Abstract
Industrial sensor networks produce complex signals with nonlinear structure and shifting distributions. We propose RIE-SenseNet, a novel geometry-aware Transformer model that embeds sensor data in a Riemannian manifold to tackle these challenges. By leveraging hyperbolic geometry for sequence modeling and introducing a manifold-based augmentation technique, RIE-SenseNet preserves sensor signal structure and generates realistic synthetic samples. Experiments show RIE-SenseNet achieves >90% F1-score, far surpassing CNN and Transformer baselines. These results illustrate the benefit of combining non-Euclidean feature representations with geometry-consistent data augmentation for robust pattern recognition in industrial sensing.
Introduction
Industrial sensor networks, such as distributed fiber-optic sensors used for pipeline monitoring, operate in high-noise environments characterized by significant structural complexity and domain instability. Traditional Euclidean models, including standard Convolutional Neural Networks (CNNs) and Transformers, often fail to capture the deep, nonlinear dependencies inherent in these signals. Furthermore, the scarcity of high-quality labeled data for industrial faults limits the effectiveness of purely data-driven approaches. RIE-SenseNet addresses these issues by shifting the geometric paradigm from "flat" Euclidean space to curved Riemannian manifolds.
Methodology
Hyperbolic Mapping and Möbius Operations
RIE-SenseNet represents industrial data within hyperbolic space, specifically utilizing the Poincaré ball model. Unlike standard linear algebra, which can push data points outside the valid manifold, RIE-SenseNet employs Möbius operations. Möbius addition and scalar multiplication ensure that mathematical transformations preserve the geometric integrity and hierarchical structure of the sensor data.
Hyperbolic Attention Mechanism
The core of the Transformer architecture—the attention mechanism—is redesigned for non-Euclidean geometry. Instead of traditional dot-product similarity, RIE-SenseNet calculates feature weights using hyperbolic distances (arccosine). This allows the model to better capture the latent hierarchical relationships often found in complex sensor sequences. Additionally, the curvature of the manifold is treated as a learnable parameter, allowing the model to adapt its internal geometry to specific sensor hardware characteristics.
Riemannian Manifold Data Augmentation
To mitigate data scarcity, the framework includes a geometry-consistent augmentation pipeline. Using Normalizing Flows, raw signals are mapped into a latent manifold space where structurally faithful perturbations are applied. Mapping these back to the signal domain creates realistic synthetic samples that respect the physical constraints of the original sensors, providing a more robust alternative to simple noise injection.
Experimental Results
The model was evaluated on a public fiber-optic vibration dataset and a custom industrial dataset featuring events like impact hammers and engine idling. RIE-SenseNet demonstrated a substantial performance leap over traditional and deep learning baselines.
Key findings include:
- Superior Accuracy: RIE-SenseNet achieved a 92.8% F1-score on the industrial dataset, whereas standard CNNs and Transformers reached only approximately 63%.
- Ablation Insights: Removing hyperbolic geometry resulted in a ~2.5% drop in F1-score, while removing manifold augmentation led to a further 1% decrease. The synergy of both components is critical for bridging the performance gap.
- Robustness: The geometry-consistent augmentation significantly improved the model's ability to generalize to new environments by generating physically plausible counterfactual examples.
Conclusion
RIE-SenseNet demonstrates that incorporating Riemannian geometry into deep learning architectures provides a powerful tool for industrial signal processing. By respecting the underlying "shape" of sensor data and utilizing manifold-aware augmentation, the model achieves state-of-the-art performance in challenging, real-world conditions. Future work will explore non-constant curvature manifolds and optimization for edge-device deployment.
Summary updated