Improving Manufacturing Production and Reliability

Correcting problems in manufacturing operations is harder than it looks: a symptom such as a temperature fluctuation is usually linked to several underlying root causes, and narrow tests routinely mistake the symptom for the source. This whitepaper - authored by Michael Guilfoyle, Director of Research at ARC Advisory Group, and published by Integration Objects - makes the case for manufacturing reliability analytics built on a wider data set, modern analytics methods and automated corrective action.

The opening argument is diagnostic. Alarm management and similar operational processes are designed to return operations to a normal state, after which downstream tests attempt to explain the symptoms. Those tests fail for three recurring reasons: they are run manually by overloaded operators, introducing delays in which the event recurs; they often validate the problem but link it to a symptom or the wrong fault, so the real cause is never corrected and the problem returns; and they are too narrow in scope, with quality teams overlooking reliability contributors and operations overlooking process abuse that rule- or threshold-based monitoring cannot detect.

The remedy starts with expanding the pool of data analysed: troubleshooting documentation, maintenance histories, expert observations, safety protocols and standards, real-time and historic event and performance data, and failure codes, so that more combinations can be examined and symptoms separated from multivariate root causes, including inputs such as higher-level supply chain issues. The paper is candid about the obstacles: data arriving in formats from fully structured to unstructured, stranded across unintegrated engineering, quality, maintenance, operations, safety and planning systems, and presentation that overwhelms operators rather than supporting them.

Three steps are then described: applying an intelligent integration framework with device and data discovery, universal connectors, automated mapping and semantic modelling to replace traditional ETL; applying analytics via a modern software engine, including neural-net machine learning pattern recognition and statistical models with visualisation in operational context; and building a knowledge base that supports prescriptive corrective action through automated alerts, designed workflows and best-practice libraries.

A refinery case study covers KnowledgeNet® Analytics applied to Reid vapour pressure analysis on a continuous catalytic reforming unit, combining unsupervised machine learning with a rules engine to predict quality and asset performance gaps and estimate remaining useful life of the CCR reboiler.