Obtaining the training data required for supervised training of neural networks is a significant challenge in various industrial applications. The manual annotation of this data is often time-consuming and requires trained personnel. This was a particular challenge for CHEKKER, which is developing a quality control tool for the precast concrete industry. The quality control includes, among other things, the inspection of built-in parts using neural networks to classify objects. This thesis presents a method to synthetically generate a training data set from 3D models of different objects. The advantage of this approach is that a large number of annotated training examples with different conditions, such as lighting conditions or backgrounds, can be generated without much additional effort. To ensure that the trained model generalizes to unpredictable input data, this thesis analyzes different image augmentation techniques and investigates how large the proportion of real data in the training dataset should be. Since CHEKKER uses stereo cameras in its setup and the spatial information of the objects should be preserved, the neural network receives the gray scale image and the depth image calculated from the stereo images as input. This thesis analyzes the generalization ability of the classification model based on synthetic training data. In addition, suitable augmentation methods are evaluated and the relationship between synthetic data and real data in training is examined for an optimal result.