Railway track condition monitoring based on sensor data and artificial neural networks

Monitoring the condition of engineering objects is one of the urgent tasks of industry, construction, and transport infrastructure. This article describes a system for condition monitoring and diagnostics of rail tracks in real time. Compared with other similar studies, the proposed system has the advantages of compactness, usability, scalability and versatility of application. The proposed monitoring system is based on an Nvidia Jetson Nano embedded computing board and also includes inertial sensor modules, a microphone, a geolocation module, communication modules, an SSD storage device, and a battery. The prototype of the diagnostic module is a portable device that can be located inside a railway car. It allows analyzing the condition of a track in real time without taking it out of service. The article describes the system operation when the data stream from sensors, supplemented by train speed and geolocation data, is stored in the device memory. Then information is converted into a multidimensional time series dataset and transmitted to an external computer for subsequent analysis by artificial neural networks. A method for collecting and labeling data for neural network training is described. The detection and classification of rail joints are performed in this work. This methodologically corresponds to defect detection, and this approach was chosen due to the absence of rail defect datasets and the possibility to acquire datasets of rail joints in practice. The accuracy of rail joint detection using convolutional neural networks for detecting events and their classification was 87.49% for two classes and 78.13% for three classes.