Abstract
The Yerkes-Dodson law characterizes the inverted U-shaped curve between arousal and performance - as arousal increases, so does performance, but only up to an optimal point beyond which it starts to decrease again. This principle highlights the critical importance of predicting arousal in industrial human-machine interaction, where maintaining optimal arousal can not only enhance worker performance, but also safeguard their overall affective well-being. Driven by this understanding, this paper introduces a novel regression-based multimodal approach that utilizes peripheral physiological signals to predict user’s arousal. This allows for precise monitoring of arousal, enabling industrial human-machine systems to keep operators functioning at their personal peak - a balance of both productivity and user comfort. By employing a sliding window method for data augmentation, LSTM networks to learn temporal patterns, and multi-head self and cross-attention techniques to focus on relevant intra and inter-signal relationships, our model effectively processes heart rate, skin temperature and electrodermal activity to predict arousal. Our proposed model surpasses state-of-the-art performance for arousal prediction with an MSE of 0.099 and an MAE 0.214, demonstrating the effectiveness of our approach using solely peripheral signals.