Abstract: Excitation system faults pose a serious threat to the secure and stable operation of power systems, while commonly used discrete-time data-driven models exhibit clear limitations for early fault detection in multi-bridge paralleled excitation systems. Relying on fixed sampling sequences, such models struggle to capture the fine evolution of continuous electromagnetic processes and the spatial coupling among multiple measurement points, and are sensitive to noise and missing data, making weak fault signatures and multi–time-scale degradation processes difficult to detect in time. To address these issues, this paper, for the first time, introduces liquid neural networks (LNNs) into excitation system fault detection and proposes an enhanced liquid neural network model (E-LNN) tailored for multi-bridge paralleled excitation devices. The model describes neuron state evolution in continuous time via ordinary differential equations and incorporates three enhancement mechanisms spatio-temporal coupling attention, a multi-scale variable time-constant scheme, and fault-sensitive sparse connectivity—to jointly characterize electromagnetic coupling among bridges, dynamic features ranging from microsecond-level transients to hour-level slow degradation, and the distinctiveness of abnormal patterns. Experimental results show that E-LNN achieves a detection accuracy of 94.83% on a dataset covering 12 fault modes, outperforming the second-best method by 2.69%. Under heavy noise and severe missing data, it still maintains stable performance, effectively overcoming the shortcomings of traditional discrete-time models in robustness and early fault capture and demonstrating strong potential for engineering applications.