PaAno+: Multiscale Encoding and Cross-Variable Attention for Time Series Anomaly Detection

Time-series anomaly detection is critical for industrial and medical monitoring, yet existing methods often struggle to balance high detection accuracy with the computational efficiency required for edge deployment. While Transformer-based models offer strong performance, they are often too resource-intensive. This paper introduces PaAno+, a lightweight, patch-based model designed to capture complex temporal and cross-variable dependencies without the overhead of large-scale architectures.

The authors propose a three-part framework: patch preprocessing, encoder training, and anomaly inference. The core innovation is the multiscale temporal encoder, which utilizes parallel convolutional kernels of varying sizes (3, 5, and 9) to capture hierarchical temporal features. To address multivariate dependencies, the model incorporates a cross-variable fusion attention module. Furthermore, the authors introduce a self-supervised pretext task—temporal patch-window sorting—which forces the model to learn the intrinsic structural properties of time series by recovering the original order of shuffled patches. The model is optimized using a combination of triplet loss and this temporal-ordering loss to create a discriminative embedding space.

Experimental results on the TSB-AD benchmark demonstrate that PaAno+ outperforms the original PaAno model and other state-of-the-art baselines across both univariate and multivariate tasks. The model shows significant improvements in metrics such as VUS-PR. By leveraging a compact network design, PaAno+ maintains high computational efficiency, making it suitable for real-time inference on resource-constrained hardware.

This research provides a practical solution for industries that require high-precision anomaly detection but operate under strict hardware limitations. By demonstrating that sophisticated feature engineering—specifically multiscale convolution and cross-variable attention—can rival the performance of massive foundation models, the paper offers a more sustainable path for deploying AI in critical, real-time monitoring environments.