Exploring machine learning methods for the identification of production cycles and anomaly detection

This paper presents a distributed platform that collects and processes data streams from Industrial Internet of Things/Cyber–Physical Systems. First we focus on the platform’s architecture, which enables seamless integration with various data sources and supports near-real-time processing and the performance of Machine Learning based methods for detection and analysis. The platform design is optimized for both accuracy and scalability, allowing it to handle large volumes of data efficiently. We demonstrate the design and predictive performance of our platform and methods with an emphasis on accuracy and scalability in three distinct phases: production cycle detection; detected cycle analysis and identification; and anomaly detection and root cause analysis for predictive maintenance. Our results highlight the potential of our platform and the value it can bring to operational decision-making in an industrial setting. • Presenting a generic framework to collect data from a production environment • Deployment scenario for this framework in a production-like environment • Supporting predictive maintenance scenarios based on anomaly detection techniques • Comparative study for eight ML methods to detect anomalies and production cycles