Hybrid neuro-fuzzy and recurrent network for sequential modeling

Abstract

Driving drowsiness is a major cause of road accidents, underscoring the need for reliable detection systems to enhance road safety. This paper introduces a novel hybrid architecture, adaptive neuro-fuzzy inference system (ANFIS) (EM)–LSTM, which integrates the complete fuzzy inference process with recurrent modeling, combining interpretability and sequential learning. In this design, the premise parameters are initialized and refined using a Gaussian mixture model with expectation–maximization (EM-GMM), while the consequent parameters are optimized independently during training. Each fuzzy rule generates an explicit output at the frame level, which is then organized into temporal windows and processed by a long short-term memory (LSTM) to capture both short- and long-term behavioral dependencies. Applied to driver drowsiness detection on the public NTHU-DDD dataset, the proposed model achieves substantial improvements: a 37% relative gain in accuracy and 45% in F1-score compared to standalone ANFIS, as well as a 10% accuracy and 14% F1-score improvement over standalone LSTM. These results confirm the effectiveness and novelty of coupling fuzzy inference with sequential modeling, demonstrating its potential for deployment in safety-critical applications such as intelligent transportation systems.