An Early Classification Approach for Improving Structural Rotor Fault Diagnosis

Abstract

Artificial intelligence (AI)-based rotating machinery fault diagnosis has extreme importance in the industrial automation and control systems since rotating machinery constitutes approximately 40% of the overall machinery in the industry. However, the majority of AI-based solutions in rotor faults diagnosis are in an experimental stage due to: 1) inadequate and unrealistic faulty data; 2) lack of opportunities to utilize domain-specific fault features; and 3) limitations in employing deep learning and advanced learning strategies. Moreover, structural rotor fault (SRF) is one of the critical but least addressed faults in rotor faults diagnosis even though it is the root cause of the majority of rotating machinery issues. Hence, we develop an SRF diagnosis framework, which addresses the issues of industrial data acquisition by creating a subsampled data set incorporating distinctive frequency components (DFC). The data scarcity and imbalance problems are handled through an augmentation method using soft-dynamic time warping (soft-DTW), enhanced by fault information content (FIC)-based weighing scheme. At the fault classification phase, we proposed an early classification (EC) approach for SRF that predicts the faults with an acceptable tradeoff between earliness and accuracy. In this line, first, a sequential deep learning classifier is developed by considering accuracy only as an objective. Then, early decision policy is defined by taking accuracy and earliness into account. The model demonstrated exceptional performance over state-of-The-Art methods in SRF diagnosis and achieved 99.5% accuracy with 14% earliness on the Meggitt testbed data set and 98.32% accuracy with 55.68% earliness on the MaFaulDa data set. © 1963-2012 IEEE.