Time Series Classification using Deep Learning for Process Planning: A Case from the Process Industry

Nijat Mehdiyev, Johannes Lahann, Andreas Emrich, David Enke, Peter Fettke, Peter Loos

In: (Hrsg.). Procedia Computer Science. Complex Adaptive Systems (CAS-2017) United States Seiten 242-249 114 2017.


Multivariate time series classification has been broadly applied in diverse domains over the past few decades. However, before applying the classification algorithms, the vast majority of current studies extract hand-engineered features that are assumed to detect local patterns in the time series. Therefore, the efficiency and precision of these classification approaches are heavily dependent on the quality of variables defined by domain experts. Recent improvements in the deep learning domain offer opportunities to avoid such intensive hand-crafted feature engineering which is particularly important for managing the processes based on time-series data obtained from various sensor networks. In our paper, we propose a framework to extract the features in an unsupervised (or self-supervised) manner using deep learning, particularly stacked LSTM Autoencoder Networks. The compressed representation of the time-series data obtained from LSTM Autoencoders is then provided to Deep Feedforward Neural Networks for classification. We apply the proposed framework on sensor time-series data from the process industry to detect the quality of the semi-finished products and accordingly predict the next production process step. To validate the efficiency of the proposed approach, we used real-world data from the steel industry.


Weitere Links

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence