Towards Near-Real-Time Anomaly Detection in Distributed Acoustic Sensing Using Convolutional Autoencoders

Distributed Acoustic Sensing (DAS) systems generate continuous high-resolution measurements along fibre-optic cables, enabling persistent monitoring of marine environments, vessel traffic, and subsea infrastructure. However, the extreme data volumes produced by modern DAS interrogators, sometimes several terabytes per day, pose major challenges for long-term storage, data transmission, and real-time analysis. This study investigates the use of a convolutional autoencoder (CAE) for near-real-time anomaly detection and intelligent data reduction on submarine DAS cables, with a focus on on-edge deployment for operational monitoring systems.

Using data from the Great Belt submarine fibre-optic cable (7250 channels sampled at 800 Hz and a spatial resolution of 4.085m), we trained a lightweight convolutional autoencoders on normal background behaviour to identify anomalous acoustic events through reconstruction error analysis. The idea is that the model learns the normal representation of the data, i.e. background noise and uses that knowledge to detect signals not normally present in the data – transient or non-stationary signals. Two preprocessing approaches were evaluated: a signed logarithmic compression (log1p) and a robust per-channel z-score normalization based on the median absolute deviation (MAD). While both approaches successfully detected signals generated by vessel activity, vehicles and transient acoustic events, they exhibited complementary behaviour. The log1p normalization produced highly sensitive automated detections with stable thresholds, whereas the robust z-score normalization preserved stronger visual contrast for human interpretation of anomalies.

The proposed framework is specifically designed for deployment directly at the acquisition site or on-edge hardware at the DAS interrogator. Instead of storing continuous raw strain-rate data, only anomaly signals, metadata, and selected event segments are stored and saved. This enables a substantial reduction in storage requirements from several terabytes per day to approximately 50 GB/day, depending on cable environment and vessel traffic density. Heavily trafficked marine areas naturally produce larger event volumes, whereas quieter cables yield significantly lower storage demands. The signals are further fused through morphological processing, making the signals more coherent and easier to extract.

The results demonstrate that CAE-based AI models can provide reliable near-real-time detection of signals of interest while dramatically reducing data storage demand. Such approaches represent an important step toward scalable operational DAS systems capable of continuous long-duration monitoring without the need for extensive storage infrastructure. The work highlights the potential of combining unsupervised learning, adaptive thresholding, and on-edge computing to enable practical real-time DAS monitoring for marine surveillance, infrastructure monitoring, and environmental sensing applications.
However, some signal types, particularly low-frequency seismic events and ocean wakes, remain difficult to detect and extract using the proposed method, as the CAE is less sensitive to slowly varying temporal features and long-duration low-frequency signals. A possible improvement would be to incorporate explicit frequency-domain information or multi-scale temporal feature extraction into the AI model. This is left for future work