Often, multivariate analysis is wildly used to process "signal information", which includes spectrum analysis, bio-signal processing and etc. In general, Least Squares (LS) and PLS fall into overfitting problem with ill-posed condition, which means the future selections make the training data have better adaptability, but the quality of the prediction would be poor, compared with the testing data. However, the goal of these models is to have consistent prediction between testing and training data. Therefore, in this study, we present a novel MVA model, Partial Regularized Least Squares, which applies regularization algorithm (entropy regularization), to Partial Least Square (PLS) method to cope with the problem mentioned above. In this paper, we briefly introduce the conventional methods and also clearly define the model, PRLS. Then, the new approach is applied to several real world cases and the outcomes demonstrate that while calibrating data with noises, PRLS shows better noise reduction performance and lower time complexity than cross-validation (CV) technique and original PLS method which indicates that PRLS is capable of processing "Bio-signal". Finally, in the future we expect utilizing another two regularization techniques instead of the one in the paper to identify the performance differentiations.