A similarity-based diagnosability analysis method for gas turbines is proposed to aim at the problems of poor data stability and low diagnostic efficiency caused by noise interference and feature redundancy in the gas turbine signal monitoring process. Firstly, the overall trend of the operation data is obtained by polynomial fitting, the noise caused by local mutation is eliminated, and the noise threshold is calculated to solve the noise interference problem in the monitoring signal. Then, the time-series similarity algorithm is used to calculate the similarity of the data under different operating conditions of the gas turbine, obtain the independence features, construct the diagnosability evaluation matrix and the fault feature matrix, and screen the optimal features of the gas turbine to combine the fault features, to solve the problem of redundancy of the monitoring signal features. Finally, the optimal combination of filtered features is used as the input of the convolutional neural network to establish the fault diagnosis model. The proposed method is experimentally verified using typical gas turbine fault data, and the results show that the proposed method effectively removes the redundant features, has good fault isolation, and achieves an accuracy of 99.98%, which improves the diagnostic efficiency by 32.6% relative to the original features.