The leading cause of line-of-road breakdowns for CSX locomotives over the past three years has been failed batteries. Each of these failures led to a train delay that generated measurably higher costs for materials and personnel, as well as potential negative impacts to other train schedules and customer satisfaction.
The Locomotive Predictive Failure project used data science techniques to develop an algorithm that indicates when a locomotive is anticipated to have a battery problem within seven days. Working with data output from multiple locomotive systems, the team analyzed prior failure patterns to develop the algorithm. The process includes a mechanism for reviewing performance of the predictive failure model and documenting incident resolutions based on the model recommendations. This focus on the precision of the model is the basis of a feedback loop for machine learning.
Using locomotive data and the predictive algorithm, the team developed a Current Prediction Week Tool, a dashboard that shows the top 20 at-risk locomotives with additional statistics on battery age, previous incidents, temperature and voltage. In addition, the team calculated and distributed weekly data trends. With this additional visibility at its fingertips, the Mechanical Department has been able to proactively address potential battery failures prior to line-of-road breakdowns. The predictive model has resulted in reduced line-of-road battery failures and improved overall locomotive fleet performance.