A Novel Deep Learning Approach to Predict the Instantaneous NOₓ Emissions From Diesel Engine

Abstract

Accurate and stable prediction of NOx emissions from diesel vehicles plays a crucial role in the establishment of virtual NOx sensors and the development and design of diesel engines. This paper presents a method for estimating transient NOx emissions by complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and a long- and short-term memory neural network (LSTM). First, the CEEMDAN algorithm is used to reduce the non-stationarity and volatility of the transient NOx emission data to obtain multiple subseries with different frequencies. Secondly, a predictive model is developed for each subsequence using an LSTM neural network. Finally, the results of each subsequence prediction are summed to obtain the final prediction. The proposed model uses NOx emission data generated by an EU IV diesel bus during real road driving. The results show that (1) The use of CEEMDAN can effectively improve the smoothness of NOx transient emission data, as well as facilitate more effective extraction of internal characteristics and variations of the raw data. (2) LSTM has better learning and prediction capability for transient changes in NOx emissions. (3) The results of CEEMDAN-LSTM for RMSE, R2, MAE and NRMSE are 46.11,0.98, 29.82 and 2.71, respectively, which are better than the other model with improved prediction performance.

Keywords

• Diesel engine
• NOₓ emissions
• long short-term memory network
• empirical mode decomposition
• machine learning