The health of rolling bearing plays an important part in the operation of rotating machinery like a gas turbine engine. Health monitoring and fault diagnosis of rolling bearings based on vibration signals have been through great development these years. But when sensors are set on the casing instead of the bearing pedestal, and the surrounding structure is very complex, the diagnosis problem becomes much more complicated, which brings more challenges to the signal processing. In this paper, a set of signal processing methods are used to enhance and extract the impact features from casing vibration signals, and to realize the detection of rolling bearing faults. A self-adaptive decomposition method called intrinsic time-scale decomposition (ITD) is applied to decompose the vibration signal into a series of proper rotation components and a monotonic trend, helping to extract dynamic features of the signal. Teager-Kaiser energy operator is a simple algorithm calculating the energy of a signal and is very sensitive to transient impact faults. As the fault feature transmitted to the casing is relatively week, the autoregressive model (AR) and minimum entropy deconvolution (MED) enhance the non-stationary impact components. Experiments are taken on the rotor-bearing-casing test rig with minor defects in the main shaft bearing. Testing on the casing vibration signal, this fault feature enhancing and extracting method shows its remarkable ability in rolling bearing fault diagnosis.