FOR AN INTERNAL COMBUSTION ENGINE

Technical Field

The invention relates to a method for calibrating a knock detection system for an internal combustion engine.

Background Art

Knocking is a problem in internal combustion engines, e.g. car engines, that may occur depending on several factors, for example engine settings like the ignition timing. A knocking combustion is a combustion where a part of the fuel-air-mixture in a cylinder starts burning ahead of the flame front ignited by the spark plug. This self-ignition process occurs when the auto-ignition conditions (pressure, temperature and delay) are reached in the unburned gas area. As a result of the auto-ignition process, high-frequency cylinder pressure peaks occur. Apart from impairing the efficiency of the combustion engine, knocking may lead to serious damage. Therefore, modern vehicles often have a knock detection system that may detect knocking autonomously and change the engine settings in order to remove the knocking. These knock detection systems do not measure the actual cylinder pressure, but normally employ a knock sensor that measures mechanical engine block vibrations in the vicinity of the respective cylinder.

According to a presently used knock detection system, data from the knock sensor are analyzed by selecting a detection window and applying several band pass filters to restrict analysis to those frequency bands that are considered significant. For optimum performance, such a system needs to be calibrated. According to a common calibration method, a fast Fourier transform (FFT) analysis of the knock sensor data relating to knocking conditions and normal conditions is performed in order to determine the relevant frequencies and the optimum position for the detection window(s). In order to gather sufficient data, the engine is sometimes kept under knocking conditions for hours. The currently used calibration method is expensive, approximative and may also cause damage to the engine due to the extended time under knocking conditions. The filters are selected by manual analysis of the FFT results, i.e. the calibration cannot be automated. Another calibration method identifies the filters and the window positions by calculating a correlation coefficient (R ² ) between a cylinder pressure maximum peak and a knock intensity determined by an engine controller. Technical Problem

It is thus an object of the present invention to provide a systematic and time-effective calibration of a knock detection system.

This problem is solved by a method according to claim 1 .

General Description of the Invention

The invention provides a method for calibrating a knock detection system for an internal combustion engine, in particular a gasoline engine.

Such a knock detection system is configured to receive time-dependent knock sensor data representing at least one engine combustion cycle. At least some aspects of the system are normally software-implemented, in particular in the engine control unit ECU. The system is connected to or comprises at least one knock sensor of the engine. As it is known, the respective knock sensor is normally an acoustic transducer and the knock sensor data represent vibrations in the engine block or a portion of the engine block. More specifically, the at least one knock sensor is disposed so that vibrations associated with an individual cylinder can be detected. In practice, one or two knock sensors may be arranged per engine block.

The system is further configured to extract, for each combustion cycle, data representing a detection window within the combustion cycle, and, based on the extracted data, calculate a knock intensity based on at least one band intensity of at least one detection frequency band. The detection window normally corresponds to a crank angle interval within each combustion cycle, but it may also be expressed as time interval. The detection window ideally comprises only those parts of the combustion cycle in which knocking may occur. The system extracts knock sensor data representing the respective detection window, which normally means that knock sensor data outside the detection window are ignored. The extracted data are used to calculate an indicator referred to as“knock intensity” based on at least one band intensity of at least one detection frequency band. The detection frequency band is a frequency band that is selectively taken into account for calculating the band intensity. In general, the band intensity is related to the intensity of the extracted data within the respective detection frequency band. In the most simple case, there may be only one detection frequency band and the knock intensity may be equal to the band intensity of this detection frequency band.

Furthermore, the system is configured to detect knocking by comparing the knock intensity with a knock detection threshold. The knock detection threshold is a threshold for the system which is used as a (simple) criterion whether knocking occurs or not. If the knock intensity is above the knock detection threshold, this is interpreted by the system as knocking occurring inside the engine, or, more specifically, in a specific cylinder. It is understood that while the system is configured to detect knocking, this includes the possibility that sometimes knocking occurs inside the engine but remains undetected (which is hereinafter referred to as a missed detection) as well as the possibility that knocking does not occur but the system detects knocking anyway (which is hereinafter referred to as a false detection). The knock detection threshold may be determined in various ways and the present invention is not limited to any specific way. For example, the system could record the knock intensities of a plurality of engine cycles, e.g. between 5 and 100 engine cycles, and set the knock detection threshold to be the average of those knock intensity is plus a (positive) offset. This approach is based on the assumption that the majority of engine cycles do not represent a knocking condition and that all engine cycles without knocking have a knock intensity within a certain range (which is approximated by the offset). Normally, an individual knock detection threshold is applied for each cylinder, although in some cases it may be feasible to use one and the same knock detection threshold for a plurality of cylinders.

The calibration procedure of the knock detection sensor in accordance with the inventive method is now described below. As it will be understood, calibration requires operating an engine to acquire measurement data, in particular knock sensor data and pressure data for a plurality of engine operating points (or working points). The amount of measurement data is considered to be relatively low compared to other prior art methods and the calibration method is thus preferably performed off-line, although it is possible to do it online. The method comprises the following steps, which are performed for each of a plurality of engine cylinders and each of a plurality of engine operating points. Each operating point corresponds to a certain engine speed and engine torque. In a first step, cylinder pressure data and knock sensor data are received for a measurement time interval. The method may be performed by a calibration device that is connected to at least one knock sensor and at least one pressure sensor. The cylinder pressure data represent the cylinder pressure in the respective cylinder of the engine. It will be understood that these cylinder pressure data are acquired with an experimental internal combustion engine, for the purposes of calibration, but not during the normal operation of an engine e.g. in an automotive vehicle. The knock sensor data as mentioned above corresponds to data received from a knock sensor. The cylinder pressure data as well as the knock sensor data are of course time- dependent (time or crank angle). These data are received for a measurement time interval (either offline or online). The length of the measurement time interval is in general not limited within the scope of the invention. However, in order to receive acceptable results, the measurement time interval for each cylinder and each operating point may be less than 10 minutes, less than 5 minutes or even less than 1 minute. As engine cycles are concerned, the measurement time interval may correspond to between 100 and 10 000 engine cycles or between 500 and 2 000 engine cycles. The knock sensor data as well as the cylinder pressure data for the measurement time interval are recorded or stored, either temporarily or permanently.

In another step, which may be performed before or after receiving the above-mentioned data, the method comprises defining a plurality of parameter sets, each parameter set representing a detection window and at least one detection frequency band. The parameter set may further comprise weighting factor(s) when using more than one detection frequency band. Within the parameter set, the detection window can be represented by a starting crank angle and a terminating crank angle and each frequency band can be represented by cut-off frequencies, i.e. a lower limit frequency and a upper limit frequency. However, if the width of the frequency band is always the same, it could be represented by a single frequency, e.g. the lower limit frequency, the upper limit frequency or a central frequency. It is understood that no two parameter sets have all identical parameters. The number of parameter sets is not limited within the scope of the invention, but in practice the number of combinations of parameters can vary between, e.g., 50 and 10000, depending on the desired accuracy.

Another step of the method comprises, for each parameter set, calculating the knock intensity and determining a cylinder peak pressure for a plurality of engine cycles from the measurement time interval. For the method to yield optimum results, the knock intensity is advantageously calculated in the same way as by the system (when operating normally in a vehicle). The cylinder peak pressure can be directly derived from the cylinder pressure data and is normally the maximum value of the cylinder pressure within a combustion cycle. As is known in the art, cylinder peak pressure is a reliable criterion to determine whether knocking actually occurs or not. The cylinder peak pressure and the knock intensity can be regarded as coordinates associated with the respective engine combustion cycle.

A further step of the method comprises, for each parameter set, identifying engine cycles representing a knocking condition, based on comparing the cylinder peak pressure of the respective engine cycle with a predetermined pressure threshold. If the cylinder peak pressure is above the pressure threshold, this engine cycle is considered to be subject to a knocking condition. Since the relation of the cylinder peak pressure and the pressure threshold does not depend on the current parameter set, the comparison only has to be done once, in which case "identifying" the respective engine cycles normally only means looking up the engine cycles that have previously been determined by comparing the cylinder peak pressure and the pressure threshold. Of course, the comparison could also be performed for each parameter set.

It shall be appreciated that the method further includes, for each parameter set, determining a missed detection rate and/or a false detection rate by comparing the knock intensity for all combustion cycles to a predetermined intensity threshold. The intensity threshold is a threshold that the knock intensity of each engine cycle can be compared to in order to determine whether this engine cycle is considered to represent a knocking condition or a non-knocking condition. The intensity threshold is, for the purpose of calibration, considered together with the pressure threshold. In general, some combustion cycles may have a knock intensity above the intensity threshold (therefore normally considered to be in a knocking condition) but have a cylinder peak pressure below the pressure threshold (therefore corresponding to an actual non-knocking condition). These correspond to false detections. Likewise, some engine cycles may have a knock intensity below the optimum detection threshold (therefore considered to be in a non-knocking condition) but have a cylinder peak pressure above the pressure threshold (therefore corresponding to a knocking condition). These correspond to missed detections. A missed detection rate can be computed by dividing the number of missed detections by the total number of engine cycles representing a knocking condition. Similarly, a false detection rate can be computed by dividing the number of false detections by the total number of engine cycles representing a non- knocking condition. In general, of course, both the missed detection rate and the false detection rate should be kept low, while it is normally not possible to reduce both rates to zero simultaneously. Therefore, different criteria can be defined as to what is considered the optimum combination of the missed detection rate and the false detection rate, some of which will be discussed below. By applying the respective criteria, the optimum detection threshold that yields the optimum combination can be determined.

Under the calibration process, missed detection rates and/or false detection rates are thus determined for each parameter set based on the predetermined intensity threshold. When this is done for each parameter set of a given operating point, an optimum parameter set is selected, which has the lowest missed detection rate, respectively false detection rate, or the best combination of missed and false detection rates, among all parameter sets.

In other words, the optimum combination of the false detection rate and the missed detection rate of each parameter set are taken into account and are compared. The optimum parameter set (including the at least one detection window and the at least one detection frequency band) is retained for use in the knock detection system. The system can use this parameter set for the respective cylinder and operating point. The inventive method is highly efficient and can be performed in a fully automatic way that does not require inspection of any data by a human being. Also, in contrast to some methods known in the art, it does not require a proper frequency analysis of the cylinder pressure or the knock sensor data or determining a correlation function. Rather, the mathematical operations necessary for performing the method are rather simple, as will to some part come apparent below. A further benefit of the present method is that the data acquisition for the purpose of calibration can be carried in a short time frame, typically within 3 or 4-hours.

Preferably, the band intensity of each frequency band is determined by applying a band pass, taking the absolute value and integrating over the detection window. It should be noted that some pre-processing could be performed before applying the bandpass, for example applying a physical low pass filter and/or an analogue/digital converter. The application of the bandpass essentially leaves all frequency components in the detection frequency band unchanged, while frequency components outside the detection frequency band are suppressed. After applying the bandpass, the resulting signal is normally an oscillating signal, which is transformed into a non-negative signal by taking the absolute value. Finally, this absolute value is integrated over the respective detection window.

According to preferred embodiment, the system is configured to calculate the knock intensity based on a plurality of band intensities of a plurality of detection frequency bands, and each parameter set represents a detection window and a plurality of detection frequency bands. It has been found that more reliable results can be achieved if more than one detection frequency band is used. In other words, the "relevant" frequencies are normally not localized in a single frequency band . In particular, the plurality of detection frequency bands may comprise three detection frequency bands. This can be regarded as a compromise in which on the one hand several frequency bands can be taken into account and irrelevant frequency bands are largely not considered, while at the same time the method does not become too complex by taking into account too many frequency bands, thus complicating the search for the optimum parameter set.

It is also preferred that the system is configured to calculate the knock intensity by combining the band intensities according to weight factors assigned to each band intensity and each parameter set represents at least one detection window, a plurality of detection frequency bands and a plurality of weight factors. In other words, in this embodiment, not all detection frequency bands are considered equally important, wherefore each band intensity is assigned with a weight factor. The values of the weight factors may be chosen in various ways, e.g. each weight factor could be between 0 and 1 . The weight factors are also parameters for a parameter set and can be varied to find the optimum parameter set.

As mentioned above, the optimum combination of the false detection rate and the missed detection rate can be defined by various criteria. According to one embodiment, the optimum combination corresponds to the lowest false detection rate combined with a missed detection rate below a predefined missed detection limit. For example, the target missed detection limit could be 10%, whereby the predetermined intensity threshold is set at 10%. In such case, the parameter set with the lowest false detection rate would be selected as optimum parameter set.

In particular, the target for the missed detection limit may be zero, which means that no missed detections are accepted. In this case, the optimum combination corresponds to the lowest false detection rate combined with a zero missed detection rate. For such determination, the predetermined intensity threshold is set at the level of the lowest knock intensity of all combustion cycles having a cylinder peak pressure above the pressure threshold.

In other embodiments, the optimum combination may correspond to the lowest missed detection rate combined with a false detection rate below a predefined false detection limit. For example, the target false detection limit may be 2%, whereby the predetermined intensity threshold is set at 2%. In such case, the parameter set yielding the lowest missed detection rate would be chosen as the optimum parameter set.

Again, the target false detection limit could be zero. In such case, the optimum combination corresponds to the lowest missed detection rate combined with a zero false detection rate. For such determination, the intensity threshold is set at the level of the highest knock intensity of all combustion cycles having a cylinder peak pressure below the pressure threshold.

There are various ways how to perform the optimization of the parameter set, which largely depend on the variation of the parameters within the plurality of parameter sets. According to one embodiment of the inventive method, a detection frequency band optimization is performed. Such a detection frequency band optimization comprises defining a first plurality of parameter sets differing only by the at least one detection frequency band. In other words, the detection window is the same for all parameter sets in the first plurality. Likewise, the weight factors, if present, are the same for all parameter sets. In another step, the detection frequency band optimization comprises determining the optimum parameter set with corresponding optimum detection frequency bands. According to another embodiment of the inventive method, a weight factor optimization is performed. Such a weight factor optimization comprises defining a second plurality of parameter sets differing only by the weight factors. In other words, the detection window and the detection frequency bands are the same for all parameter sets in the second plurality. In another step, the weight factor optimization comprises determining the optimum parameter set with corresponding optimum weight factors.

According to yet another embodiment of the inventive method, a detection window optimization is performed. Such a detection window optimization comprises defining a third plurality of parameter sets differing only by the detection window. In other words, the at least one detection frequency band and the weight factors, if present, are the same for all parameter sets in the third plurality. In another step, the detection window optimization comprises determining the optimum parameter set with corresponding the optimum detection window.

The above-mentioned three selective optimizations can be combined in that they are performed sequentially one after another. In such an embodiment, the method comprises sequentially performing a detection frequency band optimization, a weight factor optimization and a detection window optimization. The frequency band optimization is used to find the optimum frequency bands, which are then left unchanged by the following optimizations. In the frequency band optimization, predefined default values can be used for the detection window and the weight factors. In the weight factor optimization, the optimum weight factors for the respective frequency bands are determined, which are then left unchanged by the following optimization. In the weight factor optimization the above- mentioned default value for the detection window can be used. Finally, in the detection window optimization, the optimum detection window is determined. Of course, it is conceivable to change the sequence of the three above-mentioned optimizations. Also, it is conceivable to perform this sequence of optimizations several times, e.g. because changing the detection window could also lead to different optimum frequency bands. However, repetition of the optimization sequence increases the total optimization time and it is also not sure that such repetition leads to a stabilization or "convergence" of the parameters.

Brief Description of the Drawings

Preferred embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings, in which:

Fig. 1 is a schematic view of an engine with a knock detection system;

Fig. 2 is a graph illustrating a knock sensor voltage as a function of the crank angle; Fig. 3 is a graph illustrating the frequency spectrum of a knock sensor voltage;

Fig. 4 is a flowchart illustrating an embodiment of an inventive method;

Fig. 5 is a flowchart illustrating a part of the method of Fig.4;

Fig. 6 is a graph illustrating a cylinder pressure as a function of a crank angle; and

Fig. 7 is a graph showing, for a plurality of engine cycles, the knock intensity vs. cylinder peak pressure.

Description of Preferred Embodiments

Fig.1 schematically shows an internal combustion engine 10 for an automotive vehicle with a plurality of cylinders 1 1 . Two knock sensors 12 are disposed near the cylinders 1 1 , fixed to the cylinder block. That is, one knock sensor 12 is provided for a pair of cylinders 1 1 , but this is just exemplary. Each knock sensor 12 is configured to output a voltage representative of a vibration amplitude of one cylinder 1 1 . These vibrations originate mostly from the combustion events in the individual cylinders 1 1 , which occur sequentially and therefore can be distinguished from one another. The voltage is received by a control unit 2 that is part of a knock detection system 1 .

When the engine 10 is in operation, the control unit 2 receives time-dependent knock sensor data, in this case a voltage, from each knock sensor 12. Fig. 2 shows the knock sensor voltage as a function of the crank angle. For each engine cycle, the control unit 2 extracts data representing a detection window W. Fig. 2 shows a detection window W ranging from approximately 12° to 62°, which corresponds to a typical crank angle interval where knocking may occur. Based on the extracted data, the control unit 2 calculates an indicator referred to as“knock intensity”. The knock intensity is calculated by passing the signal from the detection window W through a lowpass filter and an analogue/digital converter, whereafter three independent band passes are applied, each of which corresponds to a detection frequency band F1 , F2, F3. Fig. 3, by way of example, shows a graph with the frequency spectrum of the knock sensor data with the positions of the three detection frequency bands F1 , F2, F3. For each of the detection frequency bands F1 , F2, F3, a band intensity is calculated by taking the absolute value and integrating over the detection window W. The control unit 2 has stored weighting factors k1 , k2, k3 assigned to the respective detection bands F1 , F2, F3. The knock intensity is calculated by adding the individual band intensities weighted by their respective weighting factor k1 , k2, k3. This can be mathematically written as: KI = å k _n - BI _n , Where Kl is the knock intensity, \s the respective weighting factor and S/ the Band Intensity for the filter n.

In order to detect knocking, the control unit 2 compares the knock intensity with a knock detection threshold. If the knock intensity is above the knock detection threshold, this is interpreted as a knocking condition of the respective cylinder 1 1 . By way of example, the control unit 2 may compute the knock detection threshold by taking the average of the knock intensities of a plurality of previous combustion cycles (e.g. the last 20 or 50 combustion cycles) and adding a (positive) offset. The computation of this knock detection threshold can be done in various ways and is not the focus of the invention.

The optimal working condition of the system 1 and the control unit 2 depends on the optimum selection of the detection window W, the detection frequency bands F1 , F2, F3 and the weighting factors k1 , k2, k3. Therefore, the system 1 needs to be calibrated by an inventive method which will now be explained with reference to the flowcharts in Figs. 4 and 5. The method may be performed by a calibration unit (not shown) that is connected to an internal combustion engine 10 of same design but operated under experimental conditions. Apart from being connected to knock sensors 12 as shown in fig. 1 , the calibration unit would have to be connected to pressure sensors that measure a cylinder pressure in each of the engine cylinders.

The present calibration method requires cylinder pressure data and knock sensor data for a plurality of engine operating points (each characterized by an engine speed and an engine torque). In practice, the engine is thus operated at each of the predetermined engine operating points for a given measurement time interval of e.g. 1 or 2 minutes, in order to measure the pressure and knock sensor data for all the cylinders. The measurement time interval normally corresponds to several hundred or several thousand engine cycles. The knock sensor data as well as the cylinder pressure data for the measurement time interval are recorded. The number of operating points is selected to have a mapping of the engine operating range. The number of operating points will be selected depending on the desired accuracy.

In practice, the whole set of data is acquired before the data processing (i.e. offline). It is considered that the data acquisition may take about 2 hour. It is however also possible to start data processing as soon as a first set of data is available (i.e. online).

The data processing under the present calibration method will now be explained with reference to the flowcharts. The optimization process that will be described is carried out for each cylinder and each operating point, based on the acquired data set. Turning now to Fig.4, in step 100 of the method, a first operating point is selected from a plurality of operating points, each of which is hence characterized by an engine speed and an engine torque. Next, at 1 10, a first cylinder 1 1 to be analyzed is selected. After that, at 120, the cylinder pressure data and knock sensor data for the measurement time interval are retrieved from the calibration unit memory. These cylinder pressure data and knock sensor data are of course representative of a single cylinder 1 1 and a single operating point. Fig. 6 by way of example shows the cylinder pressure vs. crank angle.

In a next step 130, a detection frequency band optimisation is performed. For this optimisation, at 140, a first plurality of parameter sets is defined, each of which represents a detection window W, three frequency bands F1 , F2, F3 and three weighting factors k1 , k2, k3. While the detection window W and the weighting factors k1 , k2, k3 are the same for all parameter sets and may correspond to predefined default values, all parameter sets represent different combinations of frequency bands F1 , F2, F3. It may be noted here that whereas a lot of filter combinations are envisaged, filter ranges should not overlap within a parameter set. Afterwards, at 150, and optimum parameter set is found representing the best filter configuration. This consists of a plurality of steps which will now be described with reference to fig. 5.

At 151 , a first parameter set is selected, and at 152, the knock intensity is calculated and a cylinder peak pressure is determined for a plurality of combustion cycles, normally all combustion cycles, from the measurement time interval in the respective cylinder. The knock intensity is calculated as explained above with respect to the knock detection system 1 and the cylinder peak pressure can be simply derived from the cylinder pressure data. Fig. 7 shows a graph where each dot corresponds to the knock intensity of a respective combustion cycle plotted vs. the corresponding cylinder peak pressure.

In another step 153 of the method, those combustions are identified that represent a knocking condition. They can be determined one time for each operating point, by comparing the cylinder peak pressure of the respective engine cycle with a pressure threshold p _t , which is represented in Fig. 7 by the vertical dashed line. All engine cycles to the right of this line indicate an actual knocking condition of the cylinder, while all engine cycles to the left of this line indicate an actual non-knocking condition. Since the pressure is measured directly in the combustion chamber, the pressure threshold line p _t is considered as a true indicator to discriminate between presence and absence of knocking. Where the pressure is greater than p _t then knocking truly occurred in the engine. In this connection, one may refer to Fig.6 where the lower line represents a normal combustion whereas the upper line represents a combustion where knocking is present: there is a sudden increase of cylinder pressure during the combustion event itself.

In a next step 154, an intensity threshold I _t for the knock intensity is set and the knock intensities for the combustions recorded for the operating point are compared to the intensity threshold.

Setting this intensity threshold I _t yields a certain combination of a missed detection rate and a false detection rate for all measurement points.

Such intensity threshold is represented by the horizontal dashed line in Fig.7. As can be seen, the pressure threshold p _t and the intensity threshold I _t divide the graph into four regions. As already indicated, for all points on the right of vertical line p _t it can be said that knocking was present. Now, the actual detection of knocking by the knock sensor depends on the position of the intensity threshold.

From the graph we can distinguish between four cases:

Good detection: knocking combustion detected by the sensor (intensity > It) and confirmed by the cylinder pressure analysis. These points are in region (I).

Missed detection: knocking combustion not detected by the sensor (intensity <lt) but detected via the cylinder pressure analysis. These points are in region (IV).

False detection: knocking combustion detected by the sensor but seen as unacceptable by the cylinder pressure analysis. These points are in region (II)

Good non-detection: knocking combustion not detected by the sensor and confirmed by the cylinder pressure analysis. These points are in region (III).

A missed detection rate can be computed by dividing the number of missed detections by the total number of engine cycles representing a knocking condition according to pressure analysis (i.e. Good detections+Missed detections).

Similarly, a false detection rate can be computed by dividing the number of false detections by the total number of engine cycles representing a non-knocking condition according to pressure analysis (i.e. False detection+Good non detection).

In fact, a certain rate of missed detection or false detection is pre-defined, and the system will then determine the other rate. Two possibilities for setting the intensity threshold are shown in Fig.7. The lower horizontal dashed line corresponds to an intensity threshold that is chosen so that the missed detection rate is zero, while the false detection rate has the lowest possible value. That is, the optimum detection threshold I _t is set to cross the point where knocking is present (right of line Pt) and having minimum intensity.

Alternatively, the upper horizontal dashed line in fig. 7 corresponds to an optimum detection threshold that is chosen so that the false detection rate is zero, while the missed detection rate is at the lowest possible value. That is, the optimum detection threshold I _t is set to cross the point of highest intensity where no actual nocking is present (left of line Pt).

Other combinations can be considered as optimal, e.g. a combination where a non-zero missed detection rate is accepted in order to reduce the false detection rate (or vice versa).

Based on the defined intensity threshold It, the false detection rate and/or the missed detection rate are determined. Where the missed detection rate is set as zero (lower dashed line in Fig.7), then the false detection rate is computed (step 154) and will be used for assessing the performance of the respective parameter set.

At 155, it is checked whether the present parameter set was the last parameter set. If not, the next parameter set is selected at 156 and the method continues with step 152. If it was the last parameter set, the false detection rate has been determined for each parameter set in the first plurality, and at 157, and an optimum parameter set is selected for the system 1 , which has the best combination of missed detection rate and false detection rate among all parameter sets. For example, this may be the parameter set with a zero missed detection rate and the lowest false detection rate.

After the detection frequency band optimisation has been completed, a weighting factor optimisation is performed at 160. A second plurality of parameter sets is defined at 170, all of which represent the same detection window W as used during the detection frequency band optimisation and the frequency bands F1 , F2, F3 of the optimum parameter set determined during the detection frequency band optimization. However, the parameter sets in the second plurality have different combinations of weighting factors k1 , k2, k3. From these parameter sets, an optimum parameter set is determined at 180 in the same way as during the detection frequency band optimization. The optimum parameter set corresponds to an optimum combination of weighting factors k1 , k2, k3.

Finally, a detection window optimisation is performed at 190. At 200, a third plurality of parameter sets is defined, all of which have the frequency bands F 1 , F2, F3 of the optimum parameter set determined during the detection frequency band optimization and the weighting factors of the optimum parameter set determined during the weighting factor optimization. However, they differ in the detection window W which they represent. From these parameter sets, an optimum parameter set along with an optimum detection threshold is determined at 210 in the same way as during the detection frequency band optimization and the weighting factor optimization.

The optimum parameter set found in the final optimisation is considered to be optimised regarding the detection frequency bands F1 , F2, F3, the weighting factors k1 , k2, k3 and the detection window W. In step 220, this parameter set is provided to the system 1 as corresponding to the current cylinder 11 and operating point. At the 230, it is checked whether the current cylinder 11 was the last cylinder. If not, the next cylinder is selected at 240 and the method goes back to step 120 receiving knock sensor data and cylinder pressure data. If it was the last cylinder, another check is performed at 250 whether the current operating point is the last one. If not, the next operating point is selected at step 260 and the method goes back to step 110 to select the first cylinder. If it was the last operating point, the method ends.

The system 1 has now been provided with optimum parameter sets for each cylinder 11 and each (previously selected) operating point. Since the number of selected operating points is limited, the actual engine speed and engine torque during operation of the engine 10 will normally not correspond exactly to any of these operating points. However, the system 1 can either do an interpolation between the optimum parameter sets of neighbouring operating points or it may simply select the optimum parameter set for the closest operating point.

Also, in practice the found optimum parameter sets will be used in the control unit, or more generally in the ECU, as a basis to build a table defining the parameters window W, filter F 1 , F2 and F3, and weighting factor k1 , k2 and k3, in function of engine speed, engine load and cylinder number. The knock intensity function will then read its parameters from such table.