No Scale Left Behind: Multi-Scale Autoencoder with Bi-directional Attention for Time Series Anomaly Detection

Problem

Existing time series anomaly detection (TSAD) methods are constrained to a single temporal granularity, which limits their ability to capture multi-scale interactions in time series data. This paper addresses this gap by proposing a novel architecture that can effectively leverage multiple scales for improved anomaly detection. The work is presented as a preprint and has not yet undergone peer review.

Method

The authors propose the Multi-Scale Autoencoder with Cross-Scale Attention (MSCAD) architecture. Key components include:

• Parallel Autoencoder Branches: The architecture consists of multiple autoencoder branches, each designed to process different patch sizes of the input time series data, allowing for the capture of various temporal features.
• Bidirectional Cross-Scale Attention Blocks: These symmetric attention blocks facilitate information exchange between the different scales, enhancing the model's ability to integrate multi-scale information effectively.
• Training Method: The model is trained using a semi-supervised approach, which allows it to leverage both labeled and unlabeled data during training.
• Data: The model is evaluated on the TSB-AD benchmark, which includes 40 datasets comprising 530 time series.

Results

The proposed MSCAD model demonstrates significant improvements over state-of-the-art methods in time series anomaly detection:

• VUS-PR (univariate split): Achieved a score of 0.57, representing a 9.6% improvement over the previous best.
• VUS-PR (multivariate split): Achieved a score of 0.47, indicating a 9.3% enhancement compared to existing methods.

Limitations

The authors do not report any limitations in their work. However, the absence of reported limitations may suggest a lack of comprehensive evaluation across diverse datasets or scenarios, which is a common concern in anomaly detection research.

Why it matters

The introduction of the MSCAD architecture has significant implications for the field of time series anomaly detection. By effectively capturing multi-scale interactions, this approach can lead to more robust detection of anomalies in complex time series data, which is critical for applications in finance, healthcare, and industrial monitoring. The semi-supervised training method also opens avenues for utilizing large amounts of unlabeled data, potentially improving model performance in real-world scenarios.