Welding defect characterization through multiple domain phased array ultrasonic scanning data analysis

Ultrasound is a relatively old technology utilised that is still heavily used in the field of Non-Destructive Test (NDT). Technology in terms of data acquisition and presentation have greatly improved, however processing of these data still requires much human intervention. Ultrasonic Scans of welding defects usually contain vast amounts of information which have to be processed and assessed by NDT experts for classification. Hence there is at present a demand to reduce time, human errors and expertise in the overall classification process. The objective of this Final Year Report (FYP) is to fully automate the entire classification process. The 1st phase of this FYP would be to perform Feature Extraction of each welding defects acquired from Ultrasonic A-Scan data from a 2m long weld specimen belong to A-Star. The main idea behind the Feature Extraction is to isolate defects signals and subsequently process them statistically into Features. Ideally, the Features of each defect would contain self-describing information and become Training Data for the second phase of the FYP. In the second phase the Training Data gathered previously will be feed to various Machine Learning Models categorised as Kernel Machines and Neural Networks (NN). Within the Kernel Machines Category, SVM, KNN and Ensemble models will be tested and have their classification performance assessed. For NN however, only a 2 layered NN will be employed since there are no readily available Deep Neural Networks (DNN). Feature selection and elimination would also be employed to determine the which the best performing model with the available features. This study has found that the weighted KNN model is the best performing model with classification accuracy of 82.1% with current Feature and Training Data. The results however in terms of selection of the Machine Learning Model to automate the classification process is inconclusive. This is due to the mainly due to the of Training Data containing only 28 samples. Due to the meagre amount of Training data, performance stability of NNs was greatly affected, and hence cannot be properly assessed. In the future, only when more weld specimens are available to produce more Training Data can the proper machine learning model be confidently decided.Bachelor of Engineering (Mechanical Engineering