Monitoring the air–fuel ratio of internal combustion engines using a neural network

Abstract

Precise control of an internal combustion engine for the reduction of exhaust emissions benefits substantially from accurate measurement of the air–fuel or λ ratio. This paper describes the use of the spark plug as a combustion sensor. The time-varying spark-voltage profile is acquired and used to train a neural network. The neural network is able to estimate the λ ratio. Minimal additional instrumentation is required. Therefore, this method potentially provides a cheap and robust sensor for measuring the λ ratio. The work described here involved devising an appropriate data acquisition system and undertaking experiments to validate the sensor technique. Vectors representing the spark-voltage profile were acquired at three different λ ratio operating points and used to train and test the neural network to ascertain its ability to determine the λ ratio accurately and linearly over a range. The results have supported the application of the spark-voltage profile as a means of λ ratio measurement and the use of neural networks for rapid calibration of engine management systems. Potential benefits and problems of neural network analysis of spark-voltage vectors, as a means of λ ratio estimation by virtual sensing, have been discussed.

A transient virtual-AFR sensor using the in-cylinder ion current signal

“…This use of NNs for AFR prediction has now been extensively researched [15,13,16,12,6,17]. Networks sizes and complexities have been determined through trial and error and estimation errors as low as 2% for 90% of test transients have been achieved [18].…”

“…The spark plug with appropriate diagnostic circuitry has also been extensively investigated for possible use in the measurement of in-cylinder combustion variables and as a 'virtual' lambda sensor [6].…”

“…It involves analysis of the time-varying spark voltage vectors and a lambda accuracy has been achieved to within AE0:1 in about 95% of the time over its identified operating range [6].…”

Effects of engine parameters on ionization current and modeling of excess air coefficient by artificial neural network

“…In that study, a three-layer NARX network was used and the prediction performance of AFR was 0.9724 of R 2 [2]. The use of spark plug voltage characteristics in an ANN to achieve a cheap solution for the AFR measurement was investigated by Walters et al According to results of the study, the excess air coefficient could be measured with ±0.1 accuracy using an ANN-based spark voltage characterization method [3]. Cycle-resolved AFR and incylinder pressure were estimated using an ionization currentbased stochastic estimation method with a maximum 2% of errors by Lee et al [4].…”

A new method for measurement of air–fuel ratio based on the response time of binary-type exhaust gas oxygen (BEGO) sensor for application in small spark ignition (SI) engines

“…Some of them are: using cylinder pressure data coupled with an engine-dependent computer model, a BEGO sensor model with an observer, a non-linear state estimator using manifold pressure data, linear parameter-varying control using input shaping and cylinder pressure coupled with combustion heat release data (Cesario et al, 2005;Chang et al, 1995;Powell et al, 1998;Zhang et al, 2008). Similarly, measurement of AFR using a combination of spark voltage profile and neural networks has been tested successfully (DWalters et al, 2006).…”