A hybrid DiGLPP–Kolmogorov–Arnold Network framework for robust fault detection in industrial chemical processes

The increasing adoption of Industry 5.0 paradigms, characterized by pervasive automation, large-scale sensor deployment, and tightly integrated cyber-physical structures, has significantly amplified the operational complexity of modern industrial chemical processes. These systems generate high-dimensional, nonlinear, and dynamically evolving data streams, posing substantial challenges for reliable fault detection. Conventional monitoring approaches often exhibit limited robustness when faced with nonlinear process behavior, dynamic correlations, and imbalanced operational datasets, leading to degraded detection accuracy and delayed identification of abnormal conditions. To address these challenges, this study proposes a hybrid dynamic framework for robust fault detection in complex industrial chemical processes. The anticipated methodology integrates Dynamic Inner Global–Local Preserving Projection (DiGLPP) for structure-preserving manifold learning with Kolmogorov–Arnold Networks (KAN) for adaptive nonlinear feature inference. By combining graph-guided dimensionality reduction with flexible functional approximation, the framework effectively captures both global process dynamics and local structural dependencies, enabling the extraction of informative representations for reliable fault detection. The efficiency of the anticipated (DiGLPP-KAN) framework is systematically assessed using two benchmark industrial systems: the Two-Stage Esterification process (TSEP) and the Tennessee Eastman Process (TEP). Comparative studies are conducted against established monitoring methods, including Wavelet-PCA and CWT-3D-CNN, to assess detection capability under complex operating conditions. Experimental results demonstrate that the proposed DiGLPP–KAN framework achieves improved robustness and detection performance in identifying abnormal process behavior, particularly in scenarios involving nonlinear dynamics and data imbalance. For the TSEP, the proposed method achieved average fault detection rates of 97.21% and 97.63% with average false alarm rates of 0.71% and 1.35% based on the T2 and SPE statistics, respectively. For the representative TEP fault cases 10 and 15, the framework achieved 100% fault detection rate and 0% false alarm rate for both monitoring indices. These results demonstrate that the proposed DiGLPP–KAN framework provides improved robustness, stronger detection capability, and more reliable monitoring performance under complex industrial operating conditions. These findings highlight the potential of integrating manifold-preserving projections with Kolmogorov–Arnold functional networks as a promising direction for next-generation intelligent monitoring systems in industrial chemical processes.