The rolling bearing is one of the most frequently faulty components in wind turbines. Accurate and effective bearing fault diagnosis methods can help ensure safe and stable operation. Bearing vibration signal characteristics are weak and difficult to diagnose, so a fault diagnosis method based on an improved denoising AutoEncoder is proposed. First, a one-dimensional signal imaging method to convert the original time domain signal into a two-dimensional feature grayscale image is introduced. Secondly, using the advantage of a convolutional neural network in image feature extraction, a combination model based on a stacked denoising AutoEncoder and convolutional neural network is proposed. The pooling layer in the traditional convolutional neural network is removed to ensure the robustness and generalization of extracted features. The overall diagnosis process is driven by data, reducing reliance on expert experience. Lastly, experimental results show that this method can accurately diagnose different types of bearing faults. Comparison experiments with other methods further verify the superiority of this method in fault diagnosis. This work is supported by the National Key Research and Development Program of China (No. 2017YFB0903403).