[Objective] To support steam-turbine blade  inspection without opening the cylinder during shutdown and enable deployment  on a confined-space micro-robot, this study proposes an acoustic  fingerprint-based damage identification method that combines low-energy impact  excitation with a lightweight one-dimensional convolutional neural network  (1D-CNN). The method uses the acoustic response produced by structural  excitation to distinguish damage states while addressing the constraints on  model size and computational cost associated with edge deployment. 

[Methods]  Acoustic responses were collected from healthy and damaged steel bars, steel  plates, and last-stage blades from the low-pressure rotor of a 1 000 MW steam  turbine; the blade defect was a naturally occurring crack. A  short-term-average/long-term-average (STA/LTA) trigger was used to extract the  effective response segments automatically. Separate lightweight 1D-CNN models  were developed for the three specimen types, and online data augmentation was  restricted to the training data. The training and test sets were divided by  excitation position rather than by randomly assigning individual signal  samples. All signals acquired at a given position were placed in only one  subset, and every test position remained unseen during model training. This  design evaluated recognition performance for acoustic responses acquired at  previously unseen excitation positions and reduced the possibility of  optimistic estimates caused by distributing signals from the same position  across both subsets. Test samples remained unchanged and were excluded from  model parameter updates. 

[Results] Under position-separated evaluation,  identification accuracy was 97.97% for the turbine blade, 100.00% for the  steel bar, and 96.67% for the steel plate. Recall for damaged samples reached  100.00% for all three specimen types. The network contained approximately 1.77  × 10⁴ trainable parameters and required 8.04–10.15 MFLOPs per input sample.  Numerical simulation showed that the crack altered the local absolute-velocity  amplitude and subsequent oscillation pattern, whereas the first-arrival-time  difference remained below the output time resolution. These simulation results  indicate that, at the selected monitoring point, the crack-related difference  was represented mainly by response amplitude and waveform evolution rather  than by a resolvable shift in arrival time. 

[Conclusion] The three  specimen-specific models identified damage at excitation positions that were  not used for training, indicating that the learned acoustic features were not  confined to the training positions under the tested conditions. By combining  low-energy excitation with a compact model and low computational cost, the  proposed method provides a technical basis for micro-robot-based edge  inspection of steam turbine blades during shutdown without cylinder opening.  The present validation covers the tested specimens and experimental  conditions; applicability across individual blades, turbine types, and  operating conditions should be examined using broader sample sets and field  tests.