Prediction of Brake Squeal: A Deep Learning Approach Analysis byMeans of Recurrent Neural Networks
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
Noise, vibration and harshness (NVH) is a principal field of research for the automotive industry. Current research methods of brake NVH involve the finite element method and complex eigenvalue analysis, both of which fall short in terms of their noise prediction capabilities. Lately, research has shifted towards deep learning with the application of machine learning algorithms to detect, characterise and predict noise in commercial brake systems. This thesis investigates the possibility of implementing novel data science techniques to predict the vibrational behaviour of brake structure by means of deep neural network models, more specifically recurrent neural network (RNN) architectures. Two versions of RNN with an encoder-decoder architecture were evaluated: the long short-term memory (LSTM) and the gated recurrent unit (GRU) networks. The networks were applied on two datasets of normal force between the brake pad and the disc, measured in Newton: the sinusoidal data signal that corresponds to the brake squeal and the quiet reference data. The effect of the multifeatured data on prediction accuracywas investigated aswell. The results showed that the LSTM network produced the most reliable results on the sinusoidal data signal with a prediction length of 10 ms, which generated a weighted mean absolute percent error (wMAPE) of 61.57% and a mean absolute error (MAE) of the normal force of 0.1647 N. The corresponding results obtained by the GRU model were a wMAPE of 21.77% and a MAE of 0.1804 N. The highest wMAPE and MAE values of 91.01%and 0.0442 N, respectively, were obtained by the LSTM network on the multifeatured sinusoidal data signal with a length of 2.5 ms. In general, shorter prediction lengths generated higher accuracy and lower MAE scores. Moreover, predictions based on multifeatured datasets generated overall slightly better results compared to the single featured data. Overall, the outlook of data-driven applications on friction-induced dynamics seem promising. Future work, should focus on classification of various types of brake signals for a better understanding of the brake squeal phenomenon. .
AI , ANN , brake squeal , deep learning , encoder-decoder , GRU , LSTM , RNN , time series forecasting