In the domain of Inertial Confinement Fusion (ICF), ensuring the surface quality of target shells is critically demanding, as process anomalies can lead to significant losses of both materials and time. This paper presents a scalable anomaly detection methodology tailored specifically for the specialized task of batch polishing these shells. Our approach utilizes an autoencoder model trained on normal process data, applying an Exponentially Weighted Moving Average (EWMA) control chart to the model's reconstruction loss to detect anomalies. This methodology effectively handles large datasets and meets the rigorous surface quality standards required to maximize ignition yield in ICF operations.We validate our methodology on datasets from two distinct batch polishing experiments, H200 and H202, which encompass multiple hours of operation. The paper also details the tuning of hyperparameters for the EWMA control chart and includes both an ablation study and a comparative analysis with other anomaly detection methods. Our results demonstrate that this methodology outperforms existing approaches by promptly detecting subtle anomalies with minimal delay and low rates of false positives. Specifically, it achieves an average detection delay of 1.3 seconds for the H200 dataset and 19.5 seconds for the H202 dataset, thereby contributing significantly to the advancement of efficient and effective anomaly detection in the ICF domain.