| With the rapid growth of data, the intelligent analysis and prediction for the massive data plays a more and more important role in the social life. Therefore various intelligent data analysis and prediction systems are generated. An intelligent data analysis and prediction system usually consists of data acquisition module, dimensionality reduction module and prediction module etc. In the system, the data is processed by the dimensionality reduction module to reduce the interference factors which affect the accuracy and the computation complexity of prediction module, and the prediction module is used to predict. However, in the most existed intelligent data analysis and prediction systems, the dimensionality reduction module and prediction module are constructed sequentially and independently, which may cause certain mismatch. Focusing on the mismatch of the dimensionality reduction module and prediction module, we propose a new framework called type 2 fuzzy deep belief network. Deep Belief Network is utilized as a dimensionality reduction module widely, which can learn the feature representation gradually and reduce the noise, non-linearity, and uncorrelated information of the raw data, and improve the prediction accuracy. However, in the features after dimensionality reduction, there are still certain uncertainties. And the Type 2 Fuzzy Logic System can handle thesecharacteristics conveniently. The main contents of the thesis are given as follows:1. Deep belief network is used to extract the features from the raw data. The training process of it is as follows: the restricted Boltzmann machine is pre-trained by contrastive divergence algorithm in an unsupervised way, which obtains a slightly good initial weights and biases; and the network is fine-tuned by backpropagation in a supervised way. In the condition where the label is hard to obtain, the fine-tuning can be conducted using part of samples having labels. Hence, deep belief network can be used as a semi-supervised feature extraction method. After the network is constructed, the feature can be extracted by inputting the sample to the network.2. The interval type 2 fuzzy logic system is used for prediction after the dimensionality reduction. In the constructing process of the fuzzy model, the input space is divided to generate the corresponding rules, and then the initial model is built. A fuzzy system consists of several rules, which can process the features extracted more particularly. Compared with most other modeling methods, it can produce a prediction interval.3. The fine-tuning is conducted by backpropagation for the whole network. After the optimization of fuzzy model, backpropagation is used for the tuning of the deep belief network module to improve the mismatch degree of the two parts.4. The type 2 fuzzy deep belief network is verified based on the dataset collected on a ball mill. The experiment based on the symmetrical samples is conducted. The results demonstrate that deep belief network can improve the distinction degree among the features of the high level, the fine-tuning of the whole network can overcome deflects caused by the mismatch, and in the whole fill levels, especially in the high levels, the accuracy is good. In the case where the samples are asymmetric, the deep belief network can make use of the information in the unlabeled samples fully, which helps to acquire the better initial weights for the network, and conduct an effective dimensionality reduction. As a result, the prediction accuracy is good. |
