To address the fault identification challenges caused by high-dimensional coupling and intense noise interference in chemical processes, a fault diagnosis method based on dual-stream time-frequency feature fusion and Kolmogorov-Arnold Network is proposed. First, a dual-stream parallel architecture is constructed, utilizing multi-scale convolution to capture time-domain transient features, combining an autoencoder and continuous wavelet transform to reconstruct frequency-domain energy spectrograms in a low-dimensional space, and achieving the adaptive fusion of heterogeneous features and amplification of weak signals through a squeeze-and-excitation mechanism. Second, the Kolmogorov-Arnold network (KAN) is introduced to replace fully connected layers, utilizing its edge-learnable B-spline functions to enhance nonlinear boundary fitting capabilities and decouple the complex coupling among variables. Finally, a two-stage progressive optimization strategy based on label smoothing and a weighted focal loss function is designed to execute hard example mining and boundary fine-tuning to suppress misjudgments. Experiments on the Tennessee Eastman process demonstrate that this method possesses excellent anti-noise robustness in diagnosing concealed and overlapping weak faults, with accuracy significantly outperforming mainstream deep learning models.