A Study on Quality Prediction Failure Cause Analysis in Batch Process

Quality prediction plays a key role in all kinds of manufacturing. To predict quality products or forecasting a product quality, data should contain time series properties. However, these properties often fail to be contained in the data. This usually happens in many industrial manufacturing environments. For example, lot processing often fails because of many reasons: parts having different lot numbers can be mixed up in the following step; parts are missing, broken, etc; in some cases, the long term data are required but short-term data are only gathered. Without these properties, any machine learning technique should fail to predict data quality. In this paper, we have data collected in short-term period although the total samples are 3,000 and under a condition, we fail to predict a quality of a sample with various machine learning prediction techniques: Gaussian process regression method and auto-regressive integrated moving average. Although hyper-parameters are well-optimized and both predicted results are almost the same, the predicted ones are far from the ground truth. Through these experiments, we concluded that well-designed data collection is the most important in quality prediction.