JP2019125021A - Information processing device, information processing method, computer program, control device of internal combustion engine - Google Patents

Abstract

To reduce a processing load, shorten processing time, and improve efficiency of initial condition generation in control of an internal combustion engine using model prediction control.SOLUTION: An information processing device generates time-series signals of a prediction value of a state of a control target part of an internal combustion engine by applying preliminarily prepared time-series signals of an operation amount of an actuator to a model expression modeling a change of the state of the control target part depending on a change of the operation amount of the actuator; stores the operation amount of the actuator and the generated state of the control target part as an initial condition; obtains an optimum operation amount of the actuator in the initial condition by repeating evaluation while changing an operation amount which is input for an objective function using the initial condition and the estimated state using the model expression; stores the obtained optimum operation amount in association with the initial condition; and obtains an approximate expression which shows a relation between the initial condition and the optimum operation amount.SELECTED DRAWING: Figure 3

Description

  The present invention relates to an information processing apparatus.

  In each field, a control method using model predictive control (MPC) is used. Model predictive control is a control method of optimizing (finding an optimal solution) while predicting future response at each time with respect to some control parameter. For example, Patent Document 1 relates to control of an internal combustion engine using model predictive control, and it is a control element within a finite interval by iterative calculation using a model of a control element of the internal combustion engine (eg, turbocharger, exhaust gas recirculation). It has been described to optimize the operation of. For example, Patent Document 2 describes that a table for performing vehicle control and an approximate expression for calculating a control variable of the vehicle are created using model predictive control.

JP, 2011-253536, A Patent No. 5870632

  In Patent Document 1, an iterative calculation for finding an optimal solution is performed in real time at each time. Since a large amount of computation is required for this iterative calculation, the processing load is high, processing takes time, and there is a problem that real-time processing may not be possible. In Patent Document 2, since a table or an approximate expression obtained in advance by iterative calculation is created, real-time iterative calculation is not required. However, for complex systems in which many elements (for example, dozens of elements) have associated initial conditions in order to derive an optimal solution of control parameters, such as control of an internal combustion engine, for example, A table or approximation of dimensions is required. It is difficult to create multi-dimensional tables or approximate expressions by exhaustive trials for initial conditions that include such many elements, but it is difficult in Patent Document 2 to No mention is made of

  The present invention has been made to solve the above-described problems, and in the control of an internal combustion engine using model predictive control, the processing load is reduced and the processing time is shortened, and the efficiency of initial condition generation is improved. The aim is to

  The present invention has been made to solve at least a part of the above-mentioned problems, and can be realized as the following modes.

(1) According to one aspect of the present invention, an information processing apparatus is provided. The information processing apparatus includes a model expression storage unit, an initial condition storage unit, an optimal operation amount storage unit, an initial condition generation unit, a prediction processing unit, and a learning processing unit. The model formula storage unit stores in advance a model formula that models a change in the state of the control target portion of the internal combustion engine according to the change in the operation amount of the actuator of the internal combustion engine. The initial condition storage unit stores an initial condition including at least a time-series signal of an operation amount of the actuator and a time-series signal of a state of the control target unit. The optimal operation amount storage unit stores the initial operation condition at each time in the initial condition storage unit in association with an optimal operation amount which is an optimal operation amount of the actuator under the initial condition. The initial condition generation unit applies a time series signal of the operation amount of the actuator created in advance to the model equation in the model equation storage unit to obtain a time series signal of the predicted value of the state of the control target unit. It is generated and stored in the initial condition storage unit as the initial condition. The prediction processing unit changes the operation amount to be input for an objective function using the initial condition in the initial condition storage unit and the state estimated using the model expression in the model expression storage unit. While repeating the evaluation, the optimal operation amount is determined and stored in the optimal operation amount storage unit. The learning processing unit obtains an approximate expression representing the relationship between the initial condition and the optimal operation amount in the optimal operation amount storage unit.

  According to this configuration, the prediction control unit obtains in advance the optimal operation amount of the actuator of the internal combustion engine according to each initial condition using model predictive control, and stores it in the optimal operation amount storage unit. Then, the learning processing unit can obtain in advance an approximate expression that represents the relationship between the initial condition in the optimum operation amount storage unit and the optimum operation amount. For this reason, according to the present configuration, control of an actual internal combustion engine does not require real-time repetitive calculation, and by using the approximate expression obtained by the learning processing unit, each element in the initial condition is accommodated. The optimum operation amount of the actuator of the internal combustion engine according to each actual value (or estimated value) can be obtained quickly, and the processing load can be reduced and the processing time can be shortened. Further, when generating the initial condition, the initial condition generation unit uses a model expression that models a change in the state of the control target unit of the internal combustion engine according to the change in the operation amount of the actuator of the internal combustion engine. For this reason, when generating the initial condition, the amount of calculation can be reduced as compared with the case where a trial in which each element is comprehensively combined is performed, and efficient generation of the initial condition can be achieved.

(2) In the information processing apparatus of the above aspect, the learning processing unit may obtain the approximate expression by supervised learning of a neural network in which the initial condition and the optimal operation amount are teacher data. Since neural networks can perform complex function approximation, they are suitable for control of an internal combustion engine that can include many elements as initial conditions. According to this configuration, since the learning processing unit obtains the approximate expression using such a neural network, the accuracy of the approximate expression can be improved.

(3) The information processing apparatus according to the above aspect may further include an experimental plan processing unit that generates the time-series signal of the operation amount of the actuator using an experimental design method. According to this configuration, since the experimental design processing unit generates a time series signal of the operation amount of the actuator using the experimental design method, it is possible to generate a physically reasonable time series signal as a combination.

(4) In the information processing apparatus of the above aspect, the initial condition further includes at least one of a disturbance to the internal combustion engine, an output of the control target unit, and a target value of an output of the control target unit. If the initial condition includes the disturbance, the model equation models changes in the state according to changes in the manipulated variable and the disturbance, and the output includes the initial condition. In this case, in the model equation, changes in the state and the output according to changes in the manipulated variable may be modeled. According to this configuration, various factors such as disturbance to the internal combustion engine, the output of the control target unit, and the target value of the output of the control target unit can be considered as the initial condition.

(5) In the information processing apparatus of the above aspect, a linear state equation and a non-linear state equation may be used as the model equation. According to this configuration, linear equation of state and nonlinear equation of state can be used as model equations.

(6) In the information processing apparatus of the above aspect, a non-linear equation configured using an NARX model may be used as the model equation. According to this configuration, since a non-linear equation configured using the NARX model is used as a model equation, a highly accurate prediction result can be obtained.

(7) In the information processing apparatus of the above embodiment, as the experimental design method, a first method of generating a signal represented by a combination of a step function and a ramp function, and a chirp signal whose frequency changes depending on time. Any of the second methods of generating the signal to be represented may be used. According to this configuration, the first method (for example, the APRBS method) of generating a signal represented by a combination of a step function and a ramp function as an experimental design method, and a chirp signal whose frequency changes depending on time Either of the second method (e.g., Sinusoidal Excitation method) of generating a signal to be represented can be used.

(8) According to one aspect of the present invention, there is provided an information processing method using a model expression that models a change in a state of a control target portion of an internal combustion engine according to a change in an operation amount of an actuator of the internal combustion engine Ru. In this information processing method, a time-series signal of a predicted value of the state of the control target unit is generated by applying a time-series signal of the operation amount of the actuator created in advance to the model equation, and Storing the time-series signal of the operation amount of the actuator and the generated time-series signal of the state of the control target unit as an initial condition, the initial condition, and the state estimated using the model equation And the step of obtaining an optimal operation amount, which is an optimum operation amount of the actuator under the initial condition, by repeating the evaluation while changing the input operation amount with respect to the objective function using and. And the step of correlating and storing the determined optimal operation amount, and the step of determining an approximate expression representing the relationship between the initial condition and the optimal operation amount. According to this method, in the control of the internal combustion engine using model predictive control, it is possible to reduce the processing load and shorten the processing time, and to improve the efficiency of the initial condition generation.

(9) According to an aspect of the present invention, there is provided a computer program using a model expression that models a change in the state of a control target portion of an internal combustion engine according to a change in the operation amount of an actuator of the internal combustion engine. . In this computer program, a time-series signal of the predicted value of the state of the control target portion is generated by applying a time-series signal of the operation amount of the actuator created in advance to the model equation, and the actuator A function of storing, as an initial condition, a time-series signal of an operation amount of the control amount and a generated time-series signal of the state of the control target unit, the initial condition, and the state estimated using the model equation For the objective function using, the function of determining the optimum manipulated value, which is the optimum manipulated value of the actuator under the initial condition, by repeating the evaluation while changing the inputted manipulated variable, and the initial condition And a function of correlating and storing the determined optimal operation amount, and a function of determining an approximate expression representing a relationship between the initial condition and the optimal operation amount. That. According to this computer program, in control of an internal combustion engine using model predictive control, processing load can be reduced and processing time can be shortened, and efficient generation of initial conditions can be achieved.

(10) According to one aspect of the present invention, there is provided a control device of an internal combustion engine including an actuator and a control target portion. In this control device for an internal combustion engine, with respect to initial conditions including at least a time-series signal of an operation amount of the actuator and a time-series signal of a state of the control target portion, the initial condition and the initial condition at each time are described. Storage unit for storing an approximate expression representing a relationship with the optimum operation amount which is the optimum operation amount of the actuator, information for acquiring the actual operation amount of the actuator, and the state of the control target portion The optimum operation amount corresponding to the actual operation amount and the state is determined using the acquisition unit, the acquired operation amount and the state, and the approximate expression in the storage unit, and the optimum operation amount is obtained. And a controller configured to operate the actuator according to. According to this configuration, the control unit uses the operation amount and the state (each actual value or estimated value) acquired by the information acquisition unit, and the approximate expression in the storage unit to calculate the internal combustion engine according to each actual value. The optimum operation amount of the actuator can be obtained quickly, and the processing load can be reduced and the processing time can be shortened.

  The present invention can be realized in various aspects. For example, an information processing apparatus, an information processing method, an information processing system, and a computer for obtaining an approximate expression for control of an internal combustion engine using model predictive control Program, control device for internal combustion engine using model predictive control, control method, control system, computer program, production device for control device for internal combustion engine, production method, production system, computer program, for distributing these respective computer programs The present invention can be realized in the form of a server apparatus, a non-temporary storage medium storing the computer program, or the like.

It is a block diagram of an information processor as one embodiment of the present invention. It is a figure explaining a prediction learning process. It is a flowchart which shows the procedure of the process in a prediction learning process. It is a flowchart which shows the procedure of the process in a prediction learning process. It is a figure explaining each step of a prediction learning process. It is a block diagram of a control device of an internal combustion engine as one embodiment of the present invention. It is a flowchart which shows the procedure of the process in internal combustion engine control. An example of operation by internal-combustion engine control is shown. An example of the calculation time required for internal combustion engine control is shown.

<Information processing device>
FIG. 1 is a block diagram of an information processing apparatus 1 according to an embodiment of the present invention. The information processing apparatus 1 is an apparatus for obtaining an approximate expression to be used for control of an internal combustion engine by using model predictive control (MPC: Model Predictive Control). In the present embodiment, the following five elements a1 to a5 are exemplified as elements used for control of an internal combustion engine. Since the elements a1 to a5 are used as an initial condition of model predictive control, the elements a1 to a5 are collectively referred to as "initial conditions". The information processing apparatus 1 of the present embodiment finds an approximate expression representing a relationship between an initial condition consisting of the elements a1 to a5 and an optimum operation amount (optimum operation amount) of the actuator of the internal combustion engine corresponding to the initial condition. It is an apparatus. The approximate expression obtained by the information processing device 1 is mounted on a control device for an internal combustion engine and used for control of the internal combustion engine. Details will be described later.
(A1) Operating amount u of actuator of internal combustion engine
(A2) Disturbance w for internal combustion engine
(A3) State x of the control target portion of the internal combustion engine
(A4) Output y of control target of internal combustion engine
(A5) Target value r of output of control target portion of internal combustion engine

  (A1) The manipulated variable u is a physical quantity that represents the operating condition of one or more actuators that can be operated in the internal combustion engine. For example, the throttle opening degree, the EGR valve opening degree in an exhaust gas recirculation (EGR: Exhaust Gas Recirculation) system, or the like corresponds to the operation amount u. (A2) The disturbance w is one or more physical quantities that affect the output of the internal combustion engine, excluding the manipulated variable u. For example, the engine speed, the outside air temperature, the outside air pressure, etc. correspond to the disturbance w. (A3) The state x is a physical quantity that represents the state of one or more control target parts included in the internal combustion engine. For example, the exhaust temperature, the exhaust flow rate, and the like in the EGR system correspond to the state x.

(A4) The output y is a physical quantity that represents the output of one or more control target units included in the internal combustion engine. For example, the EGR rate in the EGR system, the supercharging pressure in the turbocharger, etc. correspond to the output y. (A5) The target value r is a target value of the output (that is, the element a4) of one or more control target parts included in the internal combustion engine. Each of the five elements a1 to a5 described above may include a plurality of items. For example, three items of the engine speed, the outside air temperature, and the outside air pressure may be included as the disturbance w. In addition, the items listed in the five elements a1 to a5 described above are merely examples, and various items can be adopted.

  The information processing apparatus 1 includes a storage unit 100, an information processing unit 200, a ROM, a RAM, and a communication unit (not shown), and the units are mutually connected by a bus (not shown). The storage unit 100 includes a hard disk, a flash memory, a memory card, and the like. The storage unit 100 includes a model expression storage unit 110, an initial condition storage unit 120, an optimal operation amount storage unit 130, and an approximate expression storage unit 140.

  The model storage unit 110 models changes in the state x of the control target portion of the internal combustion engine and the output y according to the change in the operation amount u of the actuator of the internal combustion engine and the disturbance w to the internal combustion engine. Model expressions are stored in advance. The model equation is obtained in advance by experiments, and includes current and past information of the manipulated variable u, the disturbance w, the state x, and the output y. How much past information is included depends on the number of time samples in the model equation. As a model equation, for example, a nonlinear equation constructed using a linear equation of state, a nonlinear equation of state, or a nonlinear auto-regressive eXogenous (NARX) model can be used.

A model equation of a continuous time system using a linear equation of state can be expressed, for example, as follows for each variable of the manipulated variable u, the disturbance w, the state x, and the output y. Here, t is time, and A, B, C, D, E and F are constant matrices of appropriate sizes.
-Equation of state: dx / dt = Ax + Bu + Ew (1)
Output equation: y = Cx + Du + Fw (2)

A model equation of a discrete time system using a linear equation of state can be expressed, for example, as follows, using each of the above-mentioned variables and a discrete time k.
-Equation of state: x [k + 1] = Ax [k] + Bu [k] + Ew [k] (3)
Output equation: y [k] = Cx [k] + Du [k] + Fw [k] (4)

A model equation of a continuous time system using a nonlinear equation of state can be expressed, for example, as follows, using each of the variables described above. Here, f and g are non-linear vector functions giving an output of the same dimension as x and y, respectively. For this reason, it can be said that the nonlinear equation of state includes a linear equation of state.
-Equation of state: dx / dt = f (x, u, w) (5)
Output equation: y = g (x, u, w) (6)

A model equation of a discrete time system using a nonlinear equation of state can be expressed, for example, as follows, using each of the above-described variables and a discrete time k.
State equation: x [k + 1] = f (x [k], u [k], w [k]) (7)
Output equation: y [k] = g (x [k], u [k], w [k]) (8)

The model equation of the nonlinear equation constructed using the NARX model can be expressed, for example, as follows, using the above-mentioned variables and the discrete time k. Here, g is a non-linear vector function giving an output of the same dimension as y, and is different from the above-mentioned equation 8 in that plural past times are explicitly sampled.
Y [k + 1] = g (y [k], y [k-1], ..., y [k + 1- _ny ], u [k], u [k-1], ..., u [ k + 1−n _u ]) (9)

  In this way, linear equation of state, nonlinear equation of state, and NARX model can be used as model equations.

  FIG. 2 is a diagram for explaining prediction learning processing. FIG. 2A shows an example of the initial conditions stored in the initial condition storage unit 120. FIG. 2 (B) shows an example of change of each element in each iterative cycle of the predictive learning process. The initial condition storage unit 120 of FIG. 1 includes an operation amount storage unit 121, a disturbance storage unit 122, a state storage unit 123, an output storage unit 124, and a target value storage unit 125.

  When the operation amount storage unit 121 represents a change in physical quantity (in other words, a change in physical quantity over time) for a plurality of times of the operation quantity u of the actuator of the internal combustion engine by step S110 of prediction learning processing described later. Sequence signals are stored. In FIG. 2A, time series signals of two items belonging to the operation amount u are illustrated as the operation amount u1 and the operation amount u2. Similarly, in the disturbance storage unit 122, a time-series signal representing a change in physical quantity for a plurality of times of the disturbance w with respect to the internal combustion engine is stored in step S110 of prediction learning processing described later. In FIG. 2A, a time-series signal of one item belonging to the disturbance w is illustrated as the disturbance w1.

  The state storage unit 123 stores time-series signals representing changes in physical quantities for a plurality of times of the state x of the control target portion of the internal combustion engine in step S130 of the prediction learning process described later. In FIG. 2A, a time-series signal of one item belonging to the state x is illustrated as the state x1. Similarly, in the output storage unit 124, a time-series signal representing a change in physical quantity for a plurality of times of the output y of the control target portion of the internal combustion engine is stored in step S130 of prediction learning processing described later. In FIG. 2A, a time-series signal of one item belonging to the output y is illustrated as the output y1.

  The target value storage unit 125 stores time-series signals representing changes in physical quantities for a plurality of times of the target value r of the output of the control target portion of the internal combustion engine in step S110 of prediction learning processing described later. In FIG. 2A, the illustration of the time-series signal of the target value r is omitted.

  In the optimum operation amount storage unit 130, the initial conditions (operation amount u, disturbance w, state x, output y, target value r) at each time in the initial condition storage unit 120 are obtained in step S200 of prediction learning processing described later. On the other hand, the optimum operation amount (optimum operation amount) of the actuator of the internal combustion engine under the initial condition is stored in association with it. The approximate expression storage unit 140 stores the approximate expression obtained from the initial condition and the optimal operation amount in the optimal operation amount storage unit 130 in step S300 of the prediction learning process described later.

  The information processing unit 200 controls each unit of the information processing apparatus 1 by expanding a computer program stored in the ROM in the RAM and executing it. In addition, the information processing unit 200 functions as an experimental plan processing unit 210, an initial condition generation unit 220, a prediction processing unit 230, and a learning processing unit 240, and cooperates to execute prediction learning processing described later. The experimental design processing unit 210 generates a time series signal of the operation amount u, the disturbance w, and the target value r among the initial conditions using the experimental design method, and stores the time series signal in the initial condition storage unit 120. The initial condition generation unit 220 generates time-series signals of the state x and the output y among the initial conditions and causes the initial condition storage unit 120 to store the time-series signals.

  The prediction processing unit 230 uses the initial conditions in the initial condition storage unit 120 to perform model predictive control to optimize the initial conditions (operation amount u, disturbance w, state x, output y, target value r) at each time. The operation amount is obtained and stored in the optimum operation amount storage unit 130. The learning processing unit 240 obtains an approximate expression by supervised learning of a neural network (NN: Neural Network) in which the initial conditions and the optimum operation amount in the optimum operation amount storage unit 130 are supervised data, and Remember.

  FIG.3 and FIG.4 is a flowchart which shows the procedure of the process in a prediction learning process. The predictive learning process generates initial conditions (operation amount u, disturbance w, state x, output y, target value r), and the generated initial conditions and optimum operation amount of the actuator of the internal combustion engine (optimum operation amount) It is the process which calculates | requires the approximate expression showing the relationship with and. The prediction learning process is performed in the information processing device 1 at an arbitrary timing.

  FIG. 5 is a diagram for explaining each step of the predictive learning process. In FIG. 5, each element of the initial conditions stored as a time-series signal in the initial condition storage unit 120 is schematically represented. The vertical axis represents the name of each element of the initial condition, and the horizontal axis represents each time in the time-series signal. Usually, the physical quantities of the corresponding element at the corresponding time are displayed in the table, but in FIG. 5 the wording for the explanation is described by omitting the display of the physical quantities for the convenience of explanation. Furthermore, for convenience of explanation, FIGS. 3 to 5 do not distinguish among a plurality of items that each element of the initial condition may contain. For example, when the operation amount u includes two items, the operation amount u1 and the operation amount u2, the processing for the operation amount u1 and the operation amount u2 may be the same processing as the processing for “operation amount u” described below .

  In step S100, generation of various initial conditions (operation amount u, disturbance w, state x, output y, target value r) is executed. Specifically, in step S110, the experimental design processing unit 210 generates a time series signal of the operation amount u, the disturbance w, and the target value r using the experimental design method, and the operation amount storage unit 121 and the disturbance storage unit 122. , And the target value storage unit 125 respectively. The experimental design processing unit 210 can use, for example, any of the following methods b1 and b2 as the experimental design method.

(B1) A first method of generating a signal represented by a combination of a step function and a ramp function: As a first method, for example, APRBS (Amplitude modulated Pseudo Random Binary Sequences) method can be used. In the APRBS method, a continuous signal combination is generated pseudorandomly using a Latin hypersquare design or a D-optimal design which handles the signal levels continuously and efficiently generates various combinations. In the APRBS method, the length of one section of the signal and the signal change rate at the transition of the section can also be included in the plan, so in addition to the combination of the signal values, the combination of the signal change rates is also diverse. Can be secured. That is, in the APRBS method, an experimental design including transient change is possible.

(B2) A second method of generating a signal represented by a chirp signal whose frequency changes depending on time: As a second method, for example, the Sinusoidal Excitation method can be used. In the Sinusoidal Excitation method, a combination of continuous signals is generated using sinusoidal signals whose frequency changes with time. The time change rate of the signal level is more varied than the signal generated by the APRBS method.

  In step S120, the prediction calculation unit 221 of the initial condition generation unit 220 calculates the model expression stored in the model expression storage unit 110, the time series signal of the operation amount u stored in the operation amount storage unit 121, and the disturbance. The time-series signals of the disturbance w stored in the storage unit 122 are read out. In step S130, the prediction calculation unit 221 of the initial condition generation unit 220 applies (applies) the read time series signals of the operation amount u and the disturbance w to the model equation to obtain predicted values of the state x and the output y. A time-series signal is generated and stored in the state storage unit 123 and the output storage unit 124, respectively.

  As described above, in step S100 of the predictive learning process, generation of initial conditions that are unlikely to occur physically can be avoided and various initial conditions can be efficiently generated by using a model expression. As described above, the initial condition consists of five elements of the manipulated variable u, the disturbance w, the state x, the output y, and the target value r, but a simple method of generating a combination of those five elements is five elements. The upper and lower limits are set for all, and five elements are comprehensively combined within the upper and lower limits. However, the initial conditions generated by the exhaustive combination include initial conditions that are extremely unlikely to occur physically. This is because the behavior of the state x and the output y is governed by a physical causality that depends on the operation amount u and the disturbance w. In this regard, in step S100 of the prediction learning process, it is physically impossible as a combination by using the model expression reproducing this causal relationship not only from model predictive control in step S200 onwards, but also from the stage of generation of initial conditions. It is possible to generate time-series signals of state x and output y without. As a result, it is possible to efficiently generate only the initial conditions that may actually occur. For example, as in the case of control of an internal combustion engine, many elements (eg, dozens of elements are required to derive an optimal solution of control parameters). Even in a complex system with an initial condition associated with), the initial condition can be generated without causing an explosive increase in the number of combinations.

  In step S200, the optimum operation amount corresponding to the various initial conditions generated in step S100 is obtained. Specifically, in step S210, the prediction calculation unit 231 of the prediction processing unit 230 reads out the model expression stored in the model expression storage unit 110.

  In step S220, the prediction calculation unit 231 of the prediction processing unit 230 uses one time in the time-series signal of the initial condition stored in the initial condition storage unit 120 as a starting point, and is necessary for future prediction by model prediction control. Read information for time. For example, in FIG. 5A, when time t0 is the starting point, the prediction calculation unit 231 determines information on current and past times necessary for prediction, specifically, an initial condition of the current time t0. (Disturbance w, state x, output y, target value r) and an initial condition (operation amount u) of past time t-1 are read out. Here, the reason for reading out the initial condition of the past time t-1 only for the operation amount u is because the operation amount u at the current time t0 is to be predicted in the later step. In addition to the current initial conditions, the prediction calculation unit 231 may also read the initial conditions at several times in the past for the disturbance w and the like. How far back in time the information is read out depends on the model expression used by the prediction calculation unit 231.

  In step S230, the optimum operation amount corresponding to the initial condition read out in step S220 is determined. Specifically, in step S231, the prediction calculation unit 231 of the prediction processing unit 230 determines a time series of the operation amount u, the disturbance w, and the target value r in a finite interval from the current time to a predetermined future time. This finite interval corresponds to "prediction horizon" in model predictive control. For example, as shown in FIG. 5B, the prediction calculation unit 231 determines the operation amount u, the disturbance w, and the target value r in a finite interval from the current time t0 to a predetermined future time t5. In the example of FIG. 5 (B), the prediction calculation unit 231 sets a predetermined default value (FIG. 5: D) as the operation amount u, and sets future values t1 to t5 of the disturbance w and the target value r. The physical quantity (FIG. 5: present) of the current time t0 read in step S220 is set. When determining the default value of the operation amount u, the prediction calculation unit 231 may consider the operation amount u of the past time t-1 read in step S220.

  In step S232, the prediction calculation unit 231 of the prediction processing unit 230 predicts the state x and the output y in a future finite interval using a model expression. Specifically, the prediction calculation unit 231 sets the condition of the current time as the initial condition (operation amount u, disturbance w, state x, output y, target value r) read in step S220, and the model formula read in step S210 On the other hand, by applying (applying) the time-series signal of the manipulated variable u and the disturbance w within the finite interval determined in step S231, the time-series signal of the predicted value of the state x and the output y within the finite interval Generate For example, as shown in FIG. 5C, the prediction calculation unit 231 predicts the state x and the output y in a finite interval from predetermined future times t1 to t5 (FIG. 5: prediction).

  In step S233, the evaluation unit 232 of the prediction processing unit 230 determines the time series signal of the operation amount u, the disturbance w, and the target value r in the finite interval determined in step S231, and the finite interval predicted in step S232. The objective function is evaluated by inputting the time-series signal of the state x and the output y to the objective function. The objective function is a predetermined expression for quantitatively evaluating the control performance, and is stored in the storage unit 100. As the objective function, for example, the square sum of the difference between the output (output y) and the target (target value r) in a finite interval can be used.

  In step S234, the iterative processing unit 233 of the prediction processing unit 230 determines whether the value of the objective function has converged. When it converges (step S234: YES), the iterative processing unit 233 causes the process to transition to step S236.

  On the other hand, if convergence has not occurred (step S234: NO), the iterative processing unit 233 corrects the time series signal of the manipulated variable u in the finite interval so as to improve the objective function value, and shifts the processing to step S232. Let predictions and evaluations repeat. For example, as shown in FIG. 5D, the iterative processing unit 233 corrects the operation amount u in the finite interval from the current time t0 to a predetermined future time t5, and then the prediction calculation unit 231 performs the process shown in FIG. As shown in E), the state x based on the corrected operation amount u and the output y are predicted, and the evaluation unit 232 evaluates an objective function using these. For example, as shown in FIG. 2B, as the manipulated variable u1 and the manipulated variable u2 within the finite interval (prediction horizon HP) are corrected to the repetitive cycles C1, C2, and C3, corresponding finite intervals (prediction horizon (prediction horizon) The state x1 and the output y1 in HP) also change as shown in cycles C1, C2, and C3. In the third cycle C3 in which the value of the objective function converges, it can be seen that the time-series signal of the output y1 substantially matches the time-series signal of the target value r. The iterative processing unit 233 may use known methods such as a gradient method, a shooting method, and a C / GMRES method.

  In step S236, the iterative processing unit 233 of the prediction processing unit 230 extracts, from the time series signal of the operation amount u in the finite section, the operation amount u for one time as the starting point as the “optimum operation amount”. Then, the iterative processing unit 233 associates the optimal operation amount with the initial conditions (operation amount u, disturbance w, state x, output y, target value r) read in step S220, and the optimal operation amount storage unit. Make it memorize in 130. For example, as illustrated in FIG. 5F, the iterative processing unit 233 sets the operation amount u at the current time t0 as the optimum operation amount, and the initial condition read in step S220 (the operation amount u at the past time t-1, the current The disturbance w at time t0, the state x, the output y, and the target value r) are associated with each other and stored in the optimal operation amount storage unit 130.

  In step S240, the iterative processing unit 233 of the prediction processing unit 230 determines whether or not the processing for all the time has been completed for the time series of the initial conditions stored in the initial condition storage unit 120. When the processing for all the times has not been completed (step S240: NO), the repetitive processing unit 233 advances the starting time to one time in step S250, and repeats the processing after step S220. When the processing for all the times has been completed (step S240: YES), the iterative processing unit 233 causes the process to transition to step S300.

  As described above, in step S200 of the predictive learning process, model predictive control is performed on various initial conditions generated in step S100, and an optimal operation amount corresponding to the initial conditions at each time is determined. As described above, the prediction processing unit 230 moves the time serving as the starting point for reading out the initial condition by one time, and obtains the optimum operation amount for the initial condition of each time at all time of the time series signal Do it for a minute. Therefore, the prediction processing unit 230 can obtain the optimal operation amount for various initial conditions, and store the optimal operation amount in the optimal operation amount storage unit 130.

  In step S300, an approximate expression representing the relationship between the initial condition and the optimal operation amount stored in the optimal operation amount storage unit 130 is obtained by machine learning. Specifically, in step S310, the NN calculation unit 241 of the learning processing unit 240 reads all sets of the initial conditions and the optimal operation amount stored in the optimal operation amount storage unit 130. In step S320, the NN calculating unit 241 of the learning processing unit 240 initializes parameters in the neural network (NN).

  In step S330, the NN calculating unit 241 of the learning processing unit 240 obtains an approximate expression by supervised learning of the NN using the initial condition and the optimal operation amount as teacher data. Specifically, the NN calculation unit 241 gives (applies) the initial condition read out in step S310 to the NN, and obtains the output of the NN. In step S340, the evaluation unit 242 of the learning processing unit 240 inputs the output of the NN and the initial conditions and the optimal operation amount read out in step S310 into the objective function, and evaluates the objective function to obtain an optimal NN. Evaluate the approximation accuracy (error) of the manipulated variable. As this objective function, for example, the square sum of the difference between the output of the NN and the read optimum manipulated variable can be used.

  In step S350, the iterative processing unit 243 of the learning processing unit 240 determines whether the value of the objective function has converged. If the convergence has occurred (step S350: YES), the iterative processing unit 243 causes the process to transition to step S370. On the other hand, if convergence has not occurred (step S350: NO), the iterative processing unit 243 corrects the parameters of the NN so as to improve the objective function value, transitions the process to step S330, and repeats the NN output and evaluation. . The iterative processing unit 243 may use a known method such as back propagation.

  In step S370, the iterative processing unit 243 of the learning processing unit 240 stores the latest NN and the parameters of the NN in the approximate expression storage unit 140 as a learned NN, and ends the processing.

  As described above, in step S300 of the predictive learning process, an approximate expression that represents the relationship between the initial condition and the optimal operation amount stored in the optimal operation amount storage unit 130 is determined by machine learning. The relationship between the initial condition and the optimal manipulated variable generated in step S200 can be expressed as a non-linear function in the form of determining the optimal manipulated variable if the initial condition is determined. In step S300, an approximate expression is created in advance from this relationship. By using this approximate expression, it is possible to obtain the optimum manipulated value at high speed without performing iterative calculation (model predictive control) for finding the optimum solution in real time at each time in control of an actual internal combustion engine. It becomes. Here, since the relationship between the initial condition and the optimal manipulated variable is generally extremely non-linear, and many elements are associated with the initial condition, the multi-input multi-output high-dimensional relationship is obtained. Good approximation accuracy can not be expected in a general linear form or an n-th order polynomial. In this regard, in step S300 of the predictive learning process, machine learning by NN is used as a method capable of efficiently approximating multiple inputs and multiple outputs and strong non-linearity. When learning the relationship between the initial condition and the optimum manipulated value using NN, the back propagation can calculate the influence of the parameter in NN on the objective function value, and based on the information, it is nonlinear by iterative calculation based on the gradient method. It becomes possible to learn to reproduce the relationship well. Thus, the NN learned with sufficient accuracy can realize control performance (predictive performance) equivalent to real-time model predictive control with a low computational load.

  As described above, according to the information processing apparatus 1, the prediction processing unit 230 uses the model prediction control to set each initial condition (operation amount u, disturbance w, state x, output y, target value r) in advance. The optimum operation amount of the actuator of the internal combustion engine according to the request is determined and stored in the optimum operation amount storage unit 130 (FIG. 3: step S200). Then, the learning processing unit 240 can obtain in advance an approximate expression (learned NN and parameter) representing the relationship between the initial condition in the optimum operation amount storage unit 130 and the optimum operation amount (FIG. 4: step S300). ). For this reason, according to the information processing apparatus 1 of this configuration, in the control of an actual internal combustion engine, the approximation formula (learned NN and parameters) obtained by the learning processing unit 240 is used without requiring real-time repetitive calculation. By doing this, it is possible to quickly obtain the optimum operating amount of the actuator of the internal combustion engine in accordance with each actual value (or estimated value) corresponding to each element in the initial conditions, thereby reducing the processing load and shortening the processing time. Can be In addition, when generating the initial condition, the initial condition generation unit 220 determines the state x (and the output y) of the control target unit of the internal combustion engine according to the change in the operation amount u (and the disturbance w) of the actuator of the internal combustion engine. Use a model equation that models the change of. For this reason, when generating the initial condition, the amount of calculation can be reduced as compared with the case where a trial in which each element is comprehensively combined is performed, and efficient generation of the initial condition can be achieved.

<Control device for internal combustion engine>
FIG. 6 is a block diagram of a control device 3 of an internal combustion engine according to an embodiment of the present invention. The control device 3 is a device that controls an internal combustion engine by using the approximate expression created by the information processing device 1. As an internal combustion engine, a diesel engine, a gasoline engine, etc. are mentioned, for example. The control device 3 includes a driver interface (IF) 310, an ECU (Electronic Control Unit) 320, and a hardware system 330, and the respective parts are mutually connected by an in-vehicle network (not shown). The driver interface 310 is an interface for acquiring an operation signal by the driver of the internal combustion engine, and is, for example, an accelerator, a brake, or the like.

  The ECU 320 is a microcontroller (microcomputer) that controls the operation of the internal combustion engine. In addition, the ECU 320 functions as a target value determination unit 321, a control unit 322, and an information acquisition unit 323, and cooperates to execute internal combustion engine control described later. The target value determination unit 321 determines a target value r of the output of the control target portion of the internal combustion engine according to the operation signal by the driver.

  The control unit 322 obtains the optimal operation amount of the actuator 331 using the approximate expression (learned NN and parameters) obtained by the information processing device 1, and operates the actuator 331. The approximate expression may be stored in advance in a storage unit (not shown) in the control unit 322, for example, at the time of manufacture of a vehicle equipped with the control device 3. The control unit 322 may communicate with the information processing apparatus 1 via a communication network (not shown), periodically obtain an approximate expression, and store the approximate expression in the storage unit in the control unit 322.

  The information acquisition unit 323 acquires each actual value of each element (operation amount u, disturbance w, state x, output y) corresponding to the initial condition based on the detection value acquired by the sensor 333. Note that the information acquisition unit 323 obtains estimated values of the operation amount u, the disturbance w, the state x, and the output y from the actual values acquired by the sensor 333 using an observer or a Kalman filter based on the model of the control target. May be In the control of an internal combustion engine, many elements (for example, several tens) are required to obtain an optimal manipulated variable. For this reason, the information acquisition unit 323 may use an actual value acquired by the sensor 333 for a part of the elements and use estimated values estimated for the other elements. The target value r corresponding to the initial condition is separately set by the target value determination unit 321.

  Hardware system 330 is hardware installed in an internal combustion engine. The actuator 331 is one or more actuators operable in the internal combustion engine, such as a throttle and an EGR valve in an EGR system. The control target unit 332 is one or more control target units included in the internal combustion engine, and is an EGR system, a supercharger, or the like. The sensor 333 is a sensor for detecting each of the operation state (operation amount u) of the actuator 331, the state x of the control target portion 332, the output y of the control target portion 332, and the disturbance w to the internal combustion engine.

  FIG. 7 is a flowchart showing the procedure of processing in internal combustion engine control. The internal combustion engine control is processing for controlling the internal combustion engine (specifically, the actuator 331) using the optimal operation amount obtained using the approximate expression created by the information processing device 1. The internal combustion engine control shown in FIG. 7 is started, for example, when the vehicle equipped with the control device 3 is started, and is repeatedly executed at predetermined control cycles.

  In step S410, the driver interface 310 acquires an operation signal by the driver of the internal combustion engine from the accelerator opening degree and the brake opening degree, and transmits the operation signal to the target value determination unit 321. In step S420, the target value determination unit 321 of the ECU 320 uses the acquired operation signal (operation signal by the driver) and the current operation amount u acquired from the information acquisition unit 323, the disturbance w, the state x, and the output y. Then, the target value r for the output y of the control unit 322 is determined and transmitted to the control unit 322.

  In step S430, the control unit 322 of the ECU 320 approximates the acquired target value r, the current operation amount u acquired from the information acquisition unit 323, the disturbance w, the state x, and the output y (learned NN and parameters) Apply to As a result, the control unit 322 can determine an optimal operation amount (optimal operation amount) of the actuator 331 of the internal combustion engine to achieve the target value r specified by the target value determination unit 321. In step S440, the control unit 322 of the ECU 320 operates the actuator 331 with the determined optimal operation amount.

  In step S450, as a result of the operation of the actuator 331, the state x and the output y of the control target unit 332 change. In step S460, the sensor 333 detects and detects the latest information regarding the operation state (operation amount u) of the actuator 331, the state x of the control target portion 332, the output y of the control target portion 332, and the disturbance w for the internal combustion engine. The signal is transmitted to the information acquisition unit 323. In step S470, the information acquisition unit 323 of the ECU 320 calculates actual values or estimated values of the operation amount u, the disturbance w, the state x, and the output y based on the acquired latest information.

  FIG. 8 shows an example of operation by internal combustion engine control. In FIG. 8, three items of operation amounts u1 to u3 as operation amount u, four items of disturbances w1 to w4 as disturbance w, seven items of states x1 to x7 as state x, and two items of outputs y1 to y2 as output y, The specific example at the time of performing internal combustion engine control is shown using two items of the target value r1-r2 as the target value r. FIG. 8A shows time-series signals of the operation amount u and the disturbance w, and FIG. 8B shows time-series signals of the output y and the target value r. For convenience of illustration, the time-series signal of the state x is omitted. As shown in FIG. 8 (B), the actual value of the output y is the set target value r even in the complex case of multiple inputs and multiple outputs that require many elements to obtain the optimum manipulated variable. Properly follow.

  FIG. 9 shows an example of calculation time required for internal combustion engine control. The horizontal axis in FIG. 9 represents time, and the vertical axis represents time (ms) required for calculation by the ECU 320 of the control device 3. As shown in FIG. 9, it can be seen that the calculation time per one operation of the ECU 320 is 0.03 ms to 0.06 ms, and is within 0.2 ms at the maximum. For example, in the conventional example in the case of performing control of an internal combustion engine in real time using the C / GMRES method known as a high-speed solution for model predictive control, the computation time was about 60 ms (Nakata et al., " Application of C / GMRES model predictive control to diesel engine intake and exhaust system "). Compared to this conventional example, in the internal combustion engine control of this embodiment, the calculation speed can be increased by about 1000 times or more.

  As described above, according to the control device 3 of the internal combustion engine, the control unit 322 of the ECU 320 causes the actual value or estimated value of the operation amount u, the disturbance w, the state x, and the output y acquired by the information acquiring unit 323 The optimum operation amount of the actuator 331 of the internal combustion engine according to each actual value or estimated value can be quickly obtained using the approximate expression in the storage unit of the control unit 322, and the processing load is reduced and the processing time is shortened. Can be

<Modification of this embodiment>
The present invention is not limited to the above embodiment, and can be implemented in various aspects without departing from the scope of the present invention. For example, the following modifications are possible.

[Modification 1]
In the above embodiment, an example of the configuration of the information processing apparatus has been shown. However, the configuration of the information processing apparatus can be variously modified. For example, the information processing apparatus may be configured by cooperation of a plurality of information processing apparatuses arranged on a network. In this case, for example, at least a part of the experiment plan processing unit, the initial condition generation unit, the prediction processing unit, and the learning processing unit may be realized by a different information processing apparatus.

[Modification 2]
In the said embodiment, an example of the element which should be considered as an initial condition in model predictive control was mentioned. However, in model predictive control, at least a part of the elements a1 to a5 to be considered as initial conditions may be omitted, and further other elements may be considered. Specifically, at least a part of the disturbance w of the element a2, the output y of the element a4, and the target value r of the element a5 may be omitted. For example, when the disturbance w is omitted, the parameter of the disturbance w in the model equation can be omitted. Further, initial condition generation of the disturbance w in step S100 of the prediction learning process (FIG. 3) and consideration of the disturbance w as the initial condition in step S200 can be omitted. Further, consideration of the disturbance w in steps S420, S430, and S470 of the internal combustion engine control (FIG. 7) may be omitted. Similarly for the output y and the target value r, in the case of omission, consideration of the omitted elements may be omitted for each of the model equation, the prediction learning process, and the internal combustion engine control.

[Modification 3]
In the said embodiment, an example of the prediction learning process was shown (FIG. 3, FIG. 4). However, the prediction learning process can be variously modified. For example, in step S100, a time-series signal such as the operation amount u may be generated without using the experimental design method. For example, in step S300, an approximate expression may be obtained by using a means other than the NN, such as a support vector machine (SVM). For example, steps S100, S200, and S300 may not be performed as a series of processes, but may be performed individually.

  For example, in the predictive learning process, some steps described above may be omitted, and other steps may be additionally performed. Specifically, for example, in step S100, a time-series signal of at least a part of the initial conditions is generated by the first method (experimental design method), and the remaining time-series signals are generated by the second method. May be For example, a plurality of model formulas are stored in advance in the model formula storage unit, and in step S130 or step S232, the model formula to be applied is changed according to at least a part of the initial conditions and the actual value. It is also good. For example, after the end of step S300, the generated approximate expression may be distributed to the control device of the internal combustion engine.

[Modification 4]
In the said embodiment, an example of a structure of the control apparatus of an internal combustion engine was shown. However, the configuration of the control device for an internal combustion engine can be variously modified. For example, the control device may further include the function of the above-described information processing device for obtaining an approximate expression to be used for control of an internal combustion engine by using model predictive control.

[Modification 5]
In the said embodiment, an example of internal combustion engine control was shown (FIG. 7). However, the prediction learning process can be variously modified. For example, some steps described above may be omitted, and other steps may be additionally performed. Specifically, for example, after the end of step S470, actual values of the acquired operation amount u, disturbance w, state x, and output y may be transmitted to the information processing apparatus to help improve the accuracy of the approximate expression. .

  As mentioned above, although this aspect was demonstrated based on embodiment and a modification, embodiment of the above-mentioned aspect is for making an understanding of this aspect easy, and does not limit this aspect. The present embodiment can be modified and improved without departing from the spirit and the scope of the claims, and the present embodiment includes the equivalents thereof. In addition, if the technical feature is not described as essential in the present specification, it can be deleted as appropriate.

DESCRIPTION OF SYMBOLS 1 ... Information processing apparatus 3 ... Control apparatus 100 of internal combustion engine 100 ... Storage part 110 ... Model type storage part 120 ... Initial condition storage part 121 ... Operation amount storage part 122 ... Disturbance storage part 123 ... State storage part 124 ... Output storage part 125 ... Target value storage unit 130 ... Optimal operation amount storage unit 140 ... Approximate expression storage unit 200 ... Information processing unit 210 ... Experiment plan processing unit 220 ... Initial condition generation unit 221 ... Prediction calculation unit 230 ... Prediction processing unit 231 ... Prediction calculation unit 232 evaluation unit 233 iterative processing unit 240 learning processing unit 241 NN calculating unit 242 evaluation unit 243 iterative processing unit 310 driver interface 321 target value determination unit 322 control unit 323 information acquisition unit 330 hardware Wear system 331 ... actuator 332 ... control target portion 333 ... sensor

Claims (10)

An information processing apparatus,
A model expression storage unit for storing in advance a model expression that models a change in a state of a control target portion of the internal combustion engine according to a change in an operation amount of an actuator of the internal combustion engine;
An initial condition storage unit storing an initial condition including at least a time series signal of an operation amount of the actuator and a time series signal of a state of the control target unit;
An optimal operation amount storage unit that stores, in association with the initial condition at each time in the initial condition storage unit, an optimal operation amount that is an optimal operation amount of the actuator under the initial condition;
The time series signal of the predicted value of the state of the control target unit is generated by applying the previously generated time series signal of the operation amount of the actuator to the model expression in the model expression storage unit, and the initial condition An initial condition generation unit to be stored in the initial condition storage unit as
The evaluation is repeated while changing the input operation amount for an objective function using the initial condition in the initial condition storage unit and the state estimated using the model expression in the model expression storage unit. A prediction processing unit which obtains the optimal operation amount by the equation and stores the optimal operation amount in the optimal operation amount storage unit;
A learning processing unit for obtaining an approximate expression representing a relationship between the initial condition and the optimal operation amount in the optimal operation amount storage unit;
An information processing apparatus comprising: The information processing apparatus according to claim 1, wherein
The information processing apparatus, wherein the learning processing unit obtains the approximate expression by supervised learning of a neural network using the initial condition and the optimal operation amount as teacher data. The information processing apparatus according to claim 1 or 2, further comprising:
An information processing apparatus, comprising: an experiment plan processing unit that generates the time-series signal of the operation amount of the actuator using an experiment design method. The information processing apparatus according to any one of claims 1 to 3, wherein
In the initial condition, further,
Disturbance to the internal combustion engine,
An output of the control target unit;
A target value of the output of the control target unit;
Contains at least a portion of
When the initial condition includes the disturbance, the model equation models a change in the state according to a change in the operation amount and the disturbance.
The information processing apparatus according to any one of the first to third aspects, wherein when the output includes the output, the change in the state and the output according to a change in the operation amount is modeled in the model expression. The information processing apparatus according to any one of claims 1 to 4, wherein
An information processing apparatus using linear equation of state and nonlinear equation of state as the model equation. The information processing apparatus according to any one of claims 1 to 5, wherein
An information processing apparatus using a non-linear equation configured using an NARX model as the model equation. The information processing apparatus according to any one of claims 4 to 6, which is dependent on claim 3 or claim 3,
As the experimental design method, a first method of generating a signal represented by a combination of a step function and a ramp function, and a second method of generating a signal represented by a chirp signal whose frequency changes depending on time An information processing apparatus that uses either or. An information processing method using a model expression that models a change in a state of a control target portion of an internal combustion engine according to a change in an operation amount of an actuator of the internal combustion engine,
Applying a time-series signal of the operation amount of the actuator created in advance to the model equation to generate a time-series signal of the predicted value of the state of the control target unit;
Storing, as an initial condition, a time-series signal of an operation amount of the actuator and a generated time-series signal of the state of the control target unit;
For the objective function using the initial condition and the state estimated using the model equation, the evaluation operation is repeated while changing the operation amount to be input, whereby the optimum operation of the actuator under the initial condition is performed. Determining an optimal operation amount which is an amount
Storing the determined optimal operation amount in association with the initial condition;
Obtaining an approximate expression representing a relationship between the initial condition and the optimal operation amount;
An information processing method comprising: A computer program using a model formula modeling a change in a state of a control target portion of an internal combustion engine according to a change in an operation amount of an actuator of the internal combustion engine,
A function of generating a time-series signal of a predicted value of the state of the control target unit by applying a time-series signal of the operation amount of the actuator created in advance to the model equation;
A function of storing, as an initial condition, a time-series signal of an operation amount of the actuator and a generated time-series signal of the state of the control target unit;
For the objective function using the initial condition and the state estimated using the model equation, the evaluation operation is repeated while changing the operation amount to be input, whereby the optimum operation of the actuator under the initial condition is performed. A function for determining an optimal operation amount which is an amount
A function of associating and storing the determined optimal operation amount with respect to the initial condition;
A function of obtaining an approximate expression representing a relationship between the initial condition and the optimal operation amount;
, A computer program. A control device for an internal combustion engine comprising an actuator and a control target portion, the control device comprising:
The initial condition including at least the time-series signal of the operation amount of the actuator and the time-series signal of the state of the control target portion, the optimum of the actuator determined for the initial condition and the initial condition at each time A storage unit that stores an approximate expression representing a relationship with an optimal operation amount, which is a simple operation amount;
An information acquisition unit that acquires an actual operation amount of the actuator and a state of the control target unit;
The optimum operation amount corresponding to the actual operation amount and the state is determined using the acquired operation amount and the state, and the approximate expression in the storage unit, and the actuator is determined according to the optimum operation amount. A control unit to operate;
A control device for an internal combustion engine, comprising:

Images