A New Preprocessing Approach to Reduce Computational Complexity for Time Series Forecasting with Neuronal Networks: Temporal Resolution Warping

In modern manufacturing environments huge amount of sensor data from machines have to be analyzed in the time domain. Applications for process control are to observe the stability, to predict behavior of mechanical components or to detect abnormal behavior of the manufacturing process. For that, time series forecasting is important to make manual or automated decisions. In this paper, we introduce a preprocess method which we call “temporal resolution warping” (TRW). It is used for signal pre-and post-processing before and after applying the neural network. Thus, the computation complexity of the used network is reduced by compressing the time series in a certain way. We will show the computation reduction capability of our approach. For verification of our approach feed forward and convolution neural networks with residual layers are used to forecast reference time series of different applications. We will demonstrate that the training is speed up more than 26% with our pre- and post-processing technique.