Damage detection of steam turbine blades using deep neural network-based acoustic fingerprint recognition

This study develops an impact-acoustic framework for classifying predefined surface-damage states on steam-turbine blades, identifying one of three damage regions, and estimating class-dependent severity parameters. Impact signals were recorded for 5 s at 22,050 Hz under controlled impact and microphone angles. After event screening, each signal was converted into a 13-coefficient W-MFCC map that combines continuous-wavelet time-frequency localization, Mel-scale subband mapping, and cepstral compression. A hybrid ResNet-50-Transformer-PANet-CMTL network was designed to extract hierarchical local features, model long-range dependencies, fuse multi-scale information, and produce conditional multi-task outputs. The dataset contained 1,200 training samples and 300 validation samples covering crack, crack-erosion, and crack-abrasion conditions in Areas 1-3. On the reported validation split, W-MFCC correctly classified 296 of 300 samples (98.67 %), compared with 89.33 % for MFCC and 87.33 % for wavelet input, and its loss curves showed more stable convergence. The full multi-scale configuration also exhibited smoother training behavior than the ResNet-50 + Transformer baseline. These findings demonstrate the feasibility of combining W-MFCC representation learning with multi-scale conditional prediction for controlled turbine-blade condition assessment.