Identification approach for internal combustion engine mean value models

Abstract

A system or approach for identifying mean value models with a set of equations and appropriate constraints which define the model validity. A model may be used to design an algorithm for an engine system, collecting sensed data, optimizing control parameters based on the models and data, and providing control of the engine system. These processed may be reiterated for updating control of the engine system.

Description

BACKGROUND

The present disclosure pertains to engine modeling, analysis and control.

SUMMARY

The disclosure reveals a system or approach for identifying mean value models with a set of equations and appropriate constraints which define the model validity. A model may be used to design an algorithm for an engine system, collecting sensed data, optimizing control parameters based on the models and data, and providing control of the engine system. These processed may be reiterated for updating control of the engine system.

BRIEF DESCRIPTION OF THE DRAWING

FIG. 1 is a diagram of a turbocharged diesel engine scheme;

FIG. 2 is a diagram of the turbocharged diesel engine scheme revealing more information;

FIG. 3 is a diagram of an overall system layout;

FIG. 4 is a diagram of an algorithm flowchart; and

FIG. 5 is a diagram of a case having two state variables and one linear constraint active, one with a free variable.

DESCRIPTION

The present system and approach may incorporate one or more processors, computers, controllers, user interfaces, wireless and/or wire connections, and/or the like, in an implementation described and/or shown herein.

This description may provide one or more illustrative and specific examples or ways of implementing the present system and approach. There may be numerous other examples or ways of implementing the system and approach.

Mean value models may be noted. The mean value model of an internal combustion engine may describe the chemical and thermodynamic processes of the substances flowing through an engine. Such models may be useful especially for optimal control of the engine in the real time, particularly the emissions control. The model may predict the effects of changes of the control parameters including an evolution of the effects in time. The model may be used to optimize the control parameters to achieve optimality both in the steady state and during transients. For such purposes, a fast and reliable model identification approach may be devised. The numerical approaches described herein may be developed to improve the identification approaches in this respect.

Mean value models may differ from the crank angle resolved models in that the engine mass flow is idealized to be independent of the crank angle. When representing an uncontrolled engine in a fixed operating point, particularly at constant speed and load with all external conditions constant, the mean value model may approximate the mass flows, pressures, and so on by equivalent constant values. The crank angle resolved models, in contrast, also respect the intake and exhaust valve opening and closing and the non-stationary mass flows through the intake and exhaust ports. At a fixed operating point, the crank angle resolved model may predict periodic flows, pressures, and so forth. As a consequence, the mean value model may be sufficient for optimization of control parameters which do not necessarily need to be adjusted over one crank shaft revolution but rather are slowly adapted so that their values do not change significantly over one revolution. Validity of the mean value models may always be constrained to particular operating conditions. As an example, such a mean value model does not necessarily represent the engine physics if the direction of the mass flows through the model elements change. Many such constraints may be expressed as simple inequalities on model internal states, e.g., pressure at one point must be higher (lower) than pressure at other point.

A purpose of the algorithm is to perform the model identification with respect to the steady-state data in such a way that the model produces output values similar to the measured ones and, at the same time, the model signals are in the domain of its validity. This avoids a usual problem with fitting data with a meaningless model. In the approach, the model validity may be equivalent to satisfying linear inequality constraints “Ax<=b” for model internal states “x”. Such states may be called feasible.

An algorithm flowchart is visualized on FIG. 4. The flowchart may use logical matrices L1 and L2 which can be represented as logical matrices with as many rows as many “x” constraints the problem has and as many columns as there are individual operating points in the data.

Value features may incorporate: 1) A true value in the logical matrix L1 may indicate the activity of the corresponding constraint in the operating point; and 2) A true value in logical matrix L2 may indicate that: the constraint activity can be toggled in the next iteration.

Logical matrices L1 and L2 may be used to organize the calculations in an efficient way. The identification process consists of 1) inner optimization, which searches for feasible vector state for given parameters, 2) outer optimization, which updates the model parameters based on information from inner optimizations at all steady-state data points. The process starts with solving inner optimizations only once to obtain initial feasible points. Then the outer optimization may be solved by iteratively implementing a process which is equivalent to the active set approach. Only the operating points which may require a change of the constraint activity are updated.

A merit of the present approach may be shown in a flowchart. Although the model states may be constrained during the optimization of model parameters, the optimization optimizes either only parameters (in outer) or only model states for fixed parameters (in inner) at the operating points separately. Thus, no high dimensional optimization problem is necessarily solved at any point. However, the low dimensional optimization may need to be solved a number of times and also the number of linear constraints may be high. For this reason, a quadratic programming approach, which may efficiently handle this situation, can be used to implement the outer optimization.

The algorithm of the diagram in FIG. 4 may solve a quadratic approximation of the cost function related to a nonlinear identification problem. The algorithm may be repeated iteratively when the approximation is updated based on the results from the previous iteration. The updating may represent an extra iteration loop which is not necessarily visualized on the diagram. In such use, the initial block “Start with Initial Model Parameters” may change to “Start with Parameters from the Previous Iteration”.

The flow diagram 10 of FIG. 4 may begin at a start with initial model parameters at symbol 11. At symbol 12, a next step may be to optimize feasible state model vectors “x” for virtually all steady state operating points determining the active constraints on states, i.e., determining a logical matrix L1. The step of symbol 12 may solve inner optimization issues.

A logical matrix L2 may be initialized to all true at symbol 13. One may iterate over operating points with any true element in an L2 column at symbol 14. For an operating point, the active set in an L1 column may be transformed to corresponding constraints on parameters using equation (16) at symbol 15.

At symbol 16, a contribution to a cost function at the operating point may be expressed as a function of parameters just using equation (15). Then the model parameters may be optimized relative to a sum of R(k) of equation (10) respecting virtually all constraints on the parameters at symbol 17. The active constraints may define L2 elements. The step of symbol 17 may solve an outer optimization issue.

A question whether any L2 element is true at symbol 18 may be asked. If an answer is “no”, then a solution may be regarded as found, and the procedure may be terminated at symbol 19. If the answer is “yes”, then the L1 elements corresponding to true values in the L2 matrix may be inverted at symbol 20. Then the actions of symbols 14 through 18 may be repeated as an iteration loop.

There may be an additional iteration of the loop which may be repeated when an approximation is based on results from a previous iteration. Such process is known as sequential optimization. The updating the reiteration loop may begin at symbol 11 which can be restated as “Stan with Parameters from the Previous iteration”, as noted herein.

FIG. 1 is a diagram of a turbocharged diesel engine scheme 25. The mean value model may be graphically represented by a scheme similar to FIG. 1 which is an approximation of the engine structure. For various engine architectures, the scheme may be slightly different and the identification algorithm will need to cover virtually all of them. The diagram may lump the dominant thermodynamic processes into idealized components which approximate the actual processes. For, example, a cooler, intercooler 26 or other coolers may be idealized components where only heat transfer between the flowing gases and their surroundings (cooling liquid) takes place. The controlled valve 27, 34, or valves, in contrast, may model the adiabatic pressure changes (i.e., occurring without gain or loss of heat) which can be controlled by Changing a valve opening. A compressor 28 and turbine 29 may be two components coupled by a common rotating shaft 31. Compressor 28 may pressurize intake air. Turbine 29 may usually have a parameter which controls the turbine's power, either through bypassing a part of the exhaust gas through a waste gate 27 to point 36 to join exhaust gas from turbine 29 for exit from the system, controlling an aspect ratio (i.e., a variable geometry turbocharger having turbine vane position control). Pressurized air may go from compressor 28 via a cooler 26 and mixing point 32 to combustion engine 33. Mixing point 32 may also receive recirculated engine exhaust gas (EGR) via valve 34 and cooler 35. Exhaust from engine 33 may go through a flow splitter 30 which splits the exhaust to turbine 29 and valve 34, but not necessarily in a 50/50 fashion.

The components in the present system or approach revealed in the diagram of FIG. 1 may be rearranged or connected in various configurations. Components may be added to or subtracted from the system or approach. The configuration of the components revealed according to the diagram of FIG. 1 is one instance or illustrative example, among others, of the present system or approach.

FIG. 2 is a diagram of turbocharged diesel engine scheme 25 with more information. Engine scheme 25 and its components may be represented by models. Air from an ambient environment 41 may be an inflow 42 to compressor 28. Ambient pressure, ambient temperature may be sensed of environment 41. Compressor 28 may be driven by turbine 29 via a connection or shaft 31 which together can constitute a turbocharger 43. An outflow 44 from compressor 28 may go to cooler or charge-air cooler 26. An output 45 from cooler 26 may go to a mixing point or valve 32. Output 45 may be mixed with some output 46 from am EGR cooler 35. An amunt of mix of output 45 with output 46 may be controlled by a signal from an engine control module 48. An output 47 from mixing point 32 may be an inflow to engine 33.

Engine control module (ECM) 48, may contain the models and inputs from sensors 49, such as fresh air flow, temperature after compressor 28, temperature after intercooler 26, temperature in the intake manifold and pressure in the intake manifold of engine 33, and so on, along with processing to determine optimal values and control signals, for instance, for engine speed, injection quantity, start of injection, rail pressure, and so on, to engine 33. An outflow 54 from engine 33 may go to flow splitter 30 that splits outflow 54 into an outflow 55 and an outflow 56. Outflow 55 may go to an EGR valve which controls an outflow 50 of exhaust through an EGR cooler 35 and output 46 to mixing valve 32. Outflow 56 may be an inflow to turbine 29 that spins compressor 28 via connection 31. An outflow 57 from turbine 29 may go to ambient environment 41. A valve or a waste gate 27 may connect inflow 56 to outflow 57 at connection 36 to permit some of inflow 56 to bypass turbine 29. A speed of the turbine 29 and compressor 28 may be controlled by valve 27 and vane position input 58. Controlling inputs to valve 27 and vane position input 58 may be provided by engine control module 48. Controlling inputs to mixing valve 32 and EGR position input 59 may also be provided by engine control module 48.

The components in the present system or approach revealed in the diagram of FIG. 2 may be rearranged or connected in various configurations. Components may be added to or subtracted from the system or approach. The configuration of the components revealed according to the diagram of FIG. 2 is one instance or illustrative example, among others, of the present system or approach.

FIG. 3 is an overall diagram of a system layout. A set of equations 61 may make up an algorithm 62. Models 63 may follow from algorithm 62. An engine control module may incorporate models for reasons of providing optimized control signals 65 to an engine system 66. Engine system may provide information and data 67 from sensors and other sources to engine control module 64 for further optimizing control signals 65.

Equations of virtually all model components may be based on first principles and contain unknown parameter values which are to be identified from the data. The model representation may be similar to the way that actual electrical devices represented by electrical circuits built from idealized physical components, such as ideal inductance, capacitance, resistance, and so forth.

An elimination approach may be noted. A model identification problem may be stated as follows. One may manipulate the model parameters to minimize the norm of differences between the values predicted by the model and the values measured on a real engine. Herein, just steady state identification may be considered. This means that the error minimization may involve only data measured on an engine which already has settled to a steady state after the last change of any exogenous signals. Particularly, the engine steady state may be described by constant control variables, pressures, temperatures, concentrations and speeds in the mean value model sense when a dependency on the crank angle is not considered.

A prediction of the model may be calculated by solving a set of differential equations which can be derived by combining all algebraic and differential equations of virtually all idealized components and respecting a model structure. The differential equations of such a model may be usually represented by a system of first order nonlinear differential equations and nonlinear output functions (nonlinear state space model) as follows.

$\begin{array}{cc}\frac{d\hat{x}\left(t\right)}{dt}=f\left(\hat{x}\left(t\right),u\left(t\right),\theta \right)\hat{y}\left(t\right)=g\left(\hat{x}\left(t\right),u\left(t\right),\theta \right)& \left(1\right)\end{array}$ Here, {circumflex over (x)} is the vector of internal state variables; u is the vector of exogenous variables; ŷ is the vector of model outputs and θ is the vector of model parameters. When considering the steady state, the model can be described by the following set of nonlinear equations.0=f({circumflex over (x)}(t),u(t),θ)ŷ(t)=g({circumflex over (x)}(t),u(t),θ)  (2)

The identification problem may be cast as the following nonlinear least squares optimization problem, given a set of observed input and output values {u_k, y_k, k=1 . . . K}:min_θ,xkΣ_k=1^K∥g({circumflex over (x)}_k,u_k,θ)−y_k∥₂²  (3)This optimization may be constrained by the following equality constrain.f({circumflex over (x)}_k,u_k,θ)=0,∀k  (4)An approach to minimize (3) may seem to solve the equality constraint (4) with respect to {circumflex over (x)}_k and substitute a solution {circumflex over (x)}_k=ϕ(u_k, θ) to (3). This may turn the equality constrained optimization into an ordinary unconstrained optimization. In this way, the {circumflex over (x)}_k variables may be eliminated and the problem is now a dimensional optimization problem.

A solution of the nonlinear equation (4) may be obtained by either solving that algebraic equation or solving the differential equation from an initial condition x₀ by applying constant inputs u_k for a sufficiently long time. This approach may work better for stable systems as follows.

$\begin{array}{cc}{x}_k\approx \underset{i\to \infty }{\lim }{\int }_0^{t}f\left(x\left(t\right),u_k,\theta \right)dt,x\left(0\right)=x_0& \left(5\right)\end{array}$ Although the {circumflex over (x)}_k elimination may seem to solve the problem elegantly, there may be some possible issues in that approach.

The issues may be listed as P1, P2, P3 and P4 as in the following.

P1) A solution of the nonlinear equation (4) may be completely separated from the prediction error minimization (3). Therefore, even if many signals are measured on the engine, e.g., all state variables, {circumflex over (x)} would be measured, it does not necessarily help the nonlinear equation solver to identify the correct steady state out of multiple solutions. There may be multiple solutions to the algebraic equation (4). For an approach based on the differential equation solution (5), the issue with multiple solutions does not necessarily exist when the initial condition is fixed. However, this approach may fail when the model becomes unstable in the course of identification.

P2) The nonlinear identification may work well when the initial parameter values are as close as possible to the optimal values and the initial minimized objective function value is close to the optimal value. When looking at FIG. 1 and considering the feedback structure, it may happen that the inexact initial parameter estimates will drive the model state very far from the optimal values and that the initial objective function value will be large. This may be true especially if the initial parameters push the model to the stability boundary. When sufficiently close to the boundary, the initial objective function value may be arbitrarily large. This means that the nonlinear optimization may have to execute many iterations to optimize. This will not only take time, but may also make algorithm divergence more likely.

P3) The model structure is not necessarily valid universally for any {circumflex over (x)} values. When looking at FIG. 1 as an example, one may see that the compressed fresh air is mixed with the recirculated exhaust gases at the mixing point 32. This may be true as long as both compressed air and the recirculation flows have the directions indicated by the arrows. When the initial parameter values do not define the proper flow directions of either of the two, no mixing will necessarily actually happen at that point. Then, the model structure with this mixing point representing mass and energy summations in the intake will not necessarily represent the physical reality which would happen on the real engine with the same parameters. The behavior of the model may then be different from the expected behavior which is assumed to obey the laws of physics.

P4) Likewise, the model components are not necessarily valid universally for any {circumflex over (x)} values. Each component model is usually valid under certain and normally satisfied conditions. As an example, the simplified compressor flow equations may often not represent the physical reality if the pressure ratio over the compressor is less than one or if the turbocharger speed is negative.

In other words, modeling engines by mean value models may use not necessarily just equations but also constraints which ensure model validity. Identification (parameter optimization) of such models may use an algorithm which respects such constraints virtually all the time, i.e., not evaluating the model equations with values outside the allowed ranges. For instance, the algorithm will not necessarily try to evaluate a compressor model with a pressure ratio below one throughout a process of optimizing its parameters.

A new formulation may be noted. An effort to solve some of the indicated difficulties may be based on making the model structure more general and more complex to behave reasonably in broader ranQes of the model signal values, which are defined by {circumflex over (x)}. However, the present approach may be rather based on constraining the model signals explicitly, not allowing signal values which would invalidate either the component models or the model structure. This means that the model identification problem may be cast as an inequality constraints optimization as follows.

$\begin{array}{cc}{\min }_{\theta ,x_k}\sum _{k=1}^K{J}_k\left(x_k,u_k,\theta \right),c_i\left(x_k\right)\le 0& \left(6\right)\end{array}$ Here it may be assumed the constraints c_i(x_k) are defined to preserve the validity of the model and the objective function J_k penalizes the model prediction errors. Instead of considering the nonlinear equality constraint (4) explicitly, it may be proposed to turn it into a soft constraint penalizing the norm of state derivatives. Then the objective function J_k may be formulated as a convex combination of the two objectives:J_k(x_k,u_k,θ)=(1−λ)∥g({circumflex over (x)}(t),u_k,θ)−y_k∥₂²+λ∥f(x_k,u_k,θ)∥₂²  (7)The objective function J assumed form may be the second squared norm of the difference between measured data and model predictions:J(x_k,u_k,y_k,θ)=∥g(x_k,u_k,θ)−y_k∥₂²  (7a)Here, x_k is a sequence of the model state variables, u_k is sequence of model inputs, and y_k is sequence of measurements, θ is the vector of model parameters. Equation (7a) may be known as the method of least squares.

In this objective function, the state variables need to be such that their derivatives are zero if substituted to the model differential equation (1). That is why one had either add this equality constraint or at least a penalty for a norm of the time derivatives, i.e., the least squares need to be formulated as equation (7b):J(x_k,u_k,y_k,θ)=(1−λ)∥g(x_k,u_k,θ)−y_k∥₂²+λ∥f(x_k,u_k,θ)∥₂²  (7b)

Parameter λ should control the optimization process giving relative weights to the objectives. It may be assumed that the optimal model parameters can always define proper feasible equation (7) model signals when the input signals u_k are applied to the model. The inequality constraints c_i may be meant to be active only in the course of an identification process, but not necessarily for the successfully identified model.

The present formulation may address the issues of the elimination approach as follows.

P1) The solution of (4) is not necessarily separated from (3) because both objectives may be optimized simultaneously. Particularly, the optimization may have a better chance to find the physically meaningful solution when there is some information about the state variables {circumflex over (x)}_k in the measurements.

P2) The objective function values should not explode because the state variables {circumflex over (x)}_k are not necessarily driven by the differential equation (1) in this approach. They are directly controlled by the optimization algorithm. Both P3 and P4 may be covered by the inequalities c_i(x_k).

An optimization technique may be noted. An evident problem of the new formulation is that the number of optimized variables may be overwhelming. In a typical engine identification example, 500 to 1,000 steady state operating points may be measured. Considering 5 to 10 state variables {circumflex over (x)}_k, the number of optimized states may be 2,500 to 10,000. The number of model parameters θ appears negligible compared to these numbers (usually less than 50). Therefore, it is not necessarily easy to use the potential advantages of a new formulation, and a special identification algorithm which uses the model structure should be used. The presently new optimization approach may use a sequential quadratic programming technique to solve this large dimensional optimization problem. The technique may be based on approximating the objective function by a quadratic positive definite function and the inequality constraints by linear constraints in the following.

$\begin{array}{cc}{J}_k\left({\hat{x}}_k,\theta \right)=\frac{1}{2}{\hat{x}}_k^{\prime }{H}_k^{\mathrm{xx}}{x}_k+\frac{1}{2}{\theta }^{\prime }{H}_k^{\mathrm{\theta \theta }}\theta +{\hat{x}}_k^{\prime }{H}_k^{x\theta }\theta +f_k^{\prime x}{\hat{x}}_k+f_k^{\mathrm{\prime \theta }}\theta A_k{\hat{x}}_k\le b_k& \left(8\right)\end{array}$ One may propose to make use of the structure (6). The objective function may be additive, and the values of state variables valid at a particular operating point may always be present at one summand only. It may therefore be advantageous to distribute the optimization as follows.

$\begin{array}{cc}{\min }_{\theta ,x_k}\sum _{k=1}^K{J}_k\left(x_k,\theta \right)=\underset{\underset{\mathrm{outer}}{︸}}{{\min }_{\theta }\sum _{k=1}^K\underset{\underset{\mathrm{inner}}{︸}}{{\min }_{x_k}{J}_k\left(x_k,\theta \right),A_k{\hat{x}}_k\le b_k}}& \left(9\right)\end{array}$ In equation (9), the optimizations with respect to state variables x_k at particular operating points may be done separately. Then, the outer minimization may minimize the overall value with respect to θ. This may be done only if the inner optimizations can give the optimal values parameterized by the values of θ. In other words, the inner optimization needs to provide the following functions of θ as their results as in the following.R_k(θ)J_k(arg min_xkJ_k({circumflex over (x)}_k,θ),θ)  (10)

Then the outer optimization may minimize the sum Σ_kR_k(θ) with respect to θ. For the quadratic objectives (8) with linear constraints, the function R_k may be quadratic with parameters which depend on the set of active constraints in the inner optimization problems. Suppose the inner iteration issue has been solved by the quadratic programming algorithm and certain inequality constraints are weakly active at the constrained optimum; that may mean that the respective Lagrange multipliers will be non-negative as follows.A_k^a{circumflex over (x)}_k=b_k^a,μ_k^a≥0  (11)

One may define reduced dimension coordinates ξ_k in this active set using P_k^a as the matrix A_k^a pseudo-inverse and the matrix N_k^a as the A_k^a null-space matrix as follows.{circumflex over (x)}_k=P_k^ab_k^a+N_k^aξ_k  (12)

The optimal reduced coordinates may be determined by requiring that the derivatives with respect to the reduced coordinates have to be zeros. They can be found to be affine functions of the parameters:ξ_k=−(N′_k^aH_k^xxN′_k^a)⁻¹N′_k^a(f_k^x+H_k^xθθ)  (13)

Relative to ξ, if one says that certain linear constraints on the state variables are active, one in fact constrains the vector of states to certain linear subspace which borders the constrained area. The coordinates in this subspace were denoted ξ in the equations (12) and (13). FIG. 5 is a diagram 51 in a case of two state variables and one linear constraint active, which may leave one with one free variable ξ. The free variable may define a position of the state vector 52 on the “active” line.

When substituting (13) to (12) and then to (8), the R_k (θ) may be a quadratic function of the model parameters. Therefore, the inner optimization may provide the information needed. However, the inner optimization has to also provide constraints on θ which guarantee that the active set does not necessarily change in the inner problem. These constraints may be derived as follows. The Karush-Kuhn-Tucker conditions may imply that at the optimum, the gradient of the objective function must be a certain linear combination of the gradient of the active constraints as follows.

$\begin{array}{cc}\frac{\partial }{\partial {\hat{x}}_k}{J}_k\left(x_k,\theta \right)=-A_k^{\mathrm{\prime a}}{\mu }_k^a,{\mu }_k^a\ge 0& \left(14\right)\end{array}$ Together with (11), the following may be implied.

$\begin{array}{cc}\left(\begin{array}{c}{\hat{x}}_k\\ -{\mu }_k^a\end{array}\right)={\left(\begin{array}{cc}{H}_k^{\mathrm{xx}}& -A_k^{\mathrm{\prime a}}\\ A_k^a& 0\end{array}\right)}^{-1}\left(\left(\begin{array}{c}-f_k^x\\ b_k^a\end{array}\right)+\left(\begin{array}{c}{H}_k^{\theta x}\\ 0\end{array}\right)\theta \right)& \left(15\right)\end{array}$

This means that both constrained optimal {circumflex over (x)}_k and the Lagrange multipliers of the weakly active constraints μ_k^a are affine functions of the model parameters θ. Therefore, the necessary conditions on parameters to preserve the set of weakly active constraints at the inner optimization problem may be as follows.

$\begin{array}{cc}\left(\begin{array}{cc}{A}_k^i& 0\\ 0& I\end{array}\right){\left(\begin{array}{cc}{H}_k^{\mathrm{xx}}& -A_k^{\mathrm{\prime a}}\\ A_k^a& 0\end{array}\right)}^{-1}\left(\left(\begin{array}{c}-f_k^x\\ b_k^a\end{array}\right)+\left(\begin{array}{c}{H}_k^{\theta x}\\ 0\end{array}\right)\theta \right)\le \left(\begin{array}{c}{b}_k^i\\ 0\end{array}\right)& \left(16\right)\end{array}$ This may be a set of linear inequalities on the model parameters. The matrix A_kⁱ may be formed from the following inactive inequalities A_kⁱ{circumflex over (x)}_k<b_kⁱ.

The optimization scheme solving the problem (9) may be described as follows. The steps denoted with “O” may pertain to the outer optimization whereas the steps denoted with “I” may pertain to the inner optimizations.

O1. The outer optimization may set an initial estimate of the model parameters θ.

O2. The inner optimizations may be performed sequentially:

I1. The quadratic programming (QP1) may be used to optimize {circumflex over (x)}_k to find the constrained optimum and the set of strictly active constraints, i.e., the inequalities with positive LaQrange multipliers.

I2. Equations (13) to (12) and then to may be used to evaluate R_k in this active set.

I3. Equation (16) may be used to define constraints on the parameters which ensure that the previously strictly active constraints can be at least weakly active after having optimized θ.

O3. The outer optimization may aggregate the R_k (summing them) and the constraints on parameters (taking all of them).

O4. The outer optimization may use the quadratic programming (QP2) to find constrained optimal values of the parameters. They may be interpreted as optimal values which will not necessarily change the set of active constraints in any of the inner optimizations.

O5. The outer optimization may check the set of active constraints at QP2. If the set is empty, the parameters may be optimal. If there is an active constraint, it may indicate activity of which constraints on “x” at which steady state data points, needs to be toggled (changed from inactive to active or vice versa). This means that the active constraints may be changes at one inner optimization. This inner optimization may be recalculated with the new values of parameters. The other inner optimizations do not necessarily need to be recalculated because their R_k is still valid. Then the algorithm may proceed from O4.

A conclusion may be noted. The identification of the mean value model parameters based on steady state data may be formulated as an optimization problem. It has been shown that the optimization problem may be either very high dimensional or the values of state variables at the steady state need to be eliminated. This elimination may cause the state values to be no more under control and cannot be constrained in order to preserve model validity and steady state consistence with measurements. Therefore, one may propose here not to eliminate the state variables. An approach solving the high dimensional constrained optimization problem may be proposed instead, as indicated herein. The optimization scheme may make use of the problem structure. This may be done elegantly if the objective function is quadratic and the inequality constraints are linear. Although it may be true that the objective function (6) is not necessarily quadratic, the algorithm can still be used if, for example, the objective function is successively locally approximated by quadratic functions. This may be regarded as a sequential quadratic programming technique. Another approach in how to use the algorithm may be based on piece-wise linear approximations of the model nonlinearities.

The present identification technique may be very significant for commercial model and control design tools which are used by non-experts on identification and numerical optimization. For numerous reasons, the reliability and robustness of the optimizations are critical factors for the customers' acceptance and satisfaction.

To recap, a modeling mechanism for an engine system may incorporate an engine control module (ECM) connected to an engine system, and a model representing the engine system. The engine system may incorporate an engine, actuators and sensors of the engine connected to the ECM, a turbocharger having an output connected to an input of the engine, and an aftertreatment device having an input connected to an output of the engine. The model representing the engine system may be situated in the ECM. The model representing the engine system may be a mean value model that uses equations and constraints. The constraints may ensure model validity in that equations of the model are prevented from evaluation with values outside of ranges allowed for the engine system being represented.

The model representing the engine in a steady state may incorporate one or more equations consisting of terms {circumflex over (x)}, u, ŷ and θ representing internal state variables, exogenous variables, model outputs and model parameters, respectfully. The one or more equations may incorporateθ=f({circumflex over (x)}(t),u(t),θ)ŷ(t)=g({circumflex over (x)}(t),u(t),θ)

An identification of the model may be cast in as a least squares optimization for observed input and output values [u_k, y_k, k=1 . . . K] asmin_θ,xkΣ_k=1^K∥g({circumflex over (x)}_k,u_k,θ)−y_k∥₂²

The least squares optimization may be constrained byf({circumflex over (x)}_k,u_k,θ)=0,∀k where ∀ indicates that f for a sequence of k.

The ECM may receive data from the sensors of the engine system, which are processed in view of the model. The ECM may send signals, which are optimized in view of the model, to the actuators of the engine system. The ECM may provide real time control of the engine system.

An approach for an approximation of a cost function relative to nonlinear identification for a mean value model for an engine system, may incorporate selecting initial model parameters for an engine system, and optimizing state model vectors for steady state operating points determining active constraints on states, that is, determining a first logical matrix of elements. The first logical matrix may have as many rows as a number of state model vector constraints and as many columns as a number of individual operating points in data.

The approach may further incorporate initializing a second logical matrix of elements to virtually all true. The second logical matrix may have as many rows as a number of state model vector constraints and as many columns as a number of individual operating points in the data.

The approach may further incorporate iterating over steady state operating points with a true element in a column of the second logical matrix. For a steady state operating point, an active set in a column in the first logical matrix may be transformed to corresponding constraints on parameters.

The approach may further incorporate expressing a contribution to a cost function at the steady state operating point as a function of parameters only, and optimizing model parameters with respect to a sum of constraints on parameters. Active constraints may define elements of the second logical matrix.

If any element of the second logical matrix is false, then a solution may be found.

If any element of the second logical matrix is true, then elements of the first logical matrix corresponding to true values in the second logical matrix may be inverted. The approach may further incorporate repeating one or more portions of the previously mentioned approach.

A turbocharged engine model system may incorporate an engine model of an engine of an engine system; a mixing point model having an output connected to an inflow port of the engine model; a turbocharger model having an inflow turbine port connected to an outflow port of the engine model, an outflow compressor port connected to an input of the mixing point model, and having an outflow turbine port and an inflow compressor port; and a processor incorporating a plurality of the models, including the engine model, mixing point model and the turbocharger model, and being connected to the engine.

The models may be mean value models that use equations and constraints. The constraints may ensure model validity in that the equations can be prevented from evaluation with values outside allowed ranges. The processor may receive data from sensors of the engine system and, in accordance with the models and calculated equations and constraints, provide optimized signals to actuators of the engine system.

The plurality of models may further incorporate an exhaust gas recirculation valve model having an input connected to the outflow port of the engine model; and an exhaust gas recirculation cooler model having an input connected to an output of the exhaust gas recirculation valve model and an output to a second input of the nixing point model.

The plurality of models may further incorporate a charge air cooler model having an input connected to the outflow compressor port, and having an output connected to an input of the mixing point model.

The plurality of models may further incorporate a flow splitter model having an input connected to the outflow port of the engine model, a first output connected to the inflow turbine port of the turbocharger model and a second output connected to an inflow port of the exhaust gas recirculation valve model.

U.S. patent application Ser. No. 12/782,004, filed May 18, 2010, entitled “Distributed Model Identification”, is hereby incorporated by reference.

In the present specification, some of the matter may be of a hypothetical or prophetic nature although stated in another manner or tense.

Although the present system and/or approach has been described with respect to at least one illustrative example, many variations and modifications will become apparent to those skilled in the art upon reading the specification. It is therefore the intention that the appended claims be interpreted as broadly as possible in view of the related art to include all such variations and modifications.

Claims

1. A mechanism for controlling an engine system, comprising: an engine control module (ECM) connected to an engine system; anda model representing the engine system; andwherein:the engine system comprises: an engine;actuators and sensors of the engine connected to the ECM;a turbocharger having an output connected to an input of the engine; andan aftertreatment device having an input connected to an output of the engine;the model representing the engine system is situated in the ECM;the model representing the engine system is a mean value model that uses equations and constraints;the model representing the engine in a steady state comprises one or more equations consisting of terms {circumflex over (x)}, u, ŷ and θ representing internal state variables, exogenous variables, model outputs and model parameters, respectfully;the one or more equations comprise 0=f({circumflex over (x)}(t),u(t),θ)ŷ(t)=g({circumflex over (x)}(t),u(t),θ);the constraints ensure model validity in that equations of the model are prevented from evaluation with values outside of ranges allowed for the engine system being represented; andthe ECM receives data from the sensors of the engine system, which are processed in view of the model; andthe ECM sends signals, which are configured based on the model, to the actuators of the engine system.

2. The mechanism of claim 1, wherein: an identification of the model is cast in as a least squares optimization for observed input and output values {uk, yk, k=1 . . . K} as minθ,xkΣk=1K∥g({circumflex over (x)}k,uk,θ)−yk∥22;

3. The mechanism of claim 1, wherein the signals from the ECM to the actuators are optimized in view of the model.

4. The mechanism of claim 1, wherein the ECM provides real time control of the engine system.

5. A method for controlling an engine system based on an approximation of a cost function relative to nonlinear identification for a mean value model for the engine system, comprising: selecting initial model parameters for an engine system; andoptimizing state model vectors for steady state operating points determining active constraints on states, that is, determining a first logical matrix of elements;receiving data from sensors of the engine system, which are processed in view of the optimized state model vectors;sending signals, which are configured based on the optimized state model vectors, to the actuators of the engine system; andwherein the first logical matrix has as many rows as a number of state model vector constraints and as many columns as a number of individual operating points in data.

6. The method of claim 5, further comprising: initializing a second logical matrix of elements to virtually all true; andwherein the second logical matrix has as many rows as a number of state model vector constraints and as many columns as a number of individual operating points in the data.

7. The method of claim 6, further comprising: iterating over steady state operating points with a true element in a column of the second logical matrix; andwherein for a steady state operating point, an active set in a column in the first logical matrix is transformed to corresponding constraints on parameters.

8. The method of claim 5, further comprising: expressing a contribution to a cost function at the steady state operating point as a function of parameters only; andoptimizing model parameters with respect to a sum of constraints on parameters; andwherein active constraints define elements of the second logical matrix.

9. The method of claim 8, wherein if any element of the second logical matrix is false, then a solution is found.

10. The method of claim 8, wherein: if any element of the second logical matrix is true, then elements of the first logical matrix corresponding to true values in the second logical matrix are inverted; andfurther comprising repeating the method recited in claim 6.

11. A turbocharged engine model system comprising: an engine model of an engine of an engine system;a mixing point model having an output connected to an inflow port of the engine model;a turbocharger model having an inflow turbine port connected to an outflow port of the engine model, an outflow compressor port connected to an input of the mixing point model, and having an outflow turbine port and an inflow compressor port; anda processor incorporating a plurality of the models, including the engine model, mixing point model and the turbocharger model, and being connected to the engine; andwherein:the models are mean value models that use equations and constraints;the engine model, when representing the engine in a steady state, comprises one or more equations consisting of terms {circumflex over (x)}, u, ŷ and θ representing internal state variables, exogenous variables, model outputs and model parameters, respectfully;the one or more equations comprise 0=f({circumflex over (x)}(t),u(t),θ)ŷ(t)=g({circumflex over (x)}(t),u(t),θ),the constraints ensure model validity in that the equations are prevented from evaluation with values outside allowed ranges; andthe processor receives data from sensors of the engine system and, in accordance with the models and calculated equations and constraints, provides optimized signals to actuators of the engine system.

12. The system of claim 11, wherein the plurality of models further comprises: an exhaust gas recirculation valve model having an input connected to the outflow port of the engine model; andan exhaust gas recirculation cooler model having an input connected to an output of the exhaust gas recirculation valve model and an output to a second input of the mixing point model.

13. The system of claim 12, wherein the plurality of models further comprises a charge air cooler model having an input connected to the outflow compressor port and an output connected to an input of the mixing point model.

14. The system of claim 13, wherein the plurality of models further comprises a flow splitter model having an input connected to the outflow port of the engine model, a first output connected to the inflow turbine port of the turbocharger model and a second output connected to an inflow port of the exhaust gas recirculation valve model.