Abstract

Seq2Seq is a machine learning method that allows to translate sequences into other sequences. This method has been tried in hybrid simulation of machine tools. The method has been used to generate time series of energy consumption of jobs from the corresponding numerical control code that runs on a machine tool. Seq2Seq suffers from various problems. Firstly, the creation of training data is costly. Secondly, standard Seq2Seq metrics only allow for the evaluation of a prediction of one timestamp at a time, not an entire time series. Thirdly, training metrics are failing when vanilla data is used, as two identical numerical control codes can result in deviating time series. This causes confusion for the model in the training loop, as it is not clear which time series should be considered correct.
Here we propose a holistic framework to all three problems, that contains synthetic data, additional metrics for time series and dynamic time warping.