Abstract: Primary equipment of hydropower station is the core component of the power system, and its operation status directly affects the stability and security of the power grid. The traditional fault diagnosis method has limited effect in the face of complex equipment faults, and it is difficult to meet the demand of the power system. In this paper, a fault diagnosis method for primary equipment of hydropower station based on improved convolutional neural network is proposed. By introducing the Retinex algorithm to enhance the infrared image of the equipment, combining the cross-entropy function to construct a deep convolutional denoising self-encoder for data dimensionality reduction, and utilizing the convolutional neural network to determine the mapping relationship between the fault features and types. The experimental results show that the diagnostic accuracy of this paper's method for different fault types is always above 95%, and the training time is controlled within 4 minutes, which is significantly better than the traditional method.