Based On Fuzzy Partial Least Squares Feature Extraction Methods

The feature extraction is occupying a very important status in the pattern recognition, it has many methods. In this paper, based on the idea of Partial Least Squares (PLS) modeling, we deeply analysed the theory of PLS and researched it, explored into the PLS method and fuzzy PLS (FPLS) method, then they had been applied to the feature extraction in theory and methods. This paper mainly discussed the following questions:(1)We discussed the basic idea of PLS and its linear and nonlinear modeling process, compared the advantages and disadvantages of various methods, and then showed the scope of application. We combined fuzzy theory to explore the application of the PLS approach and researched its theory feasibility, discussed the difference and contact among PLS and other linear methods, such as CCA, researched the technology of feature extraction based on the linear PLS analysis and the kernel-based PLS technology in detail and throughly., then we compared with the classics subspace feature extraction methods in experimental and analysis.(2)Since the fuzzy inference system had the properties of a structured knowledge representation in the form of IF-THEN rules, we applied it into PLS methods, TSK fuzzy model was embedded into the framework of PLS regression method to overcome the disadvantages of a number of classic non-linear PLS algorithms with its representive power, and the PLS outer projection was used as a dimension reduction tool to remove collinearity, the TSK fuzzy inner model was used to capture the nonlinearity in the projected latent space, these capabilities made FPLS a promising modeling and monitoring method.(3)Study a new FPLS algorithm based on singular value decomposition (SVD). Although the classic NIPALS algorithm could gave us a clear picture of PLS outer projection, it had a major problem that it was not certainly, since the random selection of score u in iterative process. The SVD-based FPLS algorithm used the SVD to FPLS, could effectively avoid the problem.(4)Proposed a new FPLS algorithm based on sample labeling. The traditional type of labeling did not reflect the distribution of data, located in the central area and the types of categories at the junction zone contribution to the classification of samples was considered equivalent, in this paper, we proposed a sample encoding method - the FPLS algorithm based on sample label. Different from each sample to give a grade in traditional algorithm, the sample label depended on the distribution, each type of sample was no longer share a single category label, to replace the original encoding, which had been proved of better identification in experimental.