Unsupervised model adaptation for multivariate calibration by domain adaptation-regularization based kernel partial least square.

In chemometrics, calibration model adaptation is desired when training- and test-samples come from different distributions. Domain-invariant feature representation is currently a successful strategy to realize model adaptation and has received wide attention. The paper presents a nonlinear unsupervised model adaptation method termed as domain adaption regularization-based kernel partial least squares regression (DarKPLS). DarKPLS aims to minimize the source and target distributions in a low-dimensional latent space projected from the reproducing kernel Hilbert space (RKHS) generated with the labeled source data and unlabeled target data. Specially, the distributional means and variances between source and target latent variables are aligned in the RKHS. By extending existing domain invariant partial least square regression (di-PLS) with the projected maximum mean discrepancy (PMMD) to reduce the mean discrepancy in the RKHS further, DarKPLS could realize fine-grained domain alignment that further improves the adaptation performance. DarKPLS is applied to the γ-polyglutamic acid fermentation dataset, tobacco dataset and corn dataset, and it demonstrates improved prediction results in comparison with No adaptation partial least squares (PLS), null augmented regression (NAR), extended linear joint trained framework (ExtJT), scatter component analysis (SCA) and domain-invariant iterative partial least squares (DIPALS).

Self-organizing maps-based generalized feature set selection for model adaption without reference data for batch process.

When fourier transform infrared spectroscopy (FTIR) techniques combined with multivariate calibration are used to measure the key process features or analyte concentrations during batch process, model adaption is indispensable for maintaining the predictability of a primary calibration model in new secondary batches. Many model adaption methods conforming to the actual application scenario of batch process have been proposed. Here we report on a novel standard-free model adaption method without reference measurement called variable selection strategy with self-organizing maps (VSSOM). It uses self-organizing maps (SOM) to classify the whole spectral variables into multiple classes according to the spectra from primary batch and secondary batch, respectively; and the corresponding primary feature subsets and secondary feature subsets are formed firstly. Secondly, candidate feature subsets without empty elements are generated by operating intersection between any primary feature subsets and any secondary feature subsets. Thirdly, the candidate feature subset with minimum root mean square error of cross-validation (RMSECV) for the primary calibration set is selected as the optimal feature subset. In this manner, the optimal feature subset can be identified from the candidate feature subsets. In other words, VSSOM aims to create a stable and consistent feature subset across different batches provided that it selects better features within the intersection sets between primary feature subsets and any secondary feature subsets. Two batch process datasets (γ-polyglutamic acid fermentation and paeoniflorin extraction) are presented for comparing the VSSOM method with No transfer partial least squares (PLS), boxcar signal transfer (BST), successive projection algorithm (SPA), transfer component analysis (TCA) and domain-invariant iterative partial least squares (DIPALS). Experimental results show that VSSOM has superior performance and comparable prediction performance in all the scenarios.

An active learning representative subset selection method using net analyte signal.

To guarantee accurate predictions, representative samples are needed when building a calibration model for spectroscopic measurements. However, in general, it is not known whether a sample is representative prior to measuring its concentration, which is both time-consuming and expensive. In this paper, a method to determine whether a sample should be selected into a calibration set is presented. The selection is based on the difference of Euclidean norm of net analyte signal (NAS) vector between the candidate and existing samples. First, the concentrations and spectra of a group of samples are used to compute the projection matrix, NAS vector, and scalar values. Next, the NAS vectors of candidate samples are computed by multiplying projection matrix with spectra of samples. Scalar value of NAS is obtained by norm computation. The distance between the candidate set and the selected set is computed, and samples with the largest distance are added to selected set sequentially. Last, the concentration of the analyte is measured such that the sample can be used as a calibration sample. Using a validation test, it is shown that the presented method is more efficient than random selection. As a result, the amount of time and money spent on reference measurements is greatly reduced.

[Weight spectral angle mapper (WSAM) method for hyperspectral mineral mapping].

On the basis of analyzing the deficiency of spectral angle mapper (SAM), an improved similarity measure called weight SAM (WSAM) is proposed in the present paper. The fundamental idea is to set a weight in the "difference range" to increase the discriminability between the similar minerals. When we distinguish some kind of mineral, the authors can find the "difference range", in which the difference in spectral feature between the similar mineral spectrum and the reference spectrum is huge, and gives weight k to the spectrum in this range to reduce the similarity and increase the discriminability between similar minerals. The experiment results of spectra and AVIRIS data indicate that the WSAM method reduces the similarity of target mineral and its similar mineral and increases our ability of visual interpretation.