Many manufacturers rely on expensive, time-consuming lab measurements to evaluate their product quality. In some batch processes, the manufacturer must wait for lab results to proceed with further processing.
A leading adhesive manufacturer was relying on lab data to measure a key property. If the quality was found to be below a target, manufacturing is complete, otherwise the product requires adjustments. Unfortunately, adjustment was required in 86% of the batches. The goal of this work was to understand the key operating parameters that would allow them to reach the quality more consistently, and to determine if a multivariate soft-sensor could accurately predict this value without the need for the lab measurement.
The ProSensus Approach
ProSensus uses powerful multivariate data analysis tools to analyze batch data and build predictive models. In this project, the first dataset provided to ProSensus unfortunately did not result in good models. Therefore, ProSensus recommended the installation of new sensors to measure additional variables such as torque, ambient temperature/humidity in the plant, and cooling water parameters to capture more variation in the process.
Upon including the new measurements, multivariate modeling was successfully applied to the dataset. Batch alignment and outlier removal was performed prior to developing the final model. This model was then used for finding root causes of bad batches and an evaluation of using this model as a soft sensor in real-time was performed.
The developed model has an R2 of 83% and Q2 of 60% which is in fact a good fit for this type of process data. There was an observed separation between good and bad batches in the score space. This allowed ProSensus to examine the operating differences between the batches and provide the client with an operating strategy that can produce batches that are more likely to be in-spec without further processing.
A contribution plot of the trajectory variables shows the time behaviour of these measurements and how they contribute to the quality prediction. In order to have a batch with Y <-1 (good quality), the batch should have a high temperature and a high torque for the whole batch. This analysis helped inform the plant personnel on how to modify the standard operating procedure (i.e., the standard “batch recipe”) to ensure fewer batches needed re-processing.