The Research And Application Of Deep Neural Network Algorithm

In order to discover the distributed feature of the data, deep neural network algorithm combines low-level feature, which forms more abstract high-level attribute and feature. And deep neural network algorithm is a very hot research field of artificial intelligence and big data. However, there are some deficiencies in the auto-encoder of deep neural network algorithm, such as time-consuming and random zero set in training, which leads to some key information lost. Thus, the paper mainly studies three kinds of deep neural network algorithm deeply, and proposes deep auto-encoder model based on compressed sensing and adaptive noise marginalization deep auto-encoder model. The proposed methods overcome random noise injected and costing lots of training time. Firstly, in order to verify improved methods, we use the standard MNIST handwritten digital datasets and its variants datasets. Then we apply the proposed methods for analyzing the wheel wear states of high-speed train. The experimental results show that the methods can automatically extract the key feature information of the wheel wear vibration data effectively, which guarantees operation of high-speed train safely and smoothly. The main work of the thesis is as follows.1. For the disadvantage of stacking denoising auto-encoder, the random noise injected, thus we propose deep auto-encoder method based on compressed sensing. The improved method solves the random noise injected, and improves the precision of the algorithm. Through theoretical analysis for the resistance noise and stability, and we conduct comparative analysis experiment on the standard MNIST datasets, which verify the effectiveness and veracity of the algorithm.2. For the disadvantages of the auto-encoder, costing lots of training time for high-dimensional data and fixed noise injected, thus we propose adaptive noise marginalization deep auto-encoder method. The improved method solves the deficiency of the identification precision and shortens the training time of the model. The standard MNIST datasets and its variants datasets are applied for method validation experiments. Experiments show that the method not only greatly shortens the model training time, but also overcome the limitations of fixed noise injected. Those methods achieve better results in the digital recognition for the standard MNIST datasets.3. For the problem of wheel tread wear, in order to evaluating the wheel tread wear state of high-speed train, deep neural network and the improved algorithm is firstly applied for recognizing wheel wear state of high-speed train. Then, we conduct experiments for the actual wheel wear monitoring data of high-speed train. The experimental results show that those methods can effectively identify the wheel wear state of high-speed train, and can provide a new idea for the wheel wear monitoring of high-speed train.The research is supported by the Key Project of National Science Foundation of China with monitoring data of high-speed train service security state assessment of the key questions (No.61134002).