Intelligent control method for drag arm and drag head of drag suction dredger

Abstract

The invention belongs to the field of dredging engineering equipment, and particularly relates to an intelligent control method for a drag arm and a drag head of a drag suction dredger. Comprising the following steps: S1, benchmark testing; s2, based on the reference data, calculating the soil texture type of the current terrain by using a clustering algorithm; s3, constructing an MPC model; by using the reference data, inputting topographic changes as parameters, adding the parameters into topographic type knowledge, establishing an MPC model, and driving MPC control of the trailing suction dredger; and S4, the steps are repeated, and the optimal construction effect is achieved. According to the method, the information of the terrain and the soil texture composition of the dredging engineering site is integrated, the rake arm and the rake head are intelligently controlled on the basis of the model predictive control (MPC) theory, the motion characteristics of the rake head and the rake arm of the trailing suction dredger can be effectively controlled, and the production efficiency of the trailing suction dredger is improved.

Description

Technical Field

[0001] The invention belongs to the field of dredging engineering equipment, and in particular relates to an intelligent control method for a rake arm and a rake head of a trailing suction dredger. Background Art

[0002] A trailing suction hopper dredger is a dredging engineering vessel that can dredge mud at the dredging site, pump and load mud during the dredging process, and unload mud at the unloading site. At present, the user units of trailing suction hopper dredgers use the water depth map after construction to evaluate the construction effect, which often requires measurement by a survey ship and a long construction period to discover and correct the control deviation of the machine rake head and rake arm, and lacks real-time control of the machine rake head.

[0003] The use of artificial intelligence algorithms based on model predictive control (MPC) theory can improve the movement characteristics of the drag head and drag arm of the drag suction dredger, and achieve a certain degree of real-time depth control to keep the drag head changing with the tide level. However, due to the lack of information on terrain and soil composition, the improvement in actual mud dragging, that is, real-time production and cumulative production is limited. Summary of the invention

[0004] To this end, an intelligent terrain analysis algorithm was introduced based on the real-time terrain depth obtained by the multi-beam measurement radar installed at the bow and stern of the ship, as well as the terrain depth changes before and after construction. On the one hand, the soil composition under the trailing suction hopper dredger can be identified; on the other hand, the construction effects before and after the trailing suction hopper dredger can be evaluated in real time, and the corresponding model of the rake arm, rake head and soil-breaking effect of the trailing suction hopper dredger under certain soil conditions can be updated. The intelligent control system is introduced to form a real-time intelligent control algorithm that is sensitive to terrain and matches terrain changes.

[0005] Technical solution of the present invention:

[0006] A method for intelligently controlling a rake arm and a rake head of a trailing suction dredger comprises the following steps:

[0007] S1. Benchmarks

[0008] Perform benchmark tests to pre-set a range of drag head depths, drag head to ground angles and drag tube to ground angles to carry out construction operations on a specific route;

[0009] S2. Based on the benchmark data, the soil type of the current terrain is calculated using a clustering algorithm;

[0010] S3, constructing MPC model;

[0011] Using the above benchmark data, terrain changes are added as parameter inputs to terrain type knowledge, and an MPC model is established to drive the MPC control of the trailing suction hopper dredger.

[0012] S4. Repeat the above steps to achieve the best construction effect.

[0013] Beneficial Effects

[0014] The invention integrates the information of the topography and soil composition of the dredging project site, and proposes an intelligent control method for the rake arm and the rake head based on the model predictive control (MPC) theory, which can effectively control the movement characteristics of the rake head and the rake arm of the trailing suction hopper dredger, thereby improving the production efficiency of the trailing suction hopper dredger. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A comparison diagram of the intelligent control method of the present invention and the existing methods ((A) manual control; (B) MPC control; (C) terrain-based intelligent control);

[0016] Figure 2 This is a schematic diagram of a neural network classifier for soil classification in step S2 of the present invention;

[0017] Figure 3 This is a schematic diagram of the MPC state control of the present invention. DETAILED DESCRIPTION

[0018] The technical solution provided by the present application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of the present application will become more apparent with the following description.

[0019] like Figure 1 This is a comparison diagram between the intelligent control method of the present invention and the existing method, where Figure 1 (C) is a relationship diagram of the processing modules involved in the method of the present invention.

[0020] A method for intelligently controlling a rake arm and a rake head of a trailing suction dredger comprises the following steps:

[0021] S1. Benchmarks

[0022] Perform a benchmark test to pre-set a range of drag head depths, drag head to ground angles and drag tube to ground angles to carry out construction work on a specific route.

[0023] 1) Traverse the combinations of rake head depth, rake head angle to the ground and rake tube angle to the ground, and obtain the optimal rake head depth, rake head angle to the ground and rake tube angle to the ground under the specific soil conditions. By summarizing the data and analyzing it, the optimal combination of the optimal rake head depth, rake head angle to the ground and rake tube angle to the ground for one or more rake heads corresponding to the specific soil construction can be obtained.

[0024] 2) Carry out construction at different locations to obtain the optimal combination of the best rake head depth, rake head angle to the ground, and rake pipe angle to the ground for one or more rake heads for construction on different soil types.

[0025] S2. Based on the benchmark data, use the clustering algorithm to calculate the soil type of the current terrain

[0026] After accumulating the above benchmark data, clustering algorithm is used to infer the soil type of the current terrain.

[0027] S3. Build MPC model

[0028] Using the above benchmark data, the terrain change is added as a parameter input to the terrain type knowledge, and the MPC model is established to drive the MPC control of the trailing suction hopper dredger.

[0029] S4. Repeat the above steps to achieve the best construction effect.

[0030] During this period, if the best combination of actions still does not achieve the best construction effect, re-execute the benchmark test operation, add the benchmark data, and further improve the rake head action and soil action.

[0031] The data collection stage of step S1 is specifically as follows: during the operation of the trailing suction dredger, sensors are used to collect information such as the ship speed, the angle of the drag head movable cover, the drag head depth, the angle of the drag pipe to the ground, the terrain height difference, the instantaneous density, the instantaneous flow rate and the instantaneous vacuum.

[0032] Step S1.1: Set the sampling interval and sampling interval

[0033] The speed changes in the interval [2.0knot, 2.5knot], and the sampling interval is 0.1knot;

[0034] The angle of the drag head active cover is in the range of [0°, 30°], and the sampling interval is 5°;

[0035] The depth of the rake head is in the range of [-1.5m, 1.5m], and the sampling interval is 0.1m;

[0036] The rake tube’s angle to the ground is in the range of [30°, 60°], and the sampling interval is 5°;

[0037] The reason for selecting the above sampling intervals is that they are limited by the accuracy of the corresponding sensors. The conversion accuracy of ship navigation is 0.1 knot, the conversion accuracy of the drag head movable cover component to the drag head movable cover angle is 5°, the conversion accuracy of the drag head winch and the drag head winch movement to the drag head depth is 0.05m, and the conversion accuracy of the drag pipe winch to the drag pipe angle to the ground is 5°.

[0038] In the data collection phase, it is necessary to traverse 6 values ​​of speed change, 13 values ​​of rake head movable cover angle, 13 values ​​of rake head depth, and 7 values ​​of rake tube angle to the ground, for a total of 7098 data;

[0039] The trailing suction dredger starts construction along the preset track, and each action combination needs to be maintained for 60 seconds. The reason for selecting the above maintenance time is that the density meter and flow meter are located at the outlet of the rake pipe, and the mud or sand and gravel mixture is sucked from the rake mouth to the outlet of the rake pipe. The rake pipe including the rake head is 100 meters long, and the mud flow rate is about 2.5-5m / s. It takes about 40 seconds to obtain the density at most, and it takes about 10 seconds to obtain a stable density and obtain 1 density measurement value per second, a total of 10 density measurements, and take the average of these 10 density values;

[0040] If 720 minutes of effective test time per day is used for construction, then data collection can be completed in 10 days at most. The acquired data is taken as a test data group of one minute, among which the effective data is the data segment after the density stabilizes, and the flow rate is used as the judgment standard.

[0041] Table 1.1. Input data value range and sampling interval table

[0042]

[0043] Step S1.2, storing the acquired data into a structured historical database or a data table file in which each line of data is separated by commas, each line of data is in seconds, including time and 8 parameters: speed, drag head active cover angle, drag head drag head depth, drag pipe angle to the ground, terrain height difference, instantaneous density, instantaneous flow rate and instantaneous vacuum.

[0044] The J indicator is defined as follows:

[0045] J (speed, drag head active cover angle, drag head drag head depth, drag pipe angle to the ground) = k1*terrain height difference+k2*Sum (instantaneous density, instantaneous output) (Formula 1)

[0046] Among them, k1=1-1 / π; k2=1 / π.

[0047] Find the 100 data combinations with the largest J value and record them for the following purposes: 1) as a benchmark value for marking soil; 2) as nonlinear modeling controller data for the trailing suction dredger for this soil.

[0048] Step S2 uses a neural network to process the nonlinear characteristics of the above soil classification; among them, 8 physical quantities will be used as inputs of the neural network, and the soil type will be used as the target. Based on the input consisting of eight physical characteristic observations generated by dredging construction, the neural network can identify five types of soil in dredging construction, such as clay, fine silt, gravel, mixed soil and layered soil.

[0049] Step S2 is specifically as follows:

[0050] Step S2.1: accumulate different data combinations of at least five regions, and aggregate the data combinations into the five data sets to determine the corresponding soil properties and the control algorithm that matches the corresponding soil properties.

[0051] The neural network is set up to classify soil types by organizing the data into two matrices: input matrix X and target matrix Y.

[0052] Among them, each i-th column of the data input matrix will have 8 elements, including speed, drag head active cover angle, drag head drag head depth, drag tube angle to the ground, terrain height difference, instantaneous density, instantaneous flow, and instantaneous vacuum.

[0053] Each corresponding column of the target matrix will have 5 elements, with the first element set to 1 for soil type A, the second element set to 1 for soil type B, the third element set to 1 for soil type C, the fourth element set to 1 for soil type D, and the fifth element set to 1 for soil type E. (All other elements are zero).

[0054] Step S2.2: Construct a neural network classifier to create a neural network that can classify the corresponding combination of dredging data sets into one of the soil types A, B, C, D, and E. Figure 2 .

[0055] The present invention uses a two-layer feedforward neural network, including a single hidden layer with 50 neurons, which can learn input-output relationships.

[0056] Step S2.3 divides the dredging dataset samples into a training set, a validation set, and a test set in a ratio of 40%, 40%, and 20%. The training set is used to train the network. The training is maintained if and only if the neural network continues to improve on the validation set. The test set provides a completely independent measure of the network's accuracy.

[0057] The classifier performance is quantified using mean squared error, which is displayed on a logarithmic scale. The neural network is trained until the mean squared error converges.

[0058] Step S2.4, test the neural network classifier, use the test sample to test the trained neural network, and use the vector search algorithm to find the number of rows with non-zero values, that is, the categories of soil types A, B, C, D, and E.

[0059] The specific process of constructing the MPC model in step S3 is as follows:

[0060] Control system modeling phase

[0061] From the 7098 sets of data collected previously, the J index (Formula 1) was used to screen out the 100 sets of benchmark values ​​with the largest J index, and a nonlinear MPC model was established. The reason for selecting a nonlinear MPC model is that the control variables and the corresponding soil excavation conditions are in a fluid, nonlinear state.

[0062] Establish control objectives:

[0063] The nonlinear model predictive controller is set to use the upper rake pipe winch, the lower rake pipe winch and the rake head movable cover to control the attitude of the rake pipe in order to achieve the following goals:

[0064] ● Keep the terrain difference of the rake head within the set value range of [0.1, 0.1] m, the purpose is to use the benchmark data as the criterion,

[0065] Dredging production is now maximized.

[0066] ●Maintain the vacuum degree at [20, 60]% in order to reduce cavitation of the mud pump and reduce equipment losses.

[0067] During the dredging construction process, dredging workers focus on the continuity of two types of variable data. One is the continuity of output, that is, completing the benchmark construction task in the shortest construction time; the other is the continuity of equipment performance, that is, the longest continuous working time of equipment without failure. Through the coordinated realization of these two control objectives, the optimal working cycle of the bucket suction dredger is obtained.

[0068] For five typical soil types, a control system model is established based on the collected data, as follows:

[0069] Step S3.1: Prepare training and validation datasets

[0070] In order to identify the neural state space model, it is necessary to collect state and input variables from experiments, that is, previously acquired data. When this data fully explores the required state input space, a model corresponding to a soil condition can be determined, which can fully reproduce the state of the controlled object in the entire working range.

[0071] Step S3.2: Load data for neural state space training and validation.

[0072] The training data set contains 200 tests and 100 sets of benchmark data. Each test contains 2 state signals (drag head terrain difference, vacuum degree) and four input signals (navigation speed, drag head depth, upper rake tube angle, rake lip angle). Sailing speed, drag head depth, upper rake tube angle, and rake lip angle are variables.

[0073] A single test contains 20 samples, 10 sets of benchmark sub-data, and 10 sets of other data, with a total test duration of 20 minutes. In addition, a 6098-minute long test with 300 minutes of optional data is used as the test validation data to verify the generalization performance of the model.

[0074] Define a discrete-time neural network state space with 1 state and 2 inputs; the state network has 2 hidden layers, each with 60 nodes. The sampling time is 60 seconds, the training timeout period is 20, the minimum batch size is 20, and the learning rate is 0.002.

[0075] Step S3.3: Generate state function from neural network spatial model

[0076] The MPC model obtains the next state prediction x(t+1) through the current state x(t) and the control input u(t); in the present invention, the state x(t) includes [rake head terrain difference, vacuum degree], which is expressed as:

[0077] x(t)=[x1(t),x2(t)] T

[0078] The control input u(t) includes [navigation speed, rake head depth, upper rake tube angle, rake lip angle], which can be expressed as:

[0079] u(t)=[u1(t), u2(t), u3(t), u4(t)] T =[v(t),h head (t), θ low (t), θ viso (t)] T

[0080] An approximate expression for the system dynamics (which can be adjusted or generated by a trained model) is as follows:

[0081] Current rake head angle to the ground: D land (t) = x1(t)

[0082] Current vacuum degree: P vacu (t) = x2(t)

[0083] Then the nonlinear dynamic model is obtained:

[0084] f1(x,u)=x1(t+1)=x1(t)+δ*(0.5*u1(t)+0.2*u4(t)-0.3*u2(t))

[0085] f2(x,u)=x2(t+1)=x2(t)+δ*(0.1*u1(t)+0.2*u2(t)-0.3*u3(t))

[0086] Generally, δ=0.1

[0087] Step S3.4: Design multi-level nonlinear MPC using state functions

[0088] Set the forecast horizon to 10 steps (60 seconds per step);

[0089] P=10*60 / T s ;

[0090] A multi-level nonlinear MPC controller is created. In order, the first controlled object input (measured) is the ship speed, the second input is the rake head depth, the third input is the upper rake tube angle, and the fourth input is the rake lip angle.

[0091] The sampling time is set to 10 seconds; the change range of the control input is limited;

[0092] The first status signal is designated as the drag head topography difference, and the second status signal is designated as the vacuum degree.

[0093] Specify the cost function for the first stage and set the order to 1.

[0094] The target vacuum degree is as follows, where P stage For the stage parameters:

[0095] D aim =P stage (1)

[0096] P aim =P stage (2)

[0097] Objective function (terrain difference from target + vacuum degree from target + change in control energy + slack variable penalty), where e is the slack variable:

[0098] J1=(x1(t)-D aim ) 2 +(x2(t)-P aim ) 2 +0.1*(u1(t) 2 +u2(t) 2 +u3(t) 2 )+10*e;

[0099] Intermediate stage objective function:

[0100] J2=(x1(t)-D aim ) 2 +(x2(t)-P aim ) 2 +0.05*(u1(t) 2 +u2(t) 2 +u3(t) 2)+10*e;

[0101] Specify the objective function of the final stage and set the order to 1; as follows:

[0102] J3=(x1(t)-D aim ) 2 +2*(x2(t)-P aim ) 2 +0.1*(u1(t) 2 +u2(t) 2 +u3(t) 2 )+5*e;

[0103] Define constraints, including:

[0104] 1) The output constraints are as follows:

[0105] Minimum value of rake head terrain difference: D land min=-0.5

[0106] Maximum value of rake head terrain difference: D land max=0.5

[0107] Minimum vacuum degree: P vacu min=20

[0108] Maximum vacuum degree: P vacu max=60

[0109] 2) Specify the inequality constraint function g≤0, where:

[0110]

[0111] Generate the Jacobian matrix:

[0112] Partial derivatives of the state (Jacobian matrix A)

[0113]

[0114] Partial derivatives of the input (Jacobian matrix B)

[0115]

[0116] Then the objective function is:

[0117] J=w D *(x1(t)-15) 2 +w P *(x2(t)-50) 2 +w u *(u1(t) 2 +u2(t) 2 +u3(t) 2+u4(t) 2 )+w s *(s1(t) 2 +s2(t) 2 )

[0118] Among them, the weight

[0119] w D =10

[0120] w P =5

[0121] w u =0.1

[0122] w s =100

[0123] Among them, the slack variable s = [s1, s2] T .

[0124] Partial derivatives with respect to the state

[0125]

[0126] Partial derivative with respect to the input

[0127] g u =2*w u *u

[0128] Partial derivatives with respect to slack variables

[0129] g s =2*w s *s

[0130] Through similar steps, equations corresponding to five types of soil A, B, C, D, and E can be established.

[0131] Example 1

[0132] At present, the terrain data of the trailing suction hopper dredger is based on the non-real-time water depth file produced by the surveying company through the measuring tools. A self-propelled vehicle or a non-self-propelled vehicle is directly installed on the bow and stern of the trailing suction hopper dredger, and a multi-beam measuring device or a laser radar is configured to obtain the terrain data before and after the construction. The above control method is applied to the operation of the trailing suction hopper dredger.

[0133] The distance between the front and rear of the hopper dredger is 100-120 meters. The dredging speed of the hopper dredger is within the range of (2 knots, 2.5 knots). The front radar or multi-beam measurement equipment provides the hopper dredger with sufficient response time to detect the delay t from the next point of terrain to the ship's displacement to that point. delayBetween 90-100s, it also provides the control system of the trailing suction hopper dredger with the time for algorithm simulation matching. The time for calculating the terrain matching algorithm is between 1025s, and the time for switching the algorithm is between 1025s. 25s+25s≤90s=t delay Therefore, the system has sufficient response time, which can be realized by the adaptive terrain depth setting and soil adaptive yield optimization MPC controller of the trailing suction hopper dredger; specifically, the terrain fluctuations are sensed by the sensor in advance and the soil conditions are mastered through experimental excavation, the terrain changes are used as variable input, and the control algorithm of the trailing suction hopper dredger is switched through soil identification, so that the optimal control strategy can be adopted to adapt to different soil types.

[0134] The above description is only a description of the preferred embodiments of the present application, and is not intended to limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical contents shall be deemed as equivalent effective embodiments and shall fall within the scope of protection of the technical solution of the present application.

Claims

1. An intelligent control method for the rake arm and the rake head of a trailing suction dredger, characterized in that: The following steps are involved: S1. Benchmarks Perform benchmark tests to pre-set a range of drag head depths, drag head to ground angles and drag tube to ground angles to carry out construction operations on a specific route; S2. Based on the benchmark data, the soil type of the current terrain is calculated using a clustering algorithm; S3, constructing MPC model; Using the above benchmark data, terrain changes are added as parameter inputs to terrain type knowledge, and an MPC model is established to drive the MPC control of the trailing suction hopper dredger. S4. Repeat the above steps to achieve the best construction effect.

2. The intelligent control method for the rake arm and the rake head of a trailing suction dredger according to claim 1 is characterized in that: Step S1 data collection stage is specifically as follows: during the operation of the trailing suction dredger, the speed, the angle of the drag head movable cover, the drag head depth, the angle of the drag pipe to the ground, the terrain height difference, the instantaneous density, the instantaneous flow rate and the instantaneous vacuum information are collected through sensors; Step S1.1: Set the sampling interval and sampling interval The speed changes in the interval [2.0knot, 2.5knot], and the sampling interval is 0.1knot; The angle of the drag head active cover is in the range of [0°, 30°], and the sampling interval is 5°; The depth of the rake head is in the range of [-1.5m, 1.5m], and the sampling interval is 0.1m; The rake tube’s angle to the ground is in the range of [30°, 60°], and the sampling interval is 5°; In the data collection phase, it is necessary to traverse 6 values ​​of speed change, 13 values ​​of rake head movable cover angle, 13 values ​​of rake head depth, and 7 values ​​of rake tube angle to the ground, for a total of 7098 data; Step S1.2, storing the acquired data into a structured historical database or a data table file in which each line of data is separated by commas, each line of data is in seconds, including time and 8 parameters: speed, drag head active cover angle, drag head drag head depth, drag pipe angle to the ground, terrain height difference, instantaneous density, instantaneous flow rate and instantaneous vacuum; Define the J indicator: as follows J (speed, drag head active cover angle, drag head drag head depth, drag pipe angle to the ground) = k1*terrain height difference+k2*Sum (instantaneous density, instantaneous output) (Formula 1) Among them, K1=1-1 / π; K2=1 / π; Find the 100 data combinations with the largest J value and record them for the following purposes: 1) as a benchmark value for marking soil; 2) as nonlinear modeling controller data for the trailing suction dredger for this soil.

3. The intelligent control method for the rake arm and the rake head of a trailing suction dredger according to claim 1 is characterized in that: Step S2 is specifically as follows: Step S2.1, accumulating different data combinations of at least five regions, and aggregating the data combinations in the five data sets to determine the corresponding soil properties and the control algorithm matching the corresponding soil properties; The neural network is set up to classify soil types by organizing the data into two matrices, the input matrix X and the target matrix Y. Among them, each i-th column of the data input matrix will have 8 elements, including speed, drag head active cover angle, drag head drag head depth, drag tube angle to the ground, terrain height difference, instantaneous density, instantaneous flow rate, and instantaneous vacuum; Each corresponding column of the target matrix will have 5 elements, with the first element set to 1 for soil type A, the second element set to 1 for soil type B, the third element set to 1 for soil type C, the fourth element set to 1 for soil type D, the fifth element set to 1 for soil type E, and all other elements are zero; Step S2.2, construct a neural network classifier to create a neural network that can map the corresponding combination of dredged data sets to one of the soil types A, B, C, D, and E; The neural network uses a two-layer feedforward neural network, including a single hidden layer with 50 neurons; Step S2.3 divides the dredging data set samples into a training set, a validation set, and a test set in a ratio of 40%, 40%, and 20%; the training set is used to train the network; the training is maintained if and only if the neural network continues to improve on the validation set; the test set provides a completely independent measurement of the network accuracy; The classifier performance is quantified using mean square error, which is displayed as a logarithmic scale metric; the neural network is trained until the mean square error converges; Step S2.4, test the neural network classifier, use the test sample to test the trained neural network, and use the vector search algorithm to find the number of rows with non-zero values, that is, the categories of soil types A, B, C, D, and E.

4. The intelligent control method for the rake arm and the rake head of a trailing suction dredger according to claim 1 is characterized in that: Step S3 is specifically as follows: Step S3.1: Prepare training and validation datasets In order to identify the neural state space model, it is necessary to collect state and input variables from the experiment, that is, previously collected data; when these data fully explore the required state input space, a model corresponding to a soil condition can be determined, which can fully reproduce the state of the controlled object in the entire working range; Step S3.2: Load data for neural state space training and validation The training data set contains 200 tests and 100 sets of benchmark data. Each test contains 2 state signals (drag head terrain difference, vacuum degree) and four input signals (navigation speed, drag head depth, upper rake tube angle, rake lip angle). Navigation speed, drag head depth, upper rake tube angle, and rake lip angle are variables; A single test contains 20 samples, 10 sets of benchmark sub-data, and 10 sets of other data, and the total test duration is 20 minutes. In addition, 300 minutes of the remaining 6098 minutes of data are selected as test validation data to verify the generalization performance of the model. Define a discrete-time neural network state space, consisting of 1 state and 2 inputs; the state network has 2 hidden layers, each with 60 nodes. The sampling time is 60 seconds, the training timeout period is 20; the minimum batch size is 20; The learning rate is 0.002; Step S3.3: Generate state function from neural network spatial model The MPC model obtains the next state prediction x(t+1) through the current state x(t) and the control input u(t); in the present invention, the state x(t) includes [rake head terrain difference, vacuum degree], which is expressed as: x(t)=[x1(t),x2(t)] T The control input u(t) includes [navigation speed, rake head depth, upper rake tube angle, rake lip angle], which can be expressed as: u(t)=[u1(t),u2(t),u3(t),u4(t)] T =[v(t),h head (t),θ low (t),θ viso (t)] T Then the nonlinear dynamic model is obtained: f1(x,u)=x1(t+1)=x1(t)+δ*(0.5*u1(t)+0.2*u4(t)-0.3*u2(t)) f2(x,u)=x2(t+1)=x2(t)+δ*(0.1*u1(t)+0.2*u2(t)-0.3*u3(t)) Where, δ = 0.1; Step S3.4: Design multi-level nonlinear MPC using state functions Set the forecast horizon to 10 steps (60 seconds per step); P=10*60 / Ts: Create a multi-level nonlinear MPC controller. In order, the first controlled object input (measured) is the ship speed, the second input is the rake head depth, the third input is the upper rake tube angle, and the fourth input is the rake lip angle; The sampling time is set to 10 seconds; the change range of the control input is limited; The first status signal is designated as the rake head terrain difference, and the second status signal is designated as the vacuum degree; Specify the cost function for the first stage and set the order to 1; The target vacuum degree is as follows, where P stage For the stage parameters: D aim =P stage (1) P aim =p stage (2) Objective function (terrain difference from target + vacuum degree from target + change in control energy + slack variable penalty), where e is the slack variable: J1=(x1(t)-D aim ) 2 +(x2(t)-P aim ) 2 +0.1*(u1(t) 2 +u2(t) 2 +u3(t) 2 )+10*e; Intermediate stage objective function: J2=(x1(t)-D aim ) 2 +(x2(t)-P aim ) 2 +0.05*(u1(t) 2 +u2(t) 2 +u3(t) 2 )+10*e; Specify the objective function of the final stage and set the order to 1; as follows: J3=(x1(t)-D aim ) 2 +2*(x2(t)-P aim ) 2 +0.1*(u1(t) 2 +u2(t) 2 +u3(t) 2 )+5*e; Define constraints, including: 1) The output constraints are as follows: Minimum value of rake head terrain difference: D land min=-0.5 Maximum value of rake head terrain difference: D land max=0.5 Minimum vacuum degree: P vacu min=20 Maximum vacuum degree: P vacu max=60 2) Specify the inequality constraint function g≤0, where: Generate the Jacobian matrix: Partial derivatives of the state (Jacobian matrix A) Partial derivatives of the input (Jacobian matrix B) Then the objective function is: J=w D *(x1(t)-15) 2 +w P *(x2(t)-50) 2 +w u *(u1(t) 2 +u2(t) 2 +u3(t) 2 +u4(t) 2 )+w s *(s1(t) 2 +s2(t) 2 ) Among them, the weight w D ＝10 w P ＝5 w u ＝0.1 w s ＝100 Among them, the slack variable s = [s1, s2] T ; Partial derivatives with respect to the state Partial derivative with respect to the input gu=2*w u *at Partial derivatives with respect to slack variables gs=2*ws*s Through similar steps, equations corresponding to five types of soil A, B, C, D, and E can be established.

Images