Self-Gated Hierarchical Transformer for Robust Fault Detection in Chemical Processes

This study introduces a Self-Gated Hierarchical Transformer (SGHT) framework for industrial process monitoring. The proposed model combines hierarchical temporal representation learning with a self-gating mechanism that dynamically adjusts the contribution of individual variables. By selectively amplifying fault-relevant features while attenuating redundant or noise-dominated signals, the SGHT framework improves diagnostic reliability while maintaining a transparent decision structure. The effectiveness of the proposed approach is validated using the benchmark Tennessee Eastman Process (TEP). Comparative results show that SGHT achieves consistently superior performance when evaluated against conventional techniques such as principal component analysis (PCA), support vector machines (SVM), and recurrent neural network–based models. Notably, the model demonstrates clear performance gains for faults that have historically been difficult to detect, including Faults 3, 9, and 15. Beyond classification accuracy, the self-gating mechanism provides interpretable insights at the variable level, allowing practitioners to better understand which process variables contribute most strongly to specific fault conditions.