The Assessment of Deep Learning-Based Defect Detection Models in Electronic Components

Abstract

The paper presents the implementation of deep learning-based approaches from one-class classification methodologies to identify defects in a transistor in the electronics sector. No labelled data exists, and hence the present work utilizes "good" data to identify bent leads, cut leads, damage areas, and misplacement. The paper is concerned with creating an auto defect detection system capable of increasing effectiveness and precision in quality control at a more reasonable cost. Results indicate the potential to hasten defect detection in industrial practice that employs machine learning. This paper proposes a defect detection system in electronics practice through ResNet50, EfficientNetB0, and InceptionV3 models proposed in Google Colab, TensorFlow—Keras and Python programming language. Defects in the data are of diverse types such as bent lead, cut lead, damaged case, and misplaced. Performance metrics such as accuracy, precision, recall, F1-score, and AUC-ROC are observed. Results indicate that the model of EfficientNetB0 registers the most remarkable overall performance with the highest precision and generalizability, followed by ResNet50 and then by InceptionV3. All three models indicate difficulties in uncovering minority class defects due to the unbalance of the data, and hence balanced sets of data are of vital importance in creating robust auto defect detection systems. Results indicate the capability of balanced sets of data to create robust auto defect detection systems with potential accelerations of quality control in industry practices. Future studies are to rectify the data unbalance through oversampling, undersampling, and data augmentation to make these models more efficient and the auto defect detection system more trustworthy.