Leveraging the enterprise knowledge graph for predictive maintenance

Abstract

Knowledge is an important resource for manufacturing enterprises, hence the collection, storage and analysis of data is a key function within such firms. An important source of data is the Industrial Internet of Things (IIoT), where sensors embedded on industrial machines produce data that can provide insights about the operations and condition of the machines. This data brings about new challenges but also new opportunities. One such opportunity is Predictive Maintenance (PdM) that aims to reduce maintenance costs by maximising the use of the various machine parts and inventory whilst minimising unplanned machine outages. In PdM, machine failures are predicted by monitoring some machine health or performance indicators that come in the form of IIoT data. The challenge, however, is to collect and store the various IIoT data, that are heterogeneous in nature, and to exploit them for PdM. Research shows that Enterprise Knowledge Graphs (EKGs) are flexible data structures that are capable of integrating heterogeneous data, and can thus provide a solution to this challenge. In this dissertation we investigated the use of the EKG as an integration paradigm for the IIoT data generated by wire bonding machines, and as the foundations of a PdM framework for these machines. The use of the EKG for PdM is scarce in literature, which gave us motive to contribute another case study based on this coupling. We designed and built an ontology modelling the wire bonders, their states, sensors and the observations of these sensors. The ontology was then used to transform the IIoT data into an EKG. Machine Learning (ML) models were also trained to predict wire bonder faults upon the IIoT data. These were then integrated into a PdM framework that extracts IIoT data from the EKG to forecast possible faults of the wire bonders. The results obtained are promising and show that the ontology and the EKG were adequate for storing IIoT data, achieving an accuracy of 1.0. They also show that the PdM framework was able to predict wire bonder faults with an F1-score of 0.75 up to two hours in advance. The EKG served to integrate the disparate data sources that were needed for PdM, and to standardise the vocabulary of such data. This simplified the PdM framework that would have otherwise needed to cater for the different data structures and vocabularies when retrieving the IIoT data. Furthermore, the EKG was able to infer new knowledge through the ontological reasoning, thus completing the knowledge extracted from the wire bonders. These results also demonstrate that the EKG can be leveraged for PdM.