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

A new lightweight model for time-series anomaly detection, called PaAno+, achieves state-of-the-art accuracy on both univariate and multivariate tasks, according to a paper submitted to arXiv on 18 Jun 2026 [1][2]. The model is designed for deployment on resource-limited terminals for real-time inference [2]. The paper, authored by Xiangguang Xiong, details a patch-oriented representation learning paradigm that addresses limitations in current approaches [2]. Transformer- and large-model-based detection methods incur excessive computational overhead, while existing lightweight alternatives often suffer from insufficient feature extraction and poor modeling of dependencies across multivariate variables [2]. PaAno+ uses a multiscale feature-extraction backbone built with convolutional kernels of differentiated receptive fields to capture hierarchical temporal characteristics, combined with cross-scale adaptive attention aggregation and residual connection optimization [2]. A cross-variable fusion attention module explicitly characterizes inter-variable correlations, helping the model identify anomalous patterns in complex operational conditions [2]. The model also employs a novel pretext task based on temporal patch-window sorting to uncover intrinsic structural properties of time series, and triplet loss is used to optimize the patch embedding space for enhanced feature discrimination [2]. Extensive experiments on the TSB-AD benchmark demonstrate that PaAno+ yields significant performance gains across evaluation metrics, including VUS-PR, relative to the original PaAno [2]. The compact network design achieves favorable computational efficiency, enabling deployment on terminals with limited resources [2]. The paper was submitted as a 3,659 KB PDF to arXiv, an open-access repository of electronic preprints that is not peer-reviewed but moderated before posting [1][6]. arXiv hosts papers in fields including computer science, physics, and mathematics, and as of November 2024 receives about 24,000 submissions per month [6]. The repository passed the two-million-article milestone by the end of 2021 [6]. The submission includes links to experimental community tools under the arXivLabs framework, which allows third-party collaborators to develop features on the article record page while adhering to values of openness, community, excellence, and user data privacy [4][5].