An online monitoring system for the penetration states of the butt-lap joints in laser welding via EfficientNet-LECA and Candidate-Driven Control logic

In the laser welding process of aluminum alloy butt-lap joints for China Railway High-speed (CRH) trains, real-time recognition and reliable feedback control of laser welding penetration states are challenging. The limited speed of model inference and occasional misjudgments can lead to severe consequences when control actions are directly implemented. To address these issues, an online monitoring system was developed, based on a new neural network, EfficientNet-LECA, and a new control logic, Candidate-Driven Control (CDC). EfficientNet-LECA integrates a lightweight efficient channel attention (LECA) mechanism, achieving 98.83% accuracy and surpassing the baseline model EfficientNet-B0 while reducing parameters to 0.097M and inference time to 27.14 ms. CDC enhances model recognition reliability by state checking and dynamic candidate-set management, suppressing misjudgments and adding less than 1 ms latency. The system was validated on an interference test set with synthetic smoke and reflection. The penetration state transitions were recognized within 60 ms, and laser power was adjusted via a Programmable Logic Controller (PLC) to restore the target state. In closed-loop regulation, step disturbances from 4100 to 4400 W and 4700 W were accurately detected and suppressed, adjusting the power back to 4100 W. The system achieves a decision throughput of approximately 30 Hz and a closed-loop response time of 160 ms for 4400 W and 260 ms for 4700 W, meeting industrial requirements for stable feedback regulation.