← Ideas & projects

INDUSTRIAL COMPUTER VISION

Thermal Coal Intelligence

A thermal vision pipeline combining detection, foundation models and segmentation to monitor coal stockpiles and surface emerging combustion risk.

CV—02DEPLOYED WORKYOLO → DINO → SAM

A thermal vision pipeline combining detection, foundation models and segmentation to monitor coal stockpiles and surface emerging combustion risk.

The problem

Coal stockpiles self heat, and left unmonitored, localised hotspots can progress to spontaneous combustion, a safety and operational risk that's expensive to catch late and expensive to monitor manually at scale. The operational question is simple to state and hard to solve well: across large, irregularly shaped stockpiles, where exactly is the temperature rising, how fast, and does it warrant an operator response right now?

Manual thermal camera review doesn't scale, and naive thresholding on temperature images is too noisy. You need something that can find the coal against a cluttered industrial background, characterise emerging hotspots against normal background variation, and do it continuously.

The pipeline

A custom model, trained on custom labelled source image data, segments each stockpile in the camera frame. Temperature data from the thermal cameras is then extracted and fitted onto that segmentation mask, so every pixel of "hot" is tied to a specific point on a specific stockpile rather than just a hot patch somewhere in the image.

From there, the fitted thermal data surfaces the hot points on the pile, and those readings feed directly into a customer facing alert system, so a rising hotspot gets flagged as an actionable alert rather than sitting in a camera feed nobody's watching.

The segmentation model reaches 0.87 validation IoU. A meaningful part of the engineering effort went into making the pipeline usable at production scale: the original processing pipeline ran in the order of 17 hours per pass, brought down to roughly 70 minutes through batched GPU accelerated preprocessing and data parallel execution across multiple GPUs. Output feeds into a Spark Structured Streaming layer, and routing thresholds have been tuned against real operational output rather than left at theoretical defaults, so alerts reflect what actually constitutes an actionable temperature delta on site.

This was built as a work project at Genesis Energy.