|
|
Train Identification System at Railway Switches And Crossings Using Advanced Machine Learning Methods
Krč, Rostislav ; Vorel,, Jan (referee) ; Plášek, Otto (referee) ; Podroužek, Jan (advisor)
This doctoral thesis elaborates possibilities of automatic train type identification in railway S&C using accelerometer data. Current state-of-the-art was considered, including requirements stated by research projects such as S-Code, In2Track or Turnout 4.0. Conducted experiments considered different architectures of artificial neural networks (ANN) and statistically evaluated multiple use case scenarios. The resulting accuracy reached up to 89.2% for convolutional neural network (CNN), which was selected as a suitable baseline architecture for further experiments. High generalization capability was observed as models trained on data from one location were able to classify locomotive types in the other location. Further experiments evaluated the effect of signal filtering and denoising. Evaluation of allocated memory and processing time for pre-trained models proved feasibility for in-situ application with regard to hardware restrictions. Due to a limited amount of available accelerometer data, distribution grid power demand data were utilized for further refinement of the proposed CNN architecture. Deep multi-layer architecture with regularization techniques such as dropout or batch normalization provides state-of-the-art performance for time series classification problems. Class activation mapping (CAM) allowed an explanation of decisions made by the neural network. Presented results proved that train type identification directly in the S&C is possible. The CNN was selected as optimal architecture for this task due to high classification accuracy, automatic filtration, and pattern recognition capabilities, allowing for the incorporation of the end-to-end learning strategy. Moreover, direct on-site application of pre-trained models is feasible with respect to limitations of in-situ hardware. This thesis contributes to understanding the train type identification problem and provides a solid theoretical background for future research.
|
|
|
Train Identification System at Railway Switches And Crossings Using Advanced Machine Learning Methods
Krč, Rostislav ; Vorel,, Jan (referee) ; Plášek, Otto (referee) ; Podroužek, Jan (advisor)
This doctoral thesis elaborates possibilities of automatic train type identification in railway S&C using accelerometer data. Current state-of-the-art was considered, including requirements stated by research projects such as S-Code, In2Track or Turnout 4.0. Conducted experiments considered different architectures of artificial neural networks (ANN) and statistically evaluated multiple use case scenarios. The resulting accuracy reached up to 89.2% for convolutional neural network (CNN), which was selected as a suitable baseline architecture for further experiments. High generalization capability was observed as models trained on data from one location were able to classify locomotive types in the other location. Further experiments evaluated the effect of signal filtering and denoising. Evaluation of allocated memory and processing time for pre-trained models proved feasibility for in-situ application with regard to hardware restrictions. Due to a limited amount of available accelerometer data, distribution grid power demand data were utilized for further refinement of the proposed CNN architecture. Deep multi-layer architecture with regularization techniques such as dropout or batch normalization provides state-of-the-art performance for time series classification problems. Class activation mapping (CAM) allowed an explanation of decisions made by the neural network. Presented results proved that train type identification directly in the S&C is possible. The CNN was selected as optimal architecture for this task due to high classification accuracy, automatic filtration, and pattern recognition capabilities, allowing for the incorporation of the end-to-end learning strategy. Moreover, direct on-site application of pre-trained models is feasible with respect to limitations of in-situ hardware. This thesis contributes to understanding the train type identification problem and provides a solid theoretical background for future research.
|
guest ::
