Abstract: The main pollutants from automobile engines or off-road engines emission are NOx, THC, CO, PM and CO2 et al. The analyzers for measuring these exhaust pollutants need to be linearized for their non linearization characteristics. The linearization model based on polynomial stepwise regression linearization method is undesirable in respect of stability and prediction performance. The valid linearization forecast models of CO2 and CO analyzers, which are more clear in physical meaning, are presented on the basis of partial least squares regression, whose prediction accuracies are 29.1%～35.1% and 23.5%～39.3% higher than the model with least squares regression and the model with Chebyshev polynomial regression respectively. A general way to calculate the uncertainties of regression coefficients based on full cross validation and a principle of determining the model prediction accuracy with RMSE based on full cross validation and a principle of validating the significances of variables with the method whether the uncertain scope of regression coefficient is cross 0 are presented. These methods are simple, useful and effective, which can be applied to models not only with PLS method but with LS method and other regression methods as well. Engine exhaust emission analyzers can be linearized with this modeling progress to improve their measuring accuracies, especially of great advantage to a poor linearization and/or complex characteristic analyzer.

Key words: Analyzer, Chebyshev polynomial, Linearization, Partial least squares regression