베이지안 필터를 활용한 시계열 변화점 탐지 기법
Change Point Detection using Bayesian Filter
1
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This paper presents a novel Bayesian filter-based method for change point detection (CPD) in time series, targeting robotic systems operating in noisy and dynamic environments. Unlike traditional CPD methods such as PELT, Binary Segmentation, Bottom-Up, and Window-based approaches that often miss subtle or gradual transitions, the proposed method probabilistically models steady and transient states and evaluates their relative dominance using a log-posterior ratio. The method enhances signal changes by applying gradient computation, anomaly scoring based on median absolute deviation (MAD), and adaptive thresholding. Final detections are refined through clustering to suppress false positives. Extensive experiments on both synthetic (Blocks, Bumps) and real-world datasets (NAB EC2, ECG5000) confirm the method’s superior F1-score and robustness compared to baseline methods. Moreover, its lightweight and real-time-capable design enables effective deployment in embedded robotic systems.
Abstract
This paper presents a novel Bayesian filter-based method for change point detection (CPD) in time series, targeting robotic systems operating in noisy and dynamic environments. Unlike traditional CPD methods such as PELT, Binary Segmentation, Bottom-Up, and Window-based approaches that often miss subtle or gradual transitions, the proposed method probabilistically models steady and transient states and evaluates their relative dominance using a log-posterior ratio. The method enhances signal changes by applying gradient computation, anomaly scoring based on median absolute deviation (MAD), and adaptive thresholding. Final detections are refined through clustering to suppress false positives. Extensive experiments on both synthetic (Blocks, Bumps) and real-world datasets (NAB EC2, ECG5000) confirm the method’s superior F1-score and robustness compared to baseline methods. Moreover, its lightweight and real-time-capable design enables effective deployment in embedded robotic systems.