Abstract: The engine abnormal sound signal is proved to be non-stationary and carries intense background noise. In view of these characteristics, a denoising method using multilevel threshold based on the analysis of the autocorrelation of detailed coefficients is proposed. This method use discrete wavelet transform technique to decompose the signal into approximations and details. The autocorrelation sequences of the detailed coefficients are determined. According to whether the sequence reflects the autocorrelation of white noise, the threshold is determined. The time-frequency map of the denoised signal is then drawn through continuous wavelet transform. By combining the features of time domain and frequency domain, the faults can be classified. In the experimental research a simulated signal model is introduced as an example. Piston cylinder knocking and crank bearing knock. Which stand for familiar engine abnormal sounds are compared and analyzed. The result proves that the signal to noise ratio is increased and the higher- frequency useful signal is recovered. The time-frequency map can display the features corresponding to time and frequency of fault signal which offers a practical strategy for diagnosis.

Key words: Abnormal sound fault, Autocorrelation sequence, Engine, Multi-level threshold, Time-frequency map, Wavelet transform