Rail Surface Defect Detection: an AI & Machine Learning Approach

Railway Track

Railway systems for public transport take quite a bit of maintenance to ensure the best safety measures. Moreover, there are many different kinds of rail surface defects, and some could be developing long before they’re ever noticed. In the worst-case scenario, rail surface defects could lead to a train derailment, and that’s the last thing anyone wants.

Having multiple systems in place for defect detection is vital, and considering how vast railway systems are, manual detection methods aren’t able to keep up with every defect. This article will dive into what defect detection methods are in place for rail surfaces and what modern solutions seem to work the best.

The Importance of Railway Systems

Although transportation has continuously evolved for many years, railway systems have held their relevance in modern society since the 19th century. Over time, the construction and technology implemented into railways were more reliable and of higher quality.

Even today, with many more cars and other forms of public transportation available, railways are still highly relevant. Such as with any form of transportation, safety is of utmost importance.

The rail surface needs continuous inspection, not only for safety but also due to damage forming on rolling wheels and other components on train cars. If the rail surface isn’t in pristine condition, it can have rippling effects through the entire railway system and any transportation traveling on it.

Rolling contact fatigue (RCF) on a railway system is inevitable. It can cause significant damage to transportation and many other components of railway systems, which is why proper defect detection must be in place.

Types of Rail Surface Defects

Railway systems require an enormous amount of maintenance, and rail surfaces take on tons of damage over extended periods of time. As previously mentioned, RCF can cause all kinds of damage, and some are more serious than others.

Moreover, some are much more common than others and can be easily repaired, while others may take much more effort to maintain. Nevertheless, below are a few of some common rail surface defects.

  • Surface cracks
  • Squats – cracks that separate the bearing from the body of the rail
  • Shelling – a partially broken top rail or gauge
  • Head checks – small cracks of just a few millimeters

Some of these defects can be highly dangerous as they may not show any significant damage on the rail surface but are hiding years of wear and tear underneath. Again, this is another reason that modern and automated detection methods are so crucial.

It’s also essential to understand that there isn’t a single solution to maintaining rail surface defects. Some defects can be seen with the naked eye, while others are hidden below surface level and could easily harbor the most danger.

To handle all of the maintenance required takes multiple defect detection methods. Current research on this subject clearly shows some work better than others, but a majority of these technological advancements in rail surface defect detection seem to be much more accurate and efficient than previous manual methods.

Evolution of Defect Detection Methods

For the majority of the railway system’s existence, manual labor was required to detect rail surface defects. Not only is this a time-consuming project, but it’s also dangerous and leaves room for mistakes. With the advancement of modern technologies, it’s clear that automated defect detection methods should be utilized.

It’s vital to note, though, that even with the use of automated detection methods, the frequency of cracks forming on the rail’s surface is still just as consistent. Nevertheless, the detection methods utilized can be divided into a few specific categories: image processing, machine learning, geometry measuring, detecting wheel track motion, changes in the electric or magnetic fields, and using thermal sensors.

Moreover, the most prevalent research is primarily focused on machine learning and image processing. Below, you’ll find more researched information on these various detection methods.

Image Processing Defect Detection

The purpose of this method is to take photos of rail surfaces and detect visual damage by analyzing specific information obtained from the image. From here, the image can be filtered to show only the characteristics that show signs of any damage.

Information such as damage length, size, and quantity can be analyzed with this data. The computer vision inspection system collecting this information is composed of three modules: image acquisition, pre-processing, and resulting judgment.

Currently, to improve on this method, it will require the advancement of the modules needed for defect detection through image processing. Although this method can be efficient, there’s still much room for improvement.

With imaging, there are other factors that could affect the quality of the data. For instance, particular light reflections could make rail surface defects difficult to see.

There are other proposed image processing methods that look to improve image quality or more accurately analyze the image data for defects.

Geometry Measurement Defect Detection

Utilizing either laser reflection or a mechanical stylus, rail surface defects can be detected by measuring geometric data from railway systems. There are a few known limitations to this method, and there isn’t an all-in-one solution to defect detection.

Moreover, manual inspection will still be needed after the fact after geometric data is collected. The measurements from this data help create a 3D rendition of the damage on the rail surface and have shown to be effective in detecting defects.

There are still a few flaws regarding this method. For example, lasers are sensitive to other elements such as light and could cause incorrect measurements.

Luckily, there are other 3D modeling methods that can work around minor issues like this, providing more accurate data. It really comes down to which method produces the most detailed and accurate information, so proper action can be taken.

3D modeling is helpful for many reasons, such as analyzing rail and wheel trajectory for turnout defects and even recording the contour information of the railway while avoiding distortion of standard imaging.

Defect Detection Through Sound and Light Waves (Thermal Sensors)

Ultrasonic technology has been utilized for defect detection for quite some time. Rail surface defects can be measured in this way by watching parameters such as sound velocity, waveform, and echo.

Defects in the rail surface will reflect ultrasonic waves, providing the information needed to get a visual of the damage. As with each detection method, it has its own set of flaws. Any defect that’s less than 4mm from the surface of the tested area won’t be discovered using acoustic wave detection.

A more advanced approach to this method is utilizing rapid laser ultrasonic defect detection. Laser ultrasonic signals are able to compile a full inspection image of the rail without any missing defects.

Even if the original method works in its own right, it can still use some improvement, and the rapid laser method looks to be a promising step in the right direction.

Machine Learning

Over the last decade, machine learning has come a long way with many advancements. Furthermore, this technology can be divided into two detection methods: shallow learning and deep learning.

Shallow machine learning takes a more simplistic approach yet still provides decent results for rail surface defect detection and can easily trump manual detection methods. It collects data by extracting multiple image features, and you can find a few of those below.

  • Histogram of Oriented Gradient (HOG)
  • Local Binary Pattern (LBP)
  • Scale Invariant Feature Transform (SIFT)

Deep machine learning provides much more complex data and has the ability to create layer-by-layer descriptions of rail surface damage. The information this technology generates can be extremely detailed, especially useful for rail damage below surface level.

Moreover, the available research on deep learning shows its defect detection capabilities are much better than this article’s previously mentioned methods.

Deep learning can also be applied to many different defect scenarios rather than just one core focus. Furthermore, there are a handful of deep machine learning models are currently being used with image and visual recognition, such as:

  • Deep belief network (DBN)
  • Stacked auto-encoders (SAE)
  • Convolutional neural networks (CNN)

Out of these, convolutional neural networks (CNN) have seen satisfactory results in various applications. Using neural networks for defect detection has proven itself useful, as it uses deep machine learning to simulate our visual perception, improving the accuracy of classifying rail surface damage.

A CNN has the potential to provide results at 98% accuracy, which exceeds many other detection methods.

In Conclusion

Defect detection methods for railway systems have evolved immensely, more so within the last ten to twenty years, considering the boom in technological advancements.

With the many potential flaws associated with manual inspection methods, there’s no reason those risks should be taken today with the tech that’s available.

Aside from the general safety of public transportation, the longevity and durability of railway transportation systems are also important. The building and maintaining of railway systems cost millions upon millions of dollars, and it’s vital they’re maintained from every angle.

Based on the research available, there are a plethora of methods for defect detection that are not only safer but more pre-emptive, proactive, and accurate overall.