<p>Weakly supervised video anomaly detection (WSVAD) seeks to identify temporally localized anomalous events in untrimmed surveillance videos utilizing only video-level labels. Although existing methods employing multiple instance learning (MIL) have demonstrated promising results, they frequently encounter the challenge of gradually suppressing weak anomalies during training. This suppression can result in missed detections in videos that contain multiple anomalies of varying intensities. To address this issue, we propose a novel framework that integrates dynamic feature screening network (DFS-Net) with consistency constraints to maintain the representation of weak anomalies under weak supervision. Specifically, we introduce a dynamic feature screening strategy capable of identifying video segments exhibiting abnormal scores within a relative range centered around the average score value for the entire video. Subsequently, these segments are compared with those displaying both the highest and lowest outliers using cosine similarity to establish semantic consistency regularization. After weighting and encoding the selected features, a transformer module is applied to the resulting representations to capture long-range temporal dependencies. A multi-objective loss function is employed to jointly optimize the basic score, enhanced score, and semantic relationships among features. Extensive experiments show that DFS-Net achieves state-of-the-art results on ShanghaiTech with an AUC of 98.58% and the Industrial Steelmaking Anomaly dataset with an AUC of 98.30%, while delivering highly competitive performance on UCF-Crime with an AUC of 86.52%.</p>