Experimental Validation of a Neural Network Based A/F Virtual Sensor for SI Engine Control

The paper addresses the potentialities of Recurrent Neural Networks (RNN) for modeling and controlling Air-Fuel Ratio (AFR) excursions in Spark Ignited (SI) engines. Based on the indications provided by previous studies devoted to the definition of optimal training procedures, an RNN forward model has been identified and tested on a real system. The experiments have been conducted by altering the mapped injection time randomly, thus making the effect of fuel injection on AFR dynamics independent of the other operating variables, namely manifold pressure and engine speed. The reference engine has been tested by means of an integrated system of hardware and software tools for engine test automation and control strategies prototyping.

The developed forward model has been used to generate a reference AFR signal to train another RNN model aimed at simulating the inverse AFR dynamics by evaluating the fuel injection time as function of AFR, manifold pressure and engine speed.

Both forward and inverse models have been tested on new experimental AFR transients, showing the ability of the network in following the target patterns with satisfactory accuracy. Potential applications include the use of the forward RNN as virtual AFR sensor and the integration of forward and backward RNN into an adaptive open-loop controller.