Defect Recognition Of Ballastless Track Structure Based On Distributed Optical Fiber Sensing System

Ballastless track system,which is now extensively used in high-speed railways in China,can be susceptible to cyclic dynamic load and environmental impacts and is thus predisposed to structural defects such as interfacial separation and track slab concrete crack damage.Diagnosis of defects is generally done by manual work after the midnight,which can be laborious and low in efficiency.With its advantages in multi-point positioning and high sensitivity,optical fiber sensing system based on Φ-OTDR is now widely used in structural monitoring and territory security monitoring and can be a potential solution in railway structure monitoring.Optical fiber sensing system can be essential to the construction of proactive maintenance mechanism and is of great value to promote operation and maintenance informatization.Data processing techniques and defect recognition algorithms are extensively discussed in this thesis.The main contents are as follows:(1)Φ-OTDR apparatus configuration and train data processing.Based on the basic principles of Φ-OTDR the parameters are calculated and applied on the optical fiber sensing apparatus.The vibrations of trains are located and amplified with moving difference algorithm.Image binarization and iterative outlier removal techniques are employed to clean and compress data.With the help of maintenance of way branch we mapped collected data to geographical positions.The dataset is constructed on five types of defects and one normal situation.(2)Feature engineering,machine learning algorithm and its application in defect recognition.Features are extracted in various domains including time domain and frequency domain statistical values,wavelet packet energy and Mel-Frequency Cepstral Coefficients.XGBoost together with several other machine learning algorithms is employed and compared in terms of accuracy in recognition.(3)An end-to-end deep learning method is introduced to solve Φ-OTDR pattern recognition problems.Principles concerning designing neural networks for time sequence signals are discussed at length.An enlarged receptive field network is proposed to extract1-D vibration features and achieve an higher accuracy than feature engineering and machine learning method.(4)A multi-scale convolutional network is proposed to incorporate a small kernel into the big kernel convolution operation,such that it would capture different patterns with various resolutions.Features on different scales are concatenated in depth and various techniques for weighting and aggregating features are subsequently discussed and evaluated.The performance on the defect dataset is superior to plain networks in terms of accuracy.The effectiveness of Φ-OTDR on ballastless track defect recognition is thus verified.The discussions of engineering methods and designing of algorithm are of great value to further exploration and implementation of fiber optical sensing system on track structure monitoring.