In order to better diagnose and analyze faults in diesel engine valve train, this paper first simulates four working conditions: normal valve clearance, small clearance, large clearance, and coupling of strut bending by adjusting valve clearance and replacing the tappet. Cylinder head vibration signals are collected at idle speed of 700 rpm; Then, the ITD algorithm is used to decompose the signal, extract marginal spectral features, and select 12 frequency domain features to construct a fault feature vector; finally Using ITD marginal spectrum and Mahalanobis distance for fault identification, the results showed that there were significant differences in ITD marginal spectrum under different operating conditions. Among the 12 test samples, there were 0 false positives, and the diagnostic accuracy reached 100%. Research has shown that the ITD marginal spectrum can effectively characterize the fault characteristics of diesel engine valve mechanisms, providing a new approach for fault diagnosis of non-stationary vibration signals.