This code is the official PyTorch implementation of our Paper: TEDA: Time-Frequency Entropy-Aware Distribution Modeling for Time Series Anomaly Detection

Introduction

We proposed TEDA, a time-frequency distribution-aware method for unsupervised time series anomaly detection. Unlike reconstruction-based methods that mainly rely on reconstruction errors, TEDA models the latent distribution of normal time series. It introduces learnable distributional centroids and uses optimal transport to measure the deviation between test samples and normal patterns. TEDA contains two complementary branches. The patch-based temporal branch captures local temporal distributions, while the frequency-domain branch models long-term periodicity, trends, and global variations. We further design a training-inference differentiated OT strategy, where entropy regularization preserves smooth transport plans during training, and entropy-free sharp assignment produces more discriminative anomaly scores during inference. Extensive experiments show that TEDA achieves strong performance against competitive baselines. Ablation results verify the effectiveness of the key designs. Visualization results further show that TEDA assigns high scores to true anomalies and low scores to complex normal patterns.