Transformer health indexing (THI) helps rank a fleet for further investigation and can support decisions about repair, replacement and refurbishment. The score is not automatically a measure of risk, reliability or remaining life. Those questions require additional evidence and, where probabilities are claimed, a validated model. The practical question remains: which units need attention first? A discussion of RONIN health-index trends in steelmaking is also available.
The RONIN version discussed in this historical article used machine learning to estimate health-index scores from uploaded oil-condition data. The useful question is how to place those results within an engineering assessment, with their limitations made clear.
Read along to learn more!
What is Ronin AI and how does it help the transformer industry?
Seetalabs’ Ronin is an AI powered tool to predict the THI score of each transformer using condition monitoring data for predictive maintenance. Ronin AI uses machine learning (ML) algorithms to predict individual THI score for each asset or for a standalone asset over time, as necessary. In fact, the use of THI algorithms and its efficacy in improving asset management and grid performance is a hot-topic for the power industry. Other discussions are also available to showcase the benefits of using THI for original equipment manufacturers (OEM) and particularly, for non-expert industrial asset managers such as in the steel industry.
What are the benefits of using Ronin AI?
RONIN aims to reduce the effort involved in organizing data and reviewing a fleet. Cost savings, processing time and training requirements depend on the workflow, dataset and system configuration. A large or varied training dataset does not, by itself, prevent overfitting. Evaluate a model on independent transformers, identify the reference labels and error measures, and check performance when inputs are missing. Processing capacity and supported workflows should be checked for the version being used.
How to use Ronin AI for maximum efficiency?
In the historical dashboard described here, a score below 30% was labelled very poor. That was a product-specific screening band, not a universal intervention threshold. Review the measurements driving a low score: elevated gases, deteriorated oil quality and missing data require different follow-up. For gas concerns, examine both concentrations and changes, check analytical quality and consider confirmatory sampling. Online monitor data needs its own treatment of noise and rates. The asset manager should then set priorities using the equipment history and the consequences of an outage.
Best practices for using Ronin AI for transformer fleet management
A useful dashboard must make its inputs and reasoning inspectable. For an asset manager, that means seeing which measurements drive a score, which failure modes are covered and which information is missing. Older fleets often have incomplete or inconsistent records. Software can help organize those records, but missing observations remain missing evidence; they should not disappear behind a precise-looking number.
Upload the required gas data with sample dates, units and the available oil-quality and asset information. More complete, relevant data may reduce uncertainty, but it does not guarantee a fixed gain in accuracy. If values are estimated or unavailable, report that fact and assess how it could change the ranking. A confidence estimate should identify its calculation and validation population. CIGRE TB 761, section 4.2, recommends making the effect of unavailable information visible and using sensitivity analysis where appropriate.
So don’t wait and grab your free trial right away!
