Deep long short-term memory neural networks as virtual sensors for marine diesel engine NOx prediction at transient conditions

Abstract

Virtual Sensors based on deep-learning models for predicting the NOx emissions of a Diesel Engine under transient conditions were developed and verified. Raw data from laboratory experimental measurements, under marine transient loading cycles, were used for training and evaluation of the developed models. NOx prediction under transient conditions is often inaccurate by implementing conventional methods since they fail to capture the dynamic behavior of internal combustion engines. The proposed model is based on Long Short-Term Memory (LSTM) Networks. A Deep Feed-forward Neural Network (DFNN) was also developed to validate the LSTM. The LSTM input is a time sequence of past measurements of the inputs while the DFNN only uses the most recent measurements. The Bayesian Hyberband Optimization (BOHB) algorithm determined the structure and parameters of each network. Each model uses the same inputs and is directly derived from the engine ECU. The LSTM validation showed that the model can generalize and accurately predict the NOx emissions under transient loading compared to the DFNN.