Engines equipped with Dynamic Skip Fire (DSF) technology generate low frequency and high amplitude excitations that could reduce vehicles drive quality if not properly calibrated. The excitation frequency of each firing pattern depends on its length and on the rotational speed of the engine. Excitation amplitude mainly depends on the requested engine torque by the driver. During the calibration process, the torque characteristics that results in production level of noise, vibration, and harshness (NVH), must be identified, for each firing pattern and engine speed. This process is quite time consuming but necessary.
To improve our process, a novel machine learning technique is utilized to accelerate the calibration effort. The idea is to automate the vibration rating procedure such that given the relevant power-train parameters, a vibration rating associated with that driving condition can be predicted. This process is divided into two (2) prediction models. The first model is a multiple additive regression trees that predicts the seat accelerometer data based on the various engine and vehicle parameters. The predicted seat accelerometer data is used as an input to the second machine learning model which correlates, along with other relevant engine and vehicle parameters, to a final vibration rating score. The results indicate that using this machine learning approach can significantly improve capability of automating the DSF calibration process delivering commercial NVH performance.
