A Novel Approach to Railway Track Faults Detection Using Acoustic Analysis

Abstract

Regular inspection of railway track health is crucial for maintaining safe and reliable train operations. Factors, such as cracks, ballast issues, rail discontinuity, loose nuts and bolts, burnt wheels, superelevation, and misalignment developed on the rails due to non-maintenance, pre-emptive investigations and delayed detection, pose a grave danger and threats to the safe operation of rail transport. The traditional procedure of manually inspecting the rail track using a railway cart is both inefficient and prone to human error and biases. In a country like Pakistan where train accidents have taken many lives, it is not unusual to automate such approaches to avoid such accidents and save countless lives. This study aims at enhancing the traditional railway cart system to address these issues by introducing an automatic railway track fault detection system using acoustic analysis. In this regard, this study makes two important contributions: data collection on Pakistan railway tracks using acoustic signals and the application of various classification techniques to the collected data. Initially, three types of tracks are considered, including normal track, wheel burnt and superelevation, due to their common occurrence. Several well-known machine learning algorithms are applied such as support vector machines, logistic regression, random forest and decision tree classifier, in addition to deep learning models like multilayer perceptron and convolutional neural networks. Results suggest that acoustic data can help determine the track faults successfully. Results indicate that the best results are obtained by RF and DT with an accuracy of 97%.

Railway Track Inspection Using Deep Learning Based on Audio to Spectrogram Conversion: An on-the-Fly Approach

“…[23][24][25][26][27] use support vector machine (SVM), Refs. [19,28] use random forest (RF), Ref. [29] uses Adaboost, and [30,31] use principal components analysis (PCA).…”

“…Shafique et al [19] use tree-based classification models, random forest (RF) and decision tree (DT), which performed well compared to deep learning models for rail track inspection. Jie Liu et al [23] investigate different variants of SVM, such as twin SVM, LSSVM, etc.…”

“…The dataset used in this research is collected by [19]. It has 720 mono channel audio 'wav' samples and is collected in the area of Sadiq Abad Junction, Pakistan.…”

“…The dynamic range is 88 dB, the signal-to-noise ratio is 62 dB, and the operating voltage is 5.0 V. The microphones are connected through a wire and are unidirectional. For further details about the data collection process, the readers are referred to [19]. The dataset has three classes, i.e., normal, superelevation, and wheel burnt.…”

“…This research focuses on convolutional 1D, convolutional 2D, and long short-term memory (LSTM). The dataset [19] used in this research is a balanced one with three classes including normal, superelevation, and wheel burnt. The first step of preprocessing is to remove the noise of air, rain, and all the dead space in the audio by using a signal envelope function with a threshold of 100 Hz.…”