A Multi-Level Multi-Sensor Data Fusion Framework using Hybrid Machine Learning for Knowledge Mining
TL;DR
A novel hybrid framework titled AMF-TGNet (Attention-based Multi-Level Fusion with Temporal Graph Network) under a Multi-Level Multi-Sensor Data Fusion architecture that enhances knowledge mining performance by effectively modelling cross-sensor correlations and temporal dynamics while minimizing information loss is proposed.
Abstract
The rapid proliferation of Internet of Things (IoT) devices, wearable technologies, and smart healthcare systems has generated massive volumes of heterogeneous sensor data, creating the need for effective multi-sensor data fusion (MSDF) techniques for reliable human activity recognition (HAR). Existing multi-sensor data fusion approaches rely on single-level fusion or isolated models, failing to capture spatial–temporal and inter-sensor relationships while struggling with ambiguity, redundancy, and uncertainty. To address these limitations, this study proposes a novel hybrid framework titled AMF-TGNet (Attention-based Multi-Level Fusion with Temporal Graph Network) under a Multi-Level Multi-Sensor Data Fusion architecture. The proposed method integrates convolutional neural networks for spatial feature extraction, bidirectional LSTM for temporal dependency modelling, graph attention networks for inter-sensor relational learning, and an attention-driven adaptive weighting mechanism to prioritize informative sensors. A Bayesian inference layer is incorporated to reduce ambiguity and quantify predictive uncertainty. The framework is implemented using Python with TensorFlow and PyTorch libraries and evaluated on benchmark multi-sensor datasets to validate robustness and generalization capability. The proposed architecture enhances knowledge mining performance by effectively modelling cross-sensor correlations and temporal dynamics while minimizing information loss. Experiments are conducted on the HAR-PMD dataset, which contains multi-sensor smartphone and smartwatch recordings from 120 participants performing six mobility activities. The proposed AMF-TGNet achieves 98.74% accuracy, 98.20% precision, 98.50% recall, and 98.34% F1-score, outperforming CNN (95.4%) and LSTM (96.2%) by 2.5–3.3% and is validated using five-fold cross-validation. These results confirm its effectiveness for reliable activity recognition in wearable multi-sensor healthcare monitoring and assistive mobility systems.