Machine Learning in stress corrosion crack characterisation from full matrix capture ultrasonic signal

How can we monitor the growth of stress corrosion cracks (SCC) in an automated way through ultrasonic inspection? With breakthroughs in machine learning society and availability of full inspection sensor data, automated structure health monitoring with integrated machine learning models to determine current state of structure has seen growing focus. The benefits of ML methods including consistency, reliability and speed provide a promising way to assist human inspection, which could greatly reduce both operational and time cost in inspection industry. However, there are many challenges exist to find out SCC due to their complex orientations and branched geometries. Moreover, real crack dataset is rarely existed and shared among industries. The thesis focuses on these problems, by developing scalable, efficient algorithms that classify SCC severity level to alert humans in areas where SCC is predicted to affect the integrity of structure. Particularly, I focus on the use of full matrix capture (FMC) data that captures the complete ultrasonic time domain signals for each transmit and receive element of a linear array probe. The full acquisition nature of FMC allows us to leverage abundant information to develop ML models with limited datasets. Currently few research focused on developing machine learning model from FMC signal for defect detection and characterization. Only few focused on classifying images calculated from FMC signal where image resolution and computational cost is high. I aim to fulfil this gap by exploring raw FMC signals to classify crack severity through ML method without forming images. Two ML categories are explored, one is traditional models, in which feature extraction is usually required to reduce high feature dimension and extract relevant information. The other is neural network model where raw signals are learned without the need of feature extraction. First, I focus on feature extraction methods and apply extracted features to traditional machine learning models. Many of the current feature extraction approaches involving extract maximum amplitude, time-of-flight, spectral feature or other features for building models. However, they are not ideal for every problem and every type of signal. In addition, different defect mechanism often affect data in different and complex ways. These differences don't always simply translate into common features. Tailored feature engineering method is in need for specific defect, environment and measuring technology. In this thesis, two feature extraction approaches are explored for extracting FMC data, one is engineering handcrafted features approach utilizing signal coherence processing tools. Coherence analysis has been used to assess functional association of network activity in brain from multiple electrodes. This associated network activity also exist in the FMC signals; hence, a coherence-based method was proposed to extract crack-related features for ML learning. While the other is statistical approach that based on statistical feature learning using principal component analysis (PCA) method. PCA is a widely used feature extraction technique in NDT field that remove redundant and uncorrelated information in the dataset. Three popular traditional supervised ML models, which are support vector machine (SVM), random forest (RF) and logistic regression (LR) are employed to test derived feature quality. The first coherence approach reaches above 90% classification accuracy on all three ML models, in contrast PCA extracted uncorrelated features does not perform well and concluded not suitable for our data and task. Poor performance of de-correlated features suggest that correlation is an important factor that needs to be taken into account for classification. Then, I consider deep neural network, specifically convolutional neural networks, because CNN model is specialized at capturing correlation due to its unique learning structure that is designed to learn correlation in images. I proposed a novel CNN architecture by taking FMC transmitter structure into consideration and proposed a multi-view strategy to improve classification accuracy and also reduce model training parameters, as a result, the classification accuracy reaches 94% with 16% improvement on benchmark image model. To further reduce model parameters for its future deployment with minimum computational cost requirement. Next, I explore auto-correlation processing method as feature processing tool for CNN model. The merged auto-correlated features of each receiver in FMC dataset significantly reduce model parameters due to reduced input size and reaches classification accuracy of 97% that is comparable to performance of coherence feature. A series of supervised learning model have been proposed for accurate crack classification. However, supervised learning has limitations since they can only perform well on the task they are trained to. Finally, I explore self-supervised (SSL) model to learn representations to improve model generalization ability for a variety of inspection scenarios. Consider a set of structures of different thickness, the supervised learning paradigm might require training of multiple models for each thickness scenario. In contrast, a high-quality representation learned in SSL model could potentially generalize to different thickness as demonstrated by simulation crack data. SSL model developed in this thesis is a pre-exploration of selfsupervised learning strategy in NDT filed for improving generalization capability of ML model.