ABSTRACT Key Performance Indicator (KPI) anomaly detection is a crucial component of Artificial Intelligence for IT Operations (AIOps). In practical scenarios, service anomalies can lead to missing KPI values, which reduce the accuracy of anomaly detection. Moreover, existing methods frequently struggle to effectively detect subtle anomalies, thereby limiting detection performance. This study proposes a KPI anomaly detection framework that integrates missing value interpolation and adversarial training. First, a missing value interpolation model is built using an improved Transformer. The model is optimized via a mean absolute error loss constructed by randomly masking observations, while a diagonal mask in the self‐attention mechanism enhances interpolation accuracy. Second, a detection architecture is developed based on Variational Autoencoder‐Gated Recurrent Unit (VAE‐GRU) and adversarial training, which amplifies reconstruction error sensitivity to subtle anomalies. Finally, the Streaming Peaks Over Threshold (SPOT) algorithm is incorporated for adaptive thresholding based on the distribution of reconstruction errors. Experimental results demonstrate that the proposed framework achieves superior F1 scores and recall compared to existing methods on multiple benchmark datasets.