General pre-trained inertial signal feature extraction based on temporal memory fusion

Inertial sensors are widely used in smartphones, robotics, wearables, aerospace systems, and industrial automation. However, extracting universal features from inertial signals remains challenging. Inertial signal features are encoded in abstract, unreadable waveforms, lacking the visual intuitiveness of images, which makes semantic extraction difficult. The non-stationary nature and complex motion patterns further complicate the feature extraction process. Moreover, the lack of large-scale annotated inertial datasets limits deep learning models to learn universal features and generalize them across expansive applications of inertial sensors. To this end, we propose a Topology Guided Feature Extraction (TG-FE) approach for general inertial signal feature extraction. TG-FE fuses time-series information into graph representations, constructing a Memory Graph by emulating the complex network characteristics of human memory. Guided by small-world network principles, this graph integrates local and global information while sparsity constraints emphasize critical feature interactions. The Memory Graph preserves nonlinear relationships and higher-order dependencies, enabling the model to generalize across scenarios with minimal task-specific tuning. Furthermore, a Cross-Graph Feature Fusion mechanism integrates information across stacked TG-FE modules to enhance representation ability and ensure stable gradient flow. With self-supervised pre-training, the TG-FE modules require only minimal fine-tuning to adapt to various hardware configurations and task scenarios, consistently outperforming comparison methods across all evaluations. Compared to the current state-of-the-art method, our TG-FE achieves 11.7% and 20.0% error reduction in attitude and displacement estimation tasks. Notably, TG-FE achieves an order-of-magnitude advantage in stability evaluations, maintaining robust performance even under 20% noise conditions where competing methods degrade significantly. Overall, this work offers a solution for general inertial signal feature extraction and opens new avenues for applying graph-based deep learning to capture and represent sequential signal features.