Eduardo Furtado

Predictive Maintenance of Mining Trains

Featured image
Image generated with FLUX.1 [schnell]

I developed a model to reduce train breakdowns for a mining company's crucial two-way railway system connecting their mine to the port. Train failures were costly, causing significant operational disruptions and making repairs more extensive. The primary issue for on-track failures was overheating wheel bearings, which frequently required trains to collected from the tracks and sent to maintenance. Another problem was that the only garage available for repairs was located at one end of the track with limited capacity, so it couldn't handle too many trains at once. If a train broke down at the far end, delays became even more severe.

To tackle this issue, I developed a predictive maintenance model to classify overheating wheel bearings before they lead to train failures. By training on the precision metric of our predictions, we focus on sending only trains with a high likelihood of failure to be routed for maintenance, avoiding unnecessary jobs that could overwhelm the garage. Using Python, Scikit-learn and TensorFlow, several machine learning models were benchmarked using cross-validation, to select the one with the best performance for our use case scenario and ensure the model's robustness.

After development, the model was deployed into the company's server with a companion Power BI dashboard that displayed the model’s predictions. This allowed maintenance staff to easily see which trains needed their attention and helped them plan and schedule maintenance activities more effectively. After the model was implemented, the company saw a significant decrease in train delays due to this kind of breakdowns, reducing fail on-track time by nearly 30% in the first month.