Data Mining Approach for Device Detection using Power Signatures and Manufacturing Execution System Data

Abstract

Condition monitoring plays an important role in production systems, since a failure of a single device could block the whole production line. State-of-the-art technologies such as vibration monitoring, acoustic emission technique, shock pulse method or similar are often applied through a variety of sensors in order to monitor a big amount of different devices. However, such multi-sensor deployment requires not only high investment and high effort, but qualified personnel resources for the analysis of the various data, especially when it comes to a real-time condition monitoring. Additionally, retrofitting of plants and peripheral devices has its own obstacles, due to either limited resources, high complexity, or other diverse restrictions. To foster condition monitoring in industrial applications hardware-software combinations are needed which excel in low applications costs and reduced installation efforts. In this paper, an economic solution for condition monitoring is presented based on a central power measurement device, which can be applied with a 100 Hz resolution on a device group such as pumps or the exhaust system. Mainly, this paper deals with the detection of active individual devices in a group by their power signature. A spectral analysis is carried out to obtain the power signature of every device. Finally, the device detection is realised through supervised data mining models, which are trained with the data from a manufacturing execution system and the high-resolution power signal. This approach allows the detection of a specific device between similar ones, as for every device of interest an individual model is developed.