Automated unit for Refractory assessment using non-destructive testing

Refractory materials often develop latent internal defects during multi-stage thermo-mechanical processing that evade surface inspection. While conventional impact-acoustic methods detect such anomalies, their reliability is undermined by operator-dependent variability in excitation force, transducer positioining, and contact pressure. To overcome these limitations, this study presents the Automated Unit for Refractory Assessment (AURA), an integrated platform automating impact-acoustic non-destructive testing for fired refractories. AURA combines as a servo-controlled impactor with an actuator-driven clamping mechanism to standardize acoustic excitation and maintain consistent contact, achieving a signal energy coefficient of variation of 2.42%. Acquired acoustic signals are processed using a one-dimensional convolutional neural network with Gradient-weighted Class Activation Mapping to isolate relevant spectral regions and extract 66 discriminative features. These features train a lightweight artifical neural netwrok classifier, achieving a high precision of 0.9946. The results demonstrate that integrating automated hardware acquisition with explainable artificial intelligence provides a robust, interpretable framwork for at-line quality control, reducing data dimensionality by 99% while secuirng exceptional diagnostic performace.