Multi-Modal GAN-Based Anomaly Detection for Signal Reliability Assessment in Fabric-Integrated Smart Textiles

Smart textiles require reliable physiological sensing despite signal degradation caused by fabric deformation, material fatigue, and unstable textile–skin interfaces. This study presents a multi-modal GAN-based anomaly detection framework for signal reliability assessment using acceleration (ACC), electrodermal activity (EDA), and heart rate (HR). Controlled injection of baseline drift, amplitude scaling, and signal dropout generates normal/anomalous labels for supervised training. Temporal encoding and cross-modal attention distinguish textile-mimicking anomalies from physiologically plausible variations. On the 36-subject PhysioNet dataset, the framework achieves an F1-score of 0.9197 and exceeds the evaluated single-modal models by more than 35%. Using PhysioNet-trained weights without fine-tuning, zero-shot evaluation on the textile-integrated WWBS Metrics dataset achieves an F1-score of 0.8723 with ACC and derived HR. These results demonstrate cross-dataset transfer under the controlled synthetic-fault protocol; validation using physically induced textile faults remains necessary.