Characterization of Automobile Acoustic Insulation Performance based on Artificial Neural Networks via Sensitivity Analysis

The manufacturing industry faces challenges when comes to set the correct type of material and its proportions to get a reliable and economically acceptable material to serve as acoustic insulation. Most of the time the trial and error method or such a semi-empirical methodology is used to perform this material selection, this procedure leads to high up on the cost of the product. Therefore, in this sense, this paper aims to apply the Artificial Neural Networks (ANNs) to supply a simulated based methodology to analyze the effects of the density, thickness and flow resistivity on the acoustic insulation. The ANN was trained in a supervised scheme using the Backpropagation algorithm optimized by the Levenberg-Marquardt Algorithm in a Multilayer Perceptron. In addition, to evaluate the variable significance was proposed a modification of the Profile Method, used in six different topologies trained each one independently 100 times. The Linear Regression (LR) served as benchmark analytical method. The results showed confidently that the flow resistivity, thickness, and density in decreasing order of importance to get high acoustical insulation. It also showed that the non-linearity is strongly linked to the system response, leading to the LR a poor performance. Therefore, the proposed methodology was in accordance with the literature and can be extrapolated to other fields where similar analyzes are required.