[Objective] Inter-turn short circuit is the most frequent type of fault in distribution transformers, and if a minor inter-turn short circuit is not dealt with in time, it may evolve into a serious inter-turn short circuit, endangering the stable operation of the transformer. This study aims to explore a high-precision diagnostic method capable of effectively identifying incipient faults to achieve early warning. [Methods] Firstly, for the early short-circuit condition, a method using nonlinear arc characteristics to replace the impedance at the short-circuit point was proposed. A high-impedance, low-energy arc was used to simulate an incipient inter-turn short circuit in the transformer. The winding current signals before and after the fault were extracted, and the distortion characteristics of magnetic flux variation was analyzed. Subsequently, a multi-scale residual network model based on the attention mechanism was proposed, achieved efficient identification of interturn short-circuit faults. [Results] Simulation and experimental results indicated that the fault current obtained through nonlinear arc impedance simulation was significantly smaller than that from linear impedance simulation. The inter-turn short circuit led to an increase in the leakage flux density within the air gap between the core and the windings, resulted in an overall rise in electromagnetic losses. The method proposed in this paper was effectively able to distinguish between the three states: normal operation, minor inter-turn short circuit, and severe inter-turn short circuit, achieving a test accuracy of 0.975. [Conclusion] The research can effectively reveal the electromagnetic response characteristics of the transformer’s early inter-turn short circuit, and lay a theoretical reference for the early fault detection.