Seq-TNT Based on Dynamic Attention and Context Awareness Within a Multimodal Sensor Framework in Intelligent Consumer Electronics

Accurate user behavior recognition remains a critical challenge in smart consumer electronics, with existing approaches suffering from three fundamental limitations: inadequate temporal dependency modeling, suboptimal cross-modal fusion strategies, and insufficient environmental adaptability. This study presents a breakthrough multimodal fusion framework that synergistically integrates inertial measurement units (IMUs) with radio frequency identification (RFID) through a revolutionary Sequence-Transformer in Transformer (Seq-TNT) architecture. This study presents three key innovations. First, it introduces a Dynamic Multi-Scale Attention (DMSA) mechanism that adapts focus across different temporal scales, from microseconds to seconds. Second, it proposes a Context-Aware Relative Position Encoding (CARPE) method that adjusts positional representations based on environmental context. Third, it develops a Cross-Modal Dynamic Fusion Transformer (CMDFT) that enables deep interaction and integration of IMU and RFID data. The dataset is derived from real user behavior records in simulated intelligent retail environments, including seven typical shopping behaviors, totaling 23324 RFID samples and 20023 IMU samples. The experimental design adopts 10-fold cross validation, comparing baseline methods including Vision Transformer-based Video TimeSformer, Video Swin Transformer, and traditional machine learning models. Experimental validation reveals four significant breakthroughs: Seq-TNT achieves 98.7% accuracy with 34% fewer parameters, demonstrates 5.2x faster inference speed, exhibits remarkable cross-domain generalization with only 2.1% accuracy drop in unseen environments, and achieves real-time processing capability with