Identification of sensors in smart manufacturing via mutually exclusive multiple time series classification

Industry 4.0 envisions full integration of manufacturing machines with decision support tools and ultimately with business processes. Our research focuses on methodologies to achieve such integration: production devices are equipped with sensors that provide streams of data; decision support systems collect and manipulate these data. Seamless integration requires digital interfaces in between them, mapping each stream a correct semantics. This may be untrivial in scenarios like retrofitting. Currently, the lack of consensus on standardization requires that such a mapping is performed manually. We propose a methodology to automate it. Such task requires the simultaneous classification of multiple time series, with the specific constraint that each of them must be assigned a different class. This latter feature makes existing methods unsuitable. We propose a pipeline of models, including: ensemble methods for single time series classification (supervised learning), exact matching optimization (mathematical programming), and adaptive aggregation to improve accuracy over time. We experiment on real data from the textile industry. Our pipeline shows to clearly outperform approaches from the literature. Our ensemble methods show to be competitive with benchmarks from the literature also on the single time-series classification task. The combined use of machine learning and mathematical programming allows to simultaneously exploit a data driven approach and obtain a-priori guarantees of compliance with the specific mutual exclusion constraint. Furthermore, the computing load of the single components of our pipeline keeps very limited, enabling the deployment in computing environments with low resources.