TAKAN: A Multimodal Attention Fusion Framework With Temporal-Adaptive KAN for Precise Fault Diagnosis in Axial-Flow Fans

Axial flow fans, as critical components within thermal power plant cooling systems, significantly influence power generation efficiency and operational safety integrity. To address the challenges of multi-modal data fusion difficulties and insufficient precision in axial flow fan fault diagnosis, this paper proposes a fault diagnosis framework based on multi-modal attention fusion and Temporal-Adaptive KAN (TAKAN). First, a multi-scale graph convolutional network (MS-GCN) is employed to construct feature relationship representations both within and between modalities, and dynamic attention mechanisms achieve power-adaptive multi-modal fusion. Next, recursive connections, memory enhancement, and Lyapunov stability constraints are introduced into the Kolmogorov-Arnold Network (KAN) to develop a temporal dynamics adaptive KAN structure with nonlinear system modeling capabilities, effectively capturing the dynamic evolutionary characteristics of axial flow fan faults. Experimental results demonstrate that the proposed method achieves a comprehensive diagnostic accuracy of 99.63% under three typical operating conditions, representing improvements of 4.43%, 5.53%, 5.90%, 7.38%, and 7.75% compared to AKAN, KAN, BiLSTM-attention, Transformer, TCN, respectively. When external disturbances are applied to the model, the TAKAN model exhibits the lowest performance degradation at only 0.49%, indicating its exceptional robustness and practical engineering significance.