Partial least squares（PLS） is a popular method applied widely to the multi-component spectral calibration.it has been recognized that an efficient feature selection can be highly beneficial both to improve the predictive ability of the model and to greatly reduce its complexity.Considering the fact that there is a high probability of autocorrelation among the neighboring variables of spectra

in order to establish on-line spectral soft-sensing model with robustness and parsimony

an algorithm called backward interval selection partial least squares algorithm（BISPLS） is proposed.The main steps of this algorithm is to split the spectra into smaller equidistant subinterval first and

afterwards

eliminate all the subintervals step by step using the backward strategy

result in a sequence of subintervals

which make it possible to establish a final model by using a few subintervals.By a real example and the comparison with full-spectrum PLS

the algorithm’s advantages are showed in soft-sensor modeling.