A Robust Multivariate Time Series Classification Approach Based on Topological Data Analysis for Channel Fault Tolerance
Seong-Yeon Jeung, Jang-Woo Kwon
Abstract
In this study, we propose a robust artificial intelligence (AI) model for vibration monitoring of rotating equipment to support reliable operation across various industries, including manufacturing, power plants, and aerospace. The reliability and completeness of sensor data are essential for early detection of anomalies in equipment and for performing predictive maintenance. While AI-based predictive maintenance and condition-monitoring technologies have advanced in recent years, the issue of data loss caused by sensor failures remains a significant challenge that leads to performance degradation of AI models. In particular, for equipment utilizing multiple sensors, the complete loss of data from a single sensor significantly diminishes the predictive maintenance capability of AI models, thereby reducing their reliability. To address this issue, this study introduces topological data analysis (TDA) to develop a robust AI model. TDA analyzes the topological structure of sensor data to generate consistent feature vectors that capture the intrinsic characteristics of the data. This enables stable predictions even when certain channels of multi-sensor data are entirely missing. The proposed method demonstrates high performance resilience under conditions of partial sensor data loss, thereby contributing to enhanced reliability of AI-based predictive maintenance systems and the establishment of efficient maintenance strategies in the future.