Anomalieerkennung im Energiebedarfsverlauf von Maschinen mittels Verfahren des maschinellen Lernens

The application of anomaly detection techniques has not been investigated much on energy consumption data of machines or motors. In this work required preprocessing methods and anomaly detection techniques for energy data and motor current data are presented. The preprocessing of energy consumption data recorded from a manufacturing machine requires the following steps: segmentation of the energy consumption data in product specific signals, clustering of those signals and ltering of autonomic overlaying signals. The method used for segmentation is based on thresholds. In order to arrange the signals according to their product type a clustering algorithm based on cross{correlation is implemented. These two steps are successfully tested on the given dataset. An approach to filter autonomous overlaying signals by means of median calculation was devised and tested for simulated data but proved to be unsatisfactory for the real data. The preprocessing of motor currents being directly recorded from a motor involves a method which has never been used before on this kind of data: the continuos wavelet transformation. This method is used to extract features and consequently classify motor current signals to detect anomalies in motors such as broken bars and end{ring connectors. Therefore, several machine learning techniques are evaluated for classifying. Of all used classifers support vector machines reached the highest accurarcy rate of 92.81% for all 21 classes