The accurate identification of smart meter (SM) fault types is crucial for enhancing the efficiency of operation and maintenance (O&M) and the reliability of power collection systems. However, the intelligent classification of SM fault types faces significant challenges owing to the complexity of features and the imbalance between fault categories. To address these issues, this study presents a fault diagnosis method for SM incorporating three distinct modules. The first module employs a combination of standardization, data imputation, and feature extraction to enhance the data quality, thereby facilitating improved training and learning by the classifiers. To enhance the classification performance, the data imputation method considers feature correlation measurement and sequential imputation, and the feature extractor utilizes the discriminative enhanced sparse autoencoder. To tackle the interclass imbalance of data with discrete and continuous features, the second module introduces an assisted classifier generative adversarial network, which includes a discrete feature generation module. Finally, a novel Stacking ensemble classifier for SM fault diagnosis is developed. In contrast to previous studies, we construct a two-layer heuristic optimization framework to address the synchronous dynamic optimization problem of the combinations and hyperparameters of the Stacking ensemble classifier, enabling better handling of complex classification tasks using SM data. The proposed fault diagnosis method for SM via two-layer stacking ensemble optimization and data augmentation is trained and validated using SM fault data collected from 2010 to 2018 in Zhejiang Province, China. Experimental results demonstrate the effectiveness of the proposed method in improving the accuracy of SM fault diagnosis, particularly for minority classes.