Lehrstuhl Kognitive Integrierte Sensorsysteme: Knock Detection

Contributions to the Development of Advanced Knock Detection Methods in Gasoline Engines

Subject:
The research study in a cooperation with Robert Bosch GmbH had the objective to employ advanced techniques from computational intelligence for data analysis, pattern recogntion, and optimization on real-world data from gasoline combustion engines to advance knock detection. In particular, this includes also the inclusion of ISE-experience on data visualization, analysis, and accelerating and automating of the design of such intelligent systems adapted to that specific automotive domain.

Abstract:
Knock detection (KD) is an established and industrially relevant field. It is an interesting case, where available and suitable sensors, e.g., pressure sensors, are too expensive for large volume production and somewhat harder to evaluate but economically priced sensors, e.g., acoustic transducers, have to be applied in conjunction with elaborate evaluation techniques. Pressure sensor in ignition valve (left), knock sensor (right) (Courtesy Bosch):

Proven signal processing methods, e.g., based on filtering, rectifying, and integrating as well as judicious thresholding have been applied. Increasing challenges with regard to, e.g., fuel efficiency, emissions limitation, higher power density etc. aggravate the requirements on the applied methods and potential hardware embodiment. Thus, methods from computational intelligence or soft-computing field, in particular non-linear and supervised approaches, are attractive to investigate for achieving advances with regard to robustness and discrimination ability for more dependable systems. Support-Vector-Machines (SVM), SVM regression, as well as Radial-Basis-Function-Networks (RBF) are promising candidates for the investigation. This is exemplified in the following figure from raw data to domain specific correlation plot a output:

Preprocessing of the raw data from engines includes filtering, fast fourier transform, and spectral filtering. Application of automated feature selection and automated parameter determination was also investigated in the context of the study. The following picture outlines the flow of the conducted experiments and evaluation based on several three and eight cylinder engines:

The following figure illustrates one of the obtained results, comparing the established KD procedure RKI to the one proposed and investigated here SVM RB, for training with cylinder one data of one engine and generalization test with the same aged engine and an additional engine of similar mileage:

Performance in generalizisation for different aged and different engines was found satisfying. The investigations were extended to an interesting range of engine speed, load, as well as different cylinders. For this aim, a model based normalization was conceived as an extension of the approach. Summarizing, the following classification rates of knocking events could be achieved for the proposed method in comparison to the established approach:

The particular benefit of the proposed method could be perceived in situations of background noise during knock detection. More details on the extensive investigations and simulations can be found in the reports below. Additional investigations to classification of knocking events have been made in straightforward establishing of a functional relationship between available pressure sensor output and conditioned acoustic transducer output (Knock sensor, KS). The advantage of such an approach is, that the somehow artificial categorization of the quasi continuous data is not required, but the output of the reliable pressure sensor serves a function approximation method as supervised input to learn a functional mapping. This approach is commonly denoted as Virtual Sensor approach (VS) and is illustrated in the following figure for the general (left) and the specific case (right):

This relates to the overall pursued research and tool architecture for automated intelligent system design as follows:

From the common methods Multi-Layer-Perceptron with Backpropagation alforithm, RBF, and SVM regression (SVR), the latter two have been investigated and compared following a similar experimental proceeding as reported above:

Summarizing, SVR proved to be competitive or superior to RBF VS approaches at the cost of higher resource demans, which still could be detrimental for today's embedded implementation. However, in the light of the increasing parallelism (Multi/many core architectures) and performance increase of embedded processing resources, this could be of diminishing importance in the future. The following figure shows one excerpt of achieved results for VS with regard to the ''ideal'' pressure sensor for generalization investigation:

More details on the extensive investigations and simulations can, again, be found in the reports enlisted below.

The excellent cooperation with the head of the Bosch group Dipl.-Ing. Robert Sloboda, and his group members, in particular, Dipl.-Ing. Stefan Kempf, is gratefully acknowledged.

The research activitiy is currently continued by an external PhD project of Dipl.-Ing. Matthias Biehl at Bosch in the same business unit.

	Status:		concluded, duration 03/2007-04/2009, external doctoral project based on this project running from 2010-2013 (expected end)
	Contractor:		Robert Bosch GmbH, Diesel Gasoline Systems Business Unit, Schwieberdingen.
	Financing:		Robert Bosch GmbH, Diesel Gasoline Systems Business Unit, Schwieberdingen.
	Contact:		Prof. Dr.-Ing. Andreas König
	Contributors:		Kuncup Iswandy and Andreas König
	Publications:

			K. Iswandy, S. Kempf, R. Sloboda, and A. König. Robustness Investigation of A SVM-Based Knock Detection Method. In Motortechnische Zeitschrift (MTZ), 7-8/2010, Vol. 71, pp. 486-491.
			K. Iswandy and A. König. Hybrid Virtual Sensor Based on RBFN or SVR Compared for An Embedded Application. In Proc. of the 15th Ann. Conf. on Knowledge-Based and Intelligent Information & Engineering Systems (KES 2011), Kaiserlautern, Sept. 12-14, Springer, LNAI 6881, Part. II, pp. 333-342, 2011.
			M. Biehl, S. Kempf und A. König. Messtechnische Hardwareplattform zur Entwicklung neuer Motorsteuergeräte-Algorithmen und -Funktionen am Beispiel der Kloperkennung. In: Tagungsband des XXV Messtechnischen Symposiums des AHMT, S. 39-50, Karlsruhe, September, 2011. ISBN 978-3-8440-0388-8