The machine learning system in use at Teck's Fording River operation (pictured) is trained to look for specific electric and suspension faults in a subset of the fleet's Komatsu-model trucks. Courtesy of Teck Resources

Teck Resources switched on a new machine learning system for predictive haul truck maintenance at its Fording River steelmaking coal operation in British Columbia in early April. Within the first three weeks of operation, ten trucks were identified by the system as needing maintenance and when mechanics investigated, they found issues in need of repair.

The project, initiated in October 2017, is a joint effort between Teck’s Digital Systems group and the superintendents and maintenance engineers at Fording River. Teck laid groundwork for this technology back in 2011 by installing a company-wide digital platform and a suite of onboard performance Mobile Equipment Monitor systems (MEMS) sensors and computers on all the Caterpillar and Komatsu trucks at Teck’s steelmaking coal operations.

The new machine learning system was applied to a subset of Komatsu-model trucks within the fleet and is trained to look for specific electric and suspension faults. Teck estimates that machine learning systems like these have the potential to save between $2 and $4 million annually at a single site by minimizing unplanned maintenance.

Related: Loader monitoring system employs deep learning to keep operations running smoothly

Real-time MEMS data already detects more than 35 types of faults in the trucks. They are collected in a Wenco database platform and delivered to site superintendents through sensor alarms, explained Kalev Ruberg, Teck’s vice president of digital systems and chief information officer. Now, through a partnership with Canada-based tech solutions company Pythian and using Google Cloud, the same data is run through a set of custom algorithms and returned as an additional alert to Asset Health Supervisors, raising the alarm on trucks in need of maintenance.

Ruberg said Pythian wrote algorithms informed by data supplied by Teck from faults that had already happened. “One of the key things we learned is you can’t just buy a product of the shelf or have someone remotely do this,” he said. “The Pythian engineers had to come on site and realize what the data they were looking at meant.” The system now diagnoses fewer than five per cent false positives.

Ruberg said 60 per cent of the work done to finish the system was in data preparation, compared to 20 per cent developing the actual algorithms and 20 per cent setting up the production environment.

He said the next step for a number of Teck coal sites is to continue to integrate the system into the regular scheduling of truck maintenance, expand the system to more trucks and other mobile equipment, and use existing data to train the system on a wider range of fault types. The system will continue to expand across all sites, as part of Teck’s “One Teck” digital initiative.