11. Tunnel lining crack detection method using deep learning or multi-layer neural network

  • A method for detecting cracks in tunnel linings has been developed using multi-layer neural networks offering over 90% successful detection.
  • The method’s discernment capability is close to that of a human being, allowing automatic generation of maps showing damage progression.

Existing image processing programs to detect structural damage such as cracks required the fine-tuning of numerous parameters and experience-based expertise. Furthermore, with existing systems any change in picture being processed requires time-consuming readjustment of parameters. Another drawback with current image processing is the difficulty in removing any similar looking features, or noise, from the image, such as cables or masonry joints.

A method introducing a ‘learning’ based image analysis technology has been developed therefore, to distinguish cracks from images applying a deep learning process (multi-layer neural network). Two classifiers were created after learning from a large volume of pictures in two groups - either “presence of a crack” or “absence of a crack”. When these classifiers or sorters were presented with a new picture, they recognised the presence or absence of a crack with over 90% successful detection (Fig. 1, Fig. 2). A hybrid detection method was proposed using a color-coded pixelated picture to show the position of probable cracks and which is then analysed focusing on location and direction of the crack. Fig. 3 shows that the level of crack detection using this method is close to that of a human being, making it possible to generate an automatic maps showing damage progression.

This method could be used for other purposes such as detection of leaking water, etc. therefore the next step is to make this system commercially available.

Fig. 1 Mapping of likely cracks (the redder the pixels, the greater the likelihood of a crack)
Fig. 2 Image analysis detects only the cracks in the lining reinforcement material
Fig. 3 Comparison with man-made crack damage map