Answers are grounded in this paper and cite their sources.
Conventional ultrasonic non-destructive testing (NDT) processes rely on manual interpretation of ultrasonic signals, making defect identification excessively dependent on operator experience. Although artificial intelligence models enable reduced reliance on experiential knowledge, they require a large amount of labelled data. This paper presents a multi-feature-driven deep learning model (MFDDL) for phased array ultrasonic signal processing and intelligent defect identification under limited labelled data conditions. Only three sets full matrix capture (FMC) data need to be collected at each category of location for model training. The proposed MFDDL extracts multiple features from the limited labelled original FMC signals, and integrates them with the features extracted from the network, which are used for model training jointly. The MFDDL reconstructs ultrasonic images based on predicted category labels to visualise the testing results. Applied to 240,000 unlabelled signal samples for testing, the MFDDL accurately identifies six different categories of ultrasonic signals with labelled training data accounting for only 1.92% of the total dataset. Experimental results demonstrate that the MFDDL achieves an average accuracy of 80.02% on the unlabelled dataset.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: