To address the challenges of complex anomaly patterns and limited interpretability in time-series anomaly detection, this paper proposes a novel method that integrates model structure optimization with visual analysis. At the model level, a Deep Multi-Scale Fusion (DMSF) mechanism is introduced by embedding local convolutional branches into the Transformer encoder, effectively enhancing the model’s ability to capture fine-grained anomalies. To improve training stability, an Adversarial Association Difference Loss (AADL) is proposed, which mitigates KL divergence optimization issues via adversarial learning. Building upon these, we develop VAD-Vis (Visual Anomaly Diagnosis for Time Series), an interactive visual analysis system combining three core views: anomaly trend charts, control limit charts, and calendar heatmaps. The system supports multi-dimensional interpretation of model outputs from temporal, statistical, and periodic perspectives, and offers interactive features such as time filtering, channel switching, and threshold adjustment. Experiments on MSL, SMAP, SMD, and PSM datasets show that our approach outperforms state-of-the-art methods in F1 score, detection accuracy, and stability. The proposed method demonstrates strong interpretability, interactivity, and practical scalability for real-world time-series anomaly detection tasks.