Stress Classification using Vanilla Transformer Encoder on Multi-Channel Physiological Sensor Time Series

The study introduces an effective stress classification framework using the Vanilla Transformer encoder architecture to examine multi-channel physiological sensor time series data. The method is designed to tackle the issue of real-time psychological stress identification using continuous, multimodal bio signals obtained from wearable sensors. The WESAD dataset serves as the principal source of physiological recordings, including electrodermal activity,
heart rate, respiration, accelerometer data, and skin temperature signals. The signals undergo preprocessing into
time-synchronized segments, which are then input into a
positional embedding module prior to classification using selfattention layers in the Transformer encoder. In contrast to traditional recurrent models, the attention-based approach in this architecture enables the model to dynamically concentrate on the most relevant temporal patterns related to stress reactions. The suggested technique attains a classification accuracy of 97.80 percent, with a precision of 95.88 percent, a recall of 93 percent, and an F1-score of 94.42 percent. The findings underscore the Transformer model's potential to discern subtle physiological differences and enhance multi-class
stress detection, facilitating the development of adaptive,
individualized mental health monitoring systems in wearable
and edge computing contexts.