Ensuring the safety and efficiency of chemical processes requires robust Fault Detection and Diagnosis (FDD) systems capable of detecting and distinguishing abnormal conditions in complex industrial environments. Traditional model-based approaches often struggle with chemical plants nonlinear and dynamic nature, making data-driven methods a particularly appealing alternative. This thesis presents a two-stage deep learning framework for fault detection and diagnosis using the Tennessee Eastman Process as the case study. In the first stage, a Convolutional Neural Network trained on Gramian Angular Summation Fields representations is used for fault diagnosis, achieving state-of-the-art classification performance. We improved the diagnosis model s accuracy by systematically refining a single deep learning architecture instead of testing multiple alternatives. In the second stage, we extend this work to fault detection, reversing the conventional order of detection preceding diagnosis. We developed a detection model based on Siamese Neural Network architecture using transfer learning from the best-performing diagnosis model for its backbone. This approach allowed for a structured exploration of model modifications, including freezing convolutional layers to preserve transferred knowledge, optimizing hyperparameters for training stability, and analyzing gradient norms to enhance learning dynamics. The results were benchmarked against existing literature, showing the effectiveness of a focused architectural refinement strategy in industrial FDD applications and providing a foundation for further advancements in deep learning-based process monitoring.