Current assembly assistance systems use different methods for object detection. Deep learning methods occur, but are not elaborated in depth. For those methods, great amounts of individual training data are essential. The use of 3D data to generate synthetic training data is obvious, since this data is usually available for assembly processes. However, to guide through the entire assembly process not only the individual parts are to be detected, but also all intermediate steps. We present a system that uses the assembly sequence and the STEP file of the assembly as input to automatically generate synthetic training data as input for a convolutional neural network to identify the entire assembly process. By means of experimental validation it can be demonstrated, that domain randomization improves the results and that the developed system outperforms state of the art synthetic training data.