Ranked Pairwise Reconstruction Difference Enabled Time Series Anomaly Detection

Unsupervised anomaly detection in multivariate time series, particularly reconstruction-based methods have obtained increasing attention recently. However, existing methods primarily focus on optimizing the reconstruction error of normal samples during training, neglecting the core goal of anomaly detection, namely, enhancing the reconstruction error separability between normal and anomalous samples. This limitation often results in an increased false detection rate, particularly when detecting subtle anomalies. To relieve this issue, we introduce a novel ranked pairwise reconstruction error comparison method, Ra-PIR, which begins by introducing noise at varying levels into normal samples to generate a set of samples with different levels of deviation. It then adopts a triplet structure to input paired samples into the reconstruction network for error computation. Building on this, we propose a ranking-constrained loss function minimizes reconstruction error and ensuring that the error ranking aligns with the deviation level, thereby improving the model's sensitivity to anomalous patterns. Experimental results conducted on public datasets show that compared with four existing anomaly detection methods, Ra-PIR improves the average detection accuracy. © 2025 IEEE.