A Novel Chip-on-Board Defect Detection Approach Combining Infrared Thermal Evolution and Self-Supervised Transformer

The chip on printed circuit board assembly (PCBA) is developing toward miniaturization and high density, which makes it increasingly challenging to detect micro solder bump defects hidden inside the package. Here, we propose a chip-on-board defect detection method that leverages infrared thermal evolution with an improved transformer model, achieving highly efficient and accurate industrial chip-on-board defect detection. A periodic read-and-write is implemented in the chip work process and temporal infrared sequences are utilized to analyze the temperature evolution with the purpose of comparing the temperature variations between the reference chip and the defective chip. Subsequently, we develop an enhanced transformer-based classification model incorporating adaptive pooling and batch normalization, resulting in superior performance when compared to existing state-of-the-art models. Extensive experiments are conducted to assess the generalization and robustness of the proposed approach. The compelling results confirm that the performance of self-supervised representation learning exceeds that of a fully supervised method in accuracy and robustness, albeit with access to limited data. Our method indicates effectiveness in the classification of near-distributed datasets and exhibits a promising prospect for microelectronic packaging reliability analysis on industrial high-density PCBA.