Deep Learning-Based Combustion Quality Classification for Compression-Ignition Engines Using Nonintrusive Accelerometers

This article develops a deep learning (DL)-based combustion quality classification method for compression-ignition engines using acceleration data. The acceleration signal is collected using a nonintrusive accelerometer mounted on the engine block. Utilizing the acceleration data, a 4-class combustion quality classification is developed via a DL-based scheme. To this end, a DL architecture encompassing a convolutional neural network (CNN) and long short-term memory (LSTM) is trained for a regression task, and then, a grid search optimization is employed to find optimal thresholds for the classification purpose. The 4 classes considered are complete misfire, partial misfire, poor combustion, and normal combustion. The performance of the proposed method is validated with extensive experimental data involving many different operating conditions. Comparisons with three baseline methods are provided to illustrate the superiority of the new result.