<p>Anomaly detection for industrial cyber-physical systems (ICPSs) methods including data-driven approaches typically rely on multivariate time series data with complete features, which may not always be available, especially during cyberattacks or equipment failures. In response to this challenge, this paper introduces a data-augmented robust anomaly detection approach tailored for scenarios with incomplete feature data. Central to this approach is the data augmentation. It not only amplifies available data but also simulates conditions with incomplete features, training the model with heightened adaptability and resilience. Furthermore, the model incorporates robust feature learning by utilizing feature masking and shuffling. In this way, during the training stage, the input of the detection model can be made to have dynamic random abnormal feature data, and the model is forced to predict the original normal state. This helps to enhance the model’s ability to distinguish abnormal data under the condition of incomplete features. Additionally, an integrated memory-augmented component is incorporated to construct the masking and shuffling memory-augmented autoencoder (MSMA). This integration facilitates the more robust learning of patterns, thereby further improving anomaly detection accuracy. Ultimately, anomalies are detected through the analysis of prediction errors. Experimental studies were conducted on a petrochemical fractionation unit simulation testbed and the public benchmark dataset BATADAL to demonstrate the proposed approach.</p>