Distributed Acoustic Sensing: A cable-agnostic method for detecting & classifying signals on fibre optic cables

Distributed Acoustic Sensing (DAS) has seen an increase in attention in the last decade, offering the ability to convert existing fibre optic cables into dense networks of passive sensors. These cables, which span vast areas of the globe, hold immense potential for diverse sensing applications. Oil dumping, whale hunting, trawling near the cables are all huge issues to our environment and biodiversity. DAS offers 24/7 real-time sensor capabilities, which can aid in protecting these areas, by detecting signals from various sources. However, the enormous data volumes generated by DAS systems present significant challenges in terms of manual analysis, highlighting the need for an adaptable and scalable approach, capable of identifying and classifying signals of interest. 

This study introduces a new method, efficiently designed to address these challenges by detecting and classifying signals in DAS data. The proposed method enhances the detection of critical events such as earthquakes, marine mammal activity, and ship crossings, thereby expanding the scope of DAS applications.
The methodology should be cable-agnostic, and establishes a "normalcy model" that captures the typical data distribution over an extended period. Each channel along the fibre optic cable is modeled as a Gaussian distribution, representing its standard behavior. Incoming, unseen data is segmented and similarly treated as a Gaussian distribution. To detect deviations, the method uses the Kullback-Leibler (KL) divergence between the normalcy model and the observed data. A change is flagged when the divergence exceeds a threshold, which is determined empirically based on observed patterns in the data. By establishing a normalcy picture of the data, the model is not limited to only 1 specific fibre cable, but is adaptive to any. To classify signals, the method uses the spectral signatures unique to each event type, enabling automatic clustering based on the different signals. These clusters are validated using verified datasets. For example, ship signals are cross-referenced with Automatic Identification System (AIS) data and earthquake signals are compared against seismic databases. By incorporating these reference datasets, the system can reliably classify known signals and identify events of unknown origin for further analysis. 

Our results show that it is possible to not only detect signals in DAS data fast and efficiently, but also cluster the signals, through the spectral signature, into different origins of the sources.  We show that within the clusters is it possible to distinguish between different signals originating from the same source, e.g., differentiating between different ships, or earthquakes. This implies that we potentially are not only able to classify ships, but also identify which ship it is. This can tremendously enhance our capabilities in identifying and catching actors that dump oil or hunt whales in the ocean, so it can be stopped. The methodology can be used on any cable and has been shown to work on fibre cables in a very harsh arctic environment, with a lot of noise.