Fault detection in rotating machinery using spectral modeling

Fault detection and condition monitoring is crucial for a secure and economic operation of mechatronic systems such as rotating machinery. For this purpose sensors gather the physical condition of the machinery. This sensor signals are interpreted by the machineries control system which can trigger a reaction on the fault event. The global propagation of mechanical fault indicators over the casing make vibrational measurements ideal for an obtainment of the whole system. But due to the signal corruption in noisy environment it is a challenging task to process this vibrational data, especially when no explicit process knowledge is available. The process model free framework presented in this work takes advantage of specific fault signatures like frequency modulation. To identify this modulation a spectral prediction algorithm is proposed. Normalizing the prediction error of each spectral component adjust the algorithm to the machinery casing transfer characteristics. In this way also process noise can be suppressed effectively. These features are used to indicate changes in the process state and faults. In addition to the framework a sample implementation is presented and evaluated on centrifugal compressor and bearing datasets. The results show an advantage in detecting compressor surge or bearing fault in comparison to kurtosis statistics, especially in the case of a high noise level or weak fault signatures.