Real time fault detection in railway tracks using Fast Fourier Transformation and Discrete Wavelet Transformation

Learning from limited labelled data: contributions to weak, few-shot, and unsupervised learning

“…Automatic track surveying is based on patter analysis over sensed signals and images. Among sensing devices, different technologies such as thermal resistors [157], acoustic sensors [158,159], video recording [160][161][162][163][164][165]; [166,167]) or accelerators [149,[153][154][155][168][169][170][171][172][173][174][175][176][177][178][179][180][181][182][183][184][185][186][187] have been proposed. In particular, the use of acceleration sensors on axle-box has become more popular for detecting track irregularities of different wavelengths and occurrence.…”

“…In particular, the use of acceleration sensors on axle-box has become more popular for detecting track irregularities of different wavelengths and occurrence. Concretely, different applications include wheel flat [171], crossings monitoring [174,175], rail corrugation [176,185]; [188]), roughness derivation [178,179], rail joints [155], settlement and dipped joint ( [189]) and other railway elements [149,190]. In the aim of predictive maintenance, first works focused on visual description of the patterns that elements and defects produce on time-frequency domain [149, 168, 169, 177-179, 182-184, 190].…”

“…Further on, some works described a set of features based on classic image processing such as peak intensity, frequency-band relative intensity, or other statistics. Then, first classifiers were used on these features, such as SVMs [176] to predict rail corrugation, random forest for railway lifetime prediction [174,175] or simple costume decision trees for fault detection [185], or recent neural networks classifiers [189,191]. Very recent works [153][154][155]187] have proposed the use of deep learning models via CNNs to characterize acceleration spectrograms on predictive tasks.…”

“…Method Metric (µ ± σ) Accuracy F1-Score Precision Recall CNNs + BCE ( [155,189]) 0.5875 ± 0.0945) 0.6099 ± 0.1227 0.6111 ± 0.0863 0.6529 ± 0.2254 Hand-crafted Features + SVMs [152,185] 0.612 ± 0.061 0.652 ± 0.014 0.604 ± 0.022 0.651 ± 0.054 Autoencoder Features [154] 0.648 ± 0.025 0.652 ± 0.014 0.652 ± 0.014 0.720 ± 0.064 Proposed 0.715 ± 0.715 0.735 ± 0.050 0.726 ± 0.0496 0.747 ± 0.064 Table 6.1: Quantitative results on railway crossing wear detection for the proposed method and implemented baselines. Best results in bold.…”