Abstract
In this paper, we present a new diagnostic method for automobile abnormal sound using CNN. The procedure of our method is 1) calculating the autoregressive model (AR model) coefficients from the abnormal sound by using the maximum entropy method; 2) constructing the CNN whose memory patte rns become standard abnormal sound patterns; 3) making the coefficients obtained as an initial pattern and recalling one from the memory patterns, and then obtaining a diagnosis result. By using our method, the influence of the noise occurring from other normal parts can be avoided. Therefore, the automobile abnormal sound can be diagnosed by using CNN even when conventional diagnoses cannot be used. The results demonstrate the advantages of our approach.