Combining an optimisation-based frequency band identification method with historical data for novelty detection under time-varying operating conditions

Abstract

Incipient damage detection is important for critical rotating machines such as gearboxes found in the power generation, mining and aeronautical industries. However, the fault information frequently manifests in weak frequency bands in the vibration signals and the fault diagnosis process is impeded by time-varying operating conditions. Frequency band identification methods can be used to enhance the weak fault information in the vibration signal, however, this process is impeded by impulsive components unrelated to the component-of-interest and time-varying operating conditions. Hence, in this work, an optimisation-based frequency band identification method is developed to address these shortcomings. This method comprises of two steps; in the first step, a coarse informative frequency band procedure is used, whereafter a derivative-free optimisation algorithm is utilised to find the optimal frequency band for fault diagnosis. Since many rotating machines operate for long periods of time in a healthy condition, much healthy historical data are usually available when continuous monitoring is performed. Hence, this historical data are used with the proposed frequency band identification approach for automatic fault detection. The method is investigated on two experimental datasets acquired under time-varying operating conditions and compared to other existing approaches for fault diagnosis. The results indicate that the method is very capable of enhancing the fault information and can be used for automatic fault detection under time-varying operating conditions.