The invention discloses a fault diagnosis method based on deep learning and signal analysis. The method includes: acquiring data in normal and faulty states of an industrial process in advance, and dividing the data into a training set and a test set; and training model parameters offline based on the training set, detecting a model through the test set, wherein a performance index refers to the precision of fault diagnosis, and a value thereof represents the generalization performance of the model, namely the online diagnosis capability of faults. According to the method, as a variant of a neural network, physical information of a process operation variable in a time domain can be obtained, and frequency domain information of a process measurement variable can be obtained through introduction of a wavelet analysis method; besides, a depth structure adopted by the method adapts to big, fast, various, and uncertain characteristics of industrial big data, the physical information of theprocess operation variable and frequency characteristics of the process measurement variable are combined, a complex mode of a deep grade of the faults is learned, fault diagnosis can be effectively realized, and excellent generalization capability is displayed in an online diagnosis test.