CSCAD: Modeling cross-scale sequence correlations for multivariate time series anomaly detection

Abstract

Current anomaly detection methods struggle to adequately model the complex attribute correlations in multivariate time series and often overlook the heteroskedasticity within the series, resulting in the misidentification of low-amplitude noise and false alarms. This paper proposes Modeling Cross-Scale Sequence Correlations for Multivariate Time Series Anomaly Detection (CSCAD), a novel unsupervised anomaly detection method that models attribute correlations across time scales by constructing cross-scale splicing representations and multiscale interactive convolution. Additionally, the weights across time scales are adaptively adjusted to suppress noise interference and enhance the heterogeneous correlation representation by combining sequence heteroskedasticity with the attention mechanism. Inspired by Kolmogorov–Arnold networks (KANs), adaptive activation functions are introduced to enhance the model’s ability to capture complex temporal patterns. Detection experiments based on reconstruction error demonstrate that CSCAD improves the F1 score by 1.1% and recall by 2.14% compared to 19 baseline methods across five real datasets, validating its effectiveness in anomaly detection tasks.