Search Results

The paper focuses on the experimental identification and validation of recurrent neural networks (RNN) for virtual sensing of NO emissions in internal combustion engines (ICE). Suited training procedures and experimental tests are proposed to improve RNN precision and generalization in predicting NO formation dynamics. The reference Spark Ignition (SI) engine was tested by means of an integrated system of hardware and software tools for engine test automation and control strategies prototyping. A fast response analyzer was used to measure NO emissions at the exhaust valve. The accuracy of the developed RNN model is assessed by comparing simulated and experimental trajectories for a wide range of operating scenarios. The results evidence that RNN-based virtual NO sensor will offer significant opportunities for implementing on-board feedforward and feedback control strategies aimed at improving the performance of after-treatment devices.

The paper deals with the use of Recurrent Neural Networks (RNNs) for the Air-Fuel Ratio (AFR) control in Spark Ignition (SI) engines. Because of their features, Neural Networks can perform an adaptive control more efficiently than classical techniques. In the paper, a review of the most useful control schemes based on Neural Networks is presented and the potential use in the field of engine control is analyzed. A preliminary controller has been implemented making use of a Direct Inverse Modeling approach. The controller compensates for the wall wetting dynamics and estimates the right amount of fuel to be injected to meet the target AFR during engine transients. The Direct Inverse Controller has been tested within an engine/vehicle simulator. The simulation tests have been performed by imposing a set of throttle transients at different engine speeds. The results show that the Inverse Model can satisfactorily bound the AFR excursions around the target value.

An integrated system of phenomenological models is applied in conjunction with identification techniques to simulate SI engine performance and emissions. In the framework of a hierarchical model architecture, the model structure provides the steady state engine data required for the design and validation of synthetic engine models. This approach allows limiting the recourse to the experimental data and speeds up the engine control strategies prototyping. The model structure is composed of a multi-zone thermodynamic engine model linked to a 1-D commercial fluid-dynamic model for intake and exhaust gas flow and to a physical model for NOx exhaust emissions. In order to improve model accuracy and generalization, an identification methodology is applied to estimate the optimal parameters for the turbulent combustion model. Due to the built-in physical content, the proposed methodology requires a relatively limited amount of experimental data for characterizing the under-study engine.

A set of models for the prediction of mechanical efficiency as function of the operating conditions for an automotive spark ignition engine is presented. The models are embedded in an integrated system of models with hierarchical structure for the analysis and the optimal design of engine control strategies. The validation analysis has been performed over a set of more than 400 steady-state operating conditions, where classical engine variables and pressure cycles were measured. Models with different functional structures have been tested; parameter values and indices of statistical significance have been determined via non-linear and step-wise regression techniques. The Neural Network approach (Multi Layer Perceptrons with Back-Propagation) has been also used to evaluate the feasibility of using such an approach for fast black-box modelization.