Multi-task Learning for Joint Prediction of Remaining Useful Life and Failure Type of Train Wheelsets
MetadataShow full item record
The failures of train wheels account for half of all train derailments. Both remaining useful life (RUL) and failure types of wheels are critical for wheel maintenance. RUL prediction is a regression task, whereas failure type is a classification task. Most existing approaches to multi-task learning usually deal with homogeneous tasks, such as purely regression tasks, or entirely classification tasks, thus cannot utilize the intrinsic useful correlation information among different variables. In this paper, we propose a general methodology to jointly predict two tasks by using a common input space to achieve more desirable results. We formulate this problem as a convex optimization, a combination of linear regressions and logistic regressions, and model the joint sparsity as L2/L1 norm of the model parameters in order to couple feature selection across tasks. In our experiments, we use real-world data from a Class I railroad in North America. Competing risk analysis is also applied to failure time data, in order to compare failure probability of different failure types for different car kinds and wheel sizes. To validate our method, we perform support vector regression to predict RUL and support vector machine to predict failure type. And we show that our method outperforms the single task learning method.