DE10113538B4 - Regulating device and control method - Google Patents

Abstract

Control device comprising a controller (R) which receives at least one input variable (E) and from this determines at least one manipulated variable (U) which is fed to an actuator of a controlled system (S), with a neural correction transmitter (NAM) being provided which at least a process variable (P2) from the controlled system, at least some of the controller input variables (E = V1 (Msoll, Mist, P1)) and at least some of the manipulated variables (U) determined by the controller and is designed to generate a correction signal based on these input variables (K), the correction transmitter (NAM) being designed in such a way - on the one hand, to generate at least one correction signal (K) which is algebraically concatenated with the manipulated variable or variables (U) to form a corrected manipulated variable (UK), - and on the other hand, an adaptation of the internal parameters influencing the neural correction transmitter (NAM) in its mode of operation, the adaptation of the correction transmitter (NAM) with an onli ne learning algorithm is carried out in real time.

Description

The present invention relates to a control device and a control method. More particularly, the present invention relates to a real time neural adaptation controller for controlling internal combustion engines.

In control devices or control methods, at least one input variable is fed to a controller, which processes it to at least one manipulated variable, which is then supplied to an actuator in a controlled system. By means of the actuator then an adjustment is brought about, which in turn leads to a rule response (response) at the end of the controlled system. Due to various influences, such as tolerances and signs of aging, deviations in the course behavior may occur. In order to obtain an optimal control quality, a real-time adaptation of the controller to these deviations is proposed.

Although the present invention is generally directed to control devices and control methods, the relationships are explained below, in particular with reference to an example with an internal combustion engine.

Due to component tolerances and aging processes, internal combustion engines experience a certain amount of series dispersion of the actuators and the path to be controlled. Especially in the calculation of quantities for motor control, this scattering is currently compensated adaptively over certain fixed correction values which simply cause a linear correction of the manipulated variable or represent operating range-selective, constant correction values. In particular, when determining the fresh air mass supplied to an internal combustion engine, a global linear correction of the control actions is performed. The adaptation of the correction values is conventionally carried out on the basis of the return of the air mass difference between the predetermined and the measured air mass. The adaptation method is a linear procedure for throttle valve systems as well as for systems with variable valve timing, which iteratively determines the correction values by multiple integration of the air mass difference.

However, the influence of scattering on the controlled system and the actuators is generally non-linear. Therefore, even in the best case, linear correction models can only represent a general average linear correction for all operating points. On the other hand, a correction via a physical-empirical modeling is often very expensive.

As an alternative to the use of physical-empirical models in correcting operational quantities, the use of neural models is known.

Purely neural models, however, are only of limited suitability for integration into an adaptive controller sequence because of the high computational time required for learning. Therefore, they are usually taught off-line, d. H. The results are stored in the engine control and remain unchanged there. In addition, learning is only suitably controllable in off-line mode.

Hybrid models, such as from the DE 43 38 607 A1 The disadvantages of the above-mentioned approaches, that is to say of a purely physical-empirical model and of a purely neuronal model, are circumvented by providing physically-empirically calculated predictions with neuron-calculated correction values. However, such models exist so far only in non-adaptive form.

Furthermore, from the US 5,625,552 a control circuit structure is known in which a PID controller is optimized by means of a neural adjuster. For this purpose, the adjuster the control signal of the controller to be optimized and the output signal to be controlled controlled system supplied, in dependence of the neural adjuster determines new controller parameters and passes them to the adjuster.

From the US 5,877,954 a device for process modeling is known in which a controller (DSC) a controlled system model (hybrid run-time model) is controlled or regulated. In this case, the controlled system model comprises a linear component for simulating the controlled system and a neural component via which a correction of the controlled system takes place and the controlled system can be imitated more accurately.

DE 199 13 126 A1 relates to the determination of a rolling force in a roll stand for rolling metallic rolling stock. Here, an analytical rolling force model is used to determine the rolling force, which provides as a starting value a rough value for the rolling force. Furthermore, a neural network is provided which determines a correction value. This correction value is multiplied by the coarse value for determining the rolling force. The neural network is trained with values for the rolling force and values for different operating conditions for rolling mills different rolling mills. This is done, for example, by subtracting the deviation between the calculated value of the rolling force from the measured value for the actual rolling force.

Object of the present invention is to provide a control device and a control method of the type mentioned, which or which allows an online adaptation to non-linear variations in the aforementioned controllers or control methods.

This object is achieved by the features mentioned in claims 1 and 8 alternatively.

In the present case, a controller, in particular a real-time controller, is used which determines at least one input variable from at least one input variable. In this case, the controller may be a conventional controller from the prior art. whose control concept is based on a physical process model, an empirical approach or a design based on dynamic process identification.

In addition to the controller, a neural correction transmitter is provided which receives at least a portion of the controller input variables and at least a portion of the controller output variables (manipulated variables). The neural correction transmitter is designed. in order to generate at least one correction signal on the basis of its input signals, which is algebraically linked to the one or more manipulated variables of the controller to one or more corrected manipulated variables or manipulated variables. In addition, the quantities supplied to the neural correction generator are used for their own adaptation. In the process, the internal parameters influencing the neural correction transmitter are changed (learned). The introduced controller type or the proposed controller method are particularly suitable for controlling processes that are difficult to model physically-empirically.

The algebraic concatenation can be chosen arbitrarily, for example, a simple summation can be selected. At a summation, the output variable is determined from the sum of the manipulated variable from the controller and the correction component from the neural correction transmitter.

For example, when using the control device in a motor vehicle to compensate for component tolerances, all motors of the same type can be equipped with the same physical-empirical model component in the controller. The adaptive neural model from the neural corrector can then ensure the specific motor matching. The modeling of the model for the controller can also be easily extended with more precise knowledge of the disturbing influences. without having to influence the neural component. The neural component is always taught according to the given algorithm.

If necessary, the neural corrector can be switched off in a targeted manner. so that the input for the actuator of the controlled system is equal to the manipulated variable from the controller. Of course, in addition to the above-mentioned input variables for the neural correction generator also other input variables. like other process variables. in particular, measured variables and internal variables of the controller are selected and supplied to the correction transmitter as input variables.

According to one embodiment, it is advantageous to limit the values of the internal adaptable parameters of the neural correction transmitter in order to avoid implausible outputs or unwanted learning effects. As a network type for the neural corrector, for example, a so-called "LOLIMOT network" can be used (compare Nelles, Lolimot, Linear Models for the Identification of Nonlinear, Dynamic Systems, Automatisierungstechnik 45, 1997), which is characterized by particularly advantageous properties during the learning process and in Interpolation and extrapolation distinguishes. This network maps a multi-dimensional nonlinear unknown functional relationship between input variables and output variables.

An embodiment of the present invention will be explained in more detail with reference to the single accompanying drawing. The drawing shows a schematic block diagram of an embodiment of a control device according to the invention.

The regulator model of the present embodiment is applied to air mass flow adjustment of an internal combustion engine. However, it should be noted in advance that the model is generally valid and usable for many control operations.

In the present model, a regulator R is provided, which _{is intended} from the air mass flow deviation of the requested air mass flow M from the measured actual value _is M and other process measurement variables P ₁ linear values U (= controller output signal) (in this case m = slope value and b = offset value) required for the Air mass flow supplies. For this purpose, the regulator R _is supplied with the previously mentioned quantities M _soll , M and P ₁ .

In the present case, the input variables M _soll , M _ist and P _{1 are combined} in block V ₁ , which in practice is realized by a filter, into a vector E. The controller R generates from these quantities the linear values U.sub.S acting as a control signal.

Moreover, in the present case, an optionally switchable, neural adaptation model NAM is provided as a correction transmitter, which receives at least a portion of the input signals E and the controller output signal U as input variables. The selection of the concrete input data for the neural adaptation model NAM takes place in a second filter V ₂ , to which the aforementioned values E and U are fed. The selection of the input data for the neural adaptation model NAM in the filter V ₂ takes place in a fixed manner.

Optionally, further process parameters are supplied to the neural adaptation model NAM. In the present case this is in the form of P _2, the oil temperature T. Moreover, the filter internal variables of the regulator R, in this case R _l (trigger bits), are provided.

The neural adaptation model NAM which acts as a correction generator supplies a correction value K, which in the following is considered additively and coupled to the manipulated variable U to form a corrected manipulated variable U _K. The algebraic concatenation can be chosen arbitrarily in the appropriate form. Furthermore, with the aid of setpoint values for the measuring parameter P ₂ (in the present case the oil temperature T) and the air mass flow, the internal parameters of the correction transmitter NAM are adapted (taught) in real time. This is indicated by the arrow crossing the NAM correction generator.

The neural network used in the broadest sense performs a non-linear multidimensional regression and can be taught online with respect to a subset of the parameters using linear training algorithms. Of course, many types of networks with similar properties are conceivable, which are also well known in the literature.

In order to avoid implausible outputs of the corrector or unwanted learning effects, it is possible to limit the values of the internal adaptable parameters. For example, the above-mentioned LOLIMOT network is to be used, which is characterized by particularly advantageous properties during the learning process. It maps a multi-dimensional nonlinear unknown functional relationship between input quantities and additive output correction. The outputs are determined by the sum of individual linear models, which are multiplied by nonlinear normalized weighting factors. The parameters of these linear and nonlinear functions can be independently adjusted during operation in the controller step by step with minimal computational effort, without having to resort to past input values. Limiting the learning process to the parameters of the linear functions, the algorithm converges mathematically guaranteed with correct choice of the learning rate against the minimum quadratic deviation compared to the residual correction to be imaged. If this residual correction is a linear function of the input network, the network learns this connection exactly. The non-linearly calculated function parameters allow an extensive plasticity of the network, which extends the application spectrum. If necessary, this adaptation can also be made robust by a suitable choice of the corresponding learning rate.

If the network is defined in a standardized form, inputs and outputs with arbitrary value ranges can be used. This allows the models to be transferred without complex parameter adjustments. This gives the possibility of a universally usable independent adaptive module for vehicle controls.

In the present application described above, a value in the air mass flow correction of the idle with variable valve timing is modeled with an adaptive physical-neuronal model. Using an existing mass flow correction value of the requested and corrected target mass air flow and the engine oil temperature T as an input, the neural model calculates the additive air mass flow correction value K in idle mode. It thus maps a one-dimensional nonlinear unknown functional relationship between additive residual correction of air mass difference and oil temperature.

In the present case, an adaptive model for real-time control with a conventional controller, which covers, for example, a basic function, and a neural correction transmitter, which optimizes the controller output with regard to the control quality, are provided. This process runs automatically and does not generate any additional modeling or design effort.

Claims (8)

Control device comprising a controller (R), which receives at least one input variable (E) and determines from this at least one manipulated variable (U) which is fed to a controller of a controlled system (S), wherein a neural correction transmitter (NAM) is provided, at least a process variable (P ₂ ) from the controlled system, at least part of the controller input variables (E = V ₁ (M _soll , M _is , P ₁ )) and at least a part of the control variables determined by the controller (U) receives and is designed to to generate a correction signal (K) on the basis of these input variables, wherein the correction transmitter (NAM) is designed in such a way, on the one hand, to generate at least one correction signal (K) which corresponds to the manipulated variable or variables (U). to an algebraically concatenated to a corrected manipulated variable (U _K ), - and on the other hand an adaptation of the, the neuronal corrector (NAM) influence in its operation affecting internal parameters forth, the adaptation of the correction transmitter (NAM) with an online learning algorithm in Real time is performed. Control device according to claim 1, characterized in that the neuronal correction transmitter (NAM) additionally at least an internal size of the regulator (R) is supplied. Control device according to one of claims 1 to 2, characterized in that the algebraic concatenation is a summation. Control device according to one of claims 1 to 3, characterized in that the controller (R) implements a control concept, which is based on a physical process model, an empirical approach or a design based on a dynamic process identification. Control device according to one of the preceding claims, characterized in that the neural correction transmitter (NAM) can be optionally faded out. Control method comprising the steps of - that at least one input variable (E) is supplied to a controller (R), - that the controller (R) determines at least one manipulated variable (U) from the at least one input variable (E), - that a neural corrector ( NAM) at least a part of the input variables of the controller (E), at least part of the manipulated variables from the controller (U) and at least one process variable (P ₂ ) are supplied, - that the neural correction transmitter (NAM) due to the signals supplied to it at least one correction variable (K) generates, - that the at least one correction variable (K) is algebraically linked to at least one manipulated variable (U) and - that an adaptation of the internal parameters influencing the neural correction transmitter (NAM) with its input parameters is brought about, wherein the Neuronal Correction Transmitter (NAM) is taught online in real time using a specific learning algorithm. A method according to claim 6, characterized in that the neuronal correction transmitter (NAM) additionally at least one further internal variable of the controller (R) is supplied and the neural correction transmitter (NAM) also takes into account in the determination of the correction quantity (K). Method according to one of claims 6 to 7, characterized in that the neural correction transmitter (NAM) is hidden under certain conditions.

Images