Federated Machine Learning Architecture for Energy-Efficient Industrial Applications

Abstract

Due to the rise of new information and communication technologies manufacturing companies have access to huge amounts of power consumption data which are measured by sensors and processed by information systems. One of the most promising applications of extracting value out of the collected data is the detection of anomalies in process data from industrial machines and equipment. Many research and industry use cases apply machine learning (ML) techniques for anomaly detection. These techniques enable manufacturing companies to optimize their manufacturing processes but also to be more energy efficient and therefore have an impact for sustainable manufacturing. Most of the ML applications use central server infrastructures for data collection from different sources to process and analyse it for further usage. Nevertheless, privacy concerns and security risks motivate manufacturers to store the collected sensitive data from the production line locally. Therefore, suppliers of industrial machines (e.g. robots, machine tools) do not have the possibility, to store and analyse the data in the cloud, where data from all the machines of the supplier in different companies could be analysed and used for ML applications. One of the new paradigm shifts in ML is the concept of federated learning (FL) which enables local devices to use ML without sending data to a central server. This paper introduces an architecture for using the concepts of FL in manufacturing processes enabling machine suppliers to use ML for optimizing machine processes in a collaborative manner. Therefore, the more general federated learning concept is extended for industrial machinery and equipment using the industrial communication framework OPC-UA. Our architecture is tested and validated by using an industrial dataset of different compressors’ power consumption.