Setting Acceptance Specifications on Incoming Raw Materials To Ensure Product Quality
Final product recipes usually involve a combination of many raw materials from various suppliers. Therefore, setting specifications on each of these incoming materials separately is an unreliable metric since the final quality is affected by the simultaneous combination of all raw materials and their properties. ProSensus worked with Mondelez, a global food manufacturer, to establish specifications on new incoming raw material lots to avoid accepting lots that pose a high risk on the final product quality. The food product of interest was manufactured by combining and processing 9 raw materials with different properties and from different suppliers. This work was presented in a joint presentation with Mondelez at IFPAC, 2013.
Historical data was available on the properties of past raw materials lots that were used in the manufacture of this product as well as processing conditions from different sections in the process. In order to enrich the model, the data was augmented by DOE's (additional experiments shown in red) to fill any holes in the latent variable score space. The finalized data structure is shown below:
ProSensus’ Modeling Approach:
ProSensus developed a multi-block comprehensive PLS model that correlates the final quality back to changes in the raw material properties and process parameters. The model explained ~80% of the variation with only three latent (summary) variables. The coefficient plot (shown below) is one of the basic PLS plots for model interpretation. It is a very informative plot as it shows the simultaneous effect of all input variables on the final quality.
The plot revealed that two of the raw materials (RM-3 and RM -7) had no significant effect on the final quality while other raw materials (RM-1, RM-6, RM-9) had a much bigger impact on the quality. This is useful for highlighting where the manufacturer should focus their raw material specifications efforts. Additionally, the plot showed that some of the process conditions impact the quality which means that the process parameters can be manipulated to optimize the final quality.
Given a new lot of material, it was evaluated against the PLS reference model by generating 1000 random possible combinations of this raw material with existing/ past lots in the database under various process conditions and computing the predicted final quality, SPE (squared prediction error), and Hotellings’ T2 statistics. These histograms were a reliable indication if any of the new lots might lead to a large SPE, T2, or a large deviation in the final quality from its target. The specification region is set based on the probability of any of the three calculated metrics lying outside the specified confidence limits. This methodology also allowed the food manufacturer to assess the raw material lots currently in stock and avoid using certain combinations of them that would potentially lead to poor final quality.
MacGregor, J., Liu, Z., Bruwer, M., Polsky, B., & Visscher, G. (2016). Setting simultaneous specifications on multiple raw materials to ensure product quality and minimize risk. Chemometrics And Intelligent Laboratory Systems, 157, 96-103. doi: 10.1016/j.chemolab.2016.06.021