Learning Time-Series Anomaly Generators from Synthetic Supervision

Synthetic anomalies are widely used to alleviate the scarcity of labeled anomalies in time-series anomaly detection. Existing anomaly injection methods use predefined transformations to construct injected anomalies, which may limit data diversity and fail to sufficiently capture the temporal characteristics of the target data. This work proposes a generative framework for multivariate time-series anomaly synthesis based on normal-pattern pre-training and anomaly-oriented fine-tuning. The generator is first pre-trained exclusively on normal time-series data to capture target-specific temporal patterns. Then, multi-type injected anomalies are constructed to provide *synthetic supervision*, and a dimension-adaptive reconstruction objective is introduced during fine-tuning to adjust relative reconstruction penalties according to channel-wise statistics. After fine-tuning, latent perturbations broaden the sampling region of the learned latent distribution, enabling more diverse anomaly generation. Unlike conventional anomaly injection, the injected anomalies are not directly used as the final generated anomalies, but instead guide the generator in learning heterogeneous deviations from normal temporal patterns. Experiments on five multivariate time-series benchmarks and three downstream anomaly detectors show consistent improvements, with average AUPR gains of 13.28%, 11.28%, and 16.85% for CARLA, COUTA, and DACAD, respectively. Representation analysis further shows that the generated anomalies achieve stronger separability from normal patterns and greater diversity than direct anomaly injection, supporting their effectiveness for downstream training.