An area-array tenon-shaped flexible EMAT and dual-channel CNN for steam turbine blade defect detection

Restricted access, varying curvature, and similar defect echoes complicate inspection of steam turbine blade tips. This study combines an area-array tenon-shaped flexible electromagnetic acoustic transducer (EMAT) with a dual-branch convolutional neural network (FCD-CNN) for in situ inspection. Interconnected support units slide to conform to curved surfaces. One-dimensional and two-dimensional branches extract global spectral and local time-frequency features from fast Fourier transform (FFT) amplitude spectra and continuous wavelet transform (CWT) images, respectively. Feature concatenation distinguishes five defect types and the normal condition. At two blade locations, the flexible EMAT achieved higher signal amplitudes and signal-to-noise ratios than a rigid EMAT. With increasing curvature, amplitude attenuations were 15.47% and 52.30%, with relative localisation errors of 2.80% and 3.00%, respectively. On CM-96 blade and 2Cr13 plate data, FCD-CNN achieved 93.93% ± 1.06% accuracy and 93.88% ± 1.07% macro-F1 over five independent runs, exceeding DenseNet accuracy by 1.43 percentage points. Accuracy gains over FFT-only and CWT-only models were 16.19 and 2.50 percentage points, respectively. The model contains 419,214 trainable parameters and requires 7.412 MFLOPs per forward pass. Under additive white Gaussian noise, performance approached the clean-signal baseline at 20 dB but declined at 5 dB and below. Applicability to natural damage requires further validation.