Vibration can cause professional illnesses in train drivers, giving also rise to lawsuits to the employer. A possible cause may be the lack of systematic vibration estimation processes, due to operational complexities, subjectivities involved and the cost of dedicated tests. Estimation quality may be improved by using a driver seat model along with cabin floor vibration data acquired during the train dynamic approval tests. However, due to the nonlinearities present, analytical models frequently show inaccurate results. This work deals with the design of an appropriate neural network for predicting the seat-driver interface vibration, based on selected and processed cabin floor acceleration data obtained during the dynamic approval tests. Network type, input signals set and signal conditioning have considerable impact on the simulation accuracy. Results show good correlation between simulated and experimental data, even better between simulated and measured standard vibration dose indicators, being RMS errors between 3.9 % and 9.4 % and peak factor errors between 0.8 % and 9.6 %.

No abstract