VAE-Driven Feature Learning with Context-Aware Attention and ResNet Gradient Optimization: Applied to Industrial Hydrocracking Processes

This study proposes a deep learning model that combines a Variational Autoencoder (VAE), a Residual Network (ResNet), and a Squeeze-and-Excitation Network (SENet) for nonlinear modeling of complex industrial devices and, for the first time, applies the channel attention mechanism (SENet, Squeeze-and-Excitation Network) to hydrocracking modeling. This model extracts the latent low-dimensional feature space of input data through the VAE, and it combines the powerful feature extraction capability of ResNet with the adaptive feature-weight allocation mechanism of SENet to achieve efficient modeling of complex systems. Compared with traditional methods, this network structure has improved prediction accuracy by 2.4% in data-driven modeling of hydrocracking. The experimental results show that the model can effectively handle high-dimensional nonlinear data, demonstrates a superior performance in the modeling of hydrocracking units, and has wide applicability to common nonlinear conversion processes.