Explainable AI in Ultrasonic NDT

Yuyang Liu

Supervisors: Anthony Croxford and Paul Wilcox

While continuing the work to improve traditional methods of Non-Destructive Evaluation (NDE), modern engineers are increasingly interested in the automation potential of artificial intelligence (AI) techniques. From the suppression of artifacts in raw data [4] and automation in sizing [3, 7] to advanced imaging [5, 6], a series of AI models have been built and proven valuable in addressing existing research demands in the field of NDE.

However, as NDE is a field that studies the nature of failures across various engineering structures, the inherent black-box nature of many AI algorithms poses a major challenge. Thus, the trust and interpretation human operators have for existing NDE AI models are limited, despite their accuracy and power. The case is especially worsen when AI NDE models face previously unseen cases in real-world implementation.

Therefore, while various cases have proven the benefits of implementing AI in the field and as a crucial aspect of Industry 4.0 [1], the ability to build not only AI but also explainable AI (XAI) for NDE has become a crucial point to be addressed by researchers nowadays, as the next step of creating suitable and implementable AI in NDE.

Thereby, using corrosion profiling as an example case, the following objectives are drawn as future directions in actualizing reliable, stable, and interpretable AI in NDE:

Objective 1: Construction of toolkits to quantify and illustrate the decision-making processes of trained AI models that are proven to be implementable in NDE corrosion profiling.

Objective 2: Development of a consistent process to optimize the design of AI models for corrosion profiling, maximizing performance stability.

Objective 3: Development of AI models specialized in providing explainable NDE judgments, profiling, or defect characterization, enabling operators with rich NDE domain knowledge to understand and trust the AI outputs.

Under the guidance of these three objectives, a series of works has been conducted to identify potential XAI methods and algorithms applicable to the task. The current focus is on the potential and ability of layer-wise gradient-based functions [8, 9, 10, 11] in explaining 1D convolutional neural networks (1d-CNN). The 1d-CNN has been proven to be a viable alternative to conventional manual methods in solving the profiling problem [6]. Solid progress has been made in interpreting the estimations made by 1d-CNN based on particular A-scan time series in corrosion profiling on an individual X to Y level. However, much work remains to aggregate individual explanations and establish an overview of a trained 1D-CNN’s global decision-making process with respect to the entire population

Leave a Reply

Your email address will not be published. Required fields are marked *