Search

Photovoltaic (PV) wafers are a basic material for solar cells, and surface defect detection for solar PV wafers is necessary to improve the quality of solar PV. However, the complexity of the wafer manufacturing environment causes the surface defects of PV wafers to appear diverse, and it is still challenging in practical applications. In this research, a deep-learning (DL) network based on global information branches and dual-branch feature fusion (DBFF) is proposed for surface defect detection (SDD), named GFYolov7. First, for the issue of the low contrast on the wafer surface samples, the global information branching (GIB) is generated using the multiscale features of the backbone, which improves the information utilization of the backbone. Second, a DBFF module is designed for the GIB module, and a lightweight attention mechanism is used for the fused multiscale features, which serve to highlight defective regions and suppress irrelevant regions. The experimental results of GFYolov7 on NEU-DET and the GC10-DET reach 79.5 and 72.1 mAP and that on the real-world PV wafer dataset is 72.7 mAP. By comparing with other DL methods, GFYolov7 has been demonstrated that it has good detection ability and its detection speed can reach 61 FPS.

Abstract: Based on the National Climate Center (NCC) of China operational seasonal prediction model results for the period 1983–2009 and the US National Weather Service Climate Prediction Center merged analysis of precipitation in the same period, together with the 74 circulation indices of NCC Climate System Diagnostic Division and 40 climate indices of NOAA of US during 1951–2009, an analogue-dynamical technique for objective and quantitative prediction of monsoon precipitation in Northeast China is proposed and implemented. Useful information is extracted from the historical data to estimate the model forecast errors. Dominant predictors and the predictors that exhibit evolving analogues are identified through cross validating the anomaly correlation coefficients (ACC) among single predictors, meanwhile with reference of the results from the dynamic analogue bias correction using four analogue samples. Next, an optimal configuration of multiple predictors is set up and compared with historical optimal multi-predictor configurations and then dynamically adjusted. Finally, the model errors are evaluated and utilized to correct the NCC operational seasonal prediction model results, and the forecast of monsoon precipitation is obtained at last. The independent sample validation shows that this technique has effectively improved the monsoon precipitation prediction skill during 2005–2009. This study demonstrates that the analogue-dynamical approach is feasible in operational prediction of monsoon precipitation.