RandomAD: A Random Kernel-Based Anomaly Detector for Time Series

Abstract

Time series anomaly detection is a critical task with a wide range of applications including industrial monitoring, financial fraud detection, and medical diagnostics. Among existing methods, C $^{22}$ 22 MP represents the state-of-the-art by combining Matrix Profile with catch22, a hand-crafted feature set, to enhance anomaly detection performance. However, catch22 features are limited in their ability to capture a full range of temporal characteristics in time series data. Recent advances in random convolutional kernel methods, such as the ROCKET family, have demonstrated strong performance in time series classification and clustering tasks. In this work, we propose RandomAD, a semi-supervised anomaly detection approach that leverages thousands of random convolutional kernels to extract a rich set of features. Our method adopts MiniRocket’s random kernel generation strategy to produce a large pool of kernels with randomly initialized weights based on the training data. To address the lack of labeled anomalies in the semi-supervised setting, we introduce a kernel selection mechanism to retain only the most informative kernels. Additionally, we incorporate a multi-window selection strategy with an anomaly filtering module to optimize both window size and detection results. Through extensive experiments on the benchmark datasets, we demonstrate that RandomAD consistently outperforms existing state-of-the-art methods.