Using anomaly detection and classification to support failure recovery of plants

Failure recovery of machines is a great challenge for the plant operator. Usually, a stopped plant is quite costly and has to be put back into operation fast. In that case, experienced personnel are essential. Only their knowledge and broad understanding of the interaction of the various components allows a reliable identification of the underlying problem and enables a successful recovery. Digitization can help to address this problem. Knowledge can be stored and shared digitally across plants, locations and companies. In addition, an ever-increasing amount of machine and process data is digitally available which can further support the problem-solving process. We present a new approach to machine maintenance that relies on machine learning to help identify the underlying technical problem that led to the machines error. In contrast to predictive maintenance, our system is employed reactively and focusses on the identification and classification of singular failures rather than a predictive monitoring of the plants state. To support the operator in fault recovery, we use anomaly detection to identify the deviations in the data of the processes prior to the fault. This allows the system to generate hints as to where to look for the underlying cause of the error and helps the technician make an educated guess. Additionally, an error classification algorithm also allows troubleshooting information to be linked directly to the data. Thereby, when an error reoccurs, the knowledge gained in the first occurrence is immediately made available to the operator through information linked to the error class. In this article, we present the concept behind our anomaly detection system and the required software architecture to combine the different data sources. Furthermore, we discuss the results from an experiment that simulates an optical part inspection using a camera based system, where we introduce different errors into the process and are able to demonstrate the capability of our system to identify the correct source of the error.