Adaptive Sequential Probability Framework for Ultra-Low Latency Fault Detection in High-Speed Engineering Systems

Abstract

Real-time fault detection in high-speed engineering systems such as CNC machining centres, UAVs, and wind turbines requires statistical methods capable of detecting distributional shifts within milliseconds. Traditional fixed-sample monitoring approaches often fail under concept drift, where system health distributions change due to wear, thermal stress, or structural degradation, leading to delayed fault detection. This paper introduces an Adaptive Sequential Probability (ASP) framework that combines Sequential Probability Ratio Test (SPRT) with a window-limited Cumulative Sum Control Chart (WL-CUSUM), enhanced by Kullback–Leibler Divergence divergence-based adaptive thresholding. By continuously updating post-change parameters within a sliding observation window, ASP achieves asymptotically optimal detection delay inversely proportional to the divergence between healthy and drifted states. The framework satisfies the Lorden Minimax Criterion while preserving predefined Type I and Type II error bounds under both abrupt and gradual drift conditions. Closed-form solutions for adaptive thresholds and optimal window size are derived. Simulations across representative engineering scenarios show detection delays below 12 samples at a 5