As one of the most critical components of a nuclear power plant, the reactor coolant system could lead to a significant disaster if a breakdown occurs. This paper introduces a fault diagnosis approach for reactor coolant systems based on symmetrized dot pattern (SDP) images and a depthwise separable convolutional neural network (DSCNN). By converting the fault signal into an SDP image, the distinct features of the signal are extracted and then input into a depth separable convolutional neural network model to identify specific fault types. Using the Adam Optimizer enhances the accuracy and reliability of the model. Experimental results demonstrate that this method achieves a diagnostic accuracy of 98.9% using the training samples. Compared with other conventional depth learning methods, the proposed approach exhibits superior accuracy and stability.