BW-YOLO: A Lightweight Recurrence-Plot-Driven Network for In-Process Defect Identification from Multi-Band Photodiode Signals in Lithium-Ion Battery Laser Welding

Laser welding of tabs and busbars determines the electrical integrity and safety of lithium-ion battery modules, yet many weld defects leave little trace on the surface that post-weld vision inspection relies on. Photodiode signals record the weld-pool dynamics during processing, but their short, non-stationary waveforms are difficult to classify reliably with a compact model. This study presents BW-YOLO, a lightweight network designed together with recurrence-plot (RP) encoding of the photoelectric signal. The RP turns each weld signal into an image whose coarse block structure and fine recurrent texture reflect keyhole stability. BW-YOLO matches this two-scale structure with a Haar wavelet convolution module that integrates channel shuffle and one-shot aggregation (WT-ShuffleOSA) and a dual-scale channel reduction pyramid slice attention module (DCR-PSA). The evaluation used a balanced dataset of 1200 industrially produced joints monitored in the visible, infrared and reflected-laser bands, with one fixed training–validation partition. On the 240 validation joints, evaluated separately in each band, BW-YOLO reduced misclassifications by 30–38% in every band relative to the YOLO11n-cls baseline with a reported 15% reduction in operation count. In the infrared band it reached 96.67% accuracy and 98.28% precision with 1.50 M parameters and 2.8 GFLOPs, giving the highest observed accuracy among the ten networks compared on that partition. These results support further investigation of the proposed design; robustness across training runs and datasets, performance relative to direct time-series classifiers, and deployment latency remain to be established.