Defect Detection for Laser-Trimmed Thick-Film Resistors in UPS Hall Current Sensors via Scale-Attention and Fine-Grained Feature Enhancement

The production process of UPS Hall effect current sensors employs laser trimming, which utilizes a laser beam to physically cut the thick-film resistors to precisely adjust their resistance values. However, during this process, defects such as residual resistance, short circuits, serrations, and burrs can occur within the resistor kerf. These defects lead to poor long-term stability, increased noise, and reduced voltage withstand capability of the sensors. Regarding the inspection of thick-film resistor defects after laser trimming, the existing method relies on manual inspection using high-magnification microscopes. This approach faces challenges such as low efficiency, significant subjectivity, slow speed, operator fatigue, and the risk of false or missed detections. To address these issues, this study proposes an improved YOLO11-based detection method. Specifically, we introduce Space-to-Depth Convolution (SPDConv) to preserve fine-grained features and integrate an improved C3k2 module with the Dilation-wise Residual (DWR) module to capture multi-scale information. Furthermore, an additional P2 layer is designed utilizing the Four-head Adaptive Spatial Feature Fusion Head (FASFFHead) to enhance the perception of tiny targets. We term this model DIA-YOLO (Detail-Preservation, Interactive-Attention, Aggregation YOLO), an efficient object detection model specifically designed for micro-defect detection scenarios. In our experiments, we utilized Generative AI to generate defect images and compared the proposed method with mainstream models including YOLO11, Faster R-CNN, and RT-DETR. The proposed model achieved the highest mean Average Precision (mAP) of 95.2%, meeting the accuracy requirements for detecting surface kerf defects in UPS current sensing resistors. Furthermore, extensive validation on the public NEU-DET dataset demonstrates the strong generalization capability of the proposed method, where it achieves a precision of 0.802, a recall of 0.783, and an [email protected] of 0.793, outperforming all baseline models. This study provides a novel approach for the defect detection of thick-film resistor kerfs.