Optimization of a technique for the estimation of the composition of edible oil blends

The fatty acid, sterol and tocopherol contents of edible oils were used to determine the formulation of blends of oils, which are themselves complex mixtures, by means of a weighted least-squares estimator with backwards elimination. A t-test was used to determine whether an oil should be eliminated from the blend. Initially, the minimum t-value was set at 2.7, which corresponds to a probability of ca. 0.01 if normal distribution is assumed. Some compounds are present in one oil but cannot be detected in others and are valuable in differentiating between various oils. Assumptions about levels for the undetected compounds and their corresponding weightings were made earlier in order to implement the weighted least-squares technique. A simplex optimization procedure was applied to avoid the use of assumed values. An optimum for the t-value as well as for the levels and the weightings of the undetected compounds was established. Analytical results for 378 pure oils were used to develop the models. The optimized technique is slightly better for assessing the composition of unknown blends than the original technique, when tested on 93 samples which were blends of known amounts of sunflower-seed oil, groundnut oil, cottonseed oil, maize oil, olive oil and palm oil.