Improvement Of Stacked Auto-Encoders And Its Engineering Application

The field of machine learning has fueled by the power of massively parallel chips and big data sets,and achieved the development from shallow learning to deep learning.The shallow learning models not only require specific skills and prior knowledge in the application process,but also seem inadequate due to its limited model capacity under the situation of mass data and complex distributions.In comparison,deep learning models gain capacity by increasing the depth and breadth of models(e.g.the universal approximation property of NNs),and benefited from the effective representation learning algorithms,so that the learning of complex distributions becomes possible.On the other hand,with the radio frequency identification,sensors,storage media and other technical equipment widely used,the manufacturing industry has gradually entered the era of big data,and intelligent manufacturing has become an inevitable trend.The powerful deep learning will also be combined with the shallow learning,and evolves the manufacturing data mining and analysis.The research of this thesis is based on the auto-encoder,stacked auto-encoder and its applications in fault diagnosis.Firstly,stacked auto-encoder is improved to aim at the small dataset characteristic of manufacturing data.The Frobenius norms of the Jacobian matrices of the sample points on the representative surface are taken as the penalty,which is added to the cross entropy loss function of the model.The goal is to enhance the smoothness of the representative surface while reducing the training error,thereby reducing the complexity of the overall model in order to prevent overfitting.Eventually,we get a more smooth,robust contractive stacked auto-encoder.Secondly,the auto-encoder is improved to aim at the correlation between manufacturing data sets: a weighted auto-encoder model with different important features is proposed,and the feature weight vector is calculated by particle swarm optimization algorithm.Then,by stacking the weighted auto-encoders,stacked weighted auto-encoder is obtained to balance the importance of different dimensions of the data set and to transfer the validity information learned by a data set to the application of another data set.Finally,the stacked version of auto-encoder,denoising auto-encoders,and sparse autoencoder,alone with two improved models proposed in this thesis are applied to the rolling element bearing vibration signal fault diagnosis.The results showed the high performance of two improved models.At the same time,through the analysis of the contractive stacked autoencoder model and the stacked weighted auto-encoder model,the coincidence between the model performance and the design idea is verified,which also mean that two improved models can solve the datasets with the corresponding characteristics properly.