Recurrent and Time-Delay Neural Networks as Virtual Sensors for NOx Emissions in Marine Diesel Powertrains

Neural networks (NN) for marine engines, using raw measurement data from laboratory measurements, are developed and verified. These models can be utilized as virtual sensors of engine-out NOx emissions and lambda (λ). Investigations for the optimal NN configuration targeting models were carried so they can capture the dynamic behavior of a marine diesel engine, can generalize within the training range, and have the minimum complexity due to execution performance and portability reasons. Two configurations of NNs are investigated, the recurrent (RNN) and the time-delay neural network (TDNN). The resulting NN models are deployed on a prototype engine control unit (ECU) platform and are validated in real time for operating points and patterns that are not included in the training dataset. The real-time validation shows that the predicted quantities remain consistent in most operating areas and the dynamic behavior of the system is captured and reproduced accurately. The NN models were compared against a first-principle, physics-based virtual sensor and an engine map with comparable results and executed seamlessly fast.