FR3118102A1 - Method for determining a combustion knock indicator from acceleration measurements - Google Patents

Abstract

The invention relates to a method for determining a knock indicator of the combustion of an internal combustion engine comprising at least one accelerometer sensor. A knock indicator model is built (MOD) beforehand as a function of an acceleration of the engine by training a convolutional neural network on a learning base built by means of the engine further comprising a pressure sensor, and according to the following steps: (i) for a plurality of engine cycles, an acceleration signal and a pressure signal are measured; (ii) a knock indicator is determined for each engine cycle from the pressure signal, and the learning base is formed by associating a knock indicator with the corresponding acceleration signal. Then, an acceleration signal is measured (MES) and the knock indicator is determined (IND) from the acceleration signal and the knock indicator model as a function of an acceleration. Figure 6 to be published

Description

Method for determining a combustion knock indicator from acceleration measurements

The present invention relates to the field of determining a knock indicator for an internal combustion engine, in particular a positive-ignition engine.

This type of engine comprises at least one cylinder comprising a combustion chamber delimited by the inner side wall of the cylinder, by the top of the piston which slides in this cylinder and by the cylinder head. Generally, a fuel mixture is contained in this combustion chamber and undergoes a compression stage then a combustion stage under the effect of controlled ignition, by a spark plug, these stages being grouped under the term "combustion phase". in the following description.

In an internal combustion engine, the combustion phase of the air/gasoline mixture normally begins after the spark. The flame front propagates and its breath pushes part of the mixture against the walls of the cylinder and the top of the piston. The rise in pressure and temperature is sometimes large enough for the unburned mixture wedged against the walls to reach its self-ignition point and self-ignite in one or more places. This phenomenon is called “knock”. Thus knocking is above all a phenomenon of abnormal combustion in controlled ignition engines, perceptible externally by a metallic noise coming from the engine which results from the appearance of pressure waves in the combustion chamber.

These parasitic explosions produce vibrations in the audible range and beyond (of the order of 5 to 50 KHz). They are very sharp and can quickly lead to overheating of the walls. The buildup of knock damages the metal of the piston and/or cylinder walls and rings. After some time (depending on the intensity) this leads to the destruction of the piston, rings or cylinder walls.

Knock estimation allows combustion control to limit the knock effect. In addition, it can be noted that the specific consumption of the internal combustion engine as well as the exhaust temperature drop when the knocking increases. In addition, the safety margin that should be taken to move away from the knocking zone depends on the quality of the quantification of the knocking: the less reliable the indicator, the greater the safety margin must be, and the more fuel consumption is penalized, and the higher the exhaust temperature will be. Therefore, a very reliable quantifier makes it possible to reduce this safety margin and to obtain savings in consumption and reduce the risks of high exhaust temperatures.

State of the art

Generally, the knock detection phase includes a signal acquisition phase, then a phase for processing these signals making it possible to detect the appearance of knocking, to characterize it and/or to quantify it.

A widespread method consists of a measurement of the pressure in the cylinder, followed by a determination of an indicator called MAPO (from the English "Maximum Amplitude of Pressure Oscillations" which can be translated by maximum amplitude of pressure oscillations).

Other indicators have been developed, these indicators are generally based on a measurement of the local pressure in the cylinder or on the measurement by means of an accelerometer.

In general, the extraction of coherent combustion state indicators from an accelerometer is made difficult by the nature of the delivered signal. Indeed, it does not only contain information relating to combustion because the vibration response of the engine is also due to other events such as the slap of the injectors, the reversal of the piston at top dead center, the vibrations induced by distribution system, etc.

Patent EP1116946 discloses a method and a system for controlling combustion based on signals from an accelerometer according to which the measured signal is windowed (period of occurrence of combustion). This signal is then processed during each control loop, and it is compared to a reference signal in order to determine the modifications to be made to the indicators of the state of combustion. Processing the signal from the accelerometer includes three main steps: rectification, filtering and integration.

Patent application US2004267430 (WO05001263) describes a method for processing accelerometric signals resulting from the vibrations of an internal combustion engine. In particular, the signals are filtered by spectral filters and the combustion analysis curve is reconstructed by deconvolution of a transfer function identified from an experimental database. This then results in an estimation of the indicators of the state of the combustion allowing the control of the combustion. For example, the parameter called SoC (Start of Combustion) is estimated from a polynomial function of the energy release, its maximum angle and the boost pressure.

Patent DE19536110 (FR 2739414) describes a method for processing accelerometric signals resulting from vibrations for controlling the combustion of a diesel engine. In particular, the signals are filtered in two different frequency bands. The first frequency band [10 kHz, 30 kHz] makes it possible to extract components associated with the injection using a thresholding device. The second frequency band [0.5 kHz, 4 kHz] makes it possible to extract the components of the signal generated by the combustion by using an identical thresholding method.

US patent 6546328 describes a method which uses a wavelet transform , allowing a priori to locate indicators of the state of combustion like the proposed method.

Patent FR 2834789 describes a knock signal processing method. The proposed method consists in adapting the acquisition sampling frequency to the engine speed. A Fourier transform is calculated for the different sample value groups for a frequency of interest. The information on the combustion results from the summation of the results of the Fourier transforms.

However, the indicators resulting from the previous approaches cannot be used directly for applications related to the diagnosis and control of internal combustion engines. These methods are based on a temporal integration of a signal from an accelerometer. The signal from the accelerometer is not processed in real time. Moreover, these methods strongly depend on the nature of the combustion and/or the technology of the sensor used. Their scope is therefore limited.

Patent EP 1898075 (US 7467040) is also known, which relates to a method for determining indicators of the state of combustion of an internal combustion engine, in which a correlated signal is acquired in the form of a series of samples. to the phenomenon of combustion. This method is based on a time-frequency analysis, carried out sample by sample and requires very few operations because only a limited number of frequencies are of interest, contrary to what is conventionally done with the fast Fourier transform. which calculates all the frequencies of the signal. This method makes it possible to estimate in real time indicators of the progress of the combustion of an internal combustion engine (beginning of the oscillating phenomenon associated with the combustion phase, end of the oscillating phenomenon associated with the combustion phase, energy barycenter of the phenomenon oscillator, magnitude of the oscillating phenomenon) equipped with one or more sensors, whatever the sensor technology used and whatever the nature of the combustion. However, this method seems to require skills in signal processing and interpretation for the analysis of the spectral components. In addition, it is necessary to wait for N samples to be available before being able to make a very first estimate.

As regards the methods which describe the detection of knocking by measuring the pressure in the cylinder, patent applications JP2017207015, JP2017214917, WO17126304 and US2017226956 are known.

In particular, patent application JP2017214917 describes knock detection by measuring the pressure in the cylinder, and by considering different modes of vibration, as well as by calculating an amplitude ratio between the vibration modes. In addition, patent application WO17126304 discloses the decomposition of the signal by time windowing, and knock detection via a comparison of the peaks in each window.

In addition, patent application FR 3086391 describes a method for determining a knock indicator from a pressure measurement, the modal decomposition of which makes it possible to have characteristics of the overall pressure (representation of the pressure in all the combustion chamber) in the cylinder. This pressure makes it possible to determine, directly or indirectly, a representative knock indicator. However, such techniques have the drawback of requiring a pressure sensor in each cylinder of an engine of a production vehicle.

We also know the following documents:

- US6553949B (Ford) which relates to a method of controlling / preventing knocking in order to avoid knocking by means of a neural network. More specifically, this document describes a method of coordinated control of the ignition advance, the boost pressure and the volumetric compression ratio to avoid knocking. One aspect of the invention uses a function (known as discrimination) which, based on the read or estimated position of the actuators and read or estimated motor parameters, determines whether the motor is controlled in a knocking zone or in an operational zone safe. Another aspect of the invention uses the value of a knock sensor, whatever it is (cylinder pressure or knock indicator). On the other hand, this document does not describe a method for quantifying knocking.

- US2008051981A (Michigan Technological University) which concerns a knock detection method using metrics. More specifically, this document describes a method for analyzing an accelerometric signal based on various statistical techniques which, from a collection of combustion events recorded in memory (buffer), is capable of determining whether a new event is at the -above or below a critical threshold for entry into the knocking zone, this threshold being the result of a calculation of the average type, standard deviation. However, this method has the disadvantage of requiring a large number of previous values of the indicator to be stored in memory to obtain reliable statistics; in particular during the transient operation of the motor, the calculated statistic is based on a base of previous samples which therefore lags behind the current operation of the motor. Thus, this method does not make it possible to deliver a reliable knock intensity indicator from the first combustion cycle when the engine is started, without an estimation delay when the speed or the load changes.

- US5483936A (Visteon) which concerns a knock detection method using a neural network. More specifically, this document describes a method for analyzing an accelerometric signal based on a technique for extracting characteristics by conversion into the frequency domain (fast Fourier transform) of different portions of the signal (probably obtained by time slicing at the along the gas expansion phase), then a classification by neural network of the frequency spectrum of each portion which results in indicating the presence of knock or not (by comparison with known levels in a database) in each portion , and finally a weighting of the results obtained for each portion is performed to deliver a continuous intensity value. Thus, this method has many steps and many calibrations.

- US2011077846 (GM Global Technology Operations) which relates to a knock detection method using metrics. More specifically, this document describes a method for analyzing the cylinder pressure signal and reconstructing a heat release signal associated with the combustion event, from which filtering, windowing and averaging are carried out in order to estimating the amount of energy released by the auto-ignition sources (causing knocking). Then, depending on whether this quantity exceeds a threshold, the engine control is corrected to reduce this quantity of energy (and therefore move away from the conditions in which knocking occurs). Thus, this method requires a pressure sensor in each cylinder. Moreover, it is complex to implement by the necessary estimation of the energy release which requires a thermodynamic analysis, and uses multiple filtering/windowing/averaging techniques.

The documents listed below list and compare different knock indicators, different types of sensors, and their performance:

Dues, S.; Adams, J.; Shinkle, G.; Div, D. Combustion Knock Sensing: Sensor Selection and Application Issues SAE Trans. J. Engines, 1990.

Gao, X.; Stone, R.; Hudson, C.; Bradbury, I. The Detection and Quantification of Knock in Spark Ignition Engines SAE Technical Paper, 1993.

Naber, J.; Brough, J.; Frankowski, D.; et al. Analysis of Combustion Knock Metrics in Spark-Ignition Engines SAE Trans. J. Engines, 2006.

Shahlari, A.; Ghandhi, J. A Comparison of Engine Knock Metrics SAE Technical Paper, 2012.

However, the following limitations are encountered by this type of knock indicators:

• Some indicators, in particular the MAPO indicator, are dependent on the speed and/or the load of the internal combustion engine, which requires precise calibration of the indicator according to the speed and/or the load. Without this precise calibration, there is a bias in the indicator, the value of which changes according to the engine speed and/or the load in the absence of knocking;
• the choice of signal processing technique influences knock detection. Indeed, two filters with different frequency ranges generate different evaluations of knock.

The present invention aims to overcome these drawbacks. More specifically, the method according to the invention makes it possible, once a knock model has been determined, to detect and quantify the knock of an engine precisely by means of an indicator constructed from a measurement taken by an accelerometer. This design is particularly advantageous because an accelerometer sensor is conventionally present on series engines. In addition, such an indicator is independent of the speed and load of the internal combustion engine, which allows an unbiased estimation of knock intensity. The use of such an indicator makes it possible to reduce the risk of damage to the cylinder, but also to reduce the consumption of the internal combustion engine and reduce the risk of high exhaust temperatures.

The invention relates to a method for determining a combustion knock indicator in at least one cylinder of an internal combustion engine, in particular with spark ignition, said engine being provided with at least one accelerometer sensor. According to the invention, the method comprises at least the following steps:

a) a model of an indicator of said knocking is constructed as a function of an acceleration of said engine by training a convolutional neural network on a learning base, said learning base being constructed by means of said internal combustion engine comprising in addition to a pressure sensor in at least one cylinder, or of an internal combustion engine representative of said engine further comprising a pressure sensor in at least one cylinder, and according to at least the following steps:

- for at least one operating point of the internal combustion engine, a plurality of ignition advance values of said engine or of said engine representative of said engine are defined, and an acceleration signal is simultaneously measured by means of said accelerometric sensor and a pressure signal by means of said pressure sensor for at least one cycle of each ignition advance value of each operating point;

- for each cycle of each ignition advance value of each operating point, a value of said knock indicator is determined from said pressure signal, and said learning base is constituted by associating said value of said knock indicator knocking said measured acceleration signal simultaneously with said pressure signal for each cycle of each spark advance value of each operating point;

b) an acceleration signal of said motor is measured by means of said accelerometer sensor;

c) said knock indicator is determined from said acceleration signal of said engine and by means of said model of a knock indicator as a function of an acceleration of said engine.

According to one implementation of the invention, in step a), a knock indicator can be determined from said pressure signal in the following way:

i) determining at least one local extremum of said pressure signal in an interval calibrated around the maximum peak of said measured pressure, said maximum peak being determined from among said at least local extremum; And

ii) Said knock indicator is determined as at least one descriptor of said at least one local extremum.

According to an implementation of the invention, it is possible to determine at least one descriptor chosen from: a first descriptor representative of a predetermined number of pressure variations of said pressure signal of greater amplitudes, and/or a second descriptor representative of all of the pressure variations of said pressure signal, and/or a third descriptor representative of all of the pressure variations of said pressure signal, with the exception of a predetermined number of pressure variations of greater amplitude.

According to one implementation of the invention, said at least one local extremum can be determined by means of a topological data analysis method, and a line segment connecting two consecutive local extrema can be determined for each local extremum.

According to an implementation of the invention, it is possible to determine said at least one descriptor representative of a predetermined number of said segments of longest finite length determined by said method of topological analysis of the data, in particular by means of the ratio of the sum of the lengths of said longest segments, over the number of samples separating the longest segment from the last longest segment among the predetermined number of segments.

According to an implementation of the invention, said convolutional neural network can comprise at least 3 convolutional layers and a piecewise linear type activation function.

According to one implementation of the invention, said knock indicator can be determined in real time by means of an on-board computer.

According to an implementation of the invention, said internal combustion engine can be controlled in real time according to said knock indicator, preferably the ignition advance can be controlled.

According to one implementation of the invention, said knock indicator can be determined by means of a combustion analysis platform.

According to an implementation of the invention, it is possible to calibrate the control of the internal combustion engine as a function of said determined knock indicator.

According to one implementation of the invention, the method may include an additional step of determining abnormal combustion by analyzing said knock indicator.

Other characteristics and advantages of the method according to the invention will appear on reading the following description of non-limiting examples of embodiments, with reference to the appended figures and described below.

List of Figures

There illustrates the steps of the method according to a first embodiment of the invention.

There illustrates the steps of the method according to a second embodiment of the invention.

There illustrates the steps of the method according to a third embodiment of the invention.

There illustrates an internal combustion engine suitable for implementing step 1/ of the method according to the invention

There illustrates an internal combustion engine suitable for implementing step 2/ of the method according to the invention

There illustrates an acceleration signal and a cylinder pressure signal as a function of the crankshaft angle for an engine ignition advance value.

There illustrates the intake pressure P as a function of the engine speed NE for a plurality of ignition advance values defined for an example application of the method according to the invention.

There illustrates the values of the knock indicator TDAcrete determined for a plurality of ignition advance values defined for an example application of the method according to the invention.

The present invention relates to a method for determining a combustion knock indicator in a cylinder of an internal combustion engine. The present invention is particularly suitable for a spark-ignition engine (gasoline engine). According to one implementation of the invention, a knock indicator may correspond to the intensity of the knock.

The internal combustion engine can be of any type: it can in particular be supercharged or not, with direct or indirect fuel injection, and equipped with an exhaust gas recirculation system or not, etc.

According to a first embodiment of the invention, as illustrated in , the method for determining a knock indicator may comprise the following steps:

1/ construction of a knock model (MOD)

2/ acceleration measurement (MES)

3/ determining the knock indicator (IND)

According to a second embodiment of the invention, suitable for real-time control of the internal combustion engine, and as illustrated in , the method for determining a knock indicator may comprise the following steps:

1/ construction of a knock model (MOD)

2/ acceleration measurement (MES)

3/ determining the knock indicator (IND)

4/ internal combustion engine control (CON)

According to a third embodiment, suitable for the calibration of an internal combustion engine from measurements on an engine bench (in the laboratory), and as illustrated in , the method for determining a knock indicator may comprise the following steps:

1/ construction of a knock model (MOD)

2/ acceleration measurement (MES)

3/ determining the knock indicator (IND)

5/ calibration of the internal combustion engine (CAL)

The steps will be described later in the description. Steps 1 to 3 are common for the three embodiments of the invention. In a preferred embodiment, step 1 can be carried out once, beforehand, and steps 2 and 3 can then be repeated using the knock pattern determined during the single implementation of the 'Step 1.

Preferably, at least steps 2 to 3 of the method, and preferably also step 4, are carried out in real time (that is to say from cycle to cycle), so as to determine the indicator of knocking in real time, in order to be able to act on the knocking before the next ignition.

For the second embodiment, steps 2 to 4 can be implemented by the engine computer; thus the knock indicator is determined online in real time.

For the third embodiment, steps 2 and 3 can be implemented using a combustion analysis platform, which includes computer means for collecting and processing acceleration measurement signals.

Step 1/ of the method according to the invention requires a combustion engine comprising at least one accelerometric sensor, for example placed on the cylinder block and/or on the engine block, and at least one pressure sensor placed in at least one engine cylinder. An accelerometric sensor (that is to say an accelerometer) makes it possible to measure the linear acceleration of the engine block and/or of the cylinder block. Advantageously, the engine whose knocking is to be detected is equipped with several accelerometric sensors, arranged in a spaced apart manner, for example on the engine block and remote from the auxiliary sources of vibration (pump, pulley, etc.). Preferably, the accelerometric sensor can be able to measure a motor vibration on 1 axis up to an acceleration of 5000 times the acceleration of gravity, with a refresh rate of 30kHz, and can operate in a temperature range of - 55°C to +260°C, such as the ENDEVCO ® 7240C accelerometric sensor from PCB Piezotronics (France). The pressure sensor measures the local pressure in the cylinder, i.e. the gas pressure in the cylinder. The gases form an oxidizer or a fuel mixture (in the case of indirect injection), and may include in particular air at ambient pressure, supercharged air, an air mixture (supercharged or not) and burnt gases . Advantageously, the engine includes a pressure sensor arranged in each cylinder. Preferably, step 1/ can be implemented on the engine test bench.

In the preferred mode of the invention according to which step 1/ is carried out only once, upstream, step 2/ can be carried out, online in real time, by means of a motor comprising only one or several accelerometer sensors, but not necessarily a pressure sensor. This is advantageous, because if an accelerometric sensor generally equips the series engines, the series engines do not always include a pressure sensor in at least one cylinder. Consequently, once a knock pattern has been determined in step 1/ by means of an engine equipped with at least one accelerometer sensor and a cylinder pressure sensor, step 2/ can be applied to production engines , without requiring adaptation of these motors. However, according to this implementation, it is quite clear that the engine used for the implementation of step 1/ is preferably representative of the engine for which knocking is to be detected. Preferably, the engine representative of the engine for which knocking is to be detected is an engine having the same technical characteristics (or same technical definition; that is to say the same engine specification in terms of architecture, parts, materials, assembly, performance, etc.). Alternatively, the engine representative of the engine for which knocking is to be detected may have equivalent technical characteristics, for example in terms of architecture, parts, materials, assembly, performance, etc. Preferably, steps 2 / and the following can be implemented on a series internal combustion engine.

There illustrates, schematically and in a non-limiting manner, an internal combustion engine 10 suitable for implementing step 1/ of the method according to the invention. The internal combustion engine 10 illustrated is of the spark-ignition supercharged type, in particular of the gasoline type, comprises at least one cylinder 12 (four cylinders 12 for the embodiment of the ) with a combustion chamber 14 inside which the combustion of a mixture of supercharged air and fuel takes place (optional characteristics for the implementation of the method according to the invention).

The cylinder 12 comprises at least one pressurized fuel supply means 16, for example in the form of a fuel injector 18 controlled by a valve 20, which opens into the combustion chamber, at least one intake means 22 with an inlet valve 24 associated with an inlet pipe 26 terminating in a plenum 26b (not shown in the figure), at least one burnt gas exhaust means 28 with an exhaust valve 30 and an exhaust manifold 32 and at least one ignition means 34, such as a spark plug, which makes it possible to generate one or more sparks making it possible to ignite the fuel mixture present in the combustion chamber.

The pipes 32 of the exhaust means 28 of this engine are connected to an exhaust manifold 36 itself connected to an exhaust line 38. A supercharging device 40, for example a turbocharger or a volumetric compressor, is placed on this exhaust line and comprises a drive stage 42 with a turbine swept by the exhaust gases circulating in the exhaust line and a compression stage 44 which makes it possible to admit pressurized intake air into the combustion chambers 14 through the intake manifolds 26.

The engine comprises means 46a for measuring the local pressure in the cylinder, arranged within the cylinder 12 of the engine itself. These measurement means include a pressure sensor which makes it possible to generate a signal representative of the evolution of the local pressure in a cylinder.

The engine can also include means 46b for measuring the intake pressure (optional feature for the method according to the invention), arranged in the plenum 26b. These measurement means generally consist of an absolute pressure sensor, of the piezoelectric type, which makes it possible to generate a signal representative of the evolution of the intake pressure in the intake plenum.

The engine further comprises an accelerometer 47 disposed on the engine block. Such a sensor makes it possible to measure signals representative of vibrations in the combustion chambers. In particular, the appearance of knocking in the combustion chamber will generate characteristic vibrations, which can be measured by an accelerometer.

The engine also comprises a calculation and control unit 48, called the engine computer, which is connected by conductors (for some bidirectional) to the various organs and sensors of the engine so as to be able to receive the various signals emitted by these sensors, and then control the components of this engine to ensure its correct operation.

Thus, in the case of the example shown in , the spark plugs 34 are connected by conductors 50 to the engine computer 48 so as to control the moment of ignition of the fuel mixture, the accelerometer 47 is connected by a line 51 to this same engine computer 48 to send it the signals representing vibrations in the combustion chambers, the cylinder pressure sensor 46a is also connected by a line 52 to the engine computer 48 to send it the signals representative of the evolution of the pressure in the cylinder and the valves 20 for controlling the injectors 18, are connected by conductors 54 to the computer 48 to control the injection of fuel into the combustion chambers. The means 46b are also connected by a line 53 to the engine computer 48.

There illustrates, schematically and in a non-limiting manner, an internal combustion engine 10 suitable for implementing step 2/ of the method according to the invention. The motor described in this figure is identical in all respects to the motor of the , with the exception that it does not include means for measuring the local pressure in the cylinder (reference 46a in ) and intake pressure measurement means (reference 46b in ), as well as their associated connectors in the direction of the computer (references 52 and 53 in ).

1/ Construction of a knock model

During this step, it is a question of constructing a representative model of a knock indicator as a function of the measurement by one or more accelerometers of the vibrations of the engine block for which knocking is to be detected. In other words, during this step, an analytical model is constructed linking an acceleration value to a value of a knock indicator.

According to the invention, this step is implemented by means of a supervised learning algorithm, applied to a learning base constructed from a plurality of cylinder pressure measurements and acceleration measurements carried out simultaneously to different operating points (ie for different operating conditions (rpm and load)) and ignition advance values (AVA) which controls the engine spark plugs.

According to the invention, the construction of a model of a knock indicator as a function of the acceleration is carried out at least according to the following sub-steps: 1.1/ construction of a learning base; 1.2/ training of a supervised learning algorithm. These sub-steps are described below.

1.1/ Construction of a learning base

According to the invention, the construction of the learning base is carried out by means of an internal combustion engine comprising at least one accelerometer, and, in at least one of the cylinders of this internal combustion engine, a pressure sensor.

As discussed above, according to a preferred implementation of the invention, the construction of the learning base can be carried out by means of a motor distinct from the motor for which it is desired to detect knocking in fine in real time. Indeed, this step can be performed only once, upstream, by means of an engine adapted to the implementation of this step, whereas the application of the following steps can be applied to a series engine having preferably the same technical characteristics (or technical definition) as the engine used for stage 1/.

According to the invention, for at least one operating point of the internal combustion engine, a plurality of ignition advance values are defined, obviously including at least ignition advance values causing knocking. Advantageously, at least 10 operating points can be defined, so as to constitute a learning base representative of varied operating conditions. Advantageously, at least 10 ignition advance values can be defined per operating point, for example spaced apart by ignition advance jumps of between 0.1 and 2° CA (crankshaft angle), preferably 1° CA. For at least one cycle, preferably for a plurality of cycles (for example for at least a hundred cycles) of each ignition advance value of each operating point, a pressure signal is simultaneously measured in at least an engine cylinder by means of the pressure sensor and an acceleration signal by means of the engine's accelerometer sensor. The pressure sensor generates a sensor signal representative of the local pressure at a point in the cylinder (hereinafter referred to as the pressure measurement signal). The accelerometer sensor generates a sensor signal representative of the linear acceleration of the engine block (hereinafter called the acceleration measurement signal). These signals coming from the pressure sensor or from the accelerometric sensor are measured as a function of time or as a function of the angular position of the crankshaft of the internal combustion engine. By way of illustration, the presents at the top a signal for measuring the acceleration A and at the bottom a signal for measuring the pressure P as a function of the angle of the crankshaft CA, for the same cycle. These signals show a knocking phenomenon, for a crankshaft angle of 17.19 °CA.

According to an implementation of the invention, an acceleration signal and a pressure signal can further be measured for a plurality of engine temperatures, in addition to a plurality of operating points of the internal combustion engine, and in addition to a plurality of spark advance values. This makes it possible to increase the variability of the operating conditions to constitute the learning base.

According to the invention, for each pressure measurement signal, a value of the knock indicator is determined. As the pressure and acceleration measurements are performed simultaneously, for at least a plurality of ignition advance values, each pressure measurement signal corresponds to an acceleration measurement signal. Thus, by correspondence, once a value of the knock indicator has been determined for a pressure measurement signal, a value of the knock indicator is associated with the acceleration measurement signal corresponding to the pressure measurement signal measured simultaneously. Thus, the learning base consists of a plurality of acceleration signal-pinking indicator value pairs, more precisely an acceleration signal-pinking indicator value pair for each cycle of each value in advance at the ignition of each operating point.

Advantageously, the construction of a learning base is carried out on the engine bench, or in other words in the laboratory. According to one implementation of the invention, the construction of the learning base can be carried out in the following way: the engine is installed on the engine bench, on a fixed operating point (engine speed and load) and an operator performs a variation of ignition advance, by command of the spark plug. The ignition advance variation can be started on a low ignition advance value, chosen so that no knocking occurs. Advantageously, the operator listens to the noise produced by the combustion, via headphones, connected in the test cell to a microphone, in order to detect the occurrence of knocking. For each ignition advance value during the variation, recording of the cylinder pressure measurement is triggered. This recording covers several hundred combustion cycles (300 to 500) for all cylinders fitted with a cylinder pressure sensor. By way of example, for an operating point of 1750 rpm, a recording of 300 cycles lasts approximately 20 seconds. This number of cycles is considered sufficient because the information representative of the combustion can then be obtained in a reproducible way: on the one hand this allows the human ear to identify the presence and the repetition of abnormal combustions, and on the other hand the quantity of recorded data is sufficient to carry out reliable statistical calculations (maximum, minimum values, mean, standard deviation, median, last decile, etc.). Then the operator performs low amplitude advance jumps (for example 1°CA) by increasing its value; ignition then occurs earlier and earlier. In addition to recording the measurement of the cylinder pressure triggered for each value of ignition advance during the variation, the operator can continue to listen to the noise produced by the combustion. During the variation, the operator will notice the change in the engine noise produced by the combustion and can annotate the moment when the knocking is born (not very intense or fleeting phenomenon) and/or the knocking occurs (intense and repeated phenomenon) . This annotation is a qualitative evaluation, not essential to the invention, which does not make it possible to precisely quantify the knock: we do not know its intensity on each cycle taken individually. According to the invention, a knock indicator is determined automatically, from the pressure measurements taken for the plurality of AVA values defined for at least one operating point.

According to one implementation of the invention, from the measurement of the pressure, it is possible to determine a value of the knock indicator by means of a method conventionally used for analyzing the maximum amplitudes of the pressure oscillations. Such a method leads to an indicator called MAPO (from the English “Maximum Amplitude of Pressure Oscillations” which can be translated as maximum amplitude of pressure oscillations). A description of such a method can be found in the document (Naber et al., 2006).

According to a main variant of the invention, for each pressure measurement signal, a value of the knock indicator can be determined by means of a method comprising the following steps:

- at least one local extremum of the pressure measurement signal is determined in an interval calibrated around the maximum peak of the measured pressure, the maximum peak being determined from among the at least local extremum;

- the knock indicator is determined as at least one descriptor of the at least one local extremum.

This method has the advantage of providing a knock indicator value, from a pressure measurement signal, independent of the speed and the load of the internal combustion engine. Different embodiments of this main variant of the invention below are described below.

Advantageously, according to this main variant of the invention, it is possible to determine at least one descriptor, preferably three descriptors, chosen from: a first descriptor representative of a predetermined number of pressure variations of the pressure measurement signal of greater amplitudes , and/or a second descriptor representative of all the pressure variations of the pressure measurement signal, and/or a third descriptor representative of all the pressure variations of the pressure measurement signal, with the exception of a predetermined number of pressure variations of greater amplitude. This first descriptor is a local descriptor, whereas the second and the third descriptors are integral (or global) descriptors.

Advantageously, according to this main variant of the invention, it is possible to determine the at least one local extremum by means of a method of topological analysis of TDA data (from the English “Topological Data Analysis”). The TDA method of analysis is defined as the use of a collection of powerful tools capable of quantifying the shape and structure present in the data, and allowing an investigation of these data. The idea is that it is possible to choose simplified representations of these data, which are present in large quantities and complex, under a particular aspect, which form a topological signature, from which a rigorous quantification is possible in order to allow an investigation for the purpose of understanding. This TDA analysis method can make it possible to establish a representation in the form of several segments of different lengths.

Preferably, the TDA analysis method can be applied opposite to the signal from the pressure sensor (in other words to the signal from the pressure sensor for which the sign has been changed), so that the scan implemented in this method encounters the maximum pressure point first.

Advantageously, according to this main variant of the invention, by means of the TDA analysis method, a line segment connecting two consecutive local extrema is determined for each local extremum.

Advantageously, according to this main variant of the invention, it is possible to determine the at least one descriptor representative of a predetermined number of longest segments of finite length by the method of topological analysis of the TDA data, in particular by means of the ratio of the sum of the lengths of the longest segments, over the number of samples separating the longest segment from the last longest segment among the predetermined number of segments. Thereafter, we will note "TDAcrete" the indicator resulting from this descriptor obtained by a TDA method. For example, and without limitation, the predetermined number of segments can be between 1 and 5, and can be equal to 3.

At the end of this sub-step, a learning base is obtained comprising a plurality of pairs formed of acceleration signals and corresponding knocking indicator values.

1.2/ Training a convolutional neural network

During this sub-step, a supervised learning algorithm is implemented from the learning base to build a model of the knock indicator as a function of the acceleration. According to the invention, a convolutional neural network is used as a supervised learning algorithm. This supervised learning algorithm has the advantage of presenting several convolutional layers, which make it possible to identify, more and more precisely from one layer to another, particular shapes (or even particular patterns) in a signal. . In other words, the objective here is, by supervised learning, to determine a model making it possible to recognize, in an acceleration measurement signal, a pattern representative of knocking.

Advantageously, the learning base can be pre-processed before being used to train the supervised learning algorithm. Indeed, due to its location on the engine block, the acceleration measurement signal is sensitive to all combustions, whatever the cylinder. Consequently, the exploitation of the acceleration measurement signal may require correctly assigning a fraction of the signal to the combustion in the same cylinder. The allocation sequence must then take into account the firing order (triggering of the spark at the spark plug) of the cylinders. Thus, the pre-processing consists in defining angular windows making it possible to isolate the fraction of the accelerometric signal corresponding to the combustion of each of the cylinders of the engine. According to an implementation according to which the engine is a 4-stroke engine (reminder: the realization of 4-stroke requires two revolutions of the crankshaft, i.e. an angular range of 720° CA) with 3 cylinders with ignition according to the sequence cylinder 1 - cylinder 3 - cylinder 2 (cylinder 2 is at the center of the 3-cylinder in-line engine but it is the one that is ignited last), we can define three angular windows with a width of 200° V such that:

• cylinder 1: [-60°CA, 140°CA]
• cylinder 3: [180°CA, 380°CA]
• cylinder 2: [420°CA, 620°CA]

According to this implementation, the learning base can consist, over the angular range [-60, 620° CA], of 3 accelerometric signals (one per cylinder), each being defined in a window of amplitude 200° CA (i.e. 2000 values if we consider an angular resolution of 0.1° CA), and for each signal, a value of the corresponding knock indicator (obtained from pressure measurements, according to any of the variants of the sub-step 1.1 described above). The use of an angular window with a fixed width (that is to say not set on the maximum cylinder pressure point) and wide (for example greater than 100°CA), unlike the prior art in particular described in (Naber et al., 2006), is justified, on the one hand, by the fact that on an engine not equipped with a cylinder pressure sensor, it is not possible to identify the position of a maximum cylinder pressure from which one could define an angular window, and on the other hand, by the type of neural network used for the implementation of the method according to the invention, the action of the convolutional layers of which is to exhaustively explore the signal according to windows of different sizes and for different angular resolutions by filtering (convolution layer) and by sub-sampling (layer of "pooling") successively. Moreover, the property of being of fixed width is justified by the imposed format of the data that is placed at the input of the neural network which must have a fixed size. Alternatively, it is possible to use a reduced angular window size (for example less than 100° CA).

Advantageously, the learning base can be divided into 2, preferably 3, subsets:

• a first subset, comprising for example between 70% and 85% of all the data of the learning base, intended for learning the network: these data will directly determine the estimation error measurement of the network and how to adjust the values of the neural network parameters; this is the training database.
• a second subset, comprising for example between 15% and 30% of all the data of the learning base, intended to validate the learning process during its execution: these data are used to monitor the success of the learning operation during its course but do not directly influence the adjustment of the network parameters; this is the validation database.
• optionally but advantageously, a third subset, comprising a number of data from the learning base in an equivalent proportion to the validation database, intended to test the learning performance, at the end of the 'learning ; this is the test database.

According to one implementation of the invention, to form the subsets described above, a random draw can be made from all the data in the learning base.

According to the invention, this sub-step consists in searching for the values of the parameters of the convolutional neural network making it possible to minimize a difference between the learning data (values of the knock indicator) and an estimate (or even a prediction) of this data by this algorithm. This is a regression problem, which can be solved iteratively, for example by a gradient method. More specifically, according to this design, the values of the parameters of the neural network minimizing the deviation described above are iteratively sought, until a stopping criterion is reached. The stopping criterion can relate to a maximum value of the deviation (for example in the sense of least squares), a maximum number of iterations, etc. Parameters of a convolutional neural network include synaptic weights and activation function biases.

According to an implementation of the invention, one can search for the parameters of the neural network by defining:

• a measurement of the convolutional neural network estimation error (equal to the difference between a datum and an estimation by the convolutional neural network), for example by defining a loss function. According to an implementation of the invention, it is possible to define a loss function measuring an error in the sense of least squares (“mean squared error” in English); And
• the way in which the parameters of the neural network are adjusted (for example by back-propagation of the gradient) according to the measurement of the estimation error chosen, for example by choosing an optimizer and its learning rate ("learning spleen" in English). For example, one can use a Nadam type optimizer (for “Nesterov-accelerated Adaptive Moment Estimation”), for example implemented with a learning rate (“learning rate”) of 0.002 and an automatic reduction of the rate learning (for example on reaching a plateau, with the following parameters: a mean absolute error type monitor, with a factor of 0.1, and a deliberate stopping criterion ("early stopping patience" in English) of 3.
• a measurement of the quality of learning (or even a metric; "metrics" in English), carried out on the validation database when it was created. We can choose a mean absolute error type metric with an early stopping patience criterion of 30.

Advantageously, the convolutional neural network implemented in the invention may comprise:

- activation functions of the ReLU type (for "Rectified Linear Unit"), which are piecewise linear functions; and or

- at least 3 to 5 convolutional layers, so as to obtain an accurate knock model, in this case, we speak of deep architecture; and or

- an intermediate layer, called "pooling" ("grouping" in French), placed between at least two convolutional layers. Such a layer makes it possible to reduce the size of the patterns, while preserving their main characteristics.

According to a particular design of the invention, the convolutional neural network can be composed of 7 layers in total:

• 3 convolutional layers, the number of filters of which is 32 or 64, the number of "strides" (overlapping) is 3 or 5, and the activation function is of the ReLU type. These layers have the property of identifying patterns (sequence of signals of the same form) locally in the data tensor received as input; if these patterns are found at different scales and at different positions, then the convolutional layer is able to isolate them if this makes it possible to reduce the prediction error; the number of filters determines the variety of patterns the network can identify; the first layers of the network identify general patterns while the last layers identify characteristic patterns;
• each convolutional layer is immediately followed by a so-called "pooling" layer. Such a layer has the property of downsampling the data in order to allow the next convolutional layer to process the data on a larger scale. Note that the last layer of this type is a so-called "GLOBALMAXPOOLING" layer, which allows the information to be reduced by one dimension.
• the final layer is a dense (fully connected) layer that retrieves all network features and produces the quantizer estimate for the input data; the output of this layer is a real (scalar) number.

Advantageously, an eighth layer can be added between the last “pooling” layer, that is to say the “GLOBALMAXPOOLING” layer and the dense layer, in order to limit the overfitting of the network. , this is a so-called "DROPOUT" layer for which a rate can be defined (for example a rate of 0.25) which represents the proportion of units among the inputs which are randomly set to zero.

According to this design, the total number of neural network parameters to be calibrated can be on the order of 20000.

According to an implementation of the invention, a single convolutional neural network can be used for all the cylinders, which makes it possible to limit the number of networks to train and store. According to this design, it is advantageous to normalize the values measured by the accelerometric sensor, so as to homogenize the input data of the convolutional neural network and avoid excessive value deviations which could compromise the learning phase.

Alternatively, one can use a convolutional neural network for each cylinder, which can be of interest when the cylinders present heterogeneities in the quality of combustion. According to this design, it is advantageous to normalize the values measured by the accelerometric sensor, so as to homogenize the input data of the convolutional neural network and avoid excessive value deviations which could compromise the learning phase.

Advantageously, a convolutional neural network of learning data comprising accelerometric signals measured by at least two accelerometric sensors can be applied. According to this design, it is advantageous to normalize the values measured by the accelerometric sensors, so as to homogenize the input data of the convolutional neural network and avoid excessive value deviations which could compromise the learning phase.

After training a convolutional neural network as described above on the learning base as described above, values of the parameters of the convolutional neural network calibrated (or even calibrated, or even adjusted) are obtained on learning database data. These parameters are then fixed for the remainder of the process.

Thus, at the end of this sub-step, a representative model of a knock indicator is obtained as a function of the acceleration. The construction of this model can be carried out only once, in a preliminary way, and the stages 2 and 3 described below can then be repeated by means of the model of knock according to the acceleration determined during the unique setting. implementation of step 1.

2/ Acceleration measurement (MES)

This step is applied by means of the internal combustion engine for which knocking is to be detected. This internal combustion engine comprises at least one accelerometric sensor, for example arranged on the engine block. This engine does not need to include a pressure sensor in a cylinder. Preferably, this engine has the same technical characteristics (or technical definition) as the engine from which step 1/ above was carried out, so that the model built in step 1/ is suitable for the quantification of a knock for the engine considered.

During this step, a measurement of the acceleration of the engine is carried out by means of an accelerometer sensor. The accelerometer sensor generates a sensor signal representative of the linear acceleration of the engine block (also called the acceleration measurement signal). This signal is a function of time or of the angular position of the crankshaft of the combustion engine.

3/ Determination of the knock indicator (IND)

During this step, a value of the knock indicator is determined by means of the knock model as a function of the acceleration determined in step 1/ described above and of the acceleration measurement signal measured at step 2/ described above.

In other words, during this step, the convolutional neural network trained as described above is used to process acceleration measurement signals obtained in step 2/, for example recorded in real time when the combustion engine contributes to the movement of a vehicle.

In other words, an accelerometric signal is placed at the input of the trained convolutional neural network, and the latter calculates an estimate of the knock indicator which the network has learned.

Advantageously, a preprocessing similar to that applied to the accelerometric measurements used to build the learning base is applied beforehand to the accelerometric measurements used during this step, in particular an angular windowing and/or a normalization as described in sub-steps 1.1 and 1.2 above.

At the end of this step, a knocking indicator value is obtained for an acceleration measurement signal. By non-limiting example, a knock indicator close to 0 can correspond to an absence of knock, a knock indicator between 10 and 30 can correspond to incipient and intensifying knock, and a knock indicator greater than 30 can correspond to an established knock, these thresholds not depending on the operating point of the engine (speed and load).

According to one aspect of the invention, the method may include an additional step of determining abnormal combustion by analyzing the knock indicator. This determination of abnormal combustion can be implemented by comparing the knocking indicator with a predefined knocking threshold.

4/ Control of the internal combustion engine

During this optional step, the control of the internal combustion engine can be determined in real time according to the knock indicator determined in step 3, with the aim of reducing knocking while guaranteeing good performance of the engine at internal combustion. Indeed, the indicator according to the invention makes it possible to reduce the safety margin and to obtain gains in consumption of the internal combustion engine and to reduce the risks of high exhaust temperatures. The combustion control can be implemented by the computer of the internal combustion engine.

For this optional step, the knock indicator can be determined in real time by means of a control unit of the internal combustion engine. Preferably, when the internal combustion engine is used in a vehicle, the computer can be embedded.

In the event of knocking detected by means of the method according to the invention, the computer can initiate the actions necessary to control the following combustions in order to avoid the continuation of the occurrence of the knocking during the following combustions.

According to one embodiment, the control can be based on a comparison of the knock indicator with a knock threshold. The knocking threshold is advantageously a fixed value independent of the speed and the load of the internal combustion engine. Indeed, the knock indicator according to the invention has the advantage of being independent of the engine speed, so a single fixed knock threshold can be used.

In accordance with one implementation of the invention, two thresholds and consequently three zones can be defined: a first zone without knocking for which the knocking indicator is lower than a first threshold, a second zone of incipient knocking for which the the knock indicator is between the first threshold is a second threshold, and a third established knock zone for which the knock indicator is greater than the second threshold. Then the internal combustion engine can be checked, depending on the area in which the knock indicator is located.

For this step, the combustion can be controlled within the cylinder of the internal combustion engine, by controlling the actuators of the internal combustion engine, in particular the injector, the intake valves, the exhaust valves, etc. Preferably, for a spark ignition engine, the ignition advance can be controlled. Thus, the internal combustion engine can be protected from the knocking phenomenon and the noise of the internal combustion engine can be reduced.

For the implementation of the invention, in which three zones are defined, for example (in a non-limiting manner): the ignition advance can be advanced in the first zone, the advance can not be varied ignition in the second zone, and the ignition advance can be delayed in the third zone.

According to a second implementation, an ignition advance value can be defined beforehand for each operating point (speed, torque) of the internal combustion engine by means of a map. When knocking is determined by the previous steps, the spark advance value can be decreased. Then, when the knocking disappears, the value of the ignition advance can be gradually increased towards the value defined in the map, without exceeding this value. For this type of control, the invention makes it possible to define an ignition advance value closer to the knocking zone, with less safety margin, since the invention makes it possible to better define the knocking zone thanks to to the shape of the indicator.

Other controls may be considered.

5/ Calibration of the internal combustion engine

During this optional step, the knock indicator can be determined on an engine bench in the laboratory, using a combustion analysis platform, in order to know the operation of the internal combustion engine. The combustion analysis platform includes IT resources for collecting and processing the pressure measurement signal.

Then, the internal combustion engine can be calibrated according to the knock indicator determined in step 3, so as to adjust the settings of the internal combustion engine allowing the knock to be reduced while guaranteeing good performance of the engine at internal combustion. Indeed, the indicator according to the invention makes it possible to reduce the safety margin and to obtain gains in consumption of the internal combustion engine and to reduce the risks of high exhaust temperatures.

Furthermore, the invention relates to a system for controlling an internal combustion engine capable of implementing steps 2 to 3 of the method according to any one of the variants or any one of the combinations of variants described above. , in order to control the internal combustion engine by limiting knocking.

In particular, the control system may include:

• an accelerometer sensor,
• computing means for executing a convolutional neural network, from an accelerometric signal;
• a memory for the parameters of the convolutional neural network and the acceleration measurement signal,
• means for controlling at least one actuator of the internal combustion engine.

The calculation means, the memory, and the control means can be integrated within an on-board computer of a vehicle.

The invention also relates to an internal combustion engine, in particular spark ignition, which comprises such a control system.

Furthermore, the invention relates to a computer program product downloadable from a communication network and/or recorded on a computer-readable medium (on-board computer) and/or executable by a processor. This program comprises program code instructions for implementing the method as described above, in particular steps 1.2/ and 3/ described above, when the program is executed on a computer or a computer/controller.

The present invention is not limited to the embodiments described above but encompasses all variants and all equivalents.

The characteristics and advantages of the method according to the invention will appear more clearly on reading the application example below.

For this application example, a 3-cylinder 4-stroke engine is considered with ignition according to the cylinder 1-cylinder 3-cylinder 2 sequence described above.

In order to cover a wide variety of operating conditions, in particular to cover knocking phenomena that are absent, incipient or episodic or very strong and repeated, 13 operating points are defined and, on average, 9 ignition advance values AVA per operating point leading to 116 triplets (rpm, load, AVA). Furthermore, for each triplet, 423 combustion cycles are recorded for cylinder 1. Thus, in total, 49068 engine cycles are used. The distribution of these operating conditions is illustrated in , which presents the intake pressure P for a spark-ignition engine as a function of the engine speed NE (in revolutions/minute, or rpm). For each of these 49068 engine cycles, a measurement is taken by means of a pressure sensor in the cylinder 1 of the engine and an accelerometer sensor placed on the engine block.

A knock indicator is then determined for each of the 49068 pressure measurements by means of the main variant of the invention described above, implementing the TDA method and using the TDAcrete type indicator defined above. At the end of this step, a knock indicator value is obtained, by correspondence, for each measurement of the acceleration. A learning base comprising 49068 data in the form of acceleration signal-value pairs of the TDAcrete type indicator is therefore obtained.

A random draw is then carried out among the 49068 data available so as to constitute the following three subsets: a training base (denoted BE hereafter) comprising 41268 data, a validation base comprising 3900 data (denoted BV by the suite), and a test database (denoted BT thereafter) containing 3900 data.

By way of illustration, the values of the TDAcrete type indicator thus determined are shown in the . More specifically, the at the top presents the value of the TDAcrete type indicator for each of the data of the BE, BV and BT subsets of the learning base, organized according to their number of samples N in the learning base; there at the bottom presents the value of the TDAcrete type indicator for each of the data of the BE, BV and BT subsets of the learning base, organized according to their number of samples N in the learning base and presented according to a logarithmic scale. It is observed that each BE, BV and BT subset of the learning base covers the same range of indicator values (with a few exceptions on the test base with a dozen values greater than 1000). The homogeneity of the distribution of the values of the TDAcrete type indicator between the BE, BV and BT subsets of the learning base makes it possible to ensure that the learning of the convolutional neural network is based on a variety examples similar to those present in the other BV and BT databases.

Then, by means of this learning base, a convolutional neural network consisting of 7 layers including 3 convolutional layers is trained, according to the particular design described above. In total, the number of neural network parameters to be determined is 20833, distributed as follows: 224 parameters for the first convolutional layer, 10304 for the second convolutional layer, 10272 for the third convolutional layer, and 33 parameters for the dense layer. Note that additional tests showed that a higher number of convolutional layers or another type of activation function did not improve the accuracy of the model. At the end of the training, a representative model of the knock indicator (here the TDAcrete type indicator) is obtained as a function of the acceleration for the engine considered.

This model is then used to determine a knock indicator, in real time, as a function of the measurement of the acceleration carried out in real time by the accelerometric sensor fitted to the engine in question.

By way of illustration, the knock indicator determined by the method according to the invention using the acceleration measurement (denoted TDAcrete-A hereafter) is compared with a knock indicator of the TDAcrete type determined by analysis of the pressure signal ( denoted TDAcrete-P thereafter), and a knock indicator of the MAPO type (local quantifier) determined according to the principles described in the document (Naber et al., 2006) applied to filtered and rectified accelerometric signals. It is considered that this MAPO type quantizer (denoted MAPO-A hereafter) constitutes the state of the prior art for the processing of the accelerometric signal making it possible to quantify the intensity of knocking.

Thus, the represents the value of the TDAcrete-A type indicator (left TDAcrete scale), the value of the TDAcrete-P type indicator (left TDAcrete scale) and the MAPO-A reference indicator (MAPO scale of right) as a function of the number of ignition advance increments NA, for different engine speeds, respectively 1750 rpm (top), 3000 rpm (middle) and 5500 rpm (bottom), for a inlet pressure of 1.2 bar. It should be noted that the values of the knock indicators presented in this figure are in fact averages of the knock indicators determined over the 423 combustion cycles. In these figures, one can observe on the one hand the progressive shift towards higher values of the quantizer MAPO-A as a function mainly of the engine speed, and on the other hand that this quantizer does not reproduce exactly the form of the quantizer TDAcrete-A or TDAcrete-P with a generally lower slope at the end of the variation, where knocking phenomena occur more strongly. The quantizer therefore has a bias and an attenuated response that is not observed in the prediction by the convolutional neural network of the quantizer TDAcrete-A or in the measurements of TDAcrete-P.

Thus the method according to the invention makes it possible, once a knock model has been determined, to detect and quantify the knock of an internal combustion engine in a precise manner, from a measurement taken by an accelerometer. This design is particularly advantageous because an accelerometer sensor is conventionally present on series engines. In addition, such an indicator is independent of the speed and load of the internal combustion engine, which allows an unbiased estimation of knock intensity. The use of such an indicator makes it possible to reduce the risk of damage to the cylinder, but also to reduce the consumption of the internal combustion engine and reduce the risk of high exhaust temperatures.

Claims (11)

Method for determining a combustion knock indicator in at least one cylinder of an internal combustion engine, in particular with controlled ignition, said engine being provided with at least one accelerometric sensor, characterized in that said method comprises at minus the following steps:
a) a model of an indicator of said knocking is constructed (MOD) as a function of an acceleration of said engine by training a convolutional neural network on a learning base, said learning base being constructed by means of said combustion engine internal combustion engine further comprising a pressure sensor in at least one cylinder, or of an internal combustion engine representative of said engine further comprising a pressure sensor in at least one cylinder, and according to at least the following steps:
- for at least one operating point of the internal combustion engine, a plurality of ignition advance values of said engine or of said engine representative of said engine are defined, and an acceleration signal is simultaneously measured by means of said accelerometric sensor and a pressure signal by means of said pressure sensor for at least one cycle of each ignition advance value of each operating point;
- for each cycle of each ignition advance value of each operating point, a value of said knock indicator is determined from said pressure signal, and said learning base is constituted by associating said value of said knock indicator knocking said measured acceleration signal simultaneously with said pressure signal for each cycle of each spark advance value of each operating point;
and in that :
b) an acceleration signal of said motor is measured (MES) by means of said accelerometer sensor;
c) said knock indicator is determined (IND) from said acceleration signal of said engine and by means of said model of a knock indicator as a function of an acceleration of said engine. Method for determining a knock indicator according to claim 1, in which, in step a), a knock indicator is determined from said pressure signal in the following way:
i) determining at least one local extremum of said pressure signal in an interval calibrated around the maximum peak of said measured pressure, said maximum peak being determined from among said at least local extremum; And
ii) Said knock indicator is determined as at least one descriptor of said at least one local extremum. Method for determining a knock indicator according to claim 2, in which at least one descriptor chosen from among: a first descriptor representative of a predetermined number of pressure variations of said pressure signal of greater amplitudes, and/or a second descriptor representative of all the pressure variations of said pressure signal, and/or a third descriptor representative of all the pressure variations of said pressure signal, with the exception of a predetermined number of pressure variations larger amplitudes. Method for determining a knock indicator according to one of Claims 2 to 3, in which the said at least one local extremum is determined by means of a method of topological analysis of the data, and for each local extremum a line segment connecting two consecutive local extrema. Method for determining a knock indicator according to claim 4, in which said at least one descriptor representative of a predetermined number of said segments of longest finite length determined by said method of topological data analysis, in particular at average of the ratio of the sum of the lengths of said longest segments, to the number of samples separating the longest segment from the last longest segment among the predetermined number of segments. Method for determining a knock indicator according to one of the preceding claims, in which the said convolutional neural network comprises at least 3 convolutional layers and a piecewise linear type activation function. Method for determining a knock indicator according to one of the preceding claims, in which said knock indicator is determined in real time by means of an on-board computer. Method for determining a knock indicator according to claim 7, in which said internal combustion engine is controlled (CON) in real time as a function of said knock indicator, preferably the ignition advance is controlled. Method for determining a knock indicator according to one of Claims 1 to 6, in which the said knock indicator is determined by means of a combustion analysis platform. Method for determining a knock indicator according to claim 9, in which the control of the internal combustion engine is calibrated (CAL) as a function of said determined knock indicator. Method for determining a knock indicator according to one of the preceding claims, in which the method includes an additional step of determining abnormal combustion by analyzing said knock indicator.