Quantitative Detection of Blade Crack Damage Based on Vibro-Acoustic Information and Multi- Dimensional Feature Fusion CNN

Aiming at the inherent risk of being misled by noise in the blade damage signal of single-mode sensors and the low accuracy of model diagnosis, a quantitative detection method for blade crack damage based on the fusion of vibro-acoustic information and multi-dimensional feature CNN is proposed. The method extracts one-dimensional and two-dimensional features from the frequency domain and wavelet time-spectrum of sound and vibration signals. Firstly, the one-dimensional and two- dimensional features of sound and vibration signals are extracted by convolutional layers with four channels respectively. Secondly, a fully connected network is applied to fuse and classify the extracted multi-sensor multi-dimensional features. Finally, to improve the computational efficiency and generalization ability of the model under variable rotate speed, the transfer learning of the model under other rotate speeds is carried out combined with the parameter transfer mechanism, which makes up for the weak generalization ability and robustness of the model. Verified by the wind turbine data, compared with the single-modal sensor and single-dimensional feature diagnosis model, the model proposed in this paper has higher accuracy and stability for quantitative detection of blade damage.