Can AI Predict Equipment Failure?

The problem, in numbers

Maintenance is a large and largely controllable cost in mining. In 2017 BHP said it spent $3.5 billion a year maintaining more than 3,000 fleet units through more than 1.7 million maintenance jobs.^{Source 1} When a haul truck, crusher or conveyor fails without warning, the cost is not only the repair. It is the lost tonnes, the scramble for parts and the crew pulled off planned work.

Failure is also a safety problem. NSW mines notified 179 fires on mobile plant in FY2025. The most common failed component was a hose, and the significant incidents the regulator describes include a burst hose spraying fluid onto hot machine parts and a failed turbo whose engine oil ignited on a truck exhaust.^{Source 7} Across five years of reports, the most common combination of failed part and cause is a hose worn out by wear and tear. Degradation of that kind is exactly what condition monitoring is meant to catch.

What improves, and why

Predictive maintenance replaces the calendar with the machine's actual condition. Vibration, temperature, oil analysis, pressure and electrical signals feed a model that learns what normal looks like and flags drift before it becomes a breakdown. Machine learning adds the ability to combine many weak signals across a fleet, rather than relying on one fixed alarm threshold per sensor.

The gain that matters most is lead time. A failure seen two weeks out can be planned: parts ordered, a crew booked and the truck brought in at shift change. BHP's account of its Yandi haul truck program shows this clearly. Planning accuracy for jobs scheduled more than two weeks ahead rose from 10% to 85%, missed parts fell from 13% to 1%, and truck availability ran "well ahead of 90 percent", with $5.5 million of savings reported for FY17. On its Liebherr T282 trucks in the Hunter Valley, mean time between failures rose from 36 to 53 hours.^{Source 1} These are company-reported figures from a single operator, and some of the gain came from better planning discipline, not prediction alone.

What the evidence actually says

A widely cited set of predictive maintenance benchmarks comes from the US Department of Energy's operations and maintenance guide: savings of 8% to 12% over a preventive program, and, citing unnamed independent surveys, 70% to 75% fewer breakdowns and 35% to 45% less downtime.^{Source 2} They are useful as a benchmark, but they date from 2010, describe condition monitoring generally rather than AI, and are not specific to mining. They should not be quoted as proof that a machine learning model will deliver the same.

The research base is large and still maturing. A 2025 systematic review of 166 studies on AI-driven predictive maintenance in mining found growing use of deep learning and digital twins, but named data standardisation, model scalability and interoperability as unresolved.^{Source 3} A broader review of 60 studies across industries noted that commonly used open datasets, including a simulated aircraft engine dataset, suffer from data heterogeneity and imbalance, and called for more real-world validation.^{Source 4} Field trials do exist: researchers monitored a 1.6 km conveyor section in Western Australia for 10 months with fibre-optic sensing and an anomaly detection algorithm that flagged defective idlers.^{Source 6} What is scarce is independent, controlled evidence of what deployed systems deliver over years.

Australian research on maintenance records explains why results vary so much between sites. A University of Western Australia study of work orders for heavy mobile equipment from several organisations found that failure mode, functional location, maintenance action and status fields were often missing or wrong, and set out a rule-based way to repair them using the free-text notes and knowledge of how each operation actually ran its maintenance.^{Source 5} That is skilled, slow work. A model is only as good as the failure history it learns from, so the maturity of the maintenance data, more than the sophistication of the algorithm, sets the ceiling on what prediction can deliver.

Where it falls short

• Data maturity decides the result. Models trained on incomplete work order history inherit its gaps, and cleaning that history is a project in its own right.

• Lab results do not transfer. Accuracy on a benchmark dataset says little about a dusty pit with mixed fleets, changing duty cycles and sensors that drop out.

• Rare failures are hard to learn. The catastrophic failures that matter most are the ones with the fewest examples, so models are least certain where the stakes are highest.

• Alerts need owners. A warning that lands in an unread dashboard prevents nothing. Value appears only when planners trust and act on the flag.

Where governance decides the outcome

A predictive model changes who decides when a machine is safe to run. That decision needs a named owner, a documented threshold for pulling equipment from service, and a record of every alert and what was done about it. Suppressing or deferring warnings to protect production targets should be visible, not quiet. Sensor drift and model updates need change control, because a model that has silently stopped seeing a failure mode is worse than no model at all. In its 2025 report on mobile plant fires, the NSW regulator told operators to ensure stringent monitoring and quality control of maintenance and repair work.^{Source 7} Prediction supports that duty. It does not replace it.

How to think about adoption

Start with the assets where failure is costly and the signal is clear: haul truck engines and drivetrains, mill and conveyor drives, critical pumps. Before buying a platform, audit your maintenance records, because clean failure coding is the foundation every model stands on. Measure the baseline first, including availability, mean time between failures and the share of work planned ahead, so any gain can be tested on your own fleet rather than taken from a brochure. Treat vendor accuracy figures as hypotheses, and build the alert workflow before the model goes live.