IRmadillo logo white

Application note: Real-time analysis of phospholipids for the monitoring & control of the degumming process using a static-optics FTIR process analyser


Phosphorus (P) concentrations were monitored during an oil-refining process over a 5-month period using a static-optics FTIR spectrometer (the IRmadillo) and compared to off-line reference data. Calibration curves were built using various chemometric models. Analysis showed that a locally weighted regression (LWR) model proved the most accurate, with an average error and limit of detection of ± 32 ppm, and a limit of quantification of 105 ppm; it displayed good agreement with the reference data. Trustworthiness models, based on Hotelling’s T² and Q residuals, were analysed; these showed high confidence in the results.


One of the major challenges in edible oil refining is the removal of non-hydratable phospholipids in the degumming process. Whether the refinery is using traditional methods such as phosphoric acid dosing or more modern enzymatic approaches, the measurement of phosphorus (P) levels is a key bottleneck. Off-line analytical instrumentation such as ICP-OES is a fantastic and highly accurate method, but it requires taking a sample to the laboratory for analysis which is slow, expensive and in some cases dangerous as the handling of hot oils is not a trivial matter.

On-line analysis of P cannot be performed using simple measurement probes (such as pH or temperature probes) and requires more advanced instrumentation. The measurement of phospholipids using mid-infrared spectrometers (FTIR) is well known, but traditional FTIR instruments are fragile and not suited for use in industrial environments.

Here we present the use of a static-optics FTIR instrument specifically designed for industrial use, and its calibration for P measurement in non-hydratable phospholipids across a range of feedstocks. We also show how to confirm that we “trust” a measurement from FTIR instruments to give confidence in its ability to measure previously unseen data.


Spectra were acquired with an on-line installation of an IRmadillo static-optics spectrometer with a 120 s acquisition time. Each reference point was calibrated with triplicate spectra to reduce impacts of noise (assuming the rate of change of the process is < 360 s). Off-line reference data were provided by the customer using their standard process and an ICP-OES instrument for P measurement. Chemometric calibrations were built using Eigenvector Research Inc’s Solo software, using a range of partial least squares (PLS) and locally weighted regression (LWR) models for quantitative analysis and principal component analysis (PCA), partial least squares discriminant analysis (PLS-DA) and support vector classification (SV-C) models.

Figure 3: Photograph of IRmadillo static-optics FTIR inserted directly into process pipework.

Results & Discussion

Various different calibrations were built for P monitoring, and the best results to date of writing are obtained using locally weighted regression. The average error and limit of detection for this calibration is ± 32 ppm, with a limit of quantitation of 105 ppm. The measurements over a period of 5 months for this calibration are shown in Figure 1 (with a gap where the refinery was shut down for maintenance). There is generally very good agreement between the reference data and the on-line measurements, with only a few samples that do not fit the general trend – indicating the short-comings of laboratory sampling where a single measurement could be an outlier and incorrect measurement; when taken in isolation this is not easy to identify.

The most interesting aspect of these results is the slope of P throughout a processing run (the sharp changes in concentration are caused by a changeover in product/cleaning run and can be discounted). It was generally expected that the P level was constant for a given production run with a single sample taken at the beginning of the process and required to degum the product.

However, what can be seen in the results below is that a large amount of change occurs throughout a processing run, potentially with changes > 50% of the original measured value. This may occur because of stratification within the holding tank of product. The on-line and real-time measurements allow fine control of phosphoric acid dosing – which will enable the customer to make significant improvements to efficiency.

Figure 1: IRmadillo on-line measurement of P levels in a water washed oil stream prior to neutralization and degumming using an LWR calibration model overlaid with customer reference data over a period of 5 months (with refinery shut-down omitted for clarity)

Figure 2: Hotelling’s T², Q residuals and trustworthiness outputs for P calibration shown over a 4-week period


Spectroscopic calibrations can be linked to advanced statistical checks as well – enabling a real-time “trustworthiness check” to be performed on every measurement. This can give the operator significantly more confidence that a given measurement is realistic compared to the single value measurements made in a laboratory. Figure 2 shows the outputs of a Hotelling’s T² and Q residual based trustworthiness measurement.

Hotelling’s T² gives a statistical check on whether a datapoint sits within the distribution of values expected for a given measurement (i.e. that the peaks in the spectrum are within the sizes and shapes of those used to build the calibration). Q residuals give a statistical check on whether unknown or unexpected spectral features are present that were not observed during calibration.  Both Hotelling’s T² and Q residuals can be reported normalised to 95 % confidence limits (i.e. a value below 1 should be reported 95 % of the time); and this allows a simple check to confirm if a spectrum fits within the distribution of those expected from the calibration.

This is all shown in Figure 2, the Hotelling’s T² and Q residuals’ values have been normalised to their 95 % confidence limits, and these normalized values have been used to generate a trustworthiness “flag” (reporting 0 for no issues are detected and 1 when issues have been identified). It can be seen clearly that there are periods of high confidence and periods during which issues have been flagged. These issues appear when cleaning of the process pipework or product changeovers occurred. This is flagged in the Q residuals as water washing was not included in the calibration, so the presence of water in the spectrum is classified as “unexpected”, and therefore raises a high Q residual value. This approach gives the user confidence that a measurement is realistic and can be “trusted” for process control.


This work shows that static-optics FTIR spectroscopy is a viable and useful route for on-line and in-line real-time measurement of phospholipids, enabling the development of real-time phosphoric acid dosing for degumming. The instrument has been installed directly into the process pipework upstream of phosphoric acid dosing and can now be used to control the dosing in real time.

This work also shows that it is possible to implement a real-time check on every measurement to ensure that the measurement is trustworthy, and that the spectrum observed fits within the space observed during calibration. This enables better understanding of exactly how trustworthy a measurement is, even when using machine learning and “black box” calibration model techniques.

Read our full page on various aspects of how the IRmadillo can monitor your process in edible oil refining, read our FAQs and find out more.

Contact us

Want to know more? Contact us. We can arrange a trial installation on your plant for several months so you can see you yourself just what the IRmadillo can do.

Keit's IRmadillo Logo