Adaptive Multimodal LSTM with Online Learning for Evolving IoT Data Streams

Abstract

The Internet of Things (IoT) uses networked devices, dispersed sensors, and cameras to create huge, diverse data streams.Concept drift, in which the underlying data distribution shifts over time, is frequently caused by the non-stationary and multimodal character of these streams.Static machine learning models, based on fixed data distributions, reduce forecast accuracy and system reliability since they are unable to adapt to such changes.This paper proposes an Adaptive Multimodal Long Short-Term Memory (AM-LSTM) architecture to address these challenges by combining modality-specific temporal modelling, attention-based dynamic fusion, and drift-aware online learning.An attention mechanism adaptively weights informative streams to mitigate the impact of noisy or missing input, while specialist LSTM encoders capture the temporal correlations of each modality.Concept drift is detected using a sliding-window error monitoring technique, and adaptive learning rate adjustment and selective retraining are started when significant distributional changes occur.The proposed system is tested under synthetic drift conditions using the Edge-IoT and UNSW-NB15 benchmark datasets.Experimental results demonstrate that AM-LSTM achieves 88.7% accuracy and an F1-score of 0.85, adapting to drift within 620 samples while maintaining an average update latency of 47 ms per batch.Compared with static and existing adaptive baselines, the proposed approach provides improved robustness, faster drift adaptation, and computational efficiency suitable for real-time IoT environments.