ELTDF-Net: a PAUT welding defect detection model based on locally perceptual residual gating and multi-frequency convolutional attention

Accurate welding defect detection is vital for industrial safety, but traditional visual methods struggle with internal defect complexity. While Phased Array Ultrasonic Testing (PAUT) can capture internal defect features, deep learning-based detection faces issues like noise interference, weak low-light features, and overlapping spectra of grain boundary scattering and defects. This paper proposes the ELTDF-Net intelligent detection model, which enhances defect features via a triple innovation module: the Enhanced Channel Calibration and Refinement Module (ECCR): Fusion of multi-scale cavity convolution and dynamic channel calibration to enhance local gradient response of millimetre-scale defects and solve the problem of attention bias; the Luminance Control Residual Block (LCRB): Constructing a luminance-noise decoupling framework, suppressing frequency-domain aliasing noise through gated feature fusion and enhancing feature correlation in low-light areas; and the Translation-Variant Convolutional Cross Attention (TVconv-CA): Combines frequency domain analysis with spatial adaptive convolution to enhance low-frequency morphological features of defects and suppress high-frequency scattering noise. Top-1 accuracy: 86.36% on PAUT-Welds, 97.44% on NEU-CLS-64, 99.72% on NEU-CLS and 84.97% on MT. Furthermore, a dedicated intelligent classification software system was developed to validate the model’s engineering practicability and detection capability in industrial scenarios. Our code is available at https://github.com/storm9967/ELTDF-Net.