AbstractsComputer Science

Railway track circuits are electrical systems that are used for train detection. The identification of faults in these track circuits and the estimation of the severity of these faults is crucial for the safety and availability of railway networks. In this thesis a method is proposed to solve these tasks based on the commonly available measurement signals. By considering the signals from multiple track circuits in a geographic area, faults can be identified from their spatial and temporal dependencies. In this thesis, artificial neural networks are used to learn these dependencies from historical data. Although not enough measurement data was available during the writing of this thesis, it is more than likely that reasonable amounts of (unlabeled) measurement data will become available at a later time, as the required measurement equipment has already been installed. To train the networks in this thesis, the small available dataset is analyzed and used together with the currently available understanding of the fault dependencies to make a generative model. The synthetic data produced by this model are used to train and test the neural networks in this thesis. Artificial neural networks have recently achieved state of the art performance on difficult pattern recognition problems in several different fields such as image recognition and speech recognition. These recent successes can be largely attributed to the combination of large networks and large datasets. In the condition monitoring domain large datasets are generally not available. This prevents the use of the large neural networks that have become so successful in other fields. In-spite of this, some of the ideas that have become popular in other domains might still have value in the condition monitoring domain. This thesis focuses on bringing the Long-Short Term Memory architecture and the concept of end-to-end learning to the condition monitoring domain. To address the fact that only a limited amount of labeled data will be available, an unsupervised learning strategy is investigated. This strategy will use unlabeled data to pre-train a network so that it can more efficiently learn from the scarce labeled data. For the fault isolation task, it is shown in this thesis that when a large amount of labeled training data is available, the end-to-end learning strategy can detect and diagnose faults in the data from the generative model very accurately. When only a small amount of labeled data is available, it is shown that using a pre-trained network works better than using end-to-end learning.