A neural network approach for indirectly estimating farm tractors engine performances

The instant torque and brake specific fuel consumption (BSFC) of a farm-tractor engine are very interesting parameters from a technical and economical point of view and allow advancing many considerations in the Engineering and Farm-mechanization fields related to the optimization of the engine power and consumptions. A direct access to the CAN-BUS system, where present, can be difficult; as a consequence, some practical solutions (sensors, numerical methodologies) aimed to deduce continuously but indirectly the engine performances are therefore proposed and discussed. In particular, the focus of this study is to evaluate the possibility of using artificial neural networks (ANNs) trained with exhaust gas (EG) and motor oil temperature data, easy to be measured. Hence, the above-mentioned temperatures and several network architectures (different for neurons and hidden layers number, neuronal transfer functions) were evaluated in their reliability in estimating the torque and BSFC of different tractor diesel motors, giving also the readers some useful indications: determination coefficients were calculated with reference to the line "predicted values = experimental values". Lubricant temperature resulted to be totally unsuitable (very low and diversified R2). ANNs using the EG temperature for torque estimations achieved higher average R2 than ANNs predicting BSFC, both in the training (>0.996 vs. >0.889) and in the prediction phase (>0.993 vs. >0.621). Consequently, EG temperature is strongly recommended for estimating both parameters even if preliminary evaluations should be performed for BSFC (engine characteristics have a significant influence on the predictions). Finally, best R2 can be scored by using the Gaussian neuronal transfer function.