Valve-controlled hydraulic cylinder system model control method and system with independent load port control

Abstract

The application provides a valve-controlled hydraulic cylinder system model control method and system with independent load port control, comprising the following steps: S1, collecting hydraulic system operation data; S2, dividing the collected data into data sets, using a neural network to train a deep Koopman operator to upgrade the non-linear model, training the neural network to obtain an upgrading function and a high-dimensional space prediction model; S3, allowing a KMPC controller to use the trained high-dimensional space prediction model to calculate a control quantity, inputting a vector of an upgraded state parameter of a controlled object and a reference displacement, and outputting a control signal for a left main valve reference displacement and a right main valve spool reference displacement, and cyclically feeding back to complete hydraulic system position control. The application realizes fast model prediction control of a valve-controlled hydraulic cylinder system with independent load port control, and completes motion control of the valve-controlled hydraulic cylinder system with independent load port control.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial process nonlinear model linearization and model predictive control, in particular, to a valve-controlled hydraulic cylinder system model control method and system for load port independent control, and more particularly to a DeepKoopman-based valve-controlled hydraulic cylinder system model predictive control method and system for load port independent control. BACKGROUND

[0002] The valve-controlled hydraulic cylinder system is widely used in excavators, cranes, tractors and other engineering machinery due to its high power, high precision, fast response and other characteristics. Unlike the traditional valve-controlled hydraulic cylinder servo control system which uses a three-position four-way valve to control the hydraulic cylinder, the valve-controlled hydraulic cylinder system for load port independent control has the characteristic of two-cavity independent control, which realizes independent control of the hydraulic cylinder working cavity and the reversing valve pressure control cavity, thereby overcoming the low flexibility and high energy loss problems caused by the coupled mechanical structure of the traditional valve-controlled hydraulic cylinder system, and making the system have higher control precision.

[0003] At present, some researches have been carried out on the valve-controlled hydraulic cylinder system for load port independent control at home and abroad. The existing technical solution designs a load port independent control system with continuously adjustable return oil pressure, which solves the problems of air pocket of the actuator low-pressure cavity in the low-pressure regeneration mode of the control system and pressure loss in the ordinary mode by adding a pressure-adjustable electric proportional overflow valve and a one-way valve in parallel in the return oil circuit. The existing technical solution designs a hydraulic system and control method based on two-stage energy supply and load port independent valve control, which improves the control accuracy of the hydraulic drive unit and the efficiency of the system through fuzzy sliding mode variable structure control strategy. The existing technical solution designs a single-side control skill control method, in which the two cavities of the light-load actuator work at a relatively high pressure, which has the effect of improving damping. The flow of the heavy-load actuator is determined by matching the pump oil supply flow minus the flow of the light-load actuator. Through the joint control of the pump and valve, the pressure margin of the system can be reduced as much as possible.

[0004] In the valve-controlled hydraulic cylinder system for load port independent control, the control cavities of the hydraulic cylinder are independently controlled by two reversing valves, and the control cavities of each reversing valve are controlled by two high-speed on-off valves. The system includes multiple components and is a multi-stage multi-input servo control system of "component-integral valve-valve-controlled cylinder", and has high nonlinearity. The research on the valve-controlled hydraulic cylinder system for load port independent control mainly focuses on controlling the control accuracy of the system, but due to the nonlinear model, the nonlinear optimization problem in model control must be solved, which has high time cost for this controller, and lacks research on the linearization of nonlinear models.

[0005] The control method of the existing valve-controlled hydraulic cylinder system mainly improves the system performance by designing the system oil path structure for the valve-controlled hydraulic cylinder system with independent control of the load port, and generally defaults to the factory parameters for the hydraulic cylinder system parameters in the control process, and lacks research on model identification and prediction methods under data driving.

[0006] Patent document CN113653684B discloses a load port independent control system with continuously adjustable oil return pressure, specifically discloses that it comprises: a hydraulic power source (1) for providing the required flow for the system; an oil tank (2) for storing the required hydraulic oil for the system; a controller (3) for providing control signals for the system; a load port independent control valve group (4) for changing the hydraulic circuit of the system to enable the actuator to work in the required working mode, comprising a first proportional directional valve (41), a second proportional directional valve (42), and an electric proportional overflow valve (43) and a check valve (44) connected in parallel on the oil return path between the first proportional directional valve (41) and the second proportional directional valve (42), which adjusts the oil return pressure to avoid cavitation; the set oil return pressure of the electric proportional overflow valve (43) is adjusted according to the following formula: wherein is the actual flow through the second proportional directional valve (42), is the flow pressure coefficient of the first proportional directional valve (41) and the second proportional directional valve (42), is the oil return path pressure between the first proportional directional valve (41) and the second proportional directional valve (42), and is the pressure threshold value for cavitation; a hydraulic actuator (5) for converting the energy provided by the hydraulic power source (1) into mechanical energy to achieve the actual required operation; a control handle (6) for inputting a speed signal to the controller (3), which controls the first proportional directional valve (41) and the second proportional directional valve (42) through the controller (3), and in turn controls the action of the hydraulic actuator (5). However, the invention does not solve the nonlinear optimization problem in model control. SUMMARY

[0007] In view of the defects in the prior art, the purpose of the present application is to provide a valve-controlled hydraulic cylinder system model control method and system with independent control of the load port.

[0008] According to the valve-controlled hydraulic cylinder system model control method with independent control of the load port provided by the present application, the following steps are included:

[0009] Step S1: Collecting hydraulic system operation data;

[0010] Step S2: Dividing the collected data into data sets, using a neural network to train a deep Koopman operator to upgrade the nonlinear model, and training the neural network to obtain an upgrading function and a high-dimensional space prediction model;

[0011] Step S3: let the KMPC controller use the trained high-dimensional space prediction model to calculate the control quantity, input the vector of the state parameter of the controlled object after dimensionality increase and the reference displacement, output the control signals for the reference displacement of the left main valve and the reference displacement of the valve core of the right main valve, loop feedback, and complete the position control of the hydraulic system.

[0012] Preferably, the load port independent control valve controlled hydraulic cylinder system is realized by a digital hydraulic pilot programmable valve to independently control the pressures of the two cavities of the hydraulic cylinder; the pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way spool valve switches, and the main stage is composed of two identical three-position three-way spool valve switches; according to the preset load motion law, the system controller generates an instruction signal to control the opening degree of the spool valve core of the two main stages, so that the load moves according to the preset law; the load port independent control valve controlled hydraulic cylinder system uses the opening and closing dynamic characteristics of the pilot stage high-speed switch valve to independently control the opening degree of the spool valve core of the main stage, and the opening degrees of the spool valve cores of the two main stages will determine the states of the two cavities of the hydraulic cylinder inlet and outlet;

[0013] The control strategy adopts a PID-KMPC feedback structure, and under the input reference displacement, the load displacement pushed by the oil cylinder tracks the reference displacement according to the preset requirements; the first stage KMPC controller is composed of a dimensionality increase function and a linear model prediction controller, and takes the oil cylinder displacement, speed and two cavity pressures as feedback signals to generate control signals for the reference displacement of the left main valve and the reference displacement of the valve core of the right main valve; the second stage PID2 is composed of controllers PID2-1, PID2-2, PID2-3 and PID2-4, which respectively take the left valve core displacement and the right valve core displacement as feedback signals to generate control signals for controlling the opening and closing characteristics of the four pilot valves.

[0014] Preferably, in the step S2:

[0015] The Koopman operator is a linear operator acting on the observation function g(x), and the function g(x k ) converts the low-dimensional state variable space x k into the high-dimensional state variable space z k ; the linear operator completes the evolution of g(x k ) under the nonlinear mapping f; the observation function g(x k ) evolves into z k =g(x k ,u k ), wherein x k is the state quantity of the system at time k, and u k is the control quantity input to the system at time k;

[0016] z k+1 =g(x k+1 ,u k+1 )=κ(g(xk ,u k ),u k ) (1)

[0017] κ is a linear transformation equation to z k to z k+1 , expressed as z k+1 = Az k + Bu, A and B are linear matrices calculated by a neural network, used to represent a high-dimensional linear space, and the linear change of zk to zk+1;

[0018] Koopman linearization is a method of mapping a nonlinear system to a high-dimensional linear system through a g operator, and the state of the linear system is updated through a Koopman operator; z k is reduced to displacement state quantity The g operator is composed of a set of Koopman eigenfunctions, and the Koopman operator is a linear operator, and the combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2).

[0019] z k+1 = Az k + Bu (2)

[0020]

[0021] where C is a linear change matrix of z k to the actual state variable of the system in a linear high-dimensional space;

[0022] The observation function z k = g(x k , u k ), z k ∈ R m , x k ∈ R n , u k ∈ R 2 and the Koopman operator are solved by a deep neural network;

[0023] where R m is a high-dimensional linear space with m dimensions, R n is an actual system variable dimension space, and R 2 is a 2-dimensional space because the input variable is two;

[0024] The first set of hidden layer composite parameters is recorded as OFP(x k , u k ), which is part of the observation function; the first input layer contains the state quantity of the nonlinear model at time k and the input u k ;

[0025] The first output layer and the state quantity x k is the state quantity z of the high-dimensional linear model at time k k The state quantity z of the high-dimensional linear model at time k+1 is obtained through the full connection named DKO(z k , u k ) k+1 The observation function is represented as equation (3), where z′ k ∈ R m-n is the output of the first group of hidden layers; the linear system is represented as equation (4);

[0026] The first group of hidden layers is used to learn the observation function, which has an internal connection layer and an activation function, z k ′ is the output of the full connection layer; DKO(z k , u k ) represents the Deep Koopman operator, which is a Koopman operator calculated by fitting the dimensionality function g using a neural network, which is an improvement on the Koopman operator; DKO(z k , u k ) is a full connection layer without offset, which is used to learn the Koopman operator in equation (4):

[0027]

[0028] k+1 =DkO(z k ,u k )=Az k +Bu k (4)

[0029] In order to obtain the prediction of the state quantity x k of the nonlinear system, after the low-dimensional nonlinear model is upgraded to the high-dimensional linear model, dimensionality reduction needs to be completed; the deep neural network records the dimensionality reduction process of the third group of hidden layers as DRP(z k ), which is the input z k at time k in the linear space after passing through the hidden layer; the reconstructed values of the state quantity and the input quantity of the nonlinear system are obtained; the dimensionality reduction reconstruction of the system is completed, and the third group of hidden layers is a full connection layer:

[0030] [x k ,u k ] T =DRP(z k ) (5).

[0031] Preferably, the design of the deep neural network cost criterion is completed, and the deep neural network loss function is given:

[0032] A loss function loss1 measuring the loss of prediction accuracy:

[0033] loss1 = ‖(x k+1 ,u k+1 )-DRP(DKO(z k ,u k ))‖ (6)

[0034] A loss function loss2 measuring the effectiveness of reconstruction:

[0035] loss2 = ‖(x k ,u k )-DRP(z k )‖ (7)

[0036] A loss function loss3 for measuring the linearity of the system:

[0037]

[0038] Preferably, in the step S3:

[0039] Based on the calculation results of the neural network, the KMPC controller is established.

[0040] The controller solving process is represented by equations (9) and (10).

[0041]

[0042] s.t.z k+1 = Az k + Bu k , k = 0, …, N (9)

[0043] y k = z k

[0044] z0= (x0)

[0045]

[0046] Wherein, J is the cost function, Q, R are weight parameters, N is the prediction time domain, y k is the system predicted state, y r is the system reference state, u k is the system control quantity, the cost function is divided into two parts, one part is the deviation cost of the reference state quantity, and the other part is the control quantity cost; solving the optimization problem to obtain the control quantity u k with the minimum cost, let be the system predicted displacement, x0 is the state of the system at the time of solving, x kis the predicted state of the system at the kth step given by the predictor; J is the cost function to be optimized with respect to u k is the cost function to be optimized with respect to u k as the system control input for the next step.

[0047] According to the application, a load port independent control valve-controlled hydraulic cylinder system model control system is provided, comprising:

[0048] Module M1: collect hydraulic system operation data;

[0049] Module M2: divide the collected data into data sets, use a neural network to train a deep Koopman operator to upgrade the non-linear model, train the neural network to obtain an upgrading function and a high-dimensional space prediction model;

[0050] Module M3: let the KMPC controller use the trained high-dimensional space prediction model to calculate the control quantity, input the upgraded vector of the state parameters of the controlled object and the reference displacement, and output the control signals for the left main valve reference displacement and the right main valve spool reference displacement, and complete the position control of the hydraulic system through cyclic feedback.

[0051] Preferably, the load port independent control valve-controlled hydraulic cylinder system is realized by a digital hydraulic pilot programmable valve to independently control the pressures of the two cavities of the hydraulic cylinder; the pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way slide valve type on-off valves, and the main stage is composed of two identical three-position three-way slide valve type hydraulic directional control valves; according to a preset load motion law, the system controller generates an instruction signal for controlling the opening degrees of the spool cores of the two main stages, so that the load moves according to the preset law; the load port independent control valve-controlled hydraulic cylinder system uses the opening and closing dynamic characteristics of the pilot stage on-off valves to independently control the opening degrees of the spool cores of the main stages, and the opening degrees of the spool cores of the two main stages will determine the states of the two cavities of the hydraulic cylinder;

[0052] The control strategy adopts a PID-KMPC feedback structure, under the input reference displacement, the load displacement pushed by the oil cylinder tracks the reference displacement according to the preset requirements; the first stage KMPC controller is composed of an upgrading function and a linear model prediction controller, takes the oil cylinder displacement, speed and two cavity pressures as feedback signals, and generates control signals for the left main valve reference displacement and the right main valve spool reference displacement; the second stage PID2 is composed of controllers PID2-1, PID2-2, PID2-3 and PID2-4, respectively takes the left spool displacement and the right spool displacement as feedback signals, and generates control signals for controlling the opening and closing characteristics of the four pilot valves.

[0053] Preferably, in the module M2:

[0054] The Koopman operator is a linear operator acting on the observation function g(x), the function g(x k ) converts the low-dimensional state variable space x k into a high-dimensional state variable space z k ; the linear operator completes the evolution of g(x k ) under the nonlinear mapping f; the observation function g(x k ) evolves into z k =g(x k ,u k ), where x k is the state quantity of the system at time k, and u k is the control quantity input to the system at time k;

[0055] z k+1 =g(x k+1 ,u k+1 )=κ(g(x k ,u k ),u k ) (1)

[0056] κ is a linear transformation equation for z k to z k+1 , which is represented by z k+1 =Az k +Bu, A and B are linear matrices calculated by a neural network, used to represent a high-dimensional linear space, and the linear change of zk to zk+1;

[0057] Koopman linearization is a method of mapping a nonlinear system to a high-dimensional linear system through a g operator, and the state of the linear system is updated through the Koopman operator; z k is reduced to the displacement state quantity The g operator is composed of a set of Koopman eigenfunctions, and the Koopman operator is a linear operator, and the combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2).

[0058] z k+1 =Az k +Bu (2)

[0059]

[0060] where C is a linear change matrix of z k to the actual state variable of the system in a linear high-dimensional space;

[0061] The observation function z k =g(x k ,u k ) is solved by a deep neural network, and zk ∈R m , x k ∈R n , u k ∈R 2 and Koopman operator;

[0062] where R m is the high-dimensional linear, with m spatial dimensions, R n is the actual system variable dimension space, R 2 is the 2-dimensional space, because the input variable is two;

[0063] The first set of hidden layer composite parameters is recorded as OFP(x k , u k ), as part of the observation function; the first input layer contains the state quantity of the nonlinear model at time k and the input u k ;

[0064] The first output layer and the state quantity x k is the state quantity z k of the high-dimensional linear model at time k, which is obtained through a fully connected layer named DKO(z k , u k ), and the state quantity z k+1 of the high-dimensional linear model at time k+1, the observation function is represented as equation (3), where z' k ∈R m-n is the output of the first set of hidden layers; the linear system is represented as equation (4);

[0065] The first set of hidden layers is used to learn the observation function, with an internal connection layer and an activation function, z k ' is the output of the full connection layer; DKO(z k , u k ) represents the Deep Koopman operator, which is a Koopman operator calculated by fitting the dimensionality function g using a neural network, which is an improvement on the Koopman operator; DKO(z k , u k ) is a full connection layer without offset, which is used to learn the Koopman operator in equation (4):

[0066]

[0067] z k+1 = DKO(z k , u k ) = Az k + Bu k (4)

[0068] In order to obtain the state quantity x kFor predictions, after upgrading from a low-dimensional nonlinear model to a high-dimensional linear model, dimensionality reduction is required. The dimensionality reduction process of the deep neural network through the third set of hidden layers is recorded as DRP(z). k The input z at time k in the linear space. k After passing through the hidden layer, the state variables of the nonlinear system are obtained. and input volume The reconstructed values; the dimensionality reduction reconstruction of the system is complete, and the third hidden layer is a fully connected layer:

[0069] [x k ,u k ] T =DRP(z k (5).

[0070] Preferably, the design of the cost criterion for the deep neural network is completed, and the loss function for the deep neural network is given:

[0071] Loss function 1, which measures prediction accuracy:

[0072] loss1 = ||(x k+1 ,u k+1 )-DRP(DKO(z k ,u k ))‖ (6)

[0073] loss function 2, which measures the effectiveness of reconstruction:

[0074] loss2=‖(x k ,u k )-DRP(z k )‖ (7)

[0075] loss function 3 used to measure the linearity of the system:

[0076]

[0077]

[0078] Preferably, in module M3:

[0079] A KMPC controller is established based on the calculation results of the neural network.

[0080] The controller solution process is represented by equations (9) and (10).

[0081]

[0082] stz k+1 =Az k +Bu kk = 0, …, N (9)

[0083] y k = Cz k

[0084] z0= g(x0)

[0085]

[0086] where J is the cost function, Q, R are weight parameters, N is the prediction horizon, y k is the system predicted state, y r is the system reference state, u k is the system control variable, the cost function is divided into two parts, one part is the deviation cost of the reference state variable, and the other part is the control variable cost; the optimization problem is solved to obtain the control variable u k , let is the system predicted displacement, x0 is the state of the system when solving, x k is the system predicted state given by the predictor at the kth step; J is the objective optimization function with u k as the variable, u k is solved to minimize the cost of J under the constraint condition, as the system control variable input of the next step.

[0087] Compared with the prior art, the present application has the beneficial effects as follows:

[0088] 1. The present application proposes a data-driven model prediction method for the load port independent control valve-controlled hydraulic cylinder system, which uses neural network training deep Koopman operator to upgrade the nonlinear model, and obtains a linear prediction model in high-dimensional space.

[0089] 2. The present application designs a model predictive control method based on deep Koopman, which introduces an upgrading function to combine the nonlinear model with the linear model predictive controller, realizes the rapid model prediction control of the load port independent control valve-controlled hydraulic cylinder system, and completes the motion control of the load port independent control valve-controlled hydraulic cylinder system. BRIEF DESCRIPTION OF DRAWINGS

[0090] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0091] Figure 1 is a schematic diagram of the valve-controlled hydraulic cylinder system for load port independent control;

[0092] Figure 2 is a control block diagram of the valve-controlled hydraulic cylinder system for load port independent control;

[0093] Figure 3 is a schematic diagram of a deep neural network structure;

[0094] Figure 4 is a schematic diagram of a prediction result;

[0095] Figure 5 is a schematic diagram of a KMPC controller structure;

[0096] Figure 6 is a schematic diagram of a system predictive control method. DETAILED DESCRIPTION

[0097] The present application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These are within the scope of the present application.

[0098] Example 1:

[0099] The present application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the present application, but do not limit the present application in any form. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of changes and improvements can be made. These are within the scope of the present application.

[0100] According to the load port independent control valve hydraulic cylinder system model control method provided by the present application, as shown in Figures 1-6 , comprising:

[0101] Step S1: Collecting hydraulic system operation data;

[0102] Step S2: Dividing the collected data into data sets, using neural network to train deep Koopman operator to upgrade the non-linear model, and using neural network to train the deep Koopman operator to upgrade the non-linear model to obtain the upgrading function and high-dimensional space prediction model;

[0103] Specifically, in the step S2:

[0104] The Koopman operator is a linear operator acting on the observation function g(x). The function g(x k ) converts the low-dimensional state variable space x k into the high-dimensional state variable space z k; linear operator completes g(x k ) evolution under nonlinear mapping f; observation function g(x k ) evolves into z k = g(x k , u k ), where x k is the state quantity of the system at time k, and u k is the control quantity input to the system at time k;

[0105] z k+1 = g(x k+1 , u k+k1 ) = κ(g(x k , u k ), u k ) (1)

[0106] κ is a linear transformation equation for z k to z k+1 , expressed as z k+1 = z k + u, A and B are linear matrices calculated by a neural network, used to represent a high-dimensional linear space, and the linear change of zk to zk+1;

[0107] Koopman linearization is a method of mapping a nonlinear system to a high-dimensional linear system through a g operator, and the state of the linear system is updated through a Koopman operator; z k is reduced to displacement state quantity The g operator is composed of a set of Koopman eigenfunctions, and the Koopman operator is a linear operator, and the combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2).

[0108] z k+1 = Az k + Bu (2)

[0109]

[0110] where C is a linear change matrix of z k to the actual state variable of the system in a linear high-dimensional space;

[0111] The observation function z k = g(x k , u k ) is solved by a deep neural network, z k ∈ R m , x k ∈ R n , u k ∈ R 2and Koopman operator;

[0112] where R m is the high-dimensional linear, with m spatial dimensions, R n is the actual system variable dimension space, R 2 is the 2-dimensional space, as the input variable is two;

[0113] The first set of hidden layer composite parameters is recorded as OFP(x k , u k ), as part of the observation function; the first input layer contains the state quantity and input u k of the nonlinear model at time k;

[0114] The first output layer and state quantity x k is the state quantity z k of the high-dimensional linear model at time k; the state quantity z k of the high-dimensional linear model at time k+1 is obtained through a fully connected layer named DKO(z k , u k+1 ); the observation function is represented as equation (3), where z' k ∈ R m-n is the output of the first set of hidden layers; the linear system is represented as equation (4);

[0115] The first set of hidden layers is used to learn the observation function, with an internal connection layer and an activation function, z k ' is the output of the full connection layer; DKO(z k , u k ) represents the Deep Koopman operator, which is a Koopman operator calculated by fitting the dimension-up function g using a neural network, which is an improvement on the Koopman operator; DKO(z k , u k ) is a full connection layer without offset, used to learn the Koopman operator in equation (4):

[0116]

[0117] z k+1 = DKO(z k , u k ) = Az k + Bu k (4)

[0118] In order to obtain the prediction of the state quantity x k of the nonlinear system, after upgrading the low-dimensional nonlinear model to a high-dimensional linear model, dimension reduction needs to be completed; the deep neural network records the dimension reduction process of the third set of hidden layers as DRP(z k), input z at time k in linear space k After passing through the hidden layer, the state quantity of the nonlinear system is obtained and the reconstruction value of the input quantity ; the dimensionality reduction reconstruction of the system is completed, and the third group of hidden layers is a fully connected layer:

[0119] [x k ,u k ] t = DRP(z k ) (5).

[0120] Specifically, the design of the deep neural network cost criterion is completed, and the deep neural network loss function is given:

[0121] The loss function loss1 measures the prediction accuracy:

[0122] loss1 = ‖(x k+1 ,u k+1 )- DRP(DKO(z k ,u k ))‖ (6)

[0123] The loss function loss2 measures the reconstruction effectiveness:

[0124] loss2 = ‖(x k ,u k )- DRP(z k )‖ (7)

[0125] The loss function loss3 measures the linearity of the system:

[0126]

[0127]

[0128] Step S3: The KMPC controller uses the trained high-dimensional space prediction model to calculate the control quantity, inputs the state parameter vector of the controlled object after dimensionality increase and the reference displacement, and outputs the control signals for the left main valve reference displacement and the right main valve spool reference displacement, which are cyclically fed back to complete the position control of the hydraulic system.

[0129] Specifically, in the step S3:

[0130] Based on the calculation results of the neural network, the KMPC controller is established.

[0131] The controller solving process is represented by equations (9) and (10).

[0132]

[0133] s.t.zk+1 = Az k + Bu k , k = 0, …, N (9)

[0134] y k = Cz k

[0135] z0= (x0)

[0136]

[0137] where J is the cost function, Q, R are weight parameters, N is the prediction horizon, y k is the system predicted state, y r is the system reference state, u k is the system control variable, the cost function is divided into two parts, one part is the deviation cost of the reference state variable, and the other part is the control variable cost; solving the optimization problem obtains the control variable u k that makes the cost minimum, let be the system predicted displacement, x0 is the state of the system at the time of solving, x k is the system predicted state given by the predictor at the kth step after; J is the objective optimization function with u k as the variable, u k that makes the cost of J minimum under the constraint condition is solved as the system control variable input of the next step.

[0138] Specifically, the load port independent control valve controlled hydraulic cylinder system realizes independent control of the pressure of the two cavities of the hydraulic cylinder by a digital hydraulic pilot programmable valve; the pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way spool valve switches, and the main stage is composed of two identical three-position three-way spool valve switches; according to the preset load motion law, the system controller generates a command signal to control the opening degree of the spool valve core of the two main stages, so that the load moves according to the preset law; the load port independent control valve controlled hydraulic cylinder system utilizes the opening and closing dynamic characteristics of the pilot stage high-speed switch valve to complete independent control of the opening degree of the spool valve core of the main stage, and the opening degree of the spool valve core of the two main stages will determine the state of the two cavities of the hydraulic cylinder inlet and outlet;

[0139] The control strategy adopts a PID-KMPC feedback structure, under the input reference displacement, the load displacement pushed by the oil cylinder tracks the reference displacement according to the preset requirement; the first-stage KMPC controller is composed of an upgrading function and a linear model predictive controller, taking the oil cylinder displacement, speed and the pressure of two cavities as feedback signals to generate the control signals for the reference displacement of the left main valve and the reference displacement of the right main valve spool; the second-stage PID2 is composed of controllers PID2-1, PID2-2, PID2-3 and PID2-4, taking the left valve spool displacement and the right valve spool displacement as feedback signals to generate the control signals for controlling the opening and closing characteristics of the four pilot valves.

[0140] Embodiment 2:

[0141] Embodiment 2 is a preferred example of Embodiment 1, to more specifically illustrate the present application.

[0142] The present application also provides a load port independent control valve controlled hydraulic cylinder system model control system, which can be realized by executing the flow steps of the load port independent control valve controlled hydraulic cylinder system model control method, that is, the load port independent control valve controlled hydraulic cylinder system model control method can be understood by those skilled in the art as a preferred embodiment of the load port independent control valve controlled hydraulic cylinder system model control system.

[0143] According to the load port independent control valve controlled hydraulic cylinder system model control system provided by the present application, the system comprises:

[0144] Module M1: collecting hydraulic system operation data;

[0145] Module M2: dividing the collected data into data sets, training a deep Koopman operator using a neural network to upgrade the non-linear model to train the neural network, obtaining an upgrading function and a high-dimensional space prediction model;

[0146] Specifically, in the module M2:

[0147] The Koopman operator is a linear operator acting on the observation function g(x), the function g(x k ) converts the low-dimensional state variable space x k into the high-dimensional state variable space z k ; the linear operator completes the evolution of g(x k ) under the nonlinear mapping f; the observation function g(x k ) evolves into z k = g(x k , u k ), wherein x k is the state quantity of the system at time k, and u kthe control quantity input to the system at time k;

[0148] z k+1 = g(x k+1 , u k+1 ) = K(g(x k , u k ), u k ) (1)

[0149] K is a linear transformation equation for z k to z k+1 , expressed as z k+1 = Az k + Bu, A and B are linear matrices calculated by a neural network, used to represent a high-dimensional linear space, and the linear change of zk to zk+1;

[0150] Koopman linearization is a method of mapping a nonlinear system to a high-dimensional linear system through a g operator, and the state of the linear system is updated through a Koopman operator; z k is reduced to displacement state quantity The g operator is composed of a set of Koopman eigenfunctions, and the Koopman operator is a linear operator, and the combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2).

[0151] z k+1 = Az k + Bu (2)

[0152]

[0153] wherein C is a linear change matrix of z k to the actual state variable of the system in a linear high-dimensional space;

[0154] The observation function z k = g(x k , u k ), z k ∈ R m , x k ∈ R n , u k ∈ R 2 and the Koopman operator are solved by a deep neural network;

[0155] wherein R m is a high-dimensional linear space with a dimension of m, R n is an actual system variable dimension space, and R 2 is a 2-dimensional space because the input variable is two;

[0156] The first set of hidden layer composite parameters is recorded as OFP(x) k u k The first input layer contains the state variables of the nonlinear model at time k and the input u, which are part of the observation function. k ;

[0157] First output layer and state variable x k It is the state variable z of a high-dimensional linear model at time k. k Through a name called DKO(z) k u k The complete connection of ) yields the state variables z of the high-dimensional linear model at time k+1. k+1 The observation function is expressed as equation (3), where z′ k ∈R m-n It is the output of the first hidden layer; the linear system is represented by equation (4);

[0158] The first set of hidden layers is used to learn the observation function, and it contains connection layers and activation functions. k ′ is the output of the fully connected layer; DKO(z) k u k DKO(z) represents the Deep Koopman operator, which is an improvement on the Koopman operator obtained by fitting the increased-dimensional function g using a neural network; k u k ) is a fully connected layer without offset, used to learn the Koopman operator in equation (4):

[0159]

[0160] z k+1 =DKO(z) k ,u k ) = Az k +Bu k (4)

[0161] To obtain the state variables x of a nonlinear system k For predictions, after upgrading from a low-dimensional nonlinear model to a high-dimensional linear model, dimensionality reduction is required. The dimensionality reduction process of the deep neural network through the third set of hidden layers is recorded as DRP(z). k The input z at time k in the linear space. k After passing through the hidden layer, the state variables of the nonlinear system are obtained. and input volume The reconstructed values; the dimensionality reduction reconstruction of the system is complete, and the third hidden layer is a fully connected layer:

[0162] [x k ,u k] T = DRP(z k ) (5).

[0163] In particular, the design of the deep neural network cost criterion is completed, and the deep neural network loss function is given:

[0164] The prediction accuracy loss function loss1 is measured:

[0165] loss1 = ‖(x k+1 ,u k+1 )-DRP(DKO(z k ,u k ))‖ (6)

[0166] The loss function loss2 for measuring the effectiveness of reconstruction is:

[0167] loss2 = ‖(x k ,u k )-DRP(z k )‖ (7)

[0168] The loss function loss3 for measuring the linearity of the system is:

[0169]

[0170]

[0171] Module M3: let the KMPC controller use the trained high-dimensional space prediction model to calculate the control quantity, input the vector of the state parameter of the controlled object after dimensionality increase and the reference displacement, output the control signal for the left main valve reference displacement and the right main valve spool reference displacement, loop feedback, and complete the position control of the hydraulic system.

[0172] In particular, in the module M3:

[0173] Based on the calculation result of the neural network, the KMPC controller is established.

[0174] The controller solving process is represented by equations (9) and (10).

[0175]

[0176] s.t.z k+1 = Az k + Bu k , k = 0, …, N (9)

[0177] y k = Cz k

[0178] z0= g(x0)

[0179]

[0180] where J is the cost function, Q, R are weight parameters, N is the prediction horizon, y k is the system predicted state, y r is the system reference state, u k is the system control variable, the cost function is divided into two parts, one part is the deviation cost of the reference state variable, and the other part is the control variable cost; the control variable u k is obtained by solving the optimization problem to minimize the cost is the system predicted displacement, x0 is the state of the system at the time of solving, x k is the system predicted state given by the predictor at the kth step; J is the objective optimization function with u k as the variable, u k is obtained by solving the optimization problem to minimize the cost under the constraint condition, and is input as the system control variable of the next step.

[0181] Specifically, the load port independently controlled valve controlled hydraulic cylinder system is realized by a digital hydraulic pilot programmable valve to independently control the pressures of two cavities of the hydraulic cylinder; the pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way spool valve type on-off valves, and the main stage is composed of two identical three-position three-way spool valve type hydraulic directional control valves; according to the preset load motion law, the system controller generates an instruction signal for controlling the opening degrees of the spool valve cores of the two main stages, so that the load moves according to the preset law; the load port independently controlled valve controlled hydraulic cylinder system utilizes the opening and closing dynamic characteristics of the pilot stage high-speed on-off valve to independently control the opening degrees of the spool valve cores of the two main stages, and the opening degrees of the spool valve cores of the two main stages will determine the states of the two cavities at the inlet and outlet of the hydraulic cylinder.

[0182] The control strategy adopts a PID-KMPC feedback structure, under the input reference displacement, the load displacement pushed by the oil cylinder tracks the reference displacement according to the preset requirement; the first stage KMPC controller is composed of an ascending dimension function and a linear model predictive controller, takes the oil cylinder displacement, speed and the pressures of two cavities as feedback signals, and generates control signals for the reference displacement of the left main valve and the reference displacement of the right main valve spool; the second stage PID2 is composed of controllers PID2-1, PID2-2, PID2-3 and PID2-4, respectively takes the left valve core displacement and the right valve core displacement as feedback signals, and generates control signals for controlling the opening and closing characteristics of the four pilot valves.

[0183] Embodiment 3:

[0184] Embodiment 3 is a preferred example of Embodiment 1, which more specifically illustrates the present application.

[0185] The content of the present application is described in detail in combination with the drawings.

[0186] Figure 1 and Figure 2 The principle diagram and control block diagram of a load-oriented independent control valve-controlled hydraulic cylinder system are shown in Figures 1 and 2, respectively. The system is realized by a digital hydraulic pilot programmable valve to achieve independent control of the pressure of the two chambers of the hydraulic cylinder. The pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way high-speed on-off valves, and the main stage is composed of two identical three-position three-way slide valve type hydraulic directional control valves. For a given load motion law, the system controller generates a command signal to control the opening degree of the spool of the two main stage valves, so that the load moves according to the given law. The system first uses the superior opening and closing dynamic characteristics of the pilot stage high-speed on-off valve to achieve independent control of the opening degree of the spool of the main stage valve, and then the opening degree of the spool of the two main stage valves will determine the state of the two chambers of the hydraulic cylinder, thereby meeting the motion requirements of the load.

[0187] The control strategy adopts a PID-KMPC feedback structure. Under the input reference displacement, the load displacement pushed by the oil cylinder can track the reference displacement as required. The first stage KMPC controller is composed of an upgrade function and a linear model predictive controller (MPC), and takes the oil cylinder displacement, speed and two chamber pressures as feedback signals to generate control signals for the left main valve reference displacement and the right main valve spool reference displacement. The second stage PID2 is composed of controllers PID2-1, PID2-2, PID2-3 and PID2-4, which take the left valve core displacement and the right valve core displacement as feedback signals to generate control signals for controlling the opening and closing characteristics of the four pilot valves.

[0188] The application designs a method for constructing a deep Koopman operator prediction model. The method can obtain better Koopman linearization parameters.

[0189] The Koopman operator is a data-driven nonlinear system linearization method. By evaluating a series of nonlinear functions (upgrading) once, all nonlinear terms appearing in these functions are included in the upgraded function, and the nonlinear cost function and constraints can be processed in a linear manner. Therefore, the proposed scheme can be easily used for predictive control of nonlinear dynamic systems, using a linear MPC design solver, thereby avoiding the trouble and high computational cost of non-convex optimization problems encountered in the classical nonlinear MPC format.

[0190] The Koopman operator is extended to controlled dynamics, and an extended dynamic mode decomposition (EDMD) is applied to compute a finite-dimensional approximation of the operator. Koopman theory can be seen as a method to map a nonlinear system to a high-dimensional linear system by g lifting functions. The state of this linear system is updated by the Koopman operator. The g lifting functions are constructed by a set of v eigenfunctions. In order to further improve the prediction accuracy of the linear predictor for nonlinear systems, a neural network is used to train the g function.

[0191] The g function fitted by the neural network has more accurate prediction accuracy than the least squares method, and can quickly and accurately predict the state parameter changes of the nonlinear model within a period of time. According to the prediction model, a linear MPC controller with improved dimension is designed.

[0192] The Koopman operator is a linear operator acting on the observation function g(x). The function g(x k ) converts the low-dimensional state variable space x k into the high-dimensional state variable space z k . The linear operator completes the evolution of g(x k ) under the nonlinear mapping f. The observation function g(x k ) evolves into z k = g(x k , u k ), where x k is the state of the system at time k, and u k is the control input to the system at time k. The equation is as follows:

[0193] z k+1 = g(x k+1 , u k+1 ) = κ(g(x k , u k ), u k ) (1)

[0194] κ is a linear transformation equation for z k to z k+1 , which can be represented as z k+1 = Az k + Bu. Where A, B are linear matrices calculated by a neural network, used to represent the high-dimensional linear space, the linear change of zk to zk+1

[0195] In other words, Koopman linearization can be seen as a method to map a nonlinear system to a high-dimensional linear system by g operator. The state of this linear system is updated by the Koopman operator. z k can be reduced to displacement state variable x The g operator is composed of a set of Koopman eigenfunctions, while the Koopman operator is a linear operator. The combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2).

[0196] z k+1 = Az k + Bu (2)

[0197]

[0198] where C is a linear transformation matrix from zk in the linear high-dimensional space to the actual state variables of the system.

[0199] Therefore, the solution of the two operators can be represented by the deep neural network structure shown in the upper part of Figure 3 Through this network, the observation function z k = g(x k , u k ), z k ∈ R m , x k ∈ R n , u k ∈ R 2 and the Koopman operator are solved.

[0200] where R m is the high-dimensional linear space with dimension m, R n is the actual system variable dimension space, and R 2 is the 2-dimensional space with the same dimension as the input variable.

[0201] The first set of hidden layer composite parameters is recorded as OFP(x k , u k ), which is part of the observation function. The first input layer contains the state quantity of the nonlinear model at time k and the input u k .

[0202] The first output layer and the state quantity x k are the state quantity z k of the high-dimensional linear model at time k. Through the fully connected named DKO(z k , u k ), the state quantity z k+1 of the high-dimensional linear model at time k+1 is obtained. Therefore, the observation function can be represented as equation (3), where z' k ∈ R m-n is the output of the first set of hidden layers. The linear system can be represented as equation (4).

[0203] The first set of hidden layers is used to learn the observation function to make equation (3) approximate the span of Koopman eigenfunctions. There are full connected layers and activation functions inside, and the last z k is the output of full connected layer. DKO(z k , u k ) represents DeepKoopman operator, which is the Koopman operator calculated by fitting the lifted function g using neural network, which is an improvement of Koopman operator. DKO(z k , u k ) is a full connected layer without bias, which is used to learn the Koopman operator in equation (4).

[0204]

[0205] k+1 = DKO(z k , u k ) = Az k + Bu k (4)

[0206] In order to obtain the prediction of the state quantity x k of the nonlinear system, after the low-dimensional nonlinear model is upgraded to the high-dimensional linear model, the dimension reduction still needs to be completed. If the nonlinear observation function g(x k , u k ) is approximated as a linear observation function, there will be a certain error. At the same time, when the matrix dimension is large, the pseudo-inverse operation will consume a lot of computing resources. The lifting process of deep neural network retains the advantages of deep neural network in nonlinear function fitting. As shown in Figure 3 , the deep neural network is through the dimension reduction process of the third set of hidden layers, which is recorded as DRP(z k ). After the input z k at time k in the linear space passes through the hidden layer, the reconstructed values of the state quantity and the input quantity of the nonlinear system are obtained. The dimension reduction reconstruction of the system has been completed. Therefore, the third hidden layer should be a full connected layer and approximate the inverse operation of the dimension increasing process.

[0207] [x k , u k ] T = DRP(z k ) (5)

[0208] According to the characteristics of the deep neural network as shown in Figure 3 , it is necessary to complete the design of the network cost criterion and give the loss function of the deep neural network. The specific case is as follows:

[0209] (a) Prediction accuracy. The goal of the deep Koopman predictor model is to achieve a high approximation of the low-dimensional nonlinear space, thereby improving the accuracy of the future state quantity prediction, ensuring the accuracy of the prediction. As shown in equation (6), the loss function loss1 is given to measure the prediction accuracy:

[0210] loss1 = ‖(x k+1 ,u k+1 )-DRP(DKO(z k ,u k ))‖ (6)

[0211] (b) Reconstruction effectiveness. The reconstruction process from the high-dimensional linear space to the low-dimensional nonlinear space is the inverse process of the dimensionality increase. The result should be highly similar to the original low-dimensional nonlinear space to ensure the effectiveness of the reconstruction. As shown in equation (7), the loss function loss2 is given to measure the reconstruction effectiveness:

[0212] loss2 = ‖(x k ,u k )-DRP(z k )‖ (7)

[0213] (c) System linearity. According to the Koopman theory, DKO should ensure the linearity requirement of the high-dimensional linear system. As shown in equation (8), the loss function loss3 is given to measure the linearity of the system:

[0214]

[0215]

[0216] The effect of the trained predictor is shown in Figure 4 ;

[0217] Based on the neural network calculation results, the KMPC controller is established as shown in Figure 5 .

[0218] The controller solving process under the above structure is represented by equations (9) and (10).

[0219]

[0220] s.t.z k+1 = Az k + Bu k , k = 0, …, N (9)

[0221] y k = Cz k

[0222] z0= g(x0)

[0223]

[0224] In the formula, J is the cost function, Q, R are weight parameters, N is the prediction time domain, y k is the system predicted state, y r is the system reference state, u k is the system control quantity, the cost function is divided into two parts, one part is the deviation cost of the reference state quantity, and the other part is the control quantity cost. The control quantity u k with the minimum cost is obtained by solving the optimization problem. is the system predicted displacement, x0 is the state of the system at the time of solving, x k is the system predicted state given by the predictor at the kth step. J is the objective optimization function with u k as the variable, and u k is obtained as the system control quantity input at the next step, which makes J have the minimum cost under the constraint condition.

[0225] As Figure 6 gives the working process of the valve-controlled hydraulic cylinder system model predictive control method based on Deep Koopman load port independent control. First, the hydraulic system operation data is collected, then it is divided into data sets and used for the training of neural network under data-driven, and the dimensionality function and Deep-Koopman prediction model are obtained. The dimensionality function refers to the mapping of n-dimensional state quantity to m-dimensional space through the dimensionality function. The state change of the previous n-dimensional state quantity is xk+1=f(xk,uk), and xk is mapped to zk in m-dimensional space, and zk+1=Azk+Buk can be obtained. The nonlinear function f is simplified to linear variation A and B. Each zk can be mapped back to xk through the C matrix. In this way, the prediction model is transformed from nonlinear to the first step of nonlinear dimensionality function and the following linear calculation.

[0226] The KMPC controller uses the trained high-dimensional space prediction model to calculate the control quantity, the input is the vector of the controlled object state parameter after dimensionality and the reference displacement, and the output is the control signal for the left main valve reference displacement and the right main valve spool reference displacement. Loop feedback, complete hydraulic system position control.

[0227] Those skilled in the art know that, in addition to implementing the system, device and each module thereof provided by the present application in the form of pure computer readable program code, the same program can also be implemented in the form of logic gate, switch, special integrated circuit, programmable logic controller and embedded microcontroller, etc. by logically programming the method steps. Therefore, the system, device and each module thereof provided by the present application can be considered as a hardware component, and the modules included therein for implementing various programs can also be considered as structures in the hardware component; the modules for implementing various functions can also be considered as both software programs for implementing methods and structures in the hardware component.

[0228] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily without conflict.

Claims

1. A load port independent control valve-hydraulic cylinder system model control method, characterized by, Comprising: Step S1: collecting hydraulic system operation data; Step S2: dividing the collected data into data sets, using a neural network to train a deep Koopman operator to upgrade the non-linear model, and training the neural network to obtain an upgrading function and a high-dimensional space prediction model; Step S3: using the trained high-dimensional space prediction model to calculate the control quantity of the KMPC controller, inputting the vector of the state parameters of the controlled object after upgrading and the reference displacement, and outputting the control signals for the left main valve reference displacement and the right main valve spool reference displacement, and feedbacking in a loop to complete the position control of the hydraulic system; In the step S2: The Koopman operator is a linear operator that acts on the observation function g(x), where g(x) is the linear function of the observation function g(x). k ) reduce the low-dimensional state variable space x k Transform into a high-dimensional state variable space z k The linear operator completes g(x) k Evolution of g(x) under nonlinear mapping f; observation function g(x) k ) evolved into z k =g(x k ,u k ), where x k Let u be the state variable of the system at time k. k The control input to the system at time k; For z k to z k+1 , a linear transformation equation is represented by z k+1 = Az k + Bu, A and B are linear matrices calculated by a neural network, used to represent a high-dimensional linear space, and the linear change from zk to zk+1; Koopman linearization is a method of mapping a nonlinear system through a g operator to a high-dimensional linear system whose state is updated by a Koopman operator; z is transformed by a C linear transformation k dimensional reduction into displacement state quantities The g operator is composed of a set of Koopman eigenfunctions, the Koopman operator is a linear operator, and the combination of the two operators satisfies the conversion from the nonlinear model to the linear model represented by equation (2); where C is a linear high-dimensional space z k linear variation matrix to the actual state variables of the system; Solving an observation function z by a deep neural network k = g(x k , u k ), z k ∈ R m , x k ∈ R n , u k ∈ R 2 and a Koopman operator; where R m is high-dimensional linear, with m spatial dimensions, R n is the actual system variable dimension space, and R 2 is a 2-dimensional space, as there are two input variables. The first set of hidden layer composite parameter records as OFP(x k , u k ) as part of the observation function; the first input layer contains the state quantities and input u k of the nonlinear model at time k; First output layer and state variable x k It is the state variable z of a high-dimensional linear model at time k. k Through a name called DKO(z) k u k The complete connection of ) yields the state variables z of the high-dimensional linear model at time k+1. k+1 The observation function is expressed as equation (3), where z′ k ∈R m -n It is the output of the first hidden layer; the linear system is represented by equation (4); The first set of hidden layers is used to learn the observation function, with internal connected layers and activation functions, z k is the output of the full connected layer; DKO(z k , u k ) represents the Deep Koopman operator, which is a Koopman operator calculated by fitting the upscaling function g using a neural network, which is an improvement on the Koopman operator; DKO(z k , u k ) is a full connected layer without bias, which is used to learn the Koopman operator in equation (4): z k+1 = DKo(z k , u k ) = Az k + Bu k (4) To obtain the prediction of the state variable x k of the nonlinear system, after the low-dimensional nonlinear model is upgraded to the high-dimensional linear model, the dimension reduction needs to be completed; the dimension reduction process through the third group of hidden layers of the deep neural network is recorded as DRP(z k ), and the input z k at time k in the linear space is obtained after passing through the hidden layer; the reconstructed values of the state variable x and the input variable u of the nonlinear system are obtained; the dimension reduction reconstruction of the system is completed, and the third group of hidden layers is a fully connected layer: [x k ,u k ] T = DRP(z k ) (5).

2. The load port independent control valve controlled hydraulic cylinder system model control method according to claim 1, characterized in that: The load port independent control valve controlled hydraulic cylinder system is realized by a digital hydraulic pilot programmable valve to independently control the pressures of the two cavities of the hydraulic cylinder; the pilot stage of the digital hydraulic pilot programmable valve is composed of four identical two-position three-way spool valve switches, and the main stage is composed of two identical three-position three-way spool valve switches; according to the preset load motion law, the system controller generates an instruction signal to control the opening degrees of the spool valve cores of the two main stages, so that the load moves according to the preset law; the load port independent control valve controlled hydraulic cylinder system utilizes the opening and closing dynamic characteristics of the pilot stage high-speed switch valve to independently control the opening degrees of the spool valve cores of the two main stages, and the opening degrees of the spool valve cores of the two main stages will determine the states of the two cavities of the hydraulic cylinder; The control strategy adopts a PID-KMPC feedback structure, and under the input reference displacement, the load displacement pushed by the oil cylinder tracks the reference displacement according to the preset requirements; the first stage KMPC controller is composed of an upgrading function and a linear model prediction controller, and generates control signals for the left main valve reference displacement and the right main valve spool reference displacement based on the feedback signals of the oil cylinder displacement, speed and two cavity pressures; the second stage PID2 is composed of PID2-1, PID2-2, PID2-3 and PID2-4, which respectively take the left spool displacement and the right spool displacement as feedback signals to generate control signals for controlling the opening and closing characteristics of the four pilot valves.

3. The load port independent control valve controlled hydraulic cylinder system model control method according to claim 1, characterized in that: The design of the deep neural network cost criterion is completed, and the loss function of the deep neural network is given: The loss function loss1 for measuring the prediction accuracy: lossl = ||(x k+1 ,u k+1 ) - DRP(DKO(z k ,u k ))|| (6) The loss function loss2 for measuring the reconstruction effectiveness: loss2 = ||(x k , u k ) - DRP(z k )‖ (7) The loss function loss3 for measuring the linearity of the system:

4. The load port independent control valve-hydraulic cylinder system model control method according to claim 1, characterized by, In the step S3: Based on the calculation results of the neural network, the KMPC controller is established; The controller solving process is represented by equations (9) and (10); where J is the cost function, Q, R are weight parameters, N is the prediction horizon, y k is the system predicted state, y r is the system reference state, u k is the system control input, the cost function is divided into two parts, one part is the deviation cost of the reference state, the other part is the control cost; solving the optimization problem to get the control input u k that makes the cost minimum is the system predicted displacement, x0 is the system state at the time of solving, x k is the system predicted state given by the predictor at the kth step; J is the objective optimization function with u k as the variable, and u k that makes the cost minimum under the constraint condition is solved as the system control input of the next step.

Images