DATA EMBEDDING SCHEME FOR SENSOR FUSION: AN APPLICATION IN AFFECTIVE COMPUTING

Psychological affect monitoring holds promise as a continuous health metric for both patients and clinicians, offering insights into mental well-being and user readiness. By leveraging multimodal physiological signals—including biological and behavioral expressions—affective states can be quantified through sensor data. However, unlike domains such as autonomous systems, continuous affective computing has seen limited recent progress in computational efficiency and system design. To address this gap, we propose a modality-agnostic, unified encoder architecture that enables efficient multimodal fusion through a shared latent space. Our approach employs compressed sensing and autoencoder-based latent fusion to extract and correlate features across diverse physiological inputs. By visualizing signals in the spectral domain and encoding them using a single vector-quantized variational autoencoder (VQ-VAE), we eliminate the need for modality-specific encoders. Experimental results demonstrate that our framework significantly reduces model computation and runtime—achieving up to 64% reduction in computational complexity and a 1.4× improvement in execution time—while maintaining competitive accuracy. This positions our method as a scalable, lightweight solution for real-time affective computing on resource-constrained devices.