Continuous Emission Monitoring Systems (CEMS) for stationary pollution sources serve as the core data backbone for precision pollution control and environmental law enforcement. However, traditional data quality control (QC) primarily relies on fixed-threshold screening and manual spot checks, making it highly challenging to effectively identify sophisticated anomalies and possible artificial interference. To address the limitations of conventional methods—namely, low recognition rates for sporadic equipment failures and emerging anomalous patterns, as well as a heavy reliance on manual labels—this paper constructs an automated quality control model for industrial waste gas online monitoring data based on unsupervised learning. Utilizing multi-month, long-term, high-frequency chronological monitoring data as a foundation, the model introduces a Temporal Convolutional Network Autoencoder (TCN-AE) to capture the chronological features and spatial correlations among multi-dimensional operational parameters and pollution factors. Furthermore, by integrating Extreme Value Theory (EVT), the model achieves dynamic, self-adaptive updates of anomaly activation thresholds. Experimental results demonstrate that, under a label-free premise, the model achieves a comprehensive F1-score of 0.91 in identifying data drift, constant anomalies, and sudden-blind failures. It establishes a closed-loop workflow spanning 'data cleaning, feature correlation, dynamic quality control, and automated warning,' thereby providing robust theoretical and technical support for intelligent environmental management.