Deep Learning Experiments with Skewed Data for Defect Prediction in Plastic Injection Molding

摘要

In this work, we investigate the possibility of predicting defects in plastic injection molding by learning predictive models from time-series process data collected from a molding machine. The model of our choice is an RNN (recursive neural network) model using LSTM (long short-term memory) units in its hidden layer. This model is well known as a deep learning model specialized for processing time-series data. Since defects are rare and thus the dataset is highly skewed, we try to achieve a high average recall rather than a high classification accuracy. We give some initial results of experiments and an outlook to the direction of our future works.