Accelerometer-Based Combustion Metrics Reconstruction With Radial Basis Function Neural Network for a 9 L Diesel Engine

Accelerometer-based combustion sensing in diesel engines has the potential of providing feedback for combustion control to reduce fuel consumption and engine emissions at a lower cost than in-cylinder pressure sensors. In this work, triaxial block-mounted accelerometers were used to measure the engine vibration, and pressure transducers were installed to measure the in-cylinder pressure. The in-cylinder pressure can be further utilized to compute combustion metrics, including the apparent heat release rate (AHR). Engine tests were conducted for various speeds, torques, and start of injections, on a 9 L in-line six-cylinder diesel engine equipped with a common rail high pressure injection system. The relationship between engine block acceleration and AHR was modeled using a radial basis function neural network (RBFNN). By inputting the accelerometer signal to the fixed network, AHR and other combustion metrics were estimated. As the primary concern for radial basis network training is the hidden layer weight vector selection, two algorithms for weight vector selection (modified Gram–Schmidt orthogonalization and principal component analysis) were evaluated by examining the robustness of the resulting network. One-third of the conducted tests were utilized to train the network. The network was then applied to estimate the AHR for the remaining validation tests which were not used to train the network. Comparisons were made based on the combustion metrics estimation results and the selection efficiency among the two weight vector selection methods and the random selection method. Moreover, the capability concerning the network's tolerance for additive noise was also investigated. Results confirmed that the modified Gram–Schmidt method achieved much more accurately estimated combustion metrics with the highest efficiency. On the basis of this study, a real-time closed-loop control strategy was proposed with the feedback provided based on the application of the trained RBFNN.