Скачать или смотреть Anomaly Detection Using Self-Supervised Learning | Interactive Tutorial with Dr. Spencer Bialek

Discover how self-supervised learning is transforming underwater acoustic monitoring. In this interactive tutorial, Dr. Spencer Bialek presents a novel machine-learning framework designed to detect instrument-related anomalies in hydrophone data collected from Ocean Networks Canada (ONC).

This method tackles a major challenge in long-term ocean observing: identifying technical malfunctions and data quality issues in massive acoustic datasets where manual review is impossible and labelled failures are rare. The system focuses on common instrumentation problems—including data gaps, sensitivity drift, tonal artifacts, and other hardware-related anomalies that affect data integrity.

Using a two-stage machine learning process, hydrophone audio is converted into spectrograms for analysis.
🔹 Stage 1 – Self-supervised pretraining: Learns spatial and temporal patterns of clean hydrophone signals using pretext tasks such as patch localization and reconstruction.
🔹 Stage 2 – Supervised fine-tuning: Uses a small labelled dataset of known anomalies to classify whether a spectrogram contains an instrumentation issue.

Early results show high detection accuracy, strong performance on subtle or previously unseen anomalies, and excellent generalization—making this approach ideal for automated quality control.

This research provides a scalable, low-label solution for ensuring the quality, reliability, and scientific value of ONC’s underwater acoustic data across its national hydrophone network.

Learn more about ONC: https://www.oceannetworks.ca/