A Depth-Aware HGNN Method and Its Application in Anomaly Detection and Correction of Sparse Ocean Sensor Data

Area of Science:

• Oceanography
• Data Science
• Artificial Intelligence

Background:

• Ocean observation data is sparse due to the vast environment and limited sensors.
• Data exhibits discrete spatial distribution, discontinuous time, and vertical stratification.
• Existing methods include rule-based control, time series modeling, and traditional graph neural networks.

Purpose of the Study:

• To address the challenges of sparse ocean sensor data.
• To propose a novel method for anomaly detection and data correction.
• To improve the modeling of ocean vertical stratification features.

Main Methods:

• Developed a depth-aware heterogeneous spatiotemporal graph neural network (DAHSGNN).
• Integrated discrete data along the depth axis using local graph construction.
• Employed hierarchical feature engineering, Gaussian Hidden Markov Model for water layer segmentation, and a Transformer encoder for trend features.
• Utilized a bidirectional long short-term memory deep sequence encoder and a heterogeneous graph autoencoder for data reconstruction.

Main Results:

• DAHSGNN demonstrated good cross-variable generalization.
• Achieved higher reconstruction accuracy compared to baseline methods.
• Significantly improved anomaly detection performance.

Conclusions:

• DAHSGNN effectively models ocean vertical stratification features.
• The proposed method offers a significant advancement in processing sparse ocean sensor data.
• DAHSGNN provides a robust solution for anomaly detection and data correction in ocean observation.