Soft sensor modelling for the inline characterisation of polylactide (PLA) in a twin screw extrusion process /

Abstract

The development of a soft sensor technology to predict material properties of polylactide (PLA), extruded from a twin screw extrusion process, has been ex amined in this study. PLA is a bioresorbable (or biodegradable depending on the application) polymer used in the production of medical devices, pharmaceu ticals, food and waste packaging. Industries processing PLA face challenges in melt processing of PLA due to its poor thermal stability which is influenced by processing temperatures and process induced mechanical shearing. The characterisation of processed products currently takes place offline in laboratory environments. Scrap rates of a PLA medical grade product can be high as there is no current inline method to identify whether or not these were within specification during production. This study investigates using inline process data to make predictions of material properties, which are currently assessed offline. The properties examined are the yield stress, molecular weight and mass change of PLA. A slit die has been designed to house a number of transducers, which record the data required for the soft sensors. The transducers measure pressure, tem perature and near-infrared (NIR) spectral data. Using a slit die design also allows an estimate of the material’s shear viscosity to be made. This estimate was of interest in assessing whether the relationship between shear viscosity and the polymer’s molecular weight, (i.e. a change in molecular weight will result in a change in shear viscosity), could be useful for modelling the end properties. In-process degradation of the material will have significant impact on the final properties of the PLA product as well as its degradation behaviour. The relationships between the inline and end properties of the material are complex and non-linear and cannot realistically be derived from first principles. Machine learning algorithms pose a potential solution due to their ability to identify relationships between input and output data sets and their ability to continue to auto adapt and update over time with further observations. An initial set of experiments were designed over a range of processing condi i tions. The extruded samples underwent an accelerated degradation procedure. This allowed for nondegraded samples and also samples at various stages of degradation to be tested for material properties. The data collected from the initial experiments was used to train Principal Component Analysis Random Forest (PCA-RF) soft sensor models. A second set of experiments was then car ried out to capture data to validate the soft sensors. The yield stress soft sensor has been successfully developed for samples, which have not been degraded, and has generalised well using the validation data set, returning a root mean squared error (RMSE) of 1.24 MPa. This soft sensor has great potential for application within industry. The molecular weight and mass change soft sensor models have not had the same success and the rationale for this is discussed in detail.