Machine learning has attracted researchers from across various fields over the past few years. There are primarily two areas in the application of machine learning based on data format, namely, computer vision, as well as natural language processing. Extensive research has been conducted on both the development of more advanced machine learning algorithms, as well as expanding the scope of applications into the area of non-destructive testing (NDT). In industrial applications, the current inspection monitoring process using NDT techniques highly relies on either human- or computer-based visual inspections. Therefore, the combination of machine learning and computer vision to provide an effective and automatic process to detect defects merits further investigation.

Laser-generated Guided Wave (GW) is an effective method for long-distance and non-contact inspections. This method is among the most widely applied NDT techniques for performing an inspection on plate-like structures. Conventionally, to detect and locate the defect, an imaging algorithm referred to as ‘Delay and Sum (DAS)’ is required to locate defective areas. The difficulty in using GW to detect defects is that it receives signals which are usually dispersive, and contain multiple wave modes, making it harder to identify defect-related signals. To be able to clean the received signals so that they are free of any disturbance and in return to save the effort required to interpret the results, the inspection process is expected to be almost fully automated, with minimal human involvement. To solve this, this thesis seeks to present a novel method which can automatically detect the location of a defect using both a Continuous Wavelet Transform (CWT), and an improved Canny edge detection method. The raw received image signals were transformed to a 2.5D temporal plot. Possible edges created by the defects were identified using the aforementioned methods. Additionally, a kernel density estimation-based compensation (KDEC) method was designed to further improve the traditional DAS imaging algorithm. Compared to the conventional algorithm, when the KDEC method is added into the existing DAS method, the automatic inspection process will no longer require any baseline image. Moreover, the transmitter-receiver pairs needed for using the physical sensors, or the number of measurement points for using the non-contactable laser-generated GW, can be significantly reduced. The results obtained from the numerical analyses and experiments conducted on the aluminium plates have proven that such novel automatic methods have in the past successfully determined the location of defects.

Active or pulsed thermography has been widely used to determine surface defects in structures and materials. This is difficult, primarily because the defect image is usually invisible in the collected sequence of infrared images. Besides, the detection of internal cracks can be affected in cases which the targeted object is operating in a complicated environment. With the help of active thermography, the acquired signals involve a sequence of infrared images collected from a thermal camera after actively triggering two lamps to emit a flash on an iron pipe. The flash acts as a heat stimulus and its heat energy penetrates the pipe’s wall. It was found that part of the penetrated heat will be reflected when there is an internal crack in the pipe’s surface. Hence, the collected temporal infrared images contain the heat images of defects and show their locations. However, the appearance of defects only appeared in a few images, and could be easily ignored if verified using human vision. Therefore, automating the vision inspection process digitally is crucial. To automate the defect identification, a novel method combining an improved Canny edge detector and an unsupervised learning algorithm was introduced. The contrast of pixels in each frame was enhanced using the Canny operator, and then reconstructed using a triple-threshold signal process. Two features, namely, the mean amplitude obtained from the time waveform plot, and the maximal magnitude obtained from its frequency spectrum, were extracted from the reconstructed signals to help the process of identifying the size and location of the defects. Finally, a binary image containing the defect information was generated using the proposed unsupervised machine learning based clustering algorithm. The effectiveness of the proposed approach was tested using experiments conducted on an iron pipe which contained two internal cracks and a surface abrasion.

For the machine learning algorithm itself, a new clustering method was designed. Clustering using a fast search and identification of the density peaks (CFSFDP) has been proven to be a novel and effective algorithm. It can automatically identify the center of each classified cluster, as well as which cluster has the maximum density. However, the results obtained from the use of the CFSFDP demonstrated that it is sensitive to the cutoff distances, especially for the detection of distant non-spherical clusters. The original CFSFDP needs to rely on a typical manual interaction for selecting the appropriate cluster centers. Therefore, a new density-based clustering algorithm referred to as clustering was defined and merged with the candidates of the cluster centers via an independent affinity (CDMC-IA), which was proposed and developed. With the assistance of this novel strategy, the appropriate value of the cutoff distance for a particular application can be determined automatically. The two new quantities of ‘independence’ and ‘affinity’ were introduced to solve the problems of having multiple density peaks, by using the slow and ineffective manual selection method to select an appropriate cluster center. After implementing the proposed algorithm and testing it in benchmark datasets, the achieved results indicated its superior performance compared to prior models.

In general, this thesis reports several novel methods of computer vision and unsupervised machine learning algorithms, to construct an automated and effective inspection system. The ultimate goal is to minimize human expert dependence when using the NDT-based computer vision, while increasing the reliability and accuracy for automatic defect detection. To achieve this goal, several tailored inspection systems were designed to address the problems associated with automatic defect detection, through the use of NDT techniques. 1) A method of automatically detecting sub-surface defects in iron pipes using active thermography by combining an improved Canny edge detector, and clustering method; 2) The Canny edge detector was also applied to laser generated GWs as a feature engineering method for automatic defect imaging in plates; 3) Furthermore, a baseline free laser generated a GW defects imaging method which was proposed using a KDEC algorithm. 4) A comparison study was conducted to demonstrate the performance of these state-of-the-art clustering algorithms, especially the density-based ones. Their limitations were revealed so that a novel clustering algorithm (CDMC-IA) could be designed to overcome these limitations, and in return, attain a superior performance compared to these popular clustering algorithms.