TransformDAS: Mapping Φ-OTDR Signals to Riemannian Manifold for Robust Classification

Jiaju Kang, Puyu Han, Yang Chun, Xu Wang, Luqi Gong

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 (Φ-OTDR) used for pipeline intrusion detection, face two primary hurdles: structural complexity and domain instability. Most traditional AI models, including standard CNNs and Transformers, operate on the assumption that data resides in a "flat" Euclidean space. However, real-world industrial signals often exhibit hierarchical structures and non-linear dependencies that are better represented as "curved."

When these signals are forced into a flat model, critical geometric information is lost. Furthermore, environmental noise and sensor drift cause distribution shifts, making models that perform well in controlled environments fail in the field. RIE-SenseNet addresses these issues by abandoning traditional Euclidean math for the curved world of Riemannian manifolds.

Methods: The RIE-SenseNet Architecture

RIE-SenseNet introduces a geometry-aware approach to signal processing through three main innovations:

1. Hyperbolic Mapping via Möbius Operations

Instead of standard linear layers, the model uses Möbius scalar multiplication and addition to project sensor data into a Poincaré ball (a model of hyperbolic space). This preserves the latent hierarchical structures of the signals that Euclidean models typically squash.

Visualization of Möbius operations in the Poincaré ball model of hyperbolic space.

2. Curvature-aware Transformer

The backbone of the model is a Transformer where the attention mechanism is modified to calculate distances in hyperbolic space using the hyperbolic arccosine (arcosh). Crucially, the "curvature" of this space is a learnable parameter, allowing the model to adapt its geometry to the specific dataset during training.

3. Riemannian Manifold Data Augmentation

To solve data scarcity and improve robustness, the authors use Normalizing Flows to map signals into a latent manifold. By perturbing the data in this curved latent space and mapping it back, the model generates realistic synthetic samples that respect the physical constraints of the original signals. This ensures that counterfactual examples remain physically plausible, boosting generalization.

Experimental Results

The model was evaluated on the BJTU Public Dataset and a Custom Industrial Dataset featuring high-stakes fiber-optic sensing (10MS/s) of events like impact hammers, excavators, and engine idling.

Performance and Ablation

RIE-SenseNet achieved an overall F1-score of 92.8% with augmentation. In contrast, standard CNN and Transformer baselines struggled in complex environments, often yielding F1-scores in the ~63% range for difficult industrial signals.

Overall performance comparison of RIE-SenseNet against baseline models with and without manifold-based data augmentation.

Ablation studies confirmed the value of the geometric approach: removing the hyperbolic component caused a 2.5% drop in F1-score, while removing the manifold augmentation caused a 1% drop. The combination of these techniques accounts for a massive 30% gain over classical machine learning baselines like SVMs.

Conclusion

RIE-SenseNet demonstrates that "curvy" geometry is a powerful tool for robust industrial sensing. By leveraging learnable hyperbolic manifolds and geometry-consistent data augmentation, the model can represent complex signal relationships with significantly lower distortion than Euclidean models. Future work will explore locally adaptive curvature to further enhance performance across diverse and dynamic industrial environments.

Summary updated

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.
BibTeX
@misc{kang2025transformdas,
  author = {Jiaju Kang and Puyu Han and Yang Chun and Xu Wang and Luqi Gong},
  title = {{TransformDAS: Mapping Φ-OTDR Signals to Riemannian Manifold for Robust Classification}},
  year = {2025},
  eprint = {2502.02428},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2502.02428},
}
Figure