Fault detection and diagnosis using two-stage attention-based variational LSTM in electrolytic copper manufacturing process

In this study, a two-stage attention-based variational long short-term memory (LSTM) that allows fault detection and diagnosis in electrolytic copper manufacturing processes is proposed. As the surface quality of electrolytic copper determines the yield and quality of the product, an automated surface inspection (ASI) system has been introduced at various electrolytic copper production sites. Using the ASI system, the number of bumps on the electrolytic copper surface was converted into numerical data and the product was graded based on the data. However, when ASI is performed, it is difficult to accurately identify the variables that cause low-grade electrolytic copper. This is the first study to utilize feature data from an ASI system for fault detection and diagnosis in an electrolytic copper manufacturing process. Using a two-stage attention structure, important input and temporal features were extracted even though the noise ratio was high. In addition, the generalization performance was improved using a drop connect and variational LSTM, which can solve the overfitting problem. Experimental results using real-world data from the copper production process showed that the proposed model outperformed other time-series classification models. In addition, the proposed method was proven successful in fault diagnosis using attention weights.