Synergistic feature engineering and bidirectional long short-term memory in ultrasonic non-destructive inspection

This study introduces a novel and interpretable hybrid framework for ultrasonic non-destructive evaluation. This methodology synergistically integrates advanced feature engineering with a bidirectional long short-term memory network to accurately localise defects and identify specimen configurations using one-dimensional acoustic time-series data. Motivated by the need for data-efficient and explainable deep learning solutions in industrial inspection, the proposed approach explores multiple categories of physically-interpretable features derived from acoustic signals (amplitude spectrum, power spectrum, convolutional neural network features, statistical descriptors) and various fusion strategies (concatenation, element-wise operations). Results demonstrate consistent defect localisation precision below 0.1 mm. Moreover, the model also achieves high performance in classifying specimen configurations. This study offers a robust, scalable, and industrially relevant pathway towards high-precision ultrasonic inspection by providing a framework that prioritises both accuracy and transparency.