Virtual Sensing of SI Engines Using Recurrent Neural Networks

For engine diagnostics and fault-tolerant control system design provision of analytical models, in the form of virtual sensors, will enable more reliable system design and operation. This paper presents applications of recurrent neural network (RNN)-based architectures for the development of virtual sensors for salient SI engine variables such as manifold absolute pressure, mass airflow rate, air-fuel ratio and engine torque. The RNN architectures developed allow effective sensing of these crucial engine variables while, for computational efficiency, keeping a compact size for the network topology. A nonlinear state-space model strategy is proposed for architecting the stated recurrent neural network and is trained using variants of the real-time recurrent learning (RTRL) algorithm. Representative experimental results obtained for a 5.7 L V8 engine are listed and discussed. The application, dependency and limitations of the proposed approaches are also pointed out.