As the core component of modern electronic devices, the reliability of integrated circuits directly affects the operational safety and service life of the system. As manufacturing processes enter the 3 nanometer and below technology node, the types of faults caused by manufacturing defects, thermal aging, and radiation interference are becoming increasingly complex. Traditional diagnostic methods such as support vector machines and K-nearest neighbor algorithms, due to their weak adaptability to small sample data and limited feature extraction efficiency, are gradually unable to meet practical industrial needs. Therefore, this study designed a model that combines multi-scale convolutional neural networks with long short-term memory networks, called MCNN-LSTM. This model combines multi-scale temporal feature extraction, bidirectional temporal modeling, and data augmentation strategies to address the challenges of fault diagnosis in small sample and cross model scenarios. The experiment based on 5000 sets of test data of 14 nanometer process integrated circuits shows that the overall accuracy of the model reaches 98.7%, which is 5.2% higher than the classical convolutional neural network and 9.5% higher than the support vector machine; Especially in the recognition of single particle flipping faults caused by radiation, the F1 score reached 96.2% in small samples, and the accuracy remained at 89.3% in the migration diagnosis task across 7-nanometer process models. The above research results provide a feasible path for the automation and high robustness fault diagnosis of integrated circuits.
Wu et al. (Thu,) studied this question.