Development and Real-Time Implementation of Recurrent Neural Networks for AFR Prediction and Control

The paper focuses on the experimental identification and validation of recurrent neural networks (RNN) for real-time prediction and control of air-fuel ratio (AFR) in spark-ignited engines. Suited training procedures and experimental tests are proposed to improve RNN precision and generalization in predicting both forward and inverse AFR dynamics for a wide range of operating scenarios. 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 comparison between RNNs simulation and experimental trajectories showed the high accuracy and generalization capabilities guaranteed by RNNs in reproducing forward and inverse AFR dynamics.

Then, a fast and easy-to-handle procedure was set-up to verify the potentialities of the inverse RNN to perform feed-forward control of AFR. Preliminary experimental tests indicate how the inverse RNN controller performance are comparable and in some cases even better than those guaranteed by the commercial ECU the reference engine is equipped with. Therefore RNN-based control of AFR emerges as a high potential alternative to reduce calibration efforts and to improve control performance as compared to the currently adopted techniques.