Engine speed estimation using neural networks (vor Ort)
* Presenting author
Abstract:
The rotational speed of an internal combustion engine is an important input variable for the analysis of vehicle interior noise measurements. However, in vehicle tests the effort of applying a measurement device is not always proportionate, which is why the engine speed is not always available in acoustic measurements.In this paper a methodology is presented which estimates the rotational speed of an ICE using a regression model based on deep neural networks. A Long Short-Term Memory (LSTM) is used as the basic network architecture. The input to the LSTM is the spectrogram of a monaural airborne sound measurement. A total of 527 measurements with an average duration of 45 s are used to train the LSTM. The model shows very good stability in the learning process with a mean percent error of 6.1 % for the training data set, 6.4 % for validation and 6.79 % for the test data set. A subsequent autocorrelation analysis can improve the rotational speed estimation.