Rethinking Time Series Anomaly Detection: A Scalable Transformer-Based Framework for Large Contexts

Multivariate time series data is ubiquitous in various domains, such as finance, healthcare, and industrial systems. Identifying anomalous patterns within these time series is crucial for early warning systems, predictive maintenance, and fraud detection. While deep learning approaches have shown promise in anomaly detection, existing methods often struggle with handling large input windows and capturing long-range dependencies. We propose a novel anomaly detection framework based on a Swin Transformer U-Net architecture to address these limitations. By leveraging hierarchical attention mechanisms, our model can efficiently process large input windows, enabling it to capture long-range temporal dependencies and identify anomalies with distant temporal relationships. This capability is crucial for detecting subtle anomalies that may not be apparent within short time windows. We evaluate our model on six benchmark datasets, demonstrating superior performance in detecting anomalies, particularly when compared to state-of-the-art methods.