A comparison of statistical learning approaches for engine torque estimation

International audienceEngine torque estimation has important applications in the automotive industry: for example, automatically setting gears, optimizing engine perfor- mance, reducing emissions and designing drivelines. A methodology is described for the on-line calculation of torque values from the gear, the accelerator pedal position and the engine rotational speed. It is based on the availability of input-torque experimental signals that are pre- processed (resampled, filtered and segmented) and then learned by a statistical machine-learning method. Four methods, spanning the main learning principles, are reviewed in a uni- fied framework and compared using the torque estimation problem: linear least squares, linear and non-linear neural networks and support vector machines. It is found that a non-linear model structure is necessary for accurate torque estimation. The most efficient torque model built is a non-linear neural net- work that achieves about 2% test normalized mean square error in nominal conditions