To improve the accuracy of fault diagnosis and reduce the evaluation time， a new method of engine fault diagnosis is proposed by analyzing the acoustic signal of turbocharged four-cylinder engine. The frequency analysis of wavelet packet transformation （WPT） is conducted， and the statistical characteristics are extracted from the high and low coefficients. Then， the standard model and the principal component analysis （PCA） are analyzed and compared in combination with the Bayesian optimization model by using the extracted features. The results show that， compared to the standard model， the latter two models have higher accuracy， precision， sensitivity， specificity and F1 values. On similar levels of accuracy， the PCA combined with the Bayesian optimization model takes about 20% less total evaluation time and 8-19% less test time compared to the Bayesian optimization model. The PCA combined with the Bayesian optimization model， which reduces the computational complexity and evaluation time， ensures good accuracy and provides reference value for the real-time engine fault diagnosis.