Surface hardness prediction for laser shock peening using narrow-band MCP-PMT and deep feature fusion with key elements and key frames

Laser shock peening (LSP) is a surface modification technology that significantly enhance the mechanical properties of materials, commonly improving the fatigue life of critical components such as aero engine blades. However, the high sampling rate of laser-induced plasma spectra and the complexity of physical information challenge spatio-temporal resolution, affecting the stability and consistency of LSP quality monitoring. This study address these challenges by leverageing the ultra-high sampling rate of Narrowband Microchannel Plate Photomultiplier (Nb-MCP-PMT, NMP) GHz to monitor rapid and weak light signals from laser-induced plasma in LSP process. High-speed photography elucidates the physical mechanisms of plasma evolution behind the NMP signal, showing that the 0–1.2 μs period in the NMP signal defines the dynamic balance stage, associated with plasma generation and expansion, while the 1.2 μs–1200 μs period corresponds to plasma diffusion and decay, defined as the attenuation stage of the NMP signal. Key features such as NMP-Min, NMP-Median, NMP-Std, NMP-ADif, NMP-Ske, and NMP-Kur were extracted to capture the periodicity, trend, and non-stationarity of NMP signals. A Transformer-Fusion-Attention mechanism (TRM-F-AM) model was developed to automatically extract temporal depth information and identify key features and frames. Distinct attention weight distributions were observed during the dynamic balance and slow attenuation stages of the NMP signal. The feature set redundancy was minimized, marking critical time periods for LSP process monitoring, specifically frames 1–80 during the dynamic balance stage and frames 155–265 during the slow attenuation stage. The model achieved a prediction accuracy of 99.05 % for LSP surface hardness on TC4 and 7075Al targets, surpassing TCN, LSTM, 1D-CNN, and ensemble learning models.