WO1996005421A1

WO1996005421A1 - Method and system for controlling combustion engines

Info

Publication number: WO1996005421A1
Application number: PCT/SE1995/000914
Authority: WO
Inventors: Jan Nytomt
Original assignee: Mecel Ab
Priority date: 1994-08-11
Filing date: 1995-08-08
Publication date: 1996-02-22
Also published as: SE9402687D0; SE509805C2; SE9402687L

Abstract

The invention relates to a method for a combustion engine using a neural net controlling the primary values, preferably ignition timing or amount of fuel but also the charge pressure, and a learning system for the neural net used. By using at least engine speed and engine load as input signals supplied to a neural net, and learning the neural net to create the necessary value for the primary value in concern dependent of the input parameters, is a smoother control of the combustion engine obtained without steep transitions between different operating cases. The control based upon a neural net exhibit considerable advantages compared to the conventional method having current amount of fuels or ignition timings of interest stored in a map containing primary values thereof. The neural net is trained by using a temporary connected learning computer (10) having an additional linear lambda sensor (5), and the current value of the primary value calculated by the learning computer is used as a target value when optimising the weight factors of the neural net, which optimisation is performed in order to reduce the deviation between the output signal from the neural net and the output signal from the learning computer. When disconnecting the learning computer, which is done when the learning process of the neural net is completed and the deviation is within acceptable limits, is a switch device (12) connecting the neural net for further control of the primary value.

Description

METHOD AND SYSTEM FOR CONTROLLING COMBUSTION ENGINES

This invention relates to a method specified in accordance with the preamble of claim 1, and a system specified in accordance with the preamble of claim 9.

PRIOR ART

When controlling combustion engines, that is controlling ignition timing or amount of fuel supplied, most state of the art engines use computer controlled systems, where the ignition timing or amount of fuel supplied are control data which are obtained from a table of stored values. Different detected engine parameters are used as input data to the tables, such as number of revolutions(revs) and load(altematively throttle pedal or throttle position). In a two-dimensional table or map could revs and load constitute the X- or the Y axis respectively, and for a number of specific loads and revs having pre-stored amount of fuel, alternatively a pre-stored ignition timing. Between these specific loads and revs is each fuel amount or ignition timing of interest interpolated , in order to obtain a linear transition between the specific operating cases defined in the table. In addition to load and revs are a number of engine parameters influencing the amount of fuel or ignition timing used, increasing or decreasing the amount of fuel or advancing or retarding the ignition timing. Such engine parameters could be the engine temperature, which could constitute a third axis in a three- dimensional table, and corrective engine parameters such as derivative of throttle position , derivative of revs, inlet air pressure (often used in supercharged engines), air temperature, knock intensity, current lambda- value, or in some cases also the speed of the vehicle driven by the engine. These corrective engine parameters could give correction values for the control parameters, which correction values are stored in tables or alternatively will be given by predetermined functions. The control will hence be rather complex and demanding an extended memory capacity for storing the often very extensive tables.

The engine control using control data stored in tables for specific operating cases, with interpolation between these specific operating cases, also leads to that the control of the engine is experienced as rough, having steep transitions between different operating cases.

In automotive use a new technique for controlling different type of systems has instead employed an artificial neural net. For example in DE.A.4211556 is an artificial neural net used for modelling the reaction of the driver and the behaviour of the vehicle in real environments, alerting the driver for hazards and supporting the guidance of the vehicle. In DE,A,4300844 is an artificial neural net used for adapting the characteristics of the vehicle to the driving behaviour of the driver. In Automotive Engineering, SAE congress issue, February 1993, page's 52-55, is shown an engine suspension system, having an artificial neural net being able to learn from the detected oscillations of the engine and the control actions taken, with a view to optimise the damping capacity. In for example Bart Kosko's textbook "NEURAL NETWORKS AND FUZZY SYSTEMS, A DYNAMICAL SYSTEMS APPROACH TO MACHINE INTELLIGENCE", issued by Prentice Hall, Englewood Cliffs, NJ 07632, USA, are in an exhaustive manner described the methods which might be used when creating and optimising neural nets, obtaining control systems having artificial intelligence and being able to learn to create the desired output signals dependent of the parameters i form of input data given. For a more detailed information regarding different methods used for creating, learning and optimising neural nets, reference is made to the above mentioned textbook of Mr Kosko. This application for patent is only accounting for the new implementation of control systems founded on neural nets, and will hence not go through the theory of neural nets in detail, when it comes to adaptation of functions in the nodes, and optimising the deviation of the output signal of the neural net from the desired output signal by using different methods such as "Backpropagation", etc.

OBJECT OF THE INVENTION

The object of the invention is obtaining a smoother control of the combustion engine, and reducing the required capacity of memory for the control system of the combustion engine.

Another object is to simplify, with the right kind of equipment for authorised service-facilities and during manufacturing, re-programming of the control system of the combustion engine, in order to meet various demands put on emissions and requests concerning engine character and response. Entering the tables and modifying all of the control data contained for the different operating cases, which must be done in the state of the art type of systems, is a very time consuming process that has to be initiated when the conditions have changed.

Yet another object is to meet in an efficiently manner the stricter demands on low levels of emissions, demanding more accurate control of the combustion process and further input data from the engine i form of engine parameters detected.

Yet another object is to obtain more information for the control than could be obtained from the representation of each individual input data. Combinations of input data could also be used by the inventive control in order to obtain an optimal control.

This will be effected automatically without knowing in advance the kind of combinations of input data which might affect the quantity of the control data, and there is no need knowing the specific terms for combinations which might arise during certain operating cases Yet another object is obstructing unauthorised changes in the control system of the vehicle, often made in order to increase the effective output without taking fuel consumption or emissions into consideration, which facilitate taking due responsibilities of product liability and warranty engagements concerning levels of emissions.

SHORT DESCRIPTION OF THE INVENTION

The inventive method is characterised by the characterising clause of claim 1.

The inventive system used for training and learning of the neural net giving the desired output signal, is characterised by the characterising clause of claim 9. Other distinguishing features of the invention is evident from the characterising clauses of the dependent claims and the following description of preferred embodiments, which description is made by reference to the figures specified in the following list of figures.

FIGURES Figure 1 shows a model of a neural net used by the inventive method, controlling a combustion engine,

Figure 2 shows schematically the engine parameters required by the combustion engine control system (CPU) for controlling the control data, i.e. ignition timing or fuel amount, Figure 3 shows a control system for fuel supply, having a temporary connected control unit 10, for the learning process of the neural net.

DESCRIPTION OF EXEMPLARY EMBODIMENTS

The inventive method for controlling the control data for a combustion engine use so-called artificial intelligence in form of a neural network. A neural network implemented for controlling the control data of a combustion engine, functions in such a way that for each operating case, i.e. given by input data in form of detected parameters of the engine, is the neural net learnt to generate an output signal or output data dependent of the input data received.

In figure 1 is shown a neural net having a number of nodes(En-Ei5, E₂ι-E₂₃, E₃ι-E3₂, .. , E-, which are organised in a number n of layers. All nodes in the first layer are input data nodes, each node respectively connected to every node in the next layer being a so-called hidden layer. The first layer, having the input data nodes(E_u-Ei₅), is followed by at least one bidden laye E^-E^), and in some cases could several hidden layers follow as indicated by the nodes (E₃ι-E₃₂) having dotted contours in figure 1. The neural net is terminated by one or several output data nodes E _n! in the final layer of order n. The output signal from each node and layer is transmitted only to each node in next succeeding layer. The input data signals are sent to the input data nodesCEπ-Eis), in such a way that a unique value is received by each respective input data node. For the control of the combustion engine could for example transformed or re-scaled values, representing the engine speed, load, engin temperature, throttle position and temperature of ambient air, be fed to its respective node En -E1₅. Re-scaling is performed in order to improve the performance, such that each value could assume a value in the range between 0 to 1 , or any other suitable measuring range. The output signal from the input data nodes constitutes nothing else than the re-scaled or transformed values for the corresponding magnitude of the input data.

The output signal from each remaining node, i.e. excluding the input data nodes, constitutes the sum of the output signals from preceding nodes according a simple and preferably logical function having a weight factor W at each node. The adaptation or learning process of the neural net is performed by using a non-linear approximation of a function, and preferably is a so-called Backpropagation method used, where the weight factor W of each individual node is adjusted in such a manner that th output signal of the neural net corresponds to the desired output signal.

The number of nodes, for example input data nodes, and the number of hidden layers are adapted to each type of control method, in such a way that the least deviation between the desired output signal and the output signal from the neural net is minimised as fast as possible, with all available input data designated to an individual input data node.

The weight factor W could at the first start of the learning process be given an individual random value. When the neural net is to be updated for a slightly modified output, possibly motivated by changed emission demands or emission test cycles, could preferably the previously used factors of weight be used at start in order to shorten the learning process.

In figure 2 is shown a survey of the types of input data required for controlling a combustion engine, such that an optimal combustion and reduced fuel consumption could be obtained

The momentary engine parameters which controls the amount of fuel or the ignition timing is primarily the speed E-pm and load Th(throttle position), but also the derivative of these parameters, d/dt E-p_m and d/dt Th, the inlet air pressure P_ω, the temperature of the inlet air T_m, and in some cases also the speed of the vehicle Speedv* driven by the combustion engine. In some cases could a calculation of the air mass supplied be used in order to establish the current load, either by using the parameters P.,-, and T_m, or alternatively using hot-wire detectors arranged in the inlet manifold. These latter methods are often used in systems needing a more accurate control of the lambda value. An additional number of engine parameters are detected in order to obtain a feed back from the combustion, such that the amount of fuel or the ignition timing could be adjusted to a more optimal amount of fuel or ignition timing. Examples of such engine feed back parameters could be a knock signal, Knock, from a special knock detecting circuit, a signal from a lambda sensor, λ, indicating the residual amount of air in the exhaust gases, or a signal Ion_curre-._t obtained from a circuit detecting the degree of ionisation within the combustion chamber. A knocking condition is damaging to the engine and results in a non optimal usage of the fuel. In order to terminate the knocking condition is the ignition timing usually retarded or alternatively complemented by increasing the fuel amount supplied.

The signal from the lambda sensor is used in order to maintain a desired air-fuel mixture, specifically in Otto engines having a three-way catalyst with optimal function at a lambda value of 1.0. Another feed back signal from the combustion is the ionisation current in a measuring gap arranged in the combustion chamber. The ionisation current could in a simple manner detect a misfire, causing ionisation failure.

In figure 3 is schematically shown a control system for fuel supply to a combustion engine 1, having a neural net control implemented in a microcomputer based control unit 11 , and a temporary connected learning computer 10 used during the learning process controlling the fuel amount supplied and optimising the output signal of the neural net. The control system consists of two units during the learning process, a permanent control unit 15 and a temporary unit 13, which temporary unit only is connected during the learning process and disconnected at the interface 14, indicated with dots. The interface 14 having conventional common connector means connecting all signal lines.

An additional lambda sensor 5 is also arranged in the exhaust system 4 for the learning process. The lambda sensor 5 being a linear type of lambda sensor, which unlike a narrow banded conventional lambda sensor have an output signal being proportional to the residual amount of air in the exhaust gases. The conventional type of narrow banded lambda sensor have a distinctive transition and hence only able to detect if the lambda value is below or above 1.0. The linear type of lambda sensor is required when performing a more accurate control, which is a prerequisite for obtaining a proper result from the learning process. The linear type of lambda sensor is more expensive than a conventional narrow banded lambda sensor, which is the major reason why mass produced engines are equipped with narrow banded lambda sensors.

The fuel injectors 2 arranged in the inlet manifold of the engine, injects an amount of fuel fixed by a control pulse on line 23, respectively Oine 23 is divided into separate lines for each injector). During the learning process of the neural net is the amount of fuel supplied controlled by the learning computer 10 through control pulses at the output terminal 21, which output terminal is connected to each individual line 23 via a switching circuit 12. When the learning computer 10 is activated or connected, is the switching circuit automatically forced to switch to terminal 12b, from a preferably stable terminal position 12a, and the output signal at output terminal 22 of the neural net controlled computer 11 could not be transmitted to the fuel injector, respectively. The present engine parameters are detected by sensors 6 arranged at the engine, only one sensor shown in the figure, and these inp signals are transmitted to learning computer as well as the neural net controlled computer 11. As previously described are the input signals supplied to the input data nodes of the neural net controll computer. The neural net will thus produce an output signal dependent of the present input signals. The learning computer will simultaneously calculate the necessary amount of fuel on a basis of the detected input signals, and will supply an output signal to each individual fuel injector correspondin to the necessary amount of fuel . The output signal from the output terminal 20 of the learning computer 10 is also connected to an input terminal of the neural net controlled computer, which output signal is used as target value for the adjustment of the output signal from the neural net. When the learning computer detects the proper lambda value, i.e. the proper residual amount of oxygen in the exhaust gases is present, will the learning computer 11 send a signal 24 to the neural net controlled computer 11. When this signal is present, is the proper operating case assumed, in thi case an operating case optimised regarding emissions, and the present output signal from the neural net is compared with the output signal at output terminal 21 of the learning computer. The content of the neural net, i.e. the established weighted function for each node respectively, is thereafter optimised in such a manner that the resulting output signal from the neural net approache the target value. During this optimisation/training of the neural net are the input signals supplied to the first layer of the neural net, followed by a successive propagation of signals through each layer the neural net until an output signal is obtained at the output node of the neural net. The output sign from the neural net is thereafter compared with the output signal at output terminal 21 , and the weight factor of each node of the neural net is corrected in order to reduce the deviation between the output signal from the neural net and the output signal at output terminal 21. This correction or learning procedure for the weight factors could preferably be performed by using a so-called "Backpropagation-method", see textbook of Mr Kosko, which "Backpropagation" is performed a repeatedly number of times until the deviation is at an acceptable level. In some cases it could be necessary to repeat this procedure between 10⁶-10⁷ times during the learning process, before the output signal from the neural net in an acceptable manner corresponds to the desired target value. A predetermined acceptable level of deviation between the output signal from the neural net and the target value could be one or some percent, which should be obtainable if the right choice of neural net model is used and an adequate number of corrections of the weight factors is performed, for example by using "Backpropagation". If insufficient correspondence is obtained, then the neural net model must be revised.

The ignition timing, instead of fuel amount, could in the same manner as described with reference to figure 3, be controlled by a neural net controlled computer. The output signal from the neural net is then during the learning process compared with a signal representative for an ignition timing target value. The ignition timing target value is then preferably obtained from the learning computer 10, and in the corresponding manner is the deviation between the representative signal of the ignition timing target value and the output signal from the neural net minimised. In a supercharged combustion engine could also the charge pressure be controlled in a corresponding manner by a neural net controlled microcomputer. The charge pressure level is most often given from value stored in a table dependent of the present load and speed. Similarly to the fuel- and ignition timing control could a learning system corresponding to that shown in figure 3 be used. The line 23 will then instead control preferably a so-called waste-gate, or any similar control device for the supercharging unit.

The system shown in figure 3 could also include a narrow banded lambda sensor among the sensors being permanently attached to the engine, which sensors are only schematically represented as one common symbol in the drawing. The linear type of lambda sensor 5 will then constitute a second lambda sensor that only will be used during the learning process of the neural net. In other types of systems for smaller combustion engines, for example hand-held engines, could the conventional narrow banded lambda sensor be excluded, but the trained neural net will automatically strive in its mode of control maintaining a lambda 1.0 value, this being the base for the generation of control data in question (amount of fuel or ignition timing) from the neural net.

In a rather simple model could the fewest amounts of input parameters supplied to the first layer of the neural net be no more than two, i.e. speed and load. For combustion engines having stricter demands in aspects of control accuracy, could besides load and speed also engine temperature, derivative of throttle position and derivative of speed be supplied as input signals to individual input data nodes of the first layer of the neural net. The neural net controlled microcomputer 11 will then include means for detecting speed as well as calculation of the derivative of speed, as deduced for example from a crankshaft sensor. In a corresponding manner could the throttle position be obtained from a throttle position sensor, or if an electrically controlled throttle is used (the position of the throttle being controlled by an electric motor independently of the throttle pedal position) could the control signal sent to the throttle motor constitute the throttle signal, and the calculation of throttle derivative could be performed by the neural net controlled microcomputer.

All signals are transformed or re-scaled such that the input data signals supplied to the first layer of the neural net have similar signal range, preferably between 0-1 volts, which facilitate the learning process of the neural net. The load signal could in very simple system only be constituted by the throttle position, especially in aspirating engines, but for example in supercharged combustion engines is the load calculated from the detected amount of air supplied to the engine, which amount of air could be calculated from the detected pressure and temperature of the air supplied to the engine. This amount of air could preferably be calculated by the neural net controlled microcomputer 11 , and a signal representative for the amount of air is supplied to an input data node in the first layer of the neural net.

The neural net thus trained will for each identical combination of input signals repeatedly issue a corresponding primary value of the control data in concern, in a corresponding manner as a primary value will be given the control data from a table or map.

A possibly detected knocking condition or the output signal from a narrow banded lambda sensor arranged in the exhaust system could in a known manner cause a correction of the primary value of the control data given by the neural net, according an established corrective routine.

A knocking condition, often treated as binary condition (on off), is often controlled such that the control data is corrected in one larger corrective step at the instant of knock detection, followed by return to the ideal primary value given by the neural net or the table/map in successive incremental steps of a smaller order. Knock control could also be performed according more sophisticated routines, with a view to terminating the knocking condition more quickly at minimum deviation from the ideal condition. An initially detected knocking condition, occurring during ideal conditions i.e. when the control data is equivalent to the control data given by the neural net, could cause a corrective step having another order of size than a corrective step initiated during a recurring knocking condition which occurs when the control data is already subjected to corrective actions. Such control routines of higher intelligence could preferably be located externally of or after the neural net, i.e. the knock related signal is not supplied as input data to the neural net, and the corrective action taken of the primary value of the control data given by the neural net is performed as an additional routine. An alternative is that the knocking detection circuit is supplying and adapted input signal to a input data node in the first layer of the neural net, where the input data signal is dependent of whether or not a knocking condition have been detected or if the knocking condition is a so-called recurring knocking condition occurring during initiated corrective actions of the control data, and where the input signal supplied to the input data node will be returned to a value representative for a non- knocking condition according a predetermined function given by the knocking detection circuit.

If the ignition timing as well as the fuel amount, and in some cases also the charge pressure, should be controlled by a neural net, then an individually adapted neural net for each respective type of control data is used, where the neural nets in this manner are operating in parallel in relation to each other. Each individual neural net could as an input data signal use an output signal from one or more neural nets. Thus could, during the learning process, the signal representative for the fuel target value, i.e. the amount of fuel requested from a learning computer, be supplied as an input data signal to the neural net for the ignition timing control, and inversely could the signal representative for the ignition timing be supplied as an input data signal to the neural net for the fuel amount control. Each neural net could hence be adapted to its own function, and the complexity of the neural net, i.e. the number of layers and nodes, could be reduced.

The learning process could be performed in a laboratory as well as in an operating vehicle. It is also possible to repeat the learning process as a service action, when the combustion engine has been operating for a certain period causing wear and other changes in the operating conditions that might require modification. This is advantageously due to that real operating conditions could be tested and adjustments made in the real environment. Another advantage is that for each learning process could the operating cases of most interest be chosen, and the system could be trained more vigorously for these special operating cases.

The neural net control lead to that it will become more difficult doing unauthorised changes of the control function, while at the same time admitting during manufacturing optimisation of the control function according different conditions such as emission, response and/or fuel consumption, dependent of how and where the combustion engine should be used Unauthorised tuning is done frequently in microcomputer controlled systems, where the table containing the control data is stored and easily could be read and modified and stored in a new EPROM-memory, due to that the table is stored in a restricted area of the memory and in a relatively easy manner could be interpreted. The weight factors of the neural net could not as easily be modified without having in depth knowledge of neural nets, and a modification of any or some weight factors could obtain limited improvements of special purposes for some specific operating cases , but where other operating cases obtain changes for the worse. The need for usage of E²PROM or FLASH- type of memories, which are more difficult to copy and change as well as more expensive, is therefore reduced considerably.

There is no need for an in depth knowledge of the neural net theory for the implementation of the invention, because usage of supporting programs of a type specified in US,A,5235673 could be made, having a shell-program which selects by itself an optimal model for the neural net and the most suitable methods for the present type of input data parameters.

Claims

1. Method for controlling the primary value of a control data, preferably ignition timing or amount of fuel but also charge pressure, for a combustion engine, which primary value through feed back signals, such as a detected knocking condition or output signals from a lambda sensor, could be corrected in a manner known per se, c h a r a c t e r i s e d i n

- that a number of parameters of the engine is detected, where at least the speed(Erpπι) and loadCTh) form part of these parameters, which load parameter preferably is detected by the current throttle position, throttle pedal position or a combination of pressure and temperature of the air supplied, -that the detected parameters of the engine constitute the input data supplied to the nodes in a first layer of a neural net,

-that at least one hidden layer with nodes is included in the neural net, where each of the nodes of the hidden layer located closest to the input data nodes is connected to each input data node of the first layer respectively, and where the nodes of the hidden layer located closest to the input data nodes are transmitting an output signal dependent of the input data, which output signal from each of the nodes in the hidden layer have an unique and determinable weighted function of the input data supplied to the nodes of the first or preceding layer,

-and that a final layer included in the neural net have at least one node in turn being connected to each of the nodes in the preceding hidden layer, where the node in the final layer is transmitting an output signal dependent of an unique and determinable weighted function of the output signals from the preceding hidden layer,

-and the weighted function of each node is determinable through a learning process of the neural net in such a manner that the output signal from the node in the final layer has a least deviation from an expected primary value of the control data, dependent of the input data supplied.

2. Method according claim l c h a r a c t e r i s e d i n that besides the engine parameters speed (E -_n) and load(T_h) is also the parameters,

-engine temperature T-_ng, preferably detected as the temperature of the coolant fluid, -derivative of the throttle position, d dt Th, and -derivative of speed, d/dt E-pm, constituting input signals supplied to the nodes in the first layer of the neural net.

3. Method according claim 2 or 3 c h a r a c t e ri s e d i n that the weighted function of each node constitutes a permanently stored information of the combustion engine control system, whereby an output signal from the neural net representative for a primary value dependent of the detected prevailing parameters present is obtained in a repeatedly manner for operating cases having identical combinations of input data signals.

4. Method according claim lor3characterised in that the primary value constitutes an adaptable ignition timing dependent of the detected parameters of the engine.

5. Method according claim lor3characterised in that the primary value constitutes an adaptable amount of fuel dependent of the detected parameters of the engine.

6. Method according claim lor3characterised in that the primary value constitutes an adaptable charge pressure in a supercharged engine dependent of the detected parameters of the engine.

7. Method according claim lor3characterised in that for each primary value, ignition timing/amount of fuel/charge pressure e.t.a., to be controlled is an individually adapted neural net used for each respective primary value.

8. Method according claim 7characterised in that each primary value obtained from respective neural net, is feed as an input signal back to the neural nets used for other primary values.

9. System for training a neural net used in a permanent control system(l 1) controlling a combustion engine(l ) according claim lcharacterised in -that an additional linear lambda sensor(5) is arranged in the exhaust system(4) of the combustion engine

-that a closed loop control system (10), including the additional linear lambda sensor, controlling the current primary value of interest, is temporarily connected during a learning process of the neural net, which control system(lθ) during the learning process controls the primary value in a closed loop manner maintaining predetermined target values, and where the control system creates a target value for the output signal of the neural net corresponding to the controlled primary value for a number of operating cases, and where the weights of the nodes in the neural net is corrected when the current residual amount of air in the exhaust gases is within predetermined acceptable limits and when deviation between the output signal of the neural net and the controlled primary value exists, and corrected in such a way that the deviation is minimised,

-that the temporarily connected closed loop control system via switch means (12) could be disconnected and instead connects the output signal from the neural net at line(22) to a servo device(2) affecting the primary value in concern, which disconnection is made when the deviation has been minimised to a predetermined level.