Acoustic behavior prediction for low-frequency sound quality based on finite element method and artificial neural network

In this paper, a hybrid approach called FEM-ANN model is proposed by combining the finite element method (FEM) and artificial neural network (ANN) to predict the acoustic behavior of an auditory system. Based on the scanned point cloud data, the three-dimensional numerical models of the external auditory canal, tympanic membrane and middle ear are established by using the reverse prototyping technology, as are the FEM models. Setting the interior noises of the vehicle as excitations, the assembled FEM model is used to calculate the responses of the stapes footplate. According to the auditory perception characteristics of the human, a modified one-third octave filter bank is designed to calculate the vibration energies of stapes footplate in the critical bands, and thereby an energy-based feature matrix is established. Further, the sound quality (SQ) indices of interior noises, such as A-weighted sound pressure level (SPL), loudness and sharpness are calculated. By considering the extracted feature matrices as inputs and the calculated SQ indices as outputs, a three-layer ANN model with the radial basis function (RBF) is established for mapping the stapes footplate vibration to the human auditory perception. Verifications show that, the simulated result from the FEM model is consistent with that of the classical Ferris' model. The error percentages of A-weighted SPL, loudness and sharpness predicted by the FEM-ANN are all less than 5%, which suggests that the FEM-ANN model is accurate and effective for SQ evaluation of a low-frequency sound. The proposed hybrid approach can be used to simulate the acoustic behavior of an auditory system, which helps in revealing the mechanism of human auditory perception.