Prediction and identification of faults in track circuit are crucial for improving the safety and efficiency of railway transportation. However, the task of track circuit fault prediction through deep learning methods facing significant challenges due to the absence of reliable data. In this paper, a novel heterogeneous transfer learning method is proposed, aiming to reduce track circuit data reliance in model training by using publicly available datasets in other similar fields. An index describing the data distribution based on autoencoder feature extraction and maximum mean discrepancy is used to demonstrate the transferability between heterogeneous data firstly. Then a heterogeneous transfer learning method is constructed to accelerate track circuit fault prediction model training. Furthermore, the resulting deep learning model is compared to existing fault prediction methods. Finally, by adjusting the degree of involvement of transfer learning throughout model training, this paper comprehensively examines its effect on model training process. The simulation experimental results show that the proposed method can transfer useful knowledge in other similar fields for tasks in track circuit fault prediction, and the resulting model can correctly classify over 99% on the test dataset while reducing the amount of required track circuit data to 10% of the traditional training methods. The relevant methods proposed in this paper can significantly enhance the practical application value of fault prediction models based on deep learning methods in the field of intelligent maintenance of track circuit.

Track circuit is one of the most important parts of train operation control system. In current railway transportation research area, track circuit intelligent maintenance, including fault diagnosis and prediction is mainly achieved by using deep learning (DL). However, the application of DL models in practical has a great limitation. The training of track circuit intelligent maintenance DL models requires huge amount of labeled data. And DL models are too fragile since real data and training data can hardly meet the requirement of independent and identically distributed. In this paper, a new track circuit fault prediction model based on long short-term memory network and attention mechanism is proposed. To enhance the practical performance of this DL model, a deep data adaptation algorithm is applied to make an improved method to overcome degradation problem caused by non-identically distributed of training and real data. Moreover, it can be also applied to reduce data dependence of the DL model training. To evaluate the performance of the proposed DL model and the improved method, experiments were carried out based on real data from China Railway Shenyang Group Company Limited computer monitoring system. The experimental results show that the proposed track circuit fault prediction model can provide fault prediction results with an accuracy better than 95%, the improved method has a better performance than the DL model without deep data adaptation algorithms, and the proposed deep data adaptation algorithm has a better performance than conventional data adaptation algorithms. Compared with the conventional fault prediction DL model without transfer learning, the proposed method can achieve a much better performance, with an improvement, in the test, of over 75%. And the proposed deep data adaptation algorithm can make an accuracy improvement, in the test, of over 30% than conventional data adaptation methods.