Paper Details
Abstract
Unsupervised time series anomaly detection is critical when labeled data are scarce. Although the Anomaly Transformer achieves strong results on long sequences, our analysis reveals significant degradation on constrained, low-dimensional datasets due to its reliance on a 1D-CNN extractor. To address this, we propose a hybrid encoder combining Temporal Convolutional Network (TCN) and Convolutional Neural Network (CNN) to jointly capture global and local temporal features. Replacing the original embedding with this encoder yields more expressive representations and improves anomaly detection without supervision. Experiments on UCR, ECG, and 2D-Gesture show consistent F1 improvements over the baseline, validating the robustness and generalization of our approach.