OFC2025|Early Earthquake Detection Using Fiber Networks and Transfer Learning

Introduction

Early earthquake detection systems are crucial for mitigating the devastating humanitarian and economic impacts of seismic events, especially in densely populated areas. This paper explores an innovative approach utilizing existing terrestrial fiber networks as sensor grids for earthquake detection, combined with machine learning techniques and distributed polarization sensing technology [1].

Background and Challenges

The development of earthquake early warning systems (EEWSs) dates back to the 1960s, pioneered by Japan with the UeEDAS system for railway alerts. Traditional systems utilize seismic stations to detect P-waves (primary waves)—which arrive tens of seconds before the more destructive surface waves—allowing warnings to be issued prior to severe shaking.

However, the global density of seismic stations remains low, resulting in insufficient spatial sampling of seismic activity. Even highly developed countries face economic challenges in deploying and maintaining denser seismic station networks, prompting researchers to explore alternative methods leveraging existing infrastructure.

Innovative Earthquake Detection Method

The method presented here uses terrestrial communication fiber networks as sensor and epicenter localization grids for early earthquake detection. This machine-learning-driven approach employs distributed polarization sensing to detect P-waves across extensive geographical areas without requiring dedicated dark fiber or expensive equipment, distinguishing it from other fiber sensing technologies such as distributed acoustic sensing (DAS) and interferometric methods.

The primary advantage of this method is its cost-effectiveness and scalability by leveraging existing telecom infrastructure. Previous tests demonstrated that machine learning models accurately detected P-waves from a real magnitude 4.3 earthquake in Modena, Italy, providing a 21-second early warning window for nearby urban areas, 35 seconds for more distant regions, and up to 57 seconds for the furthest locations.

Experimental Validation Setup

The experimental setup replicated previously generated polarization state (SOP) Stoke parameters using a Python-based waveplate model. The system employed a tunable laser operating at a wavelength of 1550 nm with an output power of 6dBm.

Figure 1: Experimental setup comparing simulated, back-to-back (B2B) simulation, and propagated B2B Stokes parameters generated by strain values from an M4.3 earthquake. The optical signal was input into an optical scrambler comprising seven waveplates with random orientations, adjusted according to voltages applied to each waveplate to match the target SOP.

The scrambled signal was captured by a polarimeter, which measured the SOP and provided feedback to the scrambler. The system included two paths: Path 1 (B2B) and Path 2 (38 km propagation through deployed fiber). For Path 2, identical voltage sets were applied to the scrambler's seven waveplates, replicating the B2B simulated Stokes parameters, allowing propagation under real conditions in a 38 km single-mode fiber deployed in Turin, Italy.

To reduce complexity and minimize computational time, researchers calculated the state of polarization angular speed (SOPAS) rather than the complete SOP. This calculation allowed machine learning models to detect invariant patterns from SOPAS evolution using a single parameter instead of three parameters (S1, S2, and S3).

Machine Learning Model Architecture and Training

The pretrained model employed a deep learning architecture combining long short-term memory (LSTM) layers and attention mechanisms. The SOPAS input was analyzed by four LSTM layers returning full output sequences to handle temporal dependencies, enhancing the model’s capability to identify complex patterns and optimize sequence representations.

The attention mechanism dynamically weighted the importance of each timestep in the LSTM output sequence using a dot-product operation after generating attention probabilities with a softmax activation. This enabled the model to focus on the most informative parts of the sequence, distinguishing different seismic waves (P-waves, S-waves, and surface waves).

The model was initially trained on simulated SOPAS datasets to learn complex patterns induced by seismic events. Transfer learning was then applied through fine-tuning the pretrained model on a smaller SOPAS dataset propagated through a 38 km fiber cable, enabling the model to adapt to real-world conditions with minimal additional training.

Figure 2: Three key outcomes: comparison of B2B and propagated B2B SOPAS through 38 km fiber (left panel), confusion matrix for seismic event classification on propagated simulation (center panel), and ML model's ability to detect multiple seismic events on propagated simulation (right panel).

Results and Performance

Despite differences between simulated B2B and propagated B2B SOPAS, the model achieved outstanding performance. For simulated propagation data, it attained 98% accuracy in P-wave detection, an F1 score of 98%, a recall rate of 97%, and a precision of 98%.

The confusion matrix (Figure 2) illustrates high accuracy in detecting various wave types, including "no earthquake," "P-wave," "S-wave," and "surface wave." Out of 2156 P-wave events, the model correctly detected 2089, with minimal misclassifications: 19 instances were mislabeled as no earthquake, and 48 were misclassified as S-waves.

The model also performed exceptionally well in identifying surface waves, correctly detecting 11,017 out of 11,200 instances. Furthermore, the model demonstrated the capability to detect multiple seismic events, accurately identifying P-waves within one second.

Conclusion

This study demonstrates the effectiveness of a machine learning model in early earthquake detection by recognizing polarization change patterns despite propagation challenges through a 38 km fiber cable deployed in Turin, Italy. The results highlight the successful integration of transfer learning, where a model pretrained on simulated data was fine-tuned on experimental datasets, showcasing the potential for utilizing entire optical networks as sensor grids for real-time earthquake early detection.

Compared to traditional earthquake monitoring systems, this method offers significant advantages, including cost-effectiveness, scalability, and the capability to provide early warnings that could save lives and reduce economic damage from earthquakes. By repurposing existing telecom infrastructure, this approach presents a practical solution for the challenge of creating dense, homogeneous seismic monitoring networks.

Reference

[1] H. Awad, F. Usmani, S. Straullu, R. Bratovich, E. Virgillito, F. Aquilino, R. Proietti, and V. Curri, "Experimental Validation for Early Earthquake Detection Using Transfer Learning," in OFC 2025, 2025, pp. M1C.5.