An improved changeable size moving window partial least square applied for molecular spectroscopy

• The priority and parameter of pretreatment method can be simultaneously optimized.
• The selected variable and pretreatment method can be more reasonably explained.
• Model robustness and complexity are optimized without losing its accuracy.

When analyzing molecular spectra, optimizing the pretreatment method and the wavelength variable is always an important issue. However, currently there are unsatisfied phenomena that select the same type of pretreatment method multiple times in some results generated by previous common optimizing algorithms. Additionally, the parameters and calculation priorities of the pretreatment methods cannot be optimized. To solve those problems, an improved changeable size moving window partial least square (CSMWPLS) named pretreatment method classification and adjustable parameter changeable size moving window partial least square (CA-CSMWPLS) is presented. With regard to the chromosome construction of CA-CSMWPLS, there is a region for pretreatment method optimization and another one for wavelength variable optimization. In the former, the common pretreatment methods are classified into four different types such as smoothing, derivation, correction, and standardization, and the parameters and calculation priorities of pretreatment methods serve as genes of the CA-CSMWPLS chromosome. In the latter, there are changeable size moving windows that consist of window position genes and window width genes. Moreover, a scale factor η is designed for reducing model complexity in CA-CSMWPLS fitness function and a peculiar coding and a decoding rule are adopted in this algorithm. After testing a group of corn and a group of gasoline spectra with CA-CSMWPLS, the model accuracy was significantly improved, for the root mean square error cross validation (RMSECV) and root mean square error prediction (RMSEP) of the corn spectra were 0.0028 and 0.0032, and those of gasoline were 0.165 and 0.170, respectively. Furthermore, the optimized pretreatment methods and wavelength variables are more reasonable, the model complexity is smaller, and the model robustness is stronger than other relative methods.