Deep learning has been widely used in the field of fault diagnosis, among which convolution neural networks (CNN) and long short term memory networks (LSTM) are typical models. As a common method of multi-sensor signal fault diagnosis, CNN can obtain better diagnosis effect, but it can not realize the diagnosis of unknown composite fault. In order to solve this problem, a CNN-LSTM-FCM (fuzzy C-means) model is proposed. LSTM is more sensitive to the time signals with the connection between the front and the back. By using this feature, LSTM and CNN are combined to realize the diagnosis of unknown signals. The decoupling of composite fault is realized through the probability classification output. The CNN-LSTM-FCM model itself has optimized parameter design, which further improves the diagnosis accuracy. The chemical process fault measurement data is used for experiments, the results show that the diagnostic accuracy of CNN-LSTM-FCM model can reach 97.15%, which is superior to CNN model and LSTM model, and has a high application value.