Time Series Transformation into Images for Production State Recognition

In the industry, there is a need to automatically identify specific production states of an injection molding process. One reason for this is the necessity to distinguish cycles in which something was produced from those in which nothing was produced. Monitoring the actual runtime of an injection molding machine in this way could ensure efficiency and cost reduction. A sensor placed on the injection molding machine records 3D acceleration sensor data during a production process. In this thesis, time intervals of the acceleration sensor data are transformed into images using various image transformation methods. These resulting images are then used to train a neural network, allowing the model to assign new time intervals to individual production states. A significant advantage of using image transformation methods over classical feature analysis of time series is that images can compactly represent complex relationships. Furthermore, the use of deep learning enables the interpretation of large amounts of data in minimal time. The goal of this thesis is, on the one hand, to determine which transformation method, in conjunction with a neural network, is most suitable for recognizing production states on a single machine. On the other hand, the developed model should also be capable of recognizing production states in data from new machines. To achieve this, the search for the best generally applicable transformation method is conducted using a total of 15 datasets representing different machine and production variants.