With the rapid proliferation of devices, deep learning-based fault detection methods have attracted significant attention in industrial applications. Among these, long short-term memory network (LSTM)-based approaches have been extensively explored for fault detection in chemical processes. This article proposes a fault detection method for industrial processes using residual gated mLSTM and dynamic feature fusion. Firstly, a hybrid model integrating parallel LSTM architecture and spatio-temporal attention mechanisms is proposed to enhance the capture of temporal dependencies and spatial correlations in industrial process data. Simultaneously, residual gating is designed to mitigate the vanishing gradient problem. The proposed fault detection algorithm is validated on the TE Process benchmark, demonstrating its effectiveness through experimental results in an industrial process context.