'Beware the false positives as predictive maintenance evolves'
OPINION | Statistical models based on data science alone are not yet enough to give owners and operators the clarity they need to make the right decisions about wind turbine O&M, says Bruce Hall
With increasingly bold claims circulating about the capabilities of various predictive maintenance approaches, the industry needs to get to grips with the tools on offer and understand their limitations as well as their advantages.
Unplanned outages are a bane for every wind farm owner and operator, entailing the financial cost of repairs and maintenance, and the opportunity cost of revenues lost through asset downtime.
In an effort to reduce unnecessary O&M expenditure, the wind industry is steadily turning to innovative predictive maintenance solutions, which provide operators with a means of analysing data to pre-empt equipment failures and schedule maintenance at a time that minimises losses in production capacity and revenues.
Two distinct approaches have emerged to meet this requirement. The first builds on established engineering principles, analysing and predicting component failures using data sources that are targeted to the specifics of the turbine technology, such as vibration, inspection and lubrication data.
The second makes use of large statistical datasets that can be easier and cheaper to access, and applies processing power in the form of machine learning algorithms to detect anomalies and determine when failures might occur.
Both of these approaches are perfectly valid – but both have their limitations. The problem is that, with a flood of new players entering the market promising predictive maintenance solutions that are omniscient and driven by artificial intelligence (AI), each with their own set of buzzwords, there is a real need to understand the strengths and weaknesses of the two approaches.
Using engineering-led datasets is typically more reliable, and provides an in-depth view of the condition of critical components. However, these datasets are also expensive to gather, driving operators to focus their efforts on particular target areas. Ultimately this approach gives asset owners a deeper, but narrower perspective, which can result in blind spots.
Machine learning based approaches seek to tackle these blind spots by assessing all available data from the wind farm. Statistical datasets provide a broader, but shallower overview of the condition of the asset – and have been shown to yield reasonable anomaly detection.
But they are not a panacea, and, if divorced from the physical reality of turbine mechanics and the constant evolution in turbine technologies, may flag anomalies that, on closer inspection, pose no danger from an engineering perspective. In effect, algorithms may be learning things from these datasets that have no value in solving a real engineering problem.
In its enthusiasm to adopt innovative statistics-based solutions, the industry may be in danger of undermining the progress made in predictive maintenance to date, by driving an increase in ‘false positives’ that, in turn, will result in increased Opex costs for operators.
For example, for one operator ONYX InSight now works with, on a site where statistical analysis was used in isolation, the data analytics software employed raised over one thousand alarms, but, upon further investigation, only a handful of these were ‘real’ alarms requiring the operator to step in and carry out maintenance.
It is therefore imperative that the two approaches to predictive maintenance are not seen as opposites, and that operators’ use of emerging technologies is underpinned by advanced engineering knowledge.
"By applying the best engineering principles to the data, AI has the potential to transform predictive maintenance."
All too often, machine learning and statistical analysis has been misused as part of a data-only approach. The fundamental feature of AI is that it must be programmed to read a set of data and produce a conclusion. However, if that dataset is not large enough or does not have sufficient depth, then the conclusions will be less than optimal.
Instead, the wind industry will be best served by drawing on datasets that are grounded in an understanding of turbine engineering, then enhanced through the application of appropriate AI methods.
By applying the best engineering principles to the data, AI has the potential to transform predictive maintenance, as it will not only yield fewer anomalies and false positives for turbine owners and operators, but also equip them with a detailed understanding of their components’ condition and how they may be deteriorating. In turn, this knowledge will benefit the wider industry, enabling the development of better alternatives or upgrades.
We are confident that data analytics, AI and machine learning will have a significant impact in transforming decision-making for wind energy asset owners. These technologies will allow owners to consider a wider range of factors when assessing turbine condition, optimise maintenance programmes and bring down the levellised cost of energy.
But, for now, relying on statistics alone is not a viable option. Asset owners that take the lead in adopting approaches that combine statistical and engineering-led datasets will be the first to realise the benefits on offer.
Bruce Hall is CEO of ONYX InSight, a joint venture between Castrol and Romax Technology offering predictive maintenance solutions for asset owners in the wind energy industry and beyond