With the development of industry and science and technology, users" attention to voltage dips is increasing, and it is more and more important to identify the causes of voltage dips. In order to identify the single and composite sources of voltage dips, this paper proposes to combine the S-transform feature extraction and deep learning automatic feature extraction to establish S-CNN-LSTM model for the identification of dips. Firstly, the numerical modeling framework is used to generate single and composite transient source mathematical models, and then the transient dataset of 9 fault categories is obtained as the original data, and secondly, the original data is subjected to the improved S-transform, i.e., the standard S-transform is introduced to obtain the improved S-transform by introducing two adjusting factors to obtain the S-transformed data, and the S-transformed data is obtained, and the S-transformed data is introduced to the 16 indexes to perform the feature extraction, which are used as the indexes data, and then the CNN-LSTM is constructed. Then CNN-LSTM model is built, the above data are used as model inputs, CNN network is used to extract spatial features from the transient data, and at the same time, the data are divided into multiple one-dimensional vectors and input to Bi-LSTM to extract temporal features, and then a multi-feature fusion recognition model of indicator features, spatial features and temporal features is built. The simulation results show that the recognition accuracy without feature fusion and after multi-feature fusion are 98.36% and 99.08%, respectively, indicating that multi-feature fusion can improve the recognition accuracy.