The analysis and correct categorisation of software performance anomalies is a major challenge in current industrial Cyber-Physical Systems (CPS). The automated evaluation of runtime performance metrics provides an overview of the software behaviour of CPS and may allow discovering or even predicting the occurrence of software performance anomalies. We present an approach to automatically identify deviations from expected performance behaviour of software processes running in a distributed industrial CPS. Our approach consists of collecting process performance signatures, using regression modelling techniques. This involves determining per process signatures and signature violations, enabling system anomaly type detection and classification. We evaluate different classification algorithms and predict the type of system performance anomaly that a signature violation can provoke. We demonstrate in our experiment that our design is capable of detecting and classifying synthetically introduced performance anomalies to real execution tracing data from a real semiconductor photolithography machine. Initial results show that we can achieve up to 93% of accuracy when classifying performance anomalies.