Abstract: Power transformer is one of the key equipment to ensure the safe and stable operation of the power system, but the existing fault diagnosis methods cannot fully exploit the feature interaction within the equipment and have poor sensitivity to the changes of operating conditions, which has limited the improvement of fault diagnosis accuracy and reliability. To address the above problems, a transformer fault diagnosis method is proposed based on bi-directional long short-term memory (Bi-LSTM) considering feature coupling. Firstly, the initial transition sequence of feature state is determined based on the equipment operation mechanism; then, a deep neural network fault diagnosis model is constructed with consideration of complex dependencies to mine the feature coupling relationship for refined condition assessment; finally, the simulation results have verified the support role of the priori feature sequence for the fault diagnosis model. The proposed method improves the fault diagnosis effectiveness, and can provide a reference solution for intelligent and refined maintenance of power equipment.