t. Every interval time t, use the policy parameters ω1' of the current Action network and the value parameters θ1' of the current Critic network to update and iterate the parameters ω2 and θ2 of the target Action network and the target Critic network through a soft update method.

7. An intelligent control system for a trailing suction hopper dredger, characterized in that, The intelligent control system is applicable to the intelligent control method according to any one of claims 1-6, comprising: Data acquisition module: used to collect the working parameters of ship dredging equipment and store the collected working parameters as historical datasets; Model building module: Used to train the DDPG algorithm architecture based on the historical dataset of the data acquisition module, and to obtain the working data model of the dredging equipment; Data processing module: Used to fit and determine the optimal operating point based on the working data model obtained from the model building module; Intelligent control module: Used to control the rake lip and mud pump to move towards the optimal operating point based on the optimal operating point obtained by the data processing module.

8. The intelligent control system for a trailing suction hopper dredger according to claim 7, characterized in that, The data acquisition module includes a digging depth sensor and a ground angle sensor for the rake head movable cover, and a pump speed sensor, a mud pump flow sensor, and a mud density sensor installed in the mud pump. The data acquisition module collects data on digging depth, pump speed, mud pump flow rate, the angle of the rake head's movable cover to the ground, and mud density, and stores the collected data as a historical dataset.

9. The intelligent control system for a trailing suction hopper dredger according to claim 7, characterized in that, The DDPG algorithm architecture in the model building module includes: Experience pool unit: used to store feedback information on the operating status of dredging equipment after the execution of action commands; Environmental data unit: Used to store the data collected by the data acquisition module, output the state and reward corresponding to different actions, and then verify and correct the neural network; Current Action Network Unit: Used to interact with the environment dataset, generating the corresponding current action based on the current state as input; The current Critic network unit is used to derive the expected value corresponding to the current action based on the current state and the current action. The target Action network unit selects the action for the next time step based on the state of the next time step and interacts with the target Critic network unit. Target Critic network unit: Calculates the evaluation value of the next time step state and the next time step optimal action based on the next time step state.

Images