Detection and Classification System for Rail Surface Defects Based on Eddy Current

Abstract

The prospect of growth of a railway system impacts both the network size and its occupation. Due to the overloaded infrastructure, it is necessary to increase reliability by adopting fast maintenance services to reach economic and security conditions. In this context, one major problem is the excessive friction caused by the wheels. This contingency may cause ruptures with severe consequences. While eddy's current approaches are adequate to detect superficial damages in metal structures, there are still open challenges concerning automatic identification of rail defects. Herein, we propose an embedded system for online detection and location of rails defects based on eddy current. Moreover, we propose a new method to interpret eddy current signals by analyzing their wavelet transforms through a convolutional neural network. With this approach, the embedded system locates and classifies different types of anomalies, enabling an optimization of the railway maintenance plan. Field tests were performed, in which the rail anomalies were grouped in three classes: squids, weld and joints. The results showed a classification efficiency of ~98%, surpassing the most commonly used methods found in the literature.

Keywords: convolutional neural network; eddy current; rail grinding; rail surface defects; railway maintenance; wavelets.

Figure 1

Types of anomalies considered.

Figure 2

Rail surface defect verification system diagram.

Figure 3

Probes installed in the inspection vehicle. (a) The camera and the two probes installed. (b) The probe attachment trolley. (c) A detailed view of the trolley.

Figure 4

Vehicle track monitoring process illustration.

Figure 5

Step diagram for detecting surface defects on rails.

Figure 6

Architecture diagram of a convolutional neural network.

Figure 7

Comparison between manual (top) and vehicle acquisition (bottom). The signal behaviour is the similar for both acquisitions.

Figure 8

Example of spikes. The acquired signal can be seen in black and the filtered signal in red.

Figure 9

Example of Heaviside noise. The acquired signal can be seen in black and the filtered signal in red.

Figure 10

Heaviside noise filtering process. The acquired signal is depicted in black, the derivative in blue and the filtered signal in red.

Figure 11

An example of the rail surface detection process. On the top, we can see a signal fragment of about 134 s long sampled at 7.5 kHz, where the horizontal axis indicates the sample number. The bottom plots are zooming versions of the signal. On the right plot, the blue circle indicates the location of the GPS stamp.

Figure 12

Characteristics of the anomalies considered in this work: squat (left), joint (middle) and weld (right). From top to bottom: picture of the anomaly, the EC acquired signal mean over the training data set and the WPS of the mean signal.

Figure 13

CNN loss (a) and accuracy (b) per epoch.

Figure 14

Classification confusion matrix.

Figure 15

Classifiers’ accuracy comparison.

Figure 16

Classifiers’ precision comparison.

Figure 17

Classifiers’ confusion matrix.