With the rapid growth of the photovoltaic industry, fire incidents in photovoltaic systems are becoming increasingly concerning as they pose a serious threat to their normal operation. Research findings indicate that direct current (DC) fault arcs are the primary cause of these fires. DC arcs are characterized by high temperature, intense heat, and short duration, and they lack zero crossing or periodicity features. Detecting DC fault arcs in intricate photovoltaic systems is challenging. Hence, researching DC fault arcs in photovoltaic systems is of crucial significance. This paper discusses the application of mathematical morphology for detecting DC fault arcs. The system utilizes a multi-stage mathematical morphology filter, and experimental results have shown its effective extraction of fault arc features. Subsequently, we propose a method for detecting DC fault arcs in photovoltaic systems using a cyclic neural network, which is well-suited for time series processing tasks. By combining multiple features extracted from experiments, we trained the neural network and achieved high accuracy. This experiment demonstrates that our recurrent neural network (RNN) based scheme for DC fault arc recognition has significant reference value and implications for future research. The ROC curve on the test set approaches 1 from the initial state, and the accuracy on the test set remains at 98.24%, indicating the strong robustness of the proposed model.