Abstract: To address the difficulties in diagnosing hidden faults in power system relay protection and the inability of traditional static setting of protection parameters to adapt to system dynamic changes, this study proposes a deep learning method that integrates fault diagnosis and adaptive parameter setting. The method first constructs a convolutional neural network to achieve accurate diagnosis of hidden faults such as CT (Current Transformer) open circuits, ensuring the basic reliability of the protection system. It then utilizes a deep feedforward network to dynamically optimize protection parameters based on real-time operating conditions, thereby enhancing protection sensitivity and speed. The research establishes a transformer differential protection current comparison model and a current ratio anomaly index as auxiliary criteria. Hardware-in-the-loop testing demonstrates that the proposed method can improve the system's sensitivity to high-impedance faults and inter-turn short circuits, shorten the short-circuit response time, and effectively prevent maloperation under non-fault conditions.