FR2893982A1 - Exhaust gas temperature estimating method for e.g. supercharged diesel engine of automobile, involves implementing estimator arranged in logic controller, where estimator estimates temperature of gas downstream of post-treatment system - Google Patents

Abstract

The method involves implementing an estimator (15), arranged in a logic controller, comprising a series of networks of neurons (17, 37) mounted in cascade. Each network has a feedback loop directly or indirectly returning available quantities in an output of the network to an input of the network. Information relative to the temperature of exhaust gas downstream of a post-treatment system arranged on an exhaust line of an engine is provided to an input of the estimator, where the estimator provides the estimation of the temperature of the gas at its output. Independent claims are also included for the following: (1) a system for estimating a temperature of exhaust gas of an engine (2) an internal combustion engine comprising an estimating system.

Description

The field of the invention is that of controlling a combustion engine

  internal, and more particularly exhaust devices mounted downstream of the exhaust manifold of the engine. The invention more specifically relates to a method and a system for estimating an internal gas temperature of a post-treatment system, and an internal combustion engine equipped with such a system. To reduce the emissions of gaseous pollutants from motor vehicles, gas after-treatment systems are generally lo arranged in the exhaust line of the engines. These systems are designed to reduce emissions of carbon monoxide, unburnt hydrocarbons, particulate matter and nitrogen oxides. These systems operate discontinuously or alternately, alternating pollutant storage phases and trap regeneration phases (ie conversion of stored pollutants into non-polluting substances). In order to optimize the treatment of all the pollutants, it is necessary to better control these storage and regeneration phases. In particular, it is necessary to estimate over time the trapped masses (ie the particles in the case of the particulate filter, the nitrogen oxides in the case of the nitrogen oxide trap). Similarly, it is necessary to know the evolution over time of the masses converted during the regeneration phases. However, the evolution of these masses during the storage and regeneration phases depends directly on the temperature of the support of these traps and the gases passing through them. We therefore seek to know, if not to control, the internal gas temperature of the post-treatment systems. Moreover, the increase in the complexity of the engines and their modes of operation require electronic management means.

  more and more sophisticated and by the same token, means of measurement or estimation more and more numerous. But it turns out that many physical quantities are not directly measurable as the internal temperature that concerns us.

  It therefore appears necessary, for the control of the regeneration of the traps and for the control of the engine itself, to know the temperature of the exhaust gas of the engine. Obtaining an accurate estimate of the temperature of the exhaust gas within a post-treatment system disposed downstream of the turbine of a supercharged engine can thus be particularly useful. The invention aims to solve this problem of estimating the temperature within a post-processing system. To find the temperature of the gases in-house of a physical and / or chemical treatment device of the exhaust gases placed downstream of the turbine of an engine, today we use estimation models of this type. temperature implemented by embedded electronic computers. These numerous more or less complex estimation models are based on physical equations making it possible to estimate the gas temperature internally of a post-treatment system. It will thus be possible to refer to the international patent application WO 2004092555 filed on April 13, 2004. It proposes to use a physical model estimator to estimate the thermal state along a catalytic trap with nitrogen oxides or NOx. from the inlet temperature of the trap, determined by a sensor. The solution described in this document does not make it possible to overcome the thermal aging of the temperature sensor used at the inlet of the trap, the accuracy of the estimate of the temperature achieved being then able to be derived with it.

  It is thus necessary to remove some sensors while maintaining the ease of mapping functions such as thermal regulation or regeneration of particulate filter. On the other hand, the use of physical thermal models has drawbacks. If these models are generally faithful in steady state gas, they are rather poor transient. In addition, many parameters (physical parameters, state variables) necessary for these models are difficult to identify or to measure on the motor. In addition, these models are most of the time too greedy in terms of computing load or memory resource to be implemented in an electronic engine management computer. The present invention aims to improve the knowledge of a temperature of the exhaust gas, and in particular the temperature in-house of a device for physical treatment and / or chemical exhaust gas. For this purpose, it is proposed, according to the invention, a method for estimating a temperature of the exhaust gas of an engine, characterized in that it implements an estimator comprising a series of networks of 20 neurons cascaded. Some preferred but non-limiting aspects of the method according to the invention are as follows: the estimator comprises neural networks each having a feedback loop returning directly or indirectly at the input of the network one or more of the available quantities at the output of the network; the input of the estimator is provided with information relating to the temperature of the gases downstream of a post-treatment system disposed on the exhaust line of the engine, the estimator providing at the output 30 the estimate of the temperature of the gases exhaust system internally from said system; one of the neural networks of the estimator receives as input at least one estimate of the temperature at the output of a monolith, the size available at the output of a preceding neural network, and outputs an estimate of the temperature exhaust gases internally from a post-treatment system; the method can implement a post-treatment of the temperature estimate made by the last network of neurons of the series; the method can make a feedback of the estimated temperature, available at the output of the post-processing module; the method may implement pretreatment of one or more input variables of the estimator by performing calculations based on known physical relationships; the method may implement a reprocessing of some of the magnitudes at the output of each network before these are returned, according to a so-called indirect loopback, to the input of the network or networks; the method may comprise a preliminary step of learning the estimator using a database representative of operating areas of interest. According to another aspect, the invention relates to a system for estimating an exhaust gas temperature of an engine comprising an estimator comprising a series of neural networks characterized in that said neural networks are cascaded. . According to still other aspects, the invention relates to an internal combustion engine equipped with a system for estimating an exhaust gas temperature according to the invention. Other aspects, objects and advantages of the invention will appear on reading the following detailed description of preferred embodiments thereof, given by way of nonlimiting example and with reference to the appended drawings, in which: - Figure 1 is a schematic view of an internal combustion engine according to one aspect of the invention; FIG. 2 is a schematic view of a system for estimating an exhaust gas temperature according to one aspect of the invention; FIG. 3 is a diagram showing the steps of the method for estimating an exhaust gas temperature according to one aspect of the invention. In Figure 1, only the elements necessary for the understanding of the invention have been shown. An internal combustion engine 1 is ~ o intended to equip a vehicle such as a car. The engine 1 is for example a diesel engine turbocharged four-cylinder in-line and direct fuel injection. The engine 1 comprises an intake circuit 2 providing its air supply, an engine control computer 3, a pressurized fuel circuit 4, and an exhaust line 5 of the gases. The injection of the fuel into the cylinders is ensured by injectors, not shown, opening into the combustion chambers and controlled by the computer 3 from the circuit 4. At the output of the engine 1, the exhaust gases discharged into the engine 20 exhaust line 5 pass through one or more after-treatment devices 6 (eg particle filter, nitrogen oxide filter, Nox trap). A turbocharger 7 comprises a compressor disposed on the intake circuit 2 and a turbine disposed on the exhaust line 5. Between the compressor and the engine 1, the intake circuit 2 comprises a heat exchanger 8 for cooling the engine. compressed air at the outlet of the compressor and thus to increase its density, an air intake flap 9 controlled by the computer 3, and a pressure sensor 10 connected to the computer 3. At the output of the engine 1, in upstream of the turbine, the exhaust line 5 further comprises means 30 adapted to provide information relating to the temperature of the exhaust gas upstream of the turbine. The exhaust line 5 also comprises, downstream of the aftertreatment systems 6, means 40 adapted to provide information relating to the temperature of the exhaust gases at the outlet of the after treatment system or systems 6. Said means 30 and 40 are for example constituted by a temperature measuring probe or by an estimator provided to provide an estimate of said exhaust gas temperature lo. The engine 1 also comprises an exhaust gas recycling circuit 11 equipped with a valve 12 whose opening is controlled by the computer 3; it is thus possible to reintroduce exhaust gases into the intake circuit 2. An air flowmeter 13 is mounted in the intake circuit 2 upstream of the compressor to supply the computer 3 with information relating to the flow rate of the intake air supplying the engine. Sensors 14, for example pressure or temperature, may also be provided. The computer 3 comprises, in a conventional manner, a microprocessor or central unit, memory zones, analog / digital converters and various input and output interfaces. The microprocessor of the computer comprises electronic circuits and appropriate software for processing the signals coming from the various sensors, deriving the states of the motor 25 and generating the appropriate control signals for the various actuators controlled such as injectors. The computer 3 thus controls the fuel pressure in the circuit 4 and the opening of the injectors, and this, from the information supplied by the various sensors and in particular the air mass admitted, the engine speed, and memorized calibrations

  to achieve the desired levels of consumption and performance. The computer 3 further comprises a neural network estimator 15 provided for making an estimate of the internal temperature of the exhaust gases of the after-treatment device (s). For the sake of simplicity and sharing of a certain number of resources, in particular calculation and memory, it is particularly advantageous to have the estimator 15 within the computer 3.

  As illustrated in FIG. 2, the estimator 15 comprises a series of two cascaded neural networks 17, 37 as well as a preprocessing module 16 and a post-processing module 18. A series of neural networks 17, 37 cascaded means that each network uses as input a quantity available and calculated by the network that precedes it. Each of the networks 17, 37 has a feedback loop provided to return to the input of the neural network (s) 17, 37, one or more of the variables available at the output of said network. Each of them 17 and 37 includes an output channel relating to the desired temperature estimation, as well as possibly one or more other output channels relating to non-measurable state variables intended to be looped back, directly or indirectly, to the input of the network or networks 17 and 37. The first neural network 17 makes it possible to estimate the temperature at the output of the monolith while the second neural network 37 has an output path relating to the estimation of the the temperature of the exhaust gases internally of a post-treatment system 6. The first neural network 17 furthermore comprises an input relating to the information relating to the temperature of the exhaust gas downstream of the post-treatment systems 6, as provided for example by the means 40.

  It may also comprise one or more other entries relating to physical quantities representative of the state of the engine, such as, for example: the mass flow rate of the gases, the engine speed, the gas pressure upstream and downstream of the post-treatment system. , the vehicle speed, the ambient air temperature, to the supercharging pressure measured by the pressure sensor arranged downstream of the compressor, the heat exchanger and the air intake flap, the outlet temperature of the post-treatment system, the temperature at the output of the neural network calculated at the preceding step. The first neural network 17 can also receive as input the fuel flow, the air flow. The second neural network 37, meanwhile, receives mainly the same inputs relating to physical quantities representative of the state of the engine but also an estimate of the output temperature of the monolith calculated by the first neural network 17 and the temperature of the exhaust gas internally of a post-processing system 6 calculated in the previous step (by itself). Thus, the estimator 15 comprises a neural network 37 receiving at least at least one estimate of the temperature at the output of a monolith, a quantity available at the output of the preceding neural network 17 for outputting an estimate of the temperature of the The neural networks 17, 37 are each made up of a certain number of neurons defined by their parameters (weight, bias), and by their functions. activation. The output of a neuron is related to the inputs

  (el, e2, en) of the neuron by s = F (ei * w1 + e2 * w2 + ... + en * wn + b), where F is the activation function of the neuron, w2, ... wn the weights and b the bias. The number of neurons in a network, the values of the weights and the bias of the different neurons are calibration parameters which can in particular be determined during a learning phase which will be described later on later. Moreover, the preprocessing module 16 makes it possible to process one or more input variables of the estimator 15 by performing calculations based on known physical relationships. The pretreatment module 16 allows in particular to reduce the number of inputs of the neural networks 17 and 37 by calculating one or more variables al, a2 each representative of one or more physical quantity (s), measurable ( s) or not, absent at the input of the computer 3, from several input variables. The module 16 also makes it possible to filter certain inputs 15 that may take outliers. The post-processing module 18 makes it possible to process the available temperature estimate at the output of the last neuron network 37 of the series in order to emit an estimated temperature signal S at the output of the estimator, available to the computer 3. This processing for example, filtering can be used to make the estimator robust by limiting the return to the input of an outlier. As previously mentioned, one or more of the output variables of each network 17, 37 follow a feedback loop and are returned to the input of the neural network (s) 17, 37. Some of the output variables of the network may be processed (for example, examples filtered, delayed, or associated with other entries in the preprocessing module) before being returned to the input of the network. With reference to FIG. 2, the case where state variables / 31, fl2, / 33, 01, 02, 03 (not measurable and known only during the training) are available, respectively, in i0

  the output of the networks 17 and 37 are looped back to the input of their respective networks 17 and 37. It is also illustrated the case where the output S of the estimator 15 and more precisely the second neural network 37 (ie the estimate of the temperature of the gases internally of a post-treatment system 6 disposed on the gas exhaust line available at the output of the post-processing module 18 (and whose value is here measurable during of learning) is firstly looped back after being delayed (see retarder represented by the symbol z- ') towards the input of the network 37, and secondly looped back after having been delayed towards the input of the network 17 and to the preprocessing module 16. By implementing for each neuron network such feedback of one or more variables at the output of a neural network, said network uses as input its function of estimate one or more of the predicted values s of previous calculations by this same function of estimation, and that these values were or not measured during the learning phase. Such feedback loopback has the advantage of allowing the taking into account of highly dynamic or non-linear phenomena. The neural networks 17, 37 can thus be described as recurrent or dynamic. The use of a temperature probe at the outlet of each post-treatment system 6 instead of a probe upstream of these systems makes it possible to establish a thermal model of exhaust gas temperature known as the inverse model. This thermal model can be, thus, defined as a semi-physical neuronal inverse model. The estimator 15 finally makes it possible to obtain an estimate of the temperature of the exhaust gases internally of a post-treatment device 6 placed in the exhaust line of the engine and in different locations. The estimator can thus be seen as putting

  implementing a transfer function taking input temperature information downstream of a post processing system 6, as well as possibly one or more other information relating in particular to physical quantities, to output the estimation of desired temperature. The following description concerns the design and learning of the estimator 15. Once the estimator has been developed and implemented for example in the electronic engine management software, it is a question of choosing the networks to of neurons, in particular by determining the number of neurons and calibrating the parameters (weight, bias) of each. The learning of a neural network is performed on a computer from data collected on a vehicle. Each of the neural networks can be learned by a learning algorithm method, in particular by means of a database 19. The database 19 is fed from the results of the training. real track tests of a vehicle with tests at different gear ratios, with different accelerations and decelerations, all chosen to be representative of the normal operating conditions of the vehicle. In addition, it will be possible to split the database fed from the track tests into a learning base and a test base, so as to reduce the size of the database 19 which forms the basis of the data base. learning. The base is a separate base of the vehicle 25 and implemented during initialization operations of the vehicle or during maintenance operations. Figure 3 illustrates the design and learning steps of the neural network estimator. In step 20, tests are performed to generate the data to inform the database 19 (see arrow to the base 19 in FIG. 2), the data being representative of zones of

  interesting operation. It is indeed a question of representative travel of the variation space (or at least a part) of the inputs and outputs of the estimator. It may be useful to carry out data cleaning, by signal processing, for example to remove outliers, to remove redundant points, to re-synchronize or to filter the data. In step 21, the data is split into two parts to form a test base and a learning base. In step 22, the pretreatments to be executed by the pretreatment module 16 lo are determined and the neural networks 17, 37 perform the learnings from the training database. In step 23, the performance is tested with one or more validation criteria, on the test basis and on the learning basis. In step 24, the neural networks 17, 37 are selected and at step 25, the performance and test characterizations are performed on the vehicle. The arrow from the base 19 in FIG. 2 illustrates how the data of the base 19 is used for the learning and validation of the estimator 15. Between the steps 23 and 24, a loopback can be provided. 20 to go back upstream of step 22 to perform a certain number of iterations with possible modifications on the number of neurons, the inputs, the outputs, the division into number of networks, the architecture of the neural networks 17, the type of learning, the optimization criteria, etc. This iterative process (loopback 26) makes it possible to develop the desired estimator (precision, robustness) at a lower cost (number of inputs, number of calculations, number and difficulty of the tests, etc.). Between steps 24 and 25, loopback 27 may be provided when it turns out that the data generated in step 20 is insufficient. Then we go back to step 20 to generate new data called to

  replace the previously acquired data or to complete them in order to establish a sufficient number of measurement points. For example, there may be 5000 measurement points in the learning base.

  During the test step 23, selection criteria can be provided for selecting the set of parameters allowing the most accurate estimation (for example by removing points whose error is greater than 50 C, by searching for an average error close to 5 C, or looking for a low rolling average error).

  Finally, the implementation of the invention for estimating the temperature of the exhaust gas consumes little of the resources of the computer (computing load, memory required), uses a limited number of parameters, and can thus be easily implemented. Moreover, since the database useful for learning is sufficiently representative of all the subsequent conditions of use, the internal gas temperature of a post-processing system 6 can be accurately estimated ( of the order of a few degrees), and this whether steady state or transient.

Claims (10)

  1. A method for estimating an exhaust gas temperature of an engine (1), characterized in that it implements an estimator (15) comprising a series of neural networks (17,37). cascaded.   2. Method according to the preceding claim, characterized in that said neural networks (17,37) each have a feedback loop returning directly or indirectly network input 10 one or more of the quantities available network output .   3. Method according to one of the preceding claims, characterized in that at the input of the estimator (15) is provided information relating to the temperature of the gases downstream of a post-processing system (6) disposed on the exhaust line (5) of the engine, the estimator (15) outputting the estimation of the temperature of the exhaust gases internally of said system (6).   4. Method according to one of the preceding claims, characterized in that one of the neural networks (37) receives as input at least one estimate of the temperature at the output of a monolith, the size available at the output of a neuron network (17), and outputs an estimate of the temperature of the exhaust gases internally of a post-treatment system (6). 25   5. Method according to one of the preceding claims, characterized in that it implements a post-processing of the temperature estimate made by the last network of neurons (37) of the series.   6. Method according to the preceding claim, characterized in that it provides a feedback of the estimated temperature (S), available at the output of the post-processing module (18).   7. Method according to one of the preceding claims, characterized in that it implements a pretreatment of one or more input variables of the estimator by performing calculations based on known physical relationships.   8. Method according to one of the preceding claims, characterized in that it implements a reprocessing of some of the magnitudes at the output of each network before they are returned, according to a so-called indirect loopback, to the entrance of the network or networks.   9. System for estimating an exhaust temperature of an engine (1) comprising an estimator (15) comprising a series of neural networks (17,37) characterized in that said neural networks (17) , 37) are cascaded.   10. Internal combustion engine (1) having an exhaust gas temperature estimating system according to claim 9.25