Adaptive Multi-Scale Attention-Based LSTM Coupling for Early Detection

ICLR 2026 Conference Submission3454 Authors

Keywords: LSTM, Time Series Forecasting, Early Detection, Attention

TL;DR: We propose an adaptive, dual-path attention-coupled LSTM that fuses short-term motif detection with long-term trend modeling to enable real-time scenario recognition and prediction in automotive E/E systems.

Abstract: This paper introduces a novel adaptive, attention-coupled Long Short-Term Memory (LSTM) architecture developed specifically for real-time scenario recognition and prediction in complex automotive electrical/electronic (E/E) systems. Modern vehicles generate rapidly growing data streams from signals such as current, voltage, and temperature. We address this by monitoring critical signal patterns via a fused LSTM. The proposed dual-path methodology comprises a trend path for long-term pattern modeling and a motif path for short-term pattern recognition, coupled via a bidirectional, attention-based gating mechanism that enables dynamic information exchange. The outputs provide a reliable basis for initiating high-resolution data capture or adaptive system responses once a scenario is identified with high confidence. Experimental results demonstrate significant reductions in mean squared error compared to the individual values and interpretable attention weights that reveal information-exchange patterns. The proposed approach enables robust, noise-resilient forecasts and allows for efficient, data-driven development for future EE architectures.

Primary Area: learning on time series and dynamical systems

Submission Number: 3454