Operation and Consideration of a Pipe Corrosion Inspection System Based on Human-in-the-Loop Machine Learning

The aim of this study is to improve the efficiency of external corrosion inspection of pipes in chemical plants. Currently, the preferred method involves manual inspection of images of corroded pipes; however, this places significant workload on human experts owing to the large number of required images. Furthermore, visual assessment of corrosion levels is prone to subjective errors. To address these issues, we developed an AI (artificial intelligence)-based corrosion-diagnosis system (AI corrosion-diagnosis system) and implemented it in a factory. The proposed system architecture was based on HITL (human-in-the-loop) ML (machine learning) [1]. To overcome the difficulty of developing a highly accurate ML model during the PoC (proof-of-concept) stage, the system relies on cooperation between humans and the ML model, utilizing human expertise during operation. For instance, if the accuracy of the ML model was initially 60% during the development stage, a cooperative approach would be adopted during the operational stage, with humans supplementing the remaining 40% accuracy. The implemented system’s ML model achieved a recall rate of approximately 70%. The system’s implementation not only contributed to the efficiency of operations by supporting diagnosis through the ML model but also facilitated the transition to systematic data management, resulting in an overall workload reduction of approximately 50%. The operation based on HITL was demonstrated to be a crucial element for achieving efficient system operation through the collaboration of humans and ML models, even when the initial accuracy of the ML model was low. Future efforts will focus on improving the detection of corrosion at elevated locations by considering using video cameras to capture pipe images. The goal is to reduce the workload for inspectors and enhance the quality of inspections by identifying corrosion locations using ML models.