| With the development of micro-electronics technology electronic packaging is rapidly moving towards light, thin, tiny, and lead-free. The trend is towards ensuring the quality of the solder joint of the post-reflow printed circuit board (PCB), so it becomes more difficult to meet production requirements while depending on traditional detection methods. The image recognition-based automatic optical inspection (AOI) can detect defects of the solder joints quickly, uniformly, repeatedly, and reliably, and it has been used in practical production. However, we still face many challenges such as high false alarm rate, slow detection speed, weak operational, et al. To improve the automation of the AOI system, an intelligent solder joint inspection method was thoroughly investigated in this paper. The main contributors of the thesis are listed as follows:(1)Base on the summarization and analysis of the current solder joint inspection algorithms, the statistical feature analysis based solder joint inspection method is proposed. In the training stage, the solder joint is divided into several sub-regions. Optimal features of each region are selected and the thresholds of every feature are obtained using feature selection algorithms. In the test stage, every sub-region is incpected with the selected features, and the final solder defects are obtained with the results of sub-regions. The automation of the AOI is improved with the proposed algorithms because of the decrease of the man-made factor.(2)The features of solder joints are extracted, including color features, shape feature, position feature, et al. Bidirectional 2D Linear Discrimination feature is introduced. The method improveed conventional bidirectional 2D linear discrimination analysis by weighting different classes and samples during the calculation of covariance matrix. Taking into account the mis-alignment, virtual samples are generated and used for training to improve the robustness to mis-alignment. (3)The performance of the current AOI system is not stable because the feature selection and parameter setting depend on personal experience. AdaBoost is improved and used for feature selection. The three improvements of AdaBoost are stated as following: first, dual-threshold weak classifier are used for every sub-region; then, the relevance information among features is proposed to evaluated the selected features to decrease redundant information; finally, Boundary-SMOTE based over-sampling algorithm is used to eliminate sample imbalance. With the improved AdaBoost, the features of solder joints are automatically selected.(4)A rapid solder defect diagnosis strategy is proposed. In consideration of the imbalance of positive and negative samples, a secondary classification is proposed. First, the image comparison method is used to remove most normal solder samples which. Then, the statistical feature analysis based method is used to determine the final defect type. A CART based solder defect diagnosis decision tree is eatablished to decrease the computation. The proposed solder defect diagnosis strategy can increase the detection speed while remain the inspection precision.(5)An intelligent solder joint inspection method based on incremental cluster is proposed. The minimum distance classifier with mahalanobis distance is used to establish the initial solder joint classifier. To eliminate class overlaps, overlapped classes are divided into several sub-classes. The solder joints from different dispatches could be inspected with the classifier. When new samples which do not belong to existing classes, they are clusterd with the proposed incremental cluster algorithm. Representative samples from old samples are selected and trained with the new samples together to improve the training efficiency. The proposed method can improve the adaptability of the AOI to the production environment changes.Finally, concentration and academic future of the intelligent AOI algorithm was speculateed based on the summarization of the whole paper.
|
PDF Full Text Request