Rotor imbalances present a serious problem for wind turbines. In particular, for offshore wind turbines, aerodynamic imbalance can have a severe impact because of the large rotor size. In this study, the impact of the aerodynamic imbalance is investigated. A novel framework for detecting aerodynamic imbalance is proposed. Firstly, a model of a 3MW direct-driven wind turbine was developed. The signals were acquired to test and verify the impact of aerodynamic imbalance. Secondly, a method based on optimized maximum correlated kurtosis deconvolution was proposed for the primary detection. The intrinsic mode functions of nacelle vibration were adopted as the input variable. The weak unbalanced signals could be discerned. Moreover, the azimuth of rotor allows the unbalanced blades to be obtained. Thirdly, a convolutional neural network with a new structure was used to determine the magnitudes of aerodynamic imbalances. The first layer of the convolutional neural network is sufficiently wide for improving feature extraction, it could make nacelle acceleration as the input. This structure exhibits accuracy and robustness satisfactorily. Finally, the framework was demonstrated in a high-fidelity simulation environment. Different scenarios of aerodynamic imbalance were tested, the results demonstrate the satisfactory performance of the proposed framework.