Attachment Fault Diagnosis for Tidal Stream Turbine based on Improved ShuffleNet_v2

Tidal Stream Energy, which is clean and renewable, is an effective way to solve the depletion of nonrenewable resources. Blades of tidal stream turbine (TST) are the crucial part of power generation, but the fault rate is high. Aiming to solve blades’ attachment fault and large calculation of neural network, an attachment fault diagnosis method for TST based on improved ShuffleNet_v2 is proposed. Firstly, the blades image of TST is preprocessed by normalization and centralization. Secondly, the square convolution is decomposed by singular value decomposition (SVD) to decrease the amount of convolution calculation. Then, h-swish, whose effect is better than ReLU, is used as the activation function to improve ShuffleNet_v2, the novel structure is utilized to train the preprocessed image and the trained model is saved. Finally, the trained model is applied to diagnose the blades’ attachment. Experimental results show that the proposed method can effectively reduce parameters calculation, requirement of hardware performance, and the accuracy of blades’ attachment diagnosis is higher than some lightweight neural networks.