Explainable Riemannian Manifold Learning for Application Scene Generalization With Distributed Acoustic Sensing System

Intrusion event detection in distributed fiber optic sensing systems plays a vital role in urban emergency management and public safety. However, current AI-based systems face high false positive and false negative rates, as well as limited generalizability, due to variations in dataset quality, processing methods, and model architectures. To address these issues, we propose the Riemannian-Transformer-OTDR (RTO) model, an explainable deep-learning approach based on Riemannian manifolds. By exploiting the manifold space, our method identifies stable latent representations of dynamic data distributions from similar events, embedding intrusion detection within this latent manifold. The RTO model accurately identifies identical event types across diverse scenarios, reducing misclassification rates and enhancing model stability. Field tests using a distributed fiber optic sensing system validated the effectiveness of RTO: it achieved 98.43% accuracy in a subway construction scenario and demonstrated its cross-scenario generalization ability by maintaining 94.21% accuracy in gas pipeline monitoring, with 80.79% accuracy in a 100% noisy environment. Additionally, signal structure analysis of misclassified samples revealed insights into the model’s learning process. Our analysis of performance degradation across scenarios and noise levels further clarifies the RTO mechanisms and explainability. Ultimately, the RTO enables precise, retraining-free intrusion event classification across multiple scenarios, offering a practical solution for rapid event detection deployment in various domains.