<p>A stator winding inter-turn short-circuit fault is one of the typical faults in non-salient pole synchronous generators (NSPGs). Mild inter-turn short-circuit faults are typically difficult to detect owing to their inconspicuous fault characteristics and negligible impacts on electrical parameters. To address the challenges of early diagnosis and precise localization of mild inter-turn short-circuit faults in NSPGs, this study proposes an online monitoring method based on stator leakage magnetic field (LMF) analysis. By constructing a field-circuit coupling model, this study reveals the mapping mechanism between mild inter-turn short-circuit faults and stator LMF distortion, and the effectiveness of the fault features is verified via multi-physics simulation. For fault detection, this study adopts a magnetic field sensing system with six magnetic flux sensors uniformly arranged along the stator circumference to collect radial magnetic flux leakage (MFL) signals in real time under different operating conditions. Based on wavelet packet multi-scale decomposition (WPMD) technology, and extracts the energy of high-frequency components and the differences in time-frequency domain characteristics, and a fault sensitivity factor is constructed to achieve rapid fault detection. A multi-domain diagnosis model is established by fusing wavelet time-frequency features with a bidirectional long short-term memory (BiLSTM) network to improve the robustness of fault identification. For the fault spatial localization algorithm, a magnetic field gradient change rate analysis method is proposed for fault localization. By analyzing the amplitude-phase differences and spatial harmonic distribution features of the magnetic field from the sensor array, the quantitative relationship between the fault location and the magnetic field gradient vector (MFGV) is established, and the localization error is controlled within three stator slots. Experimental results show that the proposed method can effectively identify inter-turn short-circuit faults with a short-circuit turn ratio (SCTR) as low as 1%, and its sensitivity is 90% higher than that of the conventional harmonic analysis method. The proposed time-frequency-space multi-domain feature fusion (TFS-MDFF) method provides technical support for the online monitoring and intelligent diagnosis of inter-turn short-circuit faults in synchronous generators.</p>