Semi-Supervised Manifold Learning Theory And Applications

The machine learning is divided into learning supervision, unsupervised learning and semi-supervised learning according to whether priori information was used or not and the proportion of priori information used in reduced-rank methods. In allusion to shortcomings of unsupervised learning, for example that initial value is strongly depended on, input parameters are needed in advance and the algorithm has high complexity, this dissertation puts forward an unsupervised learning algorithm based on density expansion, and it can be seen clearly that the proposed algorithm has shorter average clustering time and higher clustering precision by simulation and experiments on typical Iris data sets and Fossil data sets.In recent years, with the great development of technology on data collection and storage, collecting large amount of data has become more and more easier. There is always a problem that large amount of data always shows prominent nonlinear performance. To solve this problem effectively, manifold learning algorithms are presented by reseachers. Manifold learning algorithms belong to unsupervised learning because it just uses the data and its inner geometric structure without considering effects of classification tag informaiton of the samples. In contrast, supervised manifold learning uses only classification tag informaiton of the samples but not informaiton of geometric structure of the data sets. But in practice, it is a hard work to get large amount of tag informaiton because it calls for great amount of manpower and resouces to finish the classification taging of the data samples. Therefore, more and more attention is paid to semi-supervised manifold learning algorithm.However, the traditional semi-supervised manifold learning algorithms require high quanlity of classification informaiton which brings certain restrictions in practice. To solve this problem, a local linear embedding algorithm based on side information is proposed. The algorithm only needs partial relationships between classification attribute(positive constraints, negative constraints) between samples without concerning about the specific type of classification information of the samples. This greatly reduces the difficulty of access to classification information of samples. Effectiveness of the algorithm is shown through the simulation results on classical face database. The proposed algorithm is applied to fault diagnosis of code precision of absolute photoelectricity encoder in practical optical communication systems.. Wavelet transform is done to extract the characteristics of the signal failure and the training samples are obtained by pre-encoder circuit simulation. After testing, the above method can accurately determine the fault type of absolute photoelectricity encoder. It has good generalization ability, simplifies the debugging on the ground, and provides a basis for the decoding circuit parameter adjustment and start of back-up signal, making it possible to improve its reliability in space applications.