This study utilizes the pre-processed fMRI data provided by the 2024 ICBHI challenge and excludes two other signal data types with missing information to construct a CNN model that can distinguish three emotion classes and their corresponding levels. However, due to the high variability and noise in the pre-processed fMRI data, a simple CNN model alone cannot achieve good classification performance. This study addresses the issue of noisy data by proposing a two-stage deep learning model training framework. In the first stage, an autoencoder method is adopted, leveraging its ability to effectively encode and decode data to extract useful signal features from the noisy data for use in the subsequent second stage. In the second stage, the effective features obtained by the encoder are transferred, and the weights of the encoding layers are combined with a fully connected layer for model retraining. This study also analyzes different methods of transferring the weights of the encoding layers. The best model for this study achieved an error rate of only 0.3383 on the official evaluation metric.