Prediction of Production Performance for Tapioca Industry using LSTM neural network

Cyber-physical system (CPS) is integrated between the virtual and physical worlds that uses service-oriented architecture (SOA) to manage service objects such as production, quality control, and maintenance engineering. Moreover, the CPS will be more effective by using the predictive ability to integrate some service object of CPS. This paper presents an intelligence function on the top CPS platform which will be the prediction of production performance to detect yield malfunction by using affect parameter with a starch yield of each production units such as production rate, starch content, loss in the process, a temperature of the flash dryer, and moisture content of starch for adjust the production process before an actual failure situation. The prediction model uses historian data from the factory with data from a mass balance model simulation and uses a Long-Short Term Memory (LSTM) processing to get the yield predicted results to indicate the trend of change in yield or irregularities that may occur, which will be the information for the design or modification of the production plan or control the production to be appropriate before abnormalities in real situations. This model has an average accuracy of 90.55% of test dataset which is predicted 1 day in advance.