The wear fault of the inner and outer race of bearing in a wind turbine generator system is vital as the performance of bearing effects the transmission efficiency. The accelerometer is not suggested be installed inside the wind turbine generator system since it would damage the structural reliability. This paper proposes a load demodulation normalization framework to detect the wear fault of bearing from electricity-related signals. First, according to the mathematical model of the generator in the three-phase stationary coordinates, this paper selects the stator current as the monitored signals. Second, synchronize the sampled signals in the time domain with synchronous re-sampling. It improves the definition of time-frequency representations (TFRs) of wear fault, and thus avoids the phenomenon of difficulty to determine the resolution of TFRs caused by load fluctuations. The short-time Fourier transform (STFT) is then applied directly to convert the angularly spaced signals into the TFRs. Finally, to improve the precision of classification, this paper proposes an adapted convolutional neural network (CNN) with dropout optimization to classify the wear of bearing. The proposed framework is verified on the motor drive-train platform. The experimental results show that the proposed method has a higher fault detection efficiency than the other methods.