TinyML-Optimised Multimodal Lstm for Unsupervised Anomaly Detection in Smart Home IoT Systems

The rapid rise in IoT smart home devices generates extensive time-series datasets, while anomaly detection solutions remain largely dependent on cloud-based processing, resulting in latency, privacy concerns, and connectivity requirements. Existing approaches often rely on a single modality and insufficiently address concept drift and edge-device constraints. This study proposes a TinyML-optimised multimodal LSTM model for unsupervised anomaly detection in smart home applications deployed on a Raspberry Pi edge device. Temperature, humidity, and power consumption modalities were combined to capture cross-modal relationships. A simulated dataset represented normal diurnal and seasonal behaviour and incorporated three anomaly types: heat spikes, power outages, and midnight appliance usage. Model training used only normal-value sequences within a 250-timestep sliding window. TinyML optimisation reduced the stored LSTM model size to 196 KB, which represents the model storage size and should not be interpreted as total runtime memory or direct evidence of real-time or low-power operation. An online learning strategy was employed to address concept drift through periodic retraining. On the simulated test set, the model achieved 98.97% recall, 8.79% precision, 8.86% F1-score, 0.294 Matthews correlation coefficient (MCC), and an AUC-ROC of 0.9984. The confusion matrix comprised 83,416 true negatives, 3,994 false positives, 4 false negatives, and 385 true positives. Although the model detected nearly all injected anomalies, its low precision indicates a substantial false-alarm burden and limits claims regarding deployment readiness.