As social media becomes increasingly prevalent, the volume of user-generated text has been rapidly increasing, and emotional expressions have become more complex and diverse, placing higher demands on sentiment analysis tasks. Traditional methods face limitations in both feature extraction and data acquisition, especially in scenarios involving implicit semantics or data scarcity, where model performance is often compromised. To tackle the aforementioned problems, this study presents a novel sentiment classification model named BERTFAN, which integrates multiple advanced techniques. The model combines GPT-driven data augmentation, multi-layer semantic feature fusion based on BERT, and an innovative feature processing module called LECE, aiming to enhance sentiment recognition from both data and model perspectives. Specifically, ChatGPT is used to automatically generate semantically rich augmented samples to expand the training set and improve generalization; then, multi-layer hidden representations from BERT are fused and refined through a multi-head attention mechanism to enhance cross-layer semantic representation; finally, the LECE module integrates the contextual modeling capability of RNN with the local feature extraction strength of CNN, further improving the depth and breadth of feature learning. Results demonstrate that BERTFAN achieves state-of-the-art performance on several benchmark sentiment analysis datasets, including online_shopping_10_cats (96.21% Acc/95.10% F1), SMP2020-EWECT (79.44% Acc/76.12% F1), and Weibo_senti_100k (98.22% Acc/98.10% F1), demonstrating the efficiency and advantages of the introduced strategies in data augmentation and feature fusion.