Field-Regularised Factorization Machines For Mining The Maintenance Logs Of Equipment

Failure prediction is very important for railway infrastructure. Traditionally, data from various sensors are collected for this task. Value of maintenance logs is often neglected. Maintenance records of equipment usually indicate equipment status. They could be valuable for prediction of equipment faults. In this paper, we propose Field-regularised Factorization Machines (FrFMs) to predict failures of railway points with maintenance logs. Factorization Machine (FM) and its variants are state-of-the-art algorithms designed for sparse data. They are widely used in click-through rate prediction and recommendation systems. Categorical variables are converted to binary features through one-hot encoding and then fed into these models. However, field information is ignored in this process. We propose Field-regularised Factorization Machines to incorporate such valuable information. Experiments on data set from railway maintenance logs and another public data set show the effectiveness of our methods.