The GIS device has been widely used in substation because it has the advantages of small pollution, small area and excellent performance. The safe operation of GIS devices is very important to ensure stability of the power grid. With the continuous expansion of GIS quantity, various GIS failures have also increased. But because the GIS key components are closed in the metal shell, it is difficult for maintenance personnel to find failures. The traditional methods for degradation identification of GIS devices are relatively complicated, and check the breakdown slowly. Single feature quantities are prone to errors and missed reports. This paper presents a method of GIS equipment degradation recognition based on multi-modal feature extraction and convolutional neural network(CNN). Firstly, extract multi -characteristic quantity of each discharge phase based on the PRPD diagram analysis is performed. According to the basic characteristics of typical defect vibrations, the multi-modal feature of mechanical defects are extracted through frequency proportion and amplitude value. Extract 720-dimensional features and compress them to 20-dimensiona,it can greatly reduces the time of fault recognition. Then, the CNN suitable for the above mentioned multi-modal feature identification is designed. Use CNN to extract features with fault recognition to achieve adaptive binding to achieve the deterioration and identification of GIS equipment, and the accuracy rate is more than 90%. Experiments show that comparison with methods such as SVM, random forests, BP neural networks, etc., this article has a strong robustness and higher recognition accuracy and efficiency for the different infusion state of GIS devices.