Lubricants operate in hostile, high-temperature environments where they are exposed to chemically reactive by-products such as partially burned hydrocarbons, soot, water, wear debris and products of combustion, including nitric acid (HNO3) and sulfuric acid (H2SO4). This highly acidic environment is detrimental to the lubrication process and must be neutralized in order to reduce corrosion. The most efficient way to neutralize is by adding basic chemical additives to the oil. However, it is important that enough of these additives are present in order for the lubricant to work efficiently.
The traditional way of monitoring the basic activity of a lubricant is by chemical titration with a mineral acid. This value is known as the base number (BN) within an engine oil, and is typically expressed in an equivalent number of milligrams of potassium hydroxide (KOH) that are required to titrate 1 gm of the lubricant. The titration is often automated, but may take up to 30 minutes per sample, making testing time-consuming in a high-throughput laboratory. In addition, the chemicals used for analysis need to be disposed of in an environmentally friendly manner. For the in-house laboratory, the cost of disposal of some process chemicals may be greater than the cost of the chemicals used in the analytical procedure. When combined with typical testing errors on used engine oils in the order of 10 percent, forward-thinking companies are beginning to view alternative methods for measuring BN levels.
It is well-accepted that infrared (IR) analysis can provide important information on the performance of a lubricant and the equipment being used. This instrumental technique relies on the principle that different molecular components in a sample have various IR wavelength-specific absorption/transmission characteristics. By comparing the intensity and shape of molecular bands, the component relating to that molecular structure can be identified and quantified. IR spectroscopy is routinely used to measure water, glycol, soot and oxidation/nitration levels in lubricating oils. Identifying these components is typically carried out by first obtaining the spectrum of the new, unused oil and then comparing it with a spectrum of the same oil that has been used to lubricate an engine, machine or transmission system. All the major components within the sample, the type of lubricant, equipment design and operating conditions have an impact on the spectral information. As a result, the spectral changes throughout the lubricant’s life cycle can be related to changes in the chemistry of the oil.
It is relatively simple to see the lubricant’s basic and acidic chemistries in the infrared spectrum, particularly with a new, unused oil. However, used oils are a complex mixture of a number of chemical compounds derived from the original base oil, additives, oil degradation products and contaminating compounds like fuel, water or antifreeze. Essentially, the infrared spectrum of a used oil sample is essentially the sum of the spectra of all of these components, consisting of a large number of overlapping peaks that are difficult to resolve, even for the experienced analyst. A typical used oil infrared spectrum of wavelength (cm-1) plotted against transmission (%T) is shown in Figure 1.
Figure 1. Typical IR Spectrum of a Used Oil Sample
It can take a great amount of effort for one to attempt to study all of these chemical differences for only one series of used oils by traditional IR comparative methods. For this reason, a statistical mathematical approach has proven to be a better method for studying spectral differences in complex used oil samples. Mathematically, one can look at a series of continually changing infrared spectra because it is essentially a matrix of x and y data sets. These changes in the infrared spectra relate to chemical changes such as the BN or acid number (AN) of the sample. The application of mathematical techniques to study chemical properties is known as chemometrics. This approach of evaluating different statistical indicators to predict a sample’s composition has been used by the scientific community for years, but has only recently been applied to the analysis of used engine oils.
Using this technique, a series of analytical results are generated for a set of control samples, which are then mathematically processed into a training model, using well-established algorithms like principal component analysis (PCA). This model then becomes a training set for the analyte or property being studied. The purpose of the PCA algorithm is to express the variation within the dataset in the simplest terms. This variation is commonly known as its “eigenvalue” and is a measurement of the principal component, which is changing throughout the dataset. Calibration models such as partial least squares (PLS) and multiple linear regression (MLR) are calculated for the property values to generate a principal component regression (PCR) model. These models are then used to draw a relationship between the principal components and the chemical property being measured, such as BN.
Sampling and Analytical Methodology
This method was used to estimate BN values for a set of used oils using the Spectrum One FTIR spectrometer with Quant+® software (Perkin-Elmer, Shelton, Conn.). With this system, the software is designed to apply chemometric algorithms to determine the sample’s properties, using either PLS, MLR or PCR prediction models. But regardless of which chemometric approach is being employed, the process works in exactly the same way. A regression model is built by collecting spectral information for a group of typical samples, then fitting a predefined mathematical relationship to the collected data. In this method, a series of samples are analyzed by both infrared analysis and conventional BN titrimetry. The spectral matrices are then processed into a PCA model, which is regressed with the BN values using either MLR or PLS calibration data. These samples become the training set for the model.
If the training samples chosen have been analyzed with precision and are within a controlled dataset, good regression correlations can be obtained, assuming that high-quality FTIR spectra has been generated for the analysis. This also means that the sample cannot be soot loaded to the point where good spectral resolution has been destroyed. It must be emphasized that the quality of the BN titrimetric data is also important, because the model cannot generate accurate results unless the original titrimetric data is of the highest quality. The number of samples used in the training model should also be large enough to represent the range of sample conditions being analyzed. In addition, to account for errors in the analysis, a number of replicates of the same samples should be analyzed.
Accurate results are achievable if a good representation of training samples has been used to develop the model. Figure 2 shows a plot of estimated BN (predicted) values versus specified (measured) values (expressed as mg KOH/g of lubricant) for a set of used oil samples. The model was developed from a series of FTIR spectra taken from a representative series of used oils that came from several engines running with the same lubricant but operating under different conditions. The PCA developed for this model was regressed against ASTM D4739 – a titrimetric method for determining BN in oil samples.
Figure 2. Predicted vs. Measured Values for a Set of Used Oil Samples
Using a Principal Component Regression (PCR) Model
The model shown represents a 94.3 percent variance statistic with a standard error of prediction of 0.5 mg KOH/g. This means similar samples predicted by this model will yield a BN value within 0.5 mg KOH/g of the measured value. For a BN of 10 mg KOH/g, this is actually better than the reported reproducibility in the ASTM method, which is achievable only if replicate analyses are available. This regression model will be valid only as long as the oil samples being analyzed have values that do not fall outside the calibration limits and their chemistries are similar to those used to develop the model.
A benefit of using this approach is that there are variables generated by the mathematics that allow the user to know if the sample being predicted is of similar chemistry to the samples used to develop the model. By generating values such as the “M-Distance” and “Residual Ratio”, the mathematical approach allows the analyst to know whether unknown samples are generating a valid prediction for the model. This is as an important capability because differences greater than the model’s standard error of prediction could indicate additional problems with the oil sample such as contamination, improper sampling, incorrect analysis or significant operational changes in the equipment.
IR Spectrum Correlated to Changes in Oil Chemistry
When using this approach to monitor changes in the oil chemistry, it’s important to understand that lubricants are formulated with a number of different additives. These additives are designed to help the final formulation meet performance specifications of the equipment it is supposed to protect. These additives improve physical properties such as dispersancy, detergency, antiwear and antifoam. For example, as a part of the detergent additive package, there is a base reserve (excess BN), which is designed to help control the formation of acidic components in the lubricant. Depending upon the application, this base reserve comes in several chemical forms, the most common being inorganic carbonates.
As mentioned previously, in the lubrication process, acids are formed from either combustion gases (oxides of nitrogen and sulfur) oxidation by-products (esters and carboxylic acids), or additive depletion reactions (phosphates, sulfates and carboxylic acids). These acids react with the base reserve to form the corresponding neutral salts. This formation of acids and their subsequent sequential reactions generate chemical changes in the lubricant that are observed within the FTIR spectrum.
A chemometric model looks for changes within the array of modeling spectra and relates these changes to the parameter being modeled (or predicted). In addition to the statistics generated, the model also contains the variations within the spectra that relates to the modeled parameter. This spectral variation, called the “regression spectrum”, contains both positive and negative spectral changes that relate to changes in chemistry in the array of modeled spectra. Figure 3 shows a typical BN, infrared regression spectrum of a series of used oils.
In this spectrum, the positive features (above the zero line) reflect chemistries that relate to higher BN values, while the negative features (below the zero line) reflect chemistries with lower BN values. For example, the most common reaction with BN will result in a loss in carbonate concentration, which is confirmed by a strong carbonate band at 1,516 cm-1. The strength of this band therefore relates to the change in BN values for the modeled samples.
Figure 3. BN Infrared Regression Spectrum of a Used Oil Series
There are many chemical changes that occur in concert with the changes in BN value, as exemplified by the different spectral bands seen in Figure 3. These changes may have lower response coefficients (relative spectral peak height) than the carbonate, but are expected to be present in the regression spectrum that gives clues as to what is happening to the used oil as the BN is depleted. An example of this is the thiophosphate band (972 cm-1) from the breakdown of dialkyldithiophosphate to the phosphate (1,007 cm-1). The regression spectrum also shows the organonitrates (1,630 cm-1) formed from the ingress of combustion gases, ester oxidation by-products (1,729 cm-1) and sulfate formation (1,151 cm-1), which are all correlated to the BN property of the oil.
Quality of Input Data
Previously mentioned, the quality of the model is directly related to the quality of the input data. For example, if the BN titrimetric measurement has precision of 1 mg KOH/g and multiple samples were obtained for each BN range, it could result in a degradation in predicted precision. In this case, the infrared spectra would be capable of differentiating BN better than the measured values, which is not a desirable analytical scenario.
Figure 4. Poorly Designed BN Regression Model, that Could Degrade Measurement Precision
This is demonstrated in Figure 4, which shows a regression model of an array of predicted BN values for each measured BN range. If this example of the PCR model is compared to the more ideal data set in Figure 2, it suggests that the infrared spectral data has the potential to degrade the measurement precision. This type of model degradation can also occur with the use of poorly generated BN titrimetric measurements. The poor measurements will increase the scatter of input data and result in a degradation of the model’s precision. For this reason, it is critical that the BN input data is as accurate as possible to ensure the regression model generates predicted values of the highest quality.
The Final Analysis
This study shows that BN values of used oil samples can be successfully estimated from an IR spectrum utilizing a chemometrics approach. By determining BN using conventional titrimetry for a set of known samples and mathematically processing the values against the FTIR spectrum, a regression model can be used to correlate the data.
The major benefit of this approach, particularly for a high workload laboratory, is that an individual sample can be analyzed in minutes compared to approximately 30 minutes using the conventional titrimetric approach. In addition, it provides the analyst additional knowledge concerning the sample, which was not previously available. There is no question that studying the infrared spectrum from a chemometric perspective has the potential to significantly lower the cost of analysis, compared to conventional oil-based testing. However, a good used oil program should always validate methodology with traditional analysis, especially when new oil types are being added to the model. Thus, by applying an optimum combination of chemometric determinations and traditional validation, high-quality results can be achieved with a cost-effective analysis program.
Technical Editor’s Note:
The use of FTIR analysis to predict changing AN and BN values has been used in various ways with widely varying degrees of effectiveness. The developmental activity noted above clearly defines attributes of the samples used to create the training model: low soot values, exactly the same brand and type of oil, high number of baseline data inputs, and high-quality baseline data input. The quality of the data used to build the model, and the close similarity between the lubricant used to create the training model and that which is from the actual testing process, both appear to be noteworthy constraints on the general applicability of this analysis technique.
The practitioner must be prepared to dedicate time to build out the test conditions for each lubricant to be tested as closely as possible to that which is considered ideal if the practitioner is to achieve the effectiveness predicted by the modeling technique. Quality verification of results through routine method calibration and verification of predicted values through conventional means (ASTM 2896/4739) should accompany the development of this analytical technique.