Knowledge Integration in Neural Networks for E-Maintenance: Application to an Oil & Gas Plant

{The paper reports the results of the KINEMA project (Knowledge Integration in Neural networks for E-Maintenance), cofunded by the BI-REX Consortium within the Competence Center Program of MIMIT. The participants to the project include Bonfiglioli Riduttori S.p.A., Eni S.p.A., Aetna Group S.p.A., I.M.A. Industria Macchine Automatiche S.p.A., Marposs S.p.A., MindIT S.r.l., NIER Ingegneria S.p.A., ALASCOM srl, University of Bologna and the MechLab department of the University of Ferrara. The paper focuses on the results achieved for the case study of Eni. The content, opinions and data reflect the point of view of the authors and do not necessarily reflect those of the BI-REX Consortium or MIMIT. The use of data-driven methods promises great improvements in the maintenance policies of mechanical systems. However, most data-driven approaches do not take full advantage of the available knowledge, such as physics-based models, design models of the equipment, technical specifications and experience of technical personnel. The KINEMA project attempts to bridge this gap by using innovative, but still industry-ready, approaches so as to maximize exploitation of all knowledge from the supply chain. The goal is achieved through the definition of a simple, but effective, methodology to integrate heterogeneous models within a neural architecture. Special emphasis is put on integrating untrainable, black-box, models that could conceivably be provided by the manufacturer of an industrial component, together with more traditional information content (e.g. specification documents or manuals). Manufacturers would have in principle the opportunity to train or calibrate their model based on benchmark tests, over a range of operating conditions that may expand or complement those encountered in the industrial plant that is the component final destination. The methodology defines a standard mathematical interface for data-driven models, which is used to define a number of integration schemes for black-box (non-trainable, potentially non-differentiable) or white-box (trainable and differentiable) models. The considered use cases include tasks such as estimation of non-measurable quantities, anomaly detection and diagnosis, and prediction of hazardous conditions. The integration schemes rely on cascade integration, data generation, and probabilistic reasoning to achieve better predictive performance, improve robustness, or enable new estimation or diagnostic tasks. The approach has been demonstrated by developing a diagnostic system on an industrial use case: a multicomponent system including the separation and stabilization units within an Oil \\& Gas plant operated by Eni. We rely on a state-of-the art commercial simulator as a proxy for testbench experiments, which are then used to investigate potential uses for externally provided black-box models. The paper will describe: - The methodology developed within the KINEMA project, capable of building the complete model of the equipment through the composition of sub-models, including non-trainable one possibly provided by external sources.- The implementation of the methodology on the separation – stabilization line of an Oil \\& Gas plant operated by Eni. }