The separation of combustion and mechanical contributions in diesel engine signals can be achieved by using Wiener filtering. The idea is to estimate in operation the transfer functions between cylinder pressure and response signals, accelerometers or microphones. This operation is particularly tricky because of the necessity to observe the response signals on very short time windows (a fraction of cycle) to isolate the contributions of each cylinder. Moreover, several mechanical excitations are occurring at the same time than the combustions, which cannot be removed by using windowing. These mechanical excitations are partly coherent with the combustion, generating a bias on estimated transfer functions. It has been shown in preceding works that a way of lowering significantly this bias is to work on the random parts of signals. The aim and the originality of this contribution is to show that the differences between transfer functions obtained from raw signals and from their random parts can be used for diagnosis purpose. This difference has indeed a physical interpretation, and can be used to improve our understanding concerning the source contributions and their transfer paths. These developments are illustrated by experimental data obtained on several diesel engines.