To solve the problem that the contact measurement method is susceptible to the limitation of the sensor acquisition channel and the additional weight, a wind turbine blade crack diagnosis method based on three-dimensional (3D) vibration information fusion and convolutional neural network is proposed. First, based on the principle of binocular photogrammetry, a multi-channel sample construction method of 3D vibration information fusion is proposed. This method can integrate the motion information of multiple measurement points on the surface of the wind turbine blade, gain the acquired signal with more abundant features, and greatly decrease additional weight interference. Secondly, in order to obtain multi-level semantic information of cracks, a new multi-scale convolutional neural network is proposed. A type of 1.5 kW wind turbine blade was selected to carry out crack diagnosis experiments, and a database of samples of different crack states was established. The prediction accuracy reached 93.4%, which verified the effectiveness of the proposed method. Comparative analysis with the classic LeNet-5 and VGG-11 networks shows that the improved convolutional neural network has higher identify precision and faster convergence speed. Multi-channel signal samples can offer a better effect in wind turbine blade crack fault diagnosis application.