Abstract: Aiming at the problem that the time-frequency domain characteristics of the acoustic emission signal of the hydraulic turbine do not change much when compared with the normal time when cavitation occurs, which leads to the low accuracy of the fault diagnostic model, a fusion diagnostic method based on spectral subtraction and convolutional neural networks (CNN) is proposed. Firstly, the dual-signal mutual subtraction of the cavitation and normal signals by spectral subtraction achieves the differential enhancement of the signals and significantly improves the separability of the band features. Then the processed normal and cavitation signals are used as samples and input into the lightweight CNN model for training, and finally the turbine cavitation fault diagnosis results are obtained. In order to verify the feasibility of the method, experiments were carried out using multi-case acoustic emission data from the top cover of the Francis turbine and the measuring points of the worm shell. The results show that the average diagnostic accuracy of the proposed model reaches 99.56% (top cover) and 99.81% (snail shell), which is a significant improvement over several other signal processing methods, and the proposed model is more accurate as well as computationally efficient. The method combines high accuracy and adaptability and provides an effective reference scheme for turbine cavitation monitoring.