Application of multi-channel ultrasonic and machine learning for defect detection in nuclear fuel assemblies

Nuclear fuel cladding integrity is critical for reactor safety, yet detecting breakage defects in densely packed fuel assemblies remains challenging due to extreme operational environments and inefficient conventional inspection techniques. This study introduces a novel multi-channel ultrasonic testing system integrated with machine learning for precise, efficient cladding defect detection in nuclear fuel assemblies. By leveraging time-division multiplexing, the system expands the number of channels of commercial ultrasonic flaw detectors from 8 to 32, reducing inspection time for a 17×17 assembly from over 40 minutes to under 10 minutes. Hilbert transform is employed to extract robust envelope features from dispersive Lamb wave signals, suppressing environmental interference while quantifying acoustic impedance changes caused by coolant leakage. Addressing the “scarce defective samples, abundant normal samples” challenge, a One-Class SVM is used to build an anomaly detection model with only normal samples. Experimental results show 100% defect detection rate (no misses) and <3% false positive rate in simulations. The algorithm screens 264 fuel rods per assembly in under 3 minutes, cutting manual review time by 66.7% (from 1.5 hours to 30 minutes). This solution overcomes traditional ultrasonic testing bottlenecks, providing a robust automated framework for in-service inspection and significantly enhancing nuclear power maintenance efficiency.