Paman Gujral

On-Line Recalibration of Spectral Measurements Using Metabolite Injections and Dynamic Orthogonal Projection

Spectrometers are enjoying increasing popularity in bioprocess monitoring due to their non-invasiveness and in situ sterilizability. Their on-line applicability and high measurement frequency create an interesting opportunity for process control and optimization tasks. However, building and maintaining a robust calibration model for the on-line estimation of key variables of interest (e.g., concentrations of selected metabolites) is time consuming and costly. One of the main drawbacks of using infrared (IR) spectrometers on-line is that IR spectra are compromised by both long-term drifts and short-term sudden shifts due to instrumental effects or process shifts that might be unseen during calibration. The effect of instrumental drifts can normally be reduced by referencing the measurements against a background solution, but this option is difficult to implement for single-beam instruments due to sterility issues. In this work, in order to maintain the robustness of calibration models for single-beam IR and to increase resistance to process and instrumental drifts, planned spikes of small amounts of analytes were injected periodically into the monitored medium. The corresponding measured difference spectra were scaled-up and used as reference measurements for updating the calibration model in real time based on dynamic orthogonal projection (DOP). Applying this technique led to a noticeable decrease in the standard error of prediction of metabolite concentrations monitored during an anaerobic fermentation of the yeast Saccharomyces cerevisiae.

Drift Correction in Multivariate Calibration Models Using On-line Reference Measurements

On-line measurements from first-order instruments such as spectrometers may be compromised by instrumental, process and operational drifts that are not seen during off-line calibration. This can render the calibration model unsuitable for prediction of key components such as analyte concentrations. In this work, infrequently available on-line reference measurements of the analytes of interest are used for drift correction. The drift-correction methods that include drift in the calibration set are referred to as implicit correction methods (ICM), while explicit correction methods (ECM) model the drift based on the reference measurements and make the calibration model orthogonal or invariant to the space spanned by the drift. Under some working assumptions such as linearity between the concentrations and the spectra, necessary and sufficient conditions for correct prediction using ICM and ECM are proposed. These so-called space-inclusion conditions can be checked on-line by monitoring the Q-statistic. Hence, violation of these conditions implies the violation of one or more of the working assumptions, which can be used, e.g. to infer the need for new reference measurements. These conditions are also valid for rank-deficient calibration data, i.e. when the concentrations of the various species are linearly dependent. A constraint on the kernel used in ECM follows from the space-inclusion condition. This kernel does not estimate the drift itself but leads to an unbiased estimate of the drift space. In a noise-free environment, it is shown that ICM and ECM are equivalent. However, in the presence of noise, a Monte Carlo simulation shows that ECM performs slightly better than ICM. A paired t-test indicates that this difference is statistically significant. When applied to experimental fermentation data, ICM and ECM lead to a significant reduction in prediction error for the concentrations of five metabolites predicted from infrared spectra.

On multivariate calibration with unlabeled data

In principal component regression (PCR) and partial least‐squares regression (PLSR), the use of unlabeled data, in addition to labeled data, helps stabilize the latent subspaces in the calibration step, typically leading to a lower prediction error. For using unlabeled data in PLSR, a non‐sequential approach based on optimal filtering (OF) has been proposed in the literature. In this work, a sequential version of the OF‐based PLSR and a PCA‐based PLSR (PLSR applied to PCA‐preprocessed data) are proposed. It is shown analytically that the sequential version of the OF‐based PLSR is equivalent to that of PCA‐based PLSR, which leads to a new interpretation of OF. Simulated and experimental data sets are used to point out the usefulness and pitfalls of using unlabeled data. Unlabeled data can replace labeled data to some extent, thereby leading to an economic benefit. However, in the presence of drift, the use of unlabeled data can result in an increase in prediction error compared to that obtained with a model based on labeled data alone. Copyright