Modeling of a Predictive Maintenance Process to Screen Railway Transit Signals for Repair Showed That a Machine Learning Algorithm Could Identify 35 Percent of Failures a Month in Advance.

Researchers in New Jersey Applied Machine Learning Algorithms to Railway Signal Maintenance Assessment Using a Training and Testing Dataset from a Transit Agency.

Date Posted

Artificial Intelligence-Aided Rail Transit Infrastructure Data Mining

Summary Information

Maintenance of transit assets can be both costly and time consuming for transit agencies. Predictive maintenance systems may offer a way for transit agencies to reduce maintenance costs while not compromising vehicle availability or safety. Predictive maintenance systems continuously collect data about transit assets and related information such as weather, and feed the data into an algorithm which makes predictions about when an asset will likely need repair or maintenance actions. This can reduce impacts related to in-service failures which can be costly and result in unpredictable delays.


To assess the effectiveness of a predictive maintenance system, a research team from Rutgers University utilized data provided by a major rail transit agency in the United States. A dataset including primary information about the agency’s rail signal equipment and records of failure histories and corresponding maintenance work were provided. The research team then added weather data from the National Oceanic and Atmospheric Administration. The team then fed all these data into a widely used machine learning algorithm, Extreme Gradient Boosting, and developed a model to predict railway signal failures. Data from May 2019 through December 2020 were used for training the algorithm and data from January 2021 through June 2021 were used to assess effectiveness.


  • The algorithm was able to successfully predict 35 percent of signal failures up to a month in advance of when actual failure occurred, through screening of the top 10 percent of signal locations.
  • Location and previous failure history were found to have predictive value for future failures.
  • Weather, especially very hot weather, was found to have predictive value for signal failure.
  • The research team noted that these approaches may be refined to include other information, such as equipment age and sensory information, for improvements in predicting potential signal failures and enabling remediation in advance.
Goal Areas
Results Type
Deployment Locations