Acoustic-optic holographic feature-decision fusion for multi-variant conditions monitoring in laser shock peening

Laser shock peening (LSP) significantly enhances the fatigue life of critical aero-engine components through its high-temperature, high-pressure, and transient characteristics. However, the complex response mechanisms induced by multi-field coupling and process parameter variations pose substantial challenges to quality consistency and monitoring reliability. This study implements synchronous monitoring of LSP processes using triple-channel AE sensors and a MCP-PMT optical sensor, with focused analysis on critical quality-influencing parameters including laser pulse energy identification, water confinement layer status detection, target material thickness estimation. Firstly, a holographic acoustic-optic feature extraction framework was developed through dimensional expansion and multi-modal fusion mechanisms, significantly enhancing signal representation capacity for complex pattern recognition. Secondly, the Dynamic Balance Multi-Task Learning (DB-MTL) architecture was established, combining modality-specific Transformer encoders with task-dedicated LSTM branches. The novel DB-MTL loss function synergistically integrates focal loss for class-imbalance mitigation and label smoothing for overfitting suppression, effectively handling heterogeneous temporal patterns across laser energy, constraint layer states, and thickness monitoring tasks. Thirdly, a cross-sensor probability-driven decision fusion framework was implemented, demonstrating three distinctive innovations: 1) Automated conflict detection/correction through probability space mapping; 2) Enhanced traceability via distribution pattern analysis; 3) Dynamic sensor reliability adaptation. This approach achieved 2–5 % accuracy improvement over conventional methods while maintaining real-time interpretability through sensor weight visualization and decision path tracking.