The paper focuses on the experimental identification and
validation of different neural networks for virtual sensing of
NOx emissions in combustion compression ignition engines
(CI). A comparison of several neural network architectures (NN,
TDNN and RNN) has been carried out in order to evaluate precision
and generalization in dynamic prediction of NOx
formation. Furthermore the model complexity (number and types of
inputs, neuron and layer number, etc.) has been considered to allow
a future ECU implementation and on line training. Suited training
procedures and experimental tests are proposed to improve the
models.
Several measurements of NOx emissions have been
performed through different devices applied to the outlet of a EURO
5 Common Rail diesel engine with EGR. The accuracy of the developed
models is assessed by comparing simulated and experimental
trajectories for a wide range of operating conditions.
The study highlights that history and proper inputs are
significant for the output estimation, and good results can be
achieved either through Recursive Neural Networks (RNN) or through
Neural Networks (NN) with input history. A virtual NOx
sensor will offer significant opportunities for implementing
on-board feed-forward and feedback control strategies in order to
improve the performance and the diagnosis of the engine and of the
after-treatment devices.