Based on spectral subtraction and convolutional neural networks (CNN)

this study develops a fusion diagnostic method to improve previous fault diagnostic model with low accuracy in the time and frequency domain characteristics of acoustic emission signals from a hydraulic turbine experiencing cavitation. Firstly

a differential enhancement of the signals through dual-signal mutual subtraction of the cavitation and normal signals was achieved by using spectral subtraction

which significantly improved the separability of band features. Then

samples of the processed normal and cavitation signals were input into a lightweight CNN model for training

and finally the turbine cavitation fault diagnosis results were obtained. To verify this method

we conduct a numerical experiment using multi-case acoustic emission data from the measuring points on the top cover and spiral case of a Francis turbine. The results showed an average diagnostic accuracy as high as 99.6% at top cover and 99.8% at spiral case

which was a significant improvement over several other signal processing methods with accuracy

adaptability and computational efficiency enhanced

putting forward an effective scheme for hydraulic turbine cavitation monitoring.