Machine Learning Algorithms And Its Application Research In Engineering

In this paper, the application of machine learning algorithms is researched inrubber-mixed process, weather forecast, genes recognition and so on.The rubber-mixed is the first step of rubber product processing process, thequality of mixed rubber directly influences the subsequent procedure and quality ofrubber product. The large measurement delay of the Mooney-viscosity, one of themost important quality indexes of mixed rubber, has long been the main bottleneckholding back further development of rubber mixing process control. According to themechanism of rubber-mixed, the method that predicts the Mooney viscosity withrheological parameters is put forward in this paper. According to the characteristics ofthe rubber-mixed process, this paper proposed a novel machine learning algorithmDiscounted-measurement Recursive Partial Least Squares-Gaussian Process (DRPLS-GP) is proposed. The introduction of discount factor overcomes the "data saturation"phenomenon, to overcome the noise and the multi-collinearity of original data,orthogonal latent variables (LVs) are extracted by Discounted-measurementRecursive Partial Least Squares (DRPLS), and then the LVs are inputted to GaussianProcess (GP) as predictors for further nonlinear regression. Moreover, the flexibilityof discounted-measurement factor of the novel method ensures the high preciseprediction of Mooney-viscosity of different mixed rubber formulas. In particular, thismethod could update Mooney-viscosity prediction model with recursive method, so itis very practical for industrial application.In order to improve the climate prediction performance, and overcome the strongnonlinear between variables, a novel regression algorithm Generalized Partial LeastSquares Gaussian Process (GPLS-GP) is developed. Profiting from the latentvariables extraction power of GPLS, the noise and co-linearity between independentvariables could be overcome successfully. More importantly, the accurate nonlinearrelationship can be obtained with the method of generalizing variables and thenonlinear GP inner model. Comparison to conventional approaches (GPLS, PLS andGP) through an example, the performance of GPLS-GP is the best.In order to promote the rapid development of gene identification method, in thispaper, eight common linear and kernel-based supervised pattern recognitiontechniques were used to identify the short coding sequences of human genes(from21to192-bp), then to assess various algorithms and find the best combination of Z-curvevariable. By measuring the prediction accuracy, the consumption time and the tradeoff between sensitivity and specificity, partial least squares (PLS) and kernel partial leastsquares (KPLS) algorithms were verified to be the most optimal linear and kernel-based classifiers respectively, it is a reliable basis for the biologists choosingclassification machine following. Through analysis,93Z-curve variables were provedto be the best combination.