Cyber-physical systems are found in production and industrial systems, as well as critical infrastructures which play a crucial role in our society. The integration of standard computing devices and IP-based technology in cyber-physical systems increases the threat of cyber-attacks. Furthermore, traditional intrusion defense strategies are often not applicable in industrial environments. This paper focuses on the widely used Siemens S7 communication protocol and presents an approach to detect anomalies in network packets by training a model with neural networks and applying the model on current network traffic. In order to stay close to practice we built an experimental setup with industry controllers, sensors and actuators. To check the applicability of the model we launched supervised S7 protocol attacks against the setup. The results show that this approach can detect anomalous network packets with satisfactory accuracy.