Automated industrial work order prioritization
Active learning
Natural language processing
Uncertainty quantification
Hierarchical modeling
The national electrical grid operator in New Zealand approached me to investigate the possibility for automatically prioritising incoming maintenance work orders from external service providers, for equipment across the national grid.
This research project turned into a sever application, web application, and a conference paper. The software runs daily, reading information from the Maximo maintenance database, running the calculations, and feeding priorities back into the database.
The web application allows Transpower to update training data with new equipment, re-train, and investigate accuracy without external intervention. This is crucial for an ML system to continue to provide value long after the original authors have left and years have passed since the creation of the original training data - entire new types of equipment can be introduced into the grid!
This system was implemented in 2018, is still in use in 2023, has prioritized tens of millions of dollars in maintenance expenditure, and annual reviews have demonstrated significantly positive all-up ROI.
For each work order (100,000+) the system estimates a probability of failure across service performance, direct cost, public safety, worker safety, and environmental impact, as well as estimated cost to remediate, which are eventually mapped to an overall work order priority.
Developing equipment and defect ontologies was key to this project. The existing equipment database was at the wrong granularity - units of equipment that fail often had a one-many relationship with parts in the equipment database. Unsupervised approaches were heavily used to explore the data and iteratively build the ontologies.
Another key piece of the project was capturing organization-wide experiential knowledge. We developed a web app for data entry along with an active learning approach to make the best use of the available input time. Among other outputs, this was used to aggregate subjective evaluations of the likelihood of failure for an asset/defect combination into a quantitative estimate
The NLP module involves a hierarchical system of Bayesian generalized linear models, quantifying uncertainty at each step, to generate an equipment and fault ontology node. Subsequent ML modules then estimate the likelihood of failure, repair cost, and calculate the final priority score.