Real-time combustion progress estimation using deep learning

This paper develops a deep learning algorithm to estimate the CA50 value of each combustion cycle from the signals of a non-intrusive accelerometer located on the engine block of a diesel engine, together with crank angle and start of injection (SOI) values. The CA50 value, which marks the crank angle at which 50% of heat release occurs, is an important feedback variable for improving the efficiency of cycle-to-cycle combustion. The proposed algorithm separates the combustion component of the accelerometer signal, utilizes a ResNet (residual convolutional neural network)-LSTM (long short-term memory) network to extract key signal features, and dense neural layers to combine the extracted time series features with start-of-injection to predict CA50. The developed algorithm is trained using extensive experimental data with a range of fuels of different cetane numbers and different operating conditions. Fuels with cetane numbers of 25, 30, 35, and 48 and a wide range of values for SOI and ignition assistance (IA) power are used. The algorithm is tested extensively using unseen data with a new ignition assistance device. For normal combustion cycles, the algorithm is found to work effectively with a root mean square error (RMSE) in CA50 prediction of 1.5 degrees. It is shown that 93.3% of all cycles have an accuracy better than 1 degree and 99.73% of cycles have an accuracy better than 2 degrees.