<p>Maintaining constant mold level variations during the continuous casting process is essential to guarantee the effectiveness and quality of steel production. An unsupervised deep learning-based mold level anomaly detection (MLAD) model for real-time monitoring of mold level fluctuations under varying operating conditions was proposed. The MLAD framework employs a two-stage encoder-decoder structure with adversarial training to accurately reconstruct time-series mold level data. In the first stage, the model learns long-term trends by reconstructing input windows, while in the second stage, it employs reconstruction errors as focus scores to capture short-term anomaly patterns. A transformer-based architecture, incorporating multi-head attention mechanisms and positional encoding, enables MLAD to capture both local and global temporal dependencies. In addition, a novel multi-threshold strategy, based on extreme value theory, is implemented to enhance the model’s ability and to adapt to varying operating conditions, including startup, steady-state, and shutdown phases. The model was validated with over 240-h real data from a steel factory. The results demonstrate its superior performance in anomaly detection compared to popular methods, with a precision of 0.9937, recall of 0.9932, and a low false alarm rate of 0.0038. MLAD represents a significant advancement in the detection of nonlinear and nonstationary anomalies in industrial processes, offering an efficient solution for smart manufacturing systems. Therefore, the established model could be used for online anomaly detection of mold level with real-time data.</p>