Regional adaptive moving average for robust anomaly detection in wiring harness manufacturing

Abstract

Wiring harnesses play a crucial role in automotive, railway, and electronic systems, with terminal crimping being a critical phase in their assembly. In this study, piezoelectric force sensors are used to monitor crimping waveforms generated during an industrial wiring-harness terminal crimping process. Recent studies have applied AI-based anomaly detection models that learn normal crimping patterns from early-stage production data. However, reliably detecting rare defects becomes increasingly challenging as waveform drift accumulates over long production runs, causing the distribution of newly acquired signals to diverge from that of the initial training data. Under these conditions, conventional AI-based monitoring and classical reference-updating methods, such as EMA-based schemes that uniformly smooth the entire signal, struggle to remain effective, particularly in the presence of extreme class imbalance. To address this limitation, we propose the Regional Adaptive Moving Average (RAMA), a reference-updating strategy that adaptively updates localized waveform regions using confidence-weighted adjustments, rather than applying uniform updates across the entire signal. This region-aware updating mechanism enables stable monitoring of localized drift while preserving sensitivity to true anomalies. The proposed method was validated using two controlled experimental datasets and one real industrial production dataset, with repeated evaluations against seven baseline algorithms. The results demonstrate that RAMA consistently achieves a superior precision-recall balance and maintains stable detection performance under progressive drift, highlighting its practical advantage for long-sequence monitoring in wiring-harness manufacturing.