Features Masked Auto-Encoder-Based Anomaly Detection in Process Industry

With the development of the modern process industry, accurate anomaly detection methods are increasingly needed. However, identifying anomalies from high-dimensional data continues to be a challenge for the process industry. In this work, a feature masked autoencoder (FMAE) method is proposed to meet this challenge. As a masked-reconstruction task, a high mask rate is first adopted to mask the feature of data. Then, the intrinsic information of unmasked features is extracted through the encoder and masked features are recovered through the decoder. This task forces the intrinsic information learned by the model to increase while alleviating the high-dimensional problem. Additionally, anomaly tends to be reconstructed to a normal by FMAE because only normal data is used for training. Thus, anomalies are detected and localized by computing input-output residuals. Finally, the effectiveness of the model is verified on the Lublin Sugar Factory dataset.