A Rapid Learning Model based on Selected Frequency Range Spectral Subtraction for the Data-Driven Fault Diagnosis of Manufacturing Systems

Abstract

Limited amount of training dataset is often a critical problem in developing a diagnosis model or computational intelligence of mechanical failures. In this paper, we propose a Selected Frequency Range Critical Information Map (SFCIM)-based data-driven diagnosis method for detecting mechanical system failure with a relatively small amount of available dataset. The algorithm determines whether there is a failure with reference to normal condition signals using the spectral subtraction method. In this process, time-domain synchronization and time-frequency representation are applied. In addition, the dominant frequency range is selected based on Fisher’s scores for the more efficient calculation. We designed this algorithm to provide users with time-frequency domain information about system failure through SFCIM instead of the only simple decision of diagnosis. The usefulness of our method has been checked with the following problems of the small amount of available dataset: diagnoses of (1) the IMS bearing faults, (2) input gear faults of a manipulator driving system, and (3) a driving system of an industrial robot by load current signals. From these case studies, we also confirmed that the proposed method was practical and useful for fault diagnosis regardless of signal types, such as stationary, non-stationary, accelerometer-based vibration, or current signal.