Adaptive marginalized stacked denoising autoencoders and its application

In this paper, we propose a modified marginalized autoencoders. Here, the noise adding way at a fixed rate in marginalized autoencoders is replaced by the adaptive noise injection. Compared with the traditional marginalized autoencoders, the proposed method obviously enlarges the recognition performance. Furthermore, the proposed method is applied to identify high-speed train wheel wear conditions. Features of high speed train wheels wear vibration signals are abstracted by using the adaptive noise marginalized autoencoders, and the features is used to realize the wheel wear characteristics of vibration signal recognitions as the input of support vector machine (SVM). The experimental results show that the accuracy of the new method for identifying high-speed train wheel wear conditions is 99.8% on average.