Data-driven prognostics for critical electronic assemblies and electromechanical components

Abstract

The industrial digitalisation enables the adoption of robust, data-driven maintenance strategies that increase safety and reliability of critical assets such as electronics. And yet, an implementation of data-driven methods which primarily address the industrialisation of diagnostic and prognostic strategies is opposed by various, application specific challenges. This thesis collates such restricting factors encountered within the oil and gas industry, in particular for the critical electrical systems and components in upstream deep drilling tools. A fleet-level, tuned machine learning approach is presented that classifies the operational state (no-failure/ failure) of downhole tool printed circuit board assemblies. It supports maintenance decision making under varying levels of failure costs and fleet reliability scenarios. Applied within a maintenance scheme it has the potential to minimise non-productive time while increasing operational reliability. Likewise, a tailored and efficient deep learning data pipeline is proposed for a component-level forecast of the end of life of electromagnetic relays. It is evaluated using high resolution life-cycle data which has been collected as a part of this thesis. In combination with a failure analysis, the proposed method improves the prognostics capabilities compared to traditional methods which have been proposed so far in order to assess the operational health of electromagnetic relays. Two case studies underpin the need for tailored prognostic methods in order to provide viable solutions that can de-risk deep drilling operations. In consequence, the proposed approaches alleviate the pressure on current maintenance strategies which can no longer meet the stringent reliability requirements of upstream assets.