Nonlinear ultrasonic testing and data analytics for damage characterization: A review

Nondestructive testing and evaluation (NDT&E) are commonly used in the industry for their ability to identify damage and assess material conditions. Ultrasonic testing (UT) is one of the most popular NDT&E techniques. A variant of ultrasonic testing known as nonlinear ultrasonic testing (NUT) has some advantages over conventional (linear) UT as it is more sensitive to damages in their early stages; even at the microscopic levels. Furthermore, the nonlinear characteristics of ultrasonic waves can be correlated to several material properties. In the last two decades, the NUT method has been investigated from two aspects, namely the direct (modeling) problem and the inverse (NUT testing) problem. The direct problem aims to establish the nonlinear mechanism and analyze the behavior of wave-damage interaction. The inverse problem is investigated under three headings: (1) data acquisition with NUT techniques, (2) signal pre-processing and feature extraction, and (3) parameter analysis for damage characterization. The conventional data analytical methods extract nonlinear features from noisy signals and build a damage index to characterize damages. However, damage index-based analyzing model can be challenging, as other factors affect the overall system nonlinearity such as complex specimen geometry, different damage characteristics, varying ambient conditions, and measurement uncertainties. To overcome these shortcomings, machine learning (ML) methods appear promising for the analysis of complex nonlinear ultrasonic signals by exploiting data mining and pattern recognition capabilities. Therefore, this paper aims to provide a comprehensive review of the state-of-the-art ML-enriched NUT for damage characterization. Other NUT-based technologies are also reviewed, including modeling of wave-damage interaction, different NUT techniques for data acquisition, signal pre-processing methods, and damage index-based parameter analysis strategies for damage characterization. Major emphasis is placed on the application of ML methods for NDT&E applications. Additionally, future research trends on data augmentation, complex damage characterization, and baseline-free methods using NUT are also discussed.