In smart manufacturing environments, Non-Destructive Testing (NDT) inspections leverage advanced human expertise to ensure high-quality defect detection. While resource intensive, these inspections benefit from skilled professionals who contribute to accurate and reliable interpretation, supporting continuous improvement and innovation in manufacturing processes. To enhance NDT inspection efficiency and reliability, we propose a novel ”AI-enabled Explainable NDT Inspection Agent”, which automates volumetric defect detection, delivers interpretable visual explanations, and combines Large Language Models (LLMs) with standards compliant inspection reports that provide transparent decision insights. The system integrates a memory-efficient 3D Convolutional Neural Network (CNN) architecture with gradient checkpoint for accurate detection of internal defects in volumetric data (e.g. from ultrasonic data), aligned, as an example, with ASME BPVC and ISO 9712 certification requirements. For explainability, we introduce a customized 3D Grad-CAM engine, which enables depth-filtered visualization of defect regions through volumetric heatmaps. These visualizations are integrated with the ParaView platform, offering intuitive exploration and validation by domain experts. To ensure regulatory compliance and decision traceability, a hybrid rule-based module incorporating guidelines such as ISO 20601 performs real-time validation of detected anomalies and maintains audit trails. Furthermore, our system employs a retrieval-augmented generation (RAG) pipeline powered by the Llama-3 (8B) model, automating the generation of customized, standards-aligned NDT reports. All components are unified within a web demonstrator built using Streamlit and React. RESTful APIs are used to integrate all its functionalities and maintain modularity. An interactive chat interface enables users to perform exploratory data analysis, query XAI visualizations, and request reports via natural language input. This approach introduces a novel combination of volumetric explainability, automated compliance check, and intelligent reporting that potentially enhances transparency, accuracy, and efficiency in NDT workflows. By reducing the reliance on manual interpretation and accelerating inspection cycles, the proposed system supports the advancement of AI-driven smart manufacturing solutions.
Vishwesh et al. (Thu,) studied this question.