A robust multi-stage information fusion framework incorporating novel IGAnet for enhanced leak detection in fuel cells with tolerance to abnormalities

Information fusion based on reliable monitoring signals represents a critical strategy for enhancing the accuracy and robustness of hydrogen leak diagnosis models. This study presents a coupled hydrogen leakage (HL) diagnosis model that integrates multi-stage information fusion and abnormal signal inspection and reconstruction using a Gaussian Process Regression (GPR) model. Thereafter, handcrafted features for the Support Vector Machine (SVM) and multi-dimensional image features—including Markov Transition Field (MTF), Recurrence Plot (RP), and Gramian Angular Summation Field (GASF)—for the proposed Improved Global Average AlexNet (IGAnet) are extracted concurrently. These image features are integrated using adaptively weighted feature-level information method to enhance representation quality while reducing model complexity. Moreover, by leveraging the complementary strengths of both SVM and IGAnet, a decision-level information fusion strategy based on Dempster–Shafer (D-S) evidence theory is employed to combine their outputs to derive the definitive diagnostic result. Experimental results show that the proposed method achieves 94.37 % diagnostic accuracy under various working conditions, even in the presence of abnormal monitoring signals, thereby confirming its robustness and effectiveness.