Information extraction, scan time reduction and corrosion quantification of reinforcing bars in concrete using an inductive sensor

摘要

Inductive scanning is one of the non-destructive testing techniques that exploits eddy current induction and detection and for inspecting and imaging steel reinforcing bars embedded within concrete. For several years, most of the research effort that has been conducted in relation to the development of this technique was devoted to improving the sensor design, the signal to noise ratio and signal and image processing and reconstruction. This was necessary to increase both the sensor depth penetration and the ability to detect surface corrosion. The area of layer separation had, at the commencement of this project, already been successfully researched within the group and provided the basis for the development of a new method for extracting bar dimensional information. However, various important areas such as the reduction of the scanning time and corrosion quantification remained untouched and required development from the fundamental level. This thesis presents three novel and significant contributions to these areas. Firstly, the thesis presents a methodology for extracting the bar dimensional information, namely, the bar depth and diameter, from the image data generated by the inductive sensor. The methodology developed is based on the fact that the sensor response parameters (the peak value and the full width at half height) are unique for a particular bar size at a particular bar depth. These parameters are used to train two neural networks to estimate the bar depth and diameter. The accuracy of the depth and diameter estimations are within the +/-10% as stated by BS1881 and +/-1bar sizes as quoted by DIN488. These estimations have been used to render three- dimensional images of the bar mesh within the scan volume. Secondly, a methodology to reduce the scanning time is described. Three image interpolation techniques have been applied to a sparsely populated data set to generate high resolution images. The method has enabled the scanning time to be reduced to a fraction of its former value yet still return accurate results from a very significantly reduced set of readings. The method requires only that the spatial impulse response should be known to optimise the method by reducing the number of samples without violating the sampling theorem. This is not difficult to measure, requiring a single line scan. Other than this, no a priori knowledge of the bars is required, either in terms of the scanning protocol or the reconstruction and measurement algorithms. It is now possible, using this method, to reduce the number of pixels in a 1m² image by a factor of forty whilst still retaining all the relevant information. Finally, the thesis describes a methodology to automatically identify and quantify corrosion generated on the steel bars in concrete using the images generated by the inductive sensor. Steel bars are first corroded using an accelerated corrosion system and scanned by the inductive sensor. Image segmentation techniques have then been used to isolate the corroded regions in the images from the noncorroded ones. These corroded regions are described by a set of features extracted during the segmentation process. These features together with the corrosion rates are then fed into neural networks, which are then trained with respect to the identification and quantification of the corroded regions. Results have shown that the method identified correctly all the corroded regions while it succeeded in quantifying 80% of the corroded regions correctly.