The photovoltaic (PV) DC system is a core component of new energy generation

yet it poses potential arc fault hazards during operation. Triggering an alarm during the initial weak stage of an arc can significantly reduce the subsequent risk of fire. However

the characteristics of weak arcs are highly concealed

and traditional static threshold methods suffer from high false alarm rates

making it difficult to achieve timely and accurate identification of early-stage arcs. To address this issue

this paper proposes an adaptive real-time arc diagnosis method based on an orthogonal perception and protection agent. The decoupled “perception-decision” architecture enables precise perception and rapid response to initial arcs under extreme operating conditions. First

an orthogonal attention convolutional neural network (OA-CNN) is employed to rapidly perceive the 15 ms current window and output the instantaneous arc probability for a single window. Second

a protection agent based on dueling double deep Q-network (D3QN) and prioritized experience replay (PER) is constructed to make adaptive tripping decisions using a continuous sequence of fault probabilities as the state input. Test results demonstrate that the proposed OA-CNN framework achieves a comprehensive accuracy of 97.94%. Furthermore

deployment tests on the Jetson Orin Nano embedded platform show that the single-window end-to-end average inference time of the overall OA-CNN+D3QN model is 2.106 1 ms

which is less than the 15 ms sampling window

indicating its potential for real-time deployment on edge devices. Simulated real-time diagnosis results indicate that this method can act before a stable arc is formed and maintain zero false alarms under typical interference

demonstrating significant engineering value for the reliable operation of PV systems.