Early Time Series Anomaly Prediction with Multi-Objective Optimization

Anomaly prediction, aiming to predict abnormal events before occurrence, plays a key role in significantly reducing costs and minimizing potential threats to mechanical devices. Monitoring machines using fixed-length time windows faces challenges in accommodating the varying characteristics of anomaly events. The lengthy and imbalanced sequence data associated with anomaly events further complicates the resolution of these challenges. Additionally, the inherent trade-off between accurate prediction and timely alarm is a crucial concern, posing difficulties in decision-making. This study puts forward a novel framework for early anomaly prediction, named early time series Anomaly Prediction with Neighbor Over-sampling and Multi-objective Optimization (APNOMO), pronounced as 'abnormal'. The framework employs three key techniques: 1) sliding windows that divide long input sequences into segments for prediction at proper intervals, 2) over-sampling and proposed neighbor over-sampling that handle imbalanced data in a novel way, and 3) multi-objective optimization that searches optimal thresholds to balance accurately prediction and timely alarm the abnormal. Experiments on a real-world dataset demonstrate APNOMO's superior performance over some state-of-the-art designs, with higher recall, F1 score, and more suitable earliness. It can predict anomalies 0.78-3.56 hours in advance, showcasing excellent early anomaly prediction capabilities for enabling predictive maintenance.