Fault Cause Assignment with Physics Informed Transfer Learning

• Published in: IFAC-PapersOnLine Vol. 54; no. 20; pp. 53 - 58
• Main Authors: Guc, Furkan, Chen, YangQuan
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 2021
• Subjects: Data-Driven Approaches; diagnosis; Dynamic Mode Decomposition; Fault detection; Health monitoring; Source Separation; Transfer Learning; Health monitoring; Source Separation; Data-Driven Approaches; Transfer Learning; Fault detection; diagnosis; Dynamic Mode Decomposition
• ISSN: 2405-8963; 2405-8963
• DOI: 10.1016/j.ifacol.2021.11.152

To maintain successful operation, the field of health monitoring, fault detection and diagnosis plays a key role. Within the scenarios of system faults, locating a fault in a complex system consisting different components are one of the key challenge in fault detection. In this context, diagnosis of system faults originated from actuator and sensor is addressed to perform fault source separation. Physics of the underlying dynamics are investigated using only input and output data streams along with a purely data-driven technique, dynamic mode decomposition with control (DMDc) without a need for system model. Then time-frequency representation of the dynamic modes are obtained using continuous wavelet transform (CWT) and utilized in deep convolutional neural network (DCNN) to classify three scenarios for the case study, namely; nominal, actuator bias and sensor bias fault scenarios. Instead of training a DCNN structure from scratch, GoogLeNet structure presented for image classification is utilized as a standard methodology of transfer learning process. Finally, results of the image classification methodology are presented.