Time Series Features Extraction Versus LSTM for Manufacturing Processes Performance Prediction

In this article is addressed the complexity of predicting the performance of manufacturing processes in cyber-physical systems in cases when the products go through hundreds of operations and when the data that is recorded when the manufacturing processes are performed is not enough in order to make accurate predictions and thus to determine those operations that might lead to low performance results. This research challenge is approached by comparing two methods namely, one method based on highly scalable hypothesis tests and machine learning predictors and one method based on a Long Short-Term Memory Recurrent Neural Network (LSTM RNN). In addition to the critical comparison of the two approaches in terms of performance, this research work addresses challenges such as the determination of the best threshold for distinguishing between performant and unperformant processes, the identification of the most frequent patterns in unperformant processes and the consideration of several techniques for replacing the missing data given the complexity of manufacturing processes.