Abstract: Use the acoustic vibration signal for the engine fault diagnosis process, some fault excitations only express a strong response in the specific vibration on the engine surface, and the vibration measurement points require high requirements with contact measurement, which is difficult to achieve under some scenarios. Therefore, a diesel engine intake fault and gear fault diagnosis method were proposed using surface radiated sound as the medium and adaptive variational mode extraction（AVME） as the preprocessing method. Bench experiments were carried out under three fault conditions of a 6-cylinder in-line heavy-duty diesel engine, namely: air filter blockage, abnormal valve clearance and timing gear damage, and the engine surface acoustic signal under different fault degrees was obtained. Based on the improved AVME method, the optimal decomposition of the intrinsic mode function（IMF） of the acoustic signal was achieved. By calculating the mutual relationship between IMF and the original signal, the highly correlated IMF was extracted to constitute the classifier input. By AVME, the fault acoustic features were effectively enhanced, and input into the support vector machine model optimized by sparrow search algorithm（SSA-SVM）, and the collaborative optimization of feature parameters and model parameters can achieve better diagnosis accuracy. The experimental verification results show that without the need for testing in a semi-anechoic chamber, only a single-channel acoustic signal is used to diagnose three types of 11 degrees of the intake system and gear faults, the accuracy rate of the front-end acoustic and the top-side acoustic signals are the highest（98.89%） and the lowest（88.78%）, respectively. After using the front, top, and rear acoustic data, the diagnostic accuracy rate can reach 99.57%. The research results provide a reference for engine fault diagnosis based on non-contact measurement methods such as acoustic signals.