Argo data anomaly detection based on transformer and Fourier transform

Abstract

Argo data is multidimensional observational data from ocean floats, which has long been plagued by data anomalies. Currently, the anomaly detection problem in Argo data still faces many challenges, as anomalies may involve complex relationships between multiple variables, leading to suboptimal performance with traditional machine learning methods. Thanks to the advancement of deep learning, it has become the predominant methodology for anomaly detection, demonstrating notable performance. The Transformer model has shown significant potential in the field of data anomaly detection, with its core Self-Attention mechanism capable of learning relationships between variables. We introduce Fast Fourier Transform (FFT) into the Transformer model, enabling the model to better capture periodic patterns and complex relationships in multivariate data, learning normal data patterns to improve the method for Argo data anomaly detection. Through experiments conducted on three public datasets and the Argo dataset, the enhanced model outperforms the original model in terms of performance. This also demonstrates the potential of FFT in multidimensional data anomaly detection, providing new insights into addressing anomaly detection challenges in real-world complex datasets.