Fleet learning-based fault detection and diagnosis in the open world

Abstract

Fault detection and diagnosis (FDD) aims to identify faults and pinpoint fault types for engineering systems using sensor-based monitoring data. Modern FDD techniques have demonstrated superior capability and benefit in improving the safety, reliability, and efficiency of industrial systems or processes, such as automotive systems, manufacturing processes, wind turbines, and chemical engineering processes. Compared to traditional physics-based FDD methods, data-driven FDD methods using machine learning techniques have been rapidly growing recently due to the increasing complexity of the system and the size of the data. Challenges occurred when facing limited data and variation coverage in single unit monitoring, obstructing the effectiveness of implementing FDD in real-world applications. We envision an immense potential of expanding and elevating FDD performance and capability from small sample-based FDD approaches to a large-scale fleet FDD, which takes advantage of rich data and information from the whole fleet. We developed a data-driven fleet FDD schema that integrates the procedure of labeling, fault detection and isolation, and novelty detection, aiming to provide an industry-level FDD solution that addresses a bunch of realistic challenges including cold start, limited model knowledge, expensive expert involvement, emerging fault types, varying operating conditions, etc. Specifically, we propose a novel data-driven fleet FDD schema to (1) infer the machine health states; (2) identify novelties automatically; (3) advance the FDD model continuously upon the arrival of new data; (4) reduce human labeling efforts. The proposed schema consists of incremental fleet learning with novelty detection capability and expert-in-the-loop active semi-supervised labeling. We evaluate the framework and algorithms using both synthetic data and real-world data for comparison studies and show their superior performance over various realistic scenarios. In summary, this dissertation develops a generic data-driven fleet FDD schema that utilizes large-scale asset monitoring data and advanced machine learning techniques and provides an industry-level FDD solution with better performance regarding effectiveness, robustness, cost-efficiency, and adaptability. --Author's abstract