2025_programme: Automated Detection of Anomalous Hydrophone Data Using Self-Supervised Machine Learning

• Day: June 19, Thursday
    Location / Time: TBA at 15:30 - 16:30
• Last minutes changes: -
• Session: Poster session
  Organiser(s): N/A
  Chairperson(s): N/A
• Lecture: Automated Detection of Anomalous Hydrophone Data Using Self-Supervised Machine Learning
  Paper ID: 2317
  Author(s): Spencer Bialek, Drew Snauffer, Alex Slonimer, Herminio Foloni Neto
  Presenter: Spencer Bialek
  Abstract: This research presents a novel approach to detecting instrument-related anomalies in underwater acoustic data collected from Ocean Networks Canada's (ONC) extensive hydrophone network. As Canada's leading ocean observing organization, ONC is responsible for maintaining high data quality standards across its continuous ocean monitoring infrastructure. The method addresses the challenging problem of identifying technical malfunctions and data quality issues in long-term marine acoustic monitoring systems, where manual inspection is impractical and labeled examples of failures are scarce. Our approach focuses specifically on detecting common instrumentation issues such as data gaps, sensitivity fluctuations, unexpected tonal artifacts, and other hardware-related anomalies that can compromise data integrity.\n\nThe framework employs a two-stage machine learning process on hydrophone audio data transformed into spectrograms The initial self-supervised pretraining phase utilizes pretext tasks such as patch localization and reconstruction to capture spatial and temporal patterns of clean hydrophone signals without requiring labeled data. This step allows the model to learn robust feature representations of properly functioning instruments. The supervised fine-tuning phase then refines these representations using a limited labeled dataset of known anomalies to predict whether a given spectrogram is anomalous or not.\n\nInitial experimental results demonstrate the effectiveness of our approach, with strong performance across both detection accuracy and reliability metrics. The model shows particular strength in detecting subtle instrumentation issues that traditional visual quality control methods might miss. Importantly, the system maintains robust performance even when encountering previously unseen types of technical anomalies, suggesting strong generalization capabilities.\n\nThis work provides a practical solution for automated quality control in ONC's underwater acoustic monitoring systems, enabling rapid identification of instrumentation issues. The self-supervised approach significantly reduces the labeling burden while maintaining high detection accuracy, making it particularly valuable for managing ONC's large-scale hydrophone network and ensuring the integrity of data distributed to the scientific community.
• Corresponding author: Dr Spencer Bialek
  Affiliation: Ocean Networks Canada
  Country: Canada