Variational Autoencoder Based Approach for Imbalance Process Fault Detection

Process fault detection has drawn growing attention from various industrial sectors. Efficient detection of process faults can help to avoid abnormal event progression and reduce productivity loss. However, in the complex process system, there are uncontrollable factors or variables which are not captured by sensors that lead to problems with the high imbalance ratio and the curse of dimensionality. It becomes more challenging for many traditional fault detection methods to diagnose the faults in the process or capture the process's hidden characteristics when the data distribution is imbalanced. Therefore, motivated by deep generative models, we proposed a variational-autoencoder (VAE) based approach which can efficiently boost the fault detection performance from imbalanced process data. The proposed approach is highly suitable for dimension reduction and feature extraction of abnormal data: fault samples with new characteristics can be generated. The prediction accuracy evaluated by state-of-the-art classification algorithms can be improved significantly. We have tested our proposed approach using one real dataset collected from a packaging production line of semiconductor integrated circuits (PoSIC), and one public dataset describes a sample of pular candidates collected during High Time Resolution Universe Survey (HTRU2). Our experimental results demonstrate that the proposed approach can be well applied to imbalance process data and significantly improve prediction accuracy.