Track Circuits Fault Diagnosis Method Based on the UNet-LSTM Network (ULN)

Table 1

Review of existing techniques.

Reference	Technique	Limitation	Advantage
Field of troubleshooting
[1]	CNN	The data used differ from real scenarios, which can easily lead to many faults being detected	Improved models such as ResNet and GAN are applied to improve model performance
[2]	EKNN	Only for small datasets	Solves the problem of inefficiency of KNN
[3]	GRU	Higher calculation costs	Convert one-dimensional data into two-dimensional images to fully utilize the temporal information of the data
[4–6]	CNN-LSTM	More complex model structure and longer training time	CNN has a denoising property that reduces the effect of noise in the learning process, and LSTM learns the long-term dependencies of time series
[7]	CNN	The amount of experimental data is large and not easily accessible	Better design model performance
Fields of track circuits
[8]	BPNN	Only one device in the track circuit is studied, not the complete track circuit system	Simulation models were designed to allow access to experimental data
[9]	Neurofuzzy	The model needs further performance enhancements	It combines the advantages of fuzzy logic and neural networks and can be learned through the neural network training process
[10]	LSTM	Larger experimental data and higher computational costs	Experimental data combining temporal and spatial features
[11]	KPCA-SAE	The accuracy is 93.04%, and further improvement of the model performance is needed	Enables fault localization
[12]	Gray’s theory and expert system	More information needs to be collected and the amount of data is larger	Diagnosis of data predicted by Gray’s theory using an expert system
[13]	SVM	Accuracy is 96%, model performance needs further improvement	Build a simulation model to obtain the data needed for the experiment
[14]	DBN	The method is more costly to calculate	Optimizing the network model by combining particle swarm algorithm to improve the robustness and accuracy of the network
[15]	Rough set and graph theory	The method is more costly to calculate	The method can effectively reduce the time and space complexity