Abstract: With the high penetration of distributed renewable energy, the fault characteristics of distribution networks have become increasingly complex, rendering traditional fixed-threshold fault detection methods inadequate. This paper proposes an Artificial Intelligence-based adaptive fault diagnosis and self-healing system for distribution networks. The system employs intelligent monitoring units deployed at key line nodes to synchronously collect transient waveforms and steady-state power frequency data, constructing a multi-source fault characteristic database. A hybrid model combining convolutional neural network (CNN) and long short-term memory (LSTM) networks is utilized for meter-level precise location and fault type identification. while an improved double-ended traveling wave ranging algorithm achieves meter-level precise location. Practical operation cases demonstrate that the system can effectively distinguish lightning strikes from non-lightning faults, accurately identify transient and permanent faults, with an average location error of less than 50 m, significantly enhancing the intelligence and power supply reliability of the distribution network.