Table of contents
Business challenge
The maintenance of power grid transformers is a complex process that relies heavily on the deep experience of senior diagnosticians. These experts must analyse a vast number of parameters, from electrical readings to physicochemical properties, to assess the state of each unit.
This manual process is both time-consuming and challenging to scale. Subtle, early signs of degradation can be missed, potentially leading to more severe failures and cascading damage to other components. Determining the correct maintenance action and its optimal timing is not straightforward. Intervening too early incurs unnecessary costs, while acting too late risks catastrophic failure and power outages.
Our client was using a set of algorithms to support this process but wanted to explore whether modern AI techniques could deliver a more accurate and reliable solution for predicting both defects and the corresponding maintenance recommendations.
Building transparent and targeted predictive models
First, we addressed the challenge of unstructured data. We developed a bespoke Optical Character Recognition (OCR) model to process thousands of historical PDF reports from 2014-2020. This “digital passport” project successfully extracted key parameters and created a unified, machine-readable dataset of over 1,000 transformer records.
The team first explored creating a single, comprehensive ML/AI model (based on neural networks) to predict all 14 types of defects and around 30 types of maintenance recommendations simultaneously. However, testing revealed that this approach generated unacceptably high rates of false positives and false negatives. The analysis showed that different types of defects were often entirely uncorrelated, making a single model ineffective.
Based on this finding, the strategy shifted. We developed a suite of individual, specialised Machine Learning models (consuming both traditional techniques like Support Vector Machines and more sophisticated solutions like Gradient Boosting), one for each specific defect and recommendation. This modular approach allowed the team to select and fine-tune the best algorithm for each unique scenario, greatly improving predictive performance.
A core requirement for the project was trust. For the system to be adopted by the client’s diagnosticians, it couldn’t be a “black box.” We implemented explainable AI (xAI) techniques using the Dalex library. Instead of just providing a probability score, the system explains its reasoning by highlighting which input parameters had the most significant impact on its prediction. This transparency allows experts to validate the system’s output against their own knowledge and make more informed, confident decisions.
Benefits of our collaboration
- Achieved over 90% in accuracy and precision, 85% in F1 Score and around of 80% in recall while predicting defects and maintenance recommendations
- Provided full transparency into the AI’s decision-making process through explainable AI (xAI), increasing diagnostician trust and adoption
- Enabled more proactive and cost-effective maintenance scheduling for critical power grid infrastructure
- Established a scalable framework capable of incorporating new defect and recommendation models as more data becomes available
Technologies used in the project
