Development and implementation of strategies for process data fusion, modelling and control

Resumen

With the emergence of Industry 4.0 and the increasing availability of sensors and data acquisition systems, modern manufacturing processes are now generating large amounts of process data on a scale as never seen before. During the past few decades, the intense development of powerful data-driven methodologies for process analytics has demonstrated the importance of multivariate data analysis for this field. Still, new strategies inspired by current methodologies and yet to be developed will continuously be required to tackle new challenges posed by the digital revolution in process analytics. This thesis has been focused on the development and application of chemometric tools for process analytical technology (PAT) and includes approaches for process monitoring, modeling and control of batch processes. All the methodology proposed has been tested on real batch processes of diverse nature monitored with sensor of different typology. The chemometric tools developed in this thesis are meant to be used in two different contexts: a) process monitoring, modeling, and control using spectroscopic probes and process sensors, and b) process monitoring using hyperspectral images. In the context of process monitoring using spectroscopic probes and process sensors, different methodologies have been designed to handle information coming from synchronized and non-synchronized batch process data. For synchronized batch process data, new strategies for offline and online Multivariate Statistical Process Control (MSPC) have been designed. Offline MSPC models, meant to control complete batches, were built based on information coming from original sensor variables or from compressed spectral information, issued from multivariate exploratory and resolution analysis outputs. Online process control methodologies were based on the use of local MSPC models built exploring the effect of different designs of process time windows onto the capacity to discriminate between observations following normal operation conditions (NOC) and showing an abnormal behavior. For non-synchronized batch data, a novel batch synchronization-free online MSPC methodology for tracking process evolution and control was proposed based on the idea of a global batch process trajectory and the use of local MSPC models. A clear improvement of the results linked to all MSPC scenarios is linked to the use of new mid-level data fusion strategies. The novel contribution in this thesis is the extension of the idea of data fusion to incorporate both diverse sensor outputs and diverse model outputs issued from the same sensor, but related to different modeling tasks. These model outputs, which are much more specific than mere compressed scores, help significantly to tune the information introduced in the MSPC models and to a better interpretation of the sources of abnormal process behavior. The chemometric solutions proposed for process monitoring using hyperspectral images (HSI) were mainly oriented to take advantage of the spatial information of the measurement for the qualitative and quantitative heterogeneity assessment in blending processes. The qualitative description of heterogeneity is linked to HSI unmixing analysis, which provides pure component distribution maps that offer a good visual representation of the evenness in the spatial distribution of the different materials in the blending formulation. The quantitative characterization of heterogeneity is obtained from the variographic analysis of the distribution maps and results in two indices: the Global Heterogeneity Index (GHI), related to the scatter of the individual pixel concentration values, and the Distributional Uniformity Index (DUI), describing the distributional heterogeneity, usually overlooked in traditional approaches, that expresses the evenness in the spatial distribution of the different materials forming a blend. These indices have been proven to be a powerful process analytical tool to characterize the heterogeneity in blending processes monitored atline and inline with NIR-HSI. For image-based inline process monitoring, an extension of this methodology, called SWiVIA (Sliding Window Variographic Image Analysis), has been adapted for the continuous assessment of heterogeneity in real-time blending process monitoring. The versatility of the SWiVIA methodology enables heterogeneity assessment at the time resolution and spatial scale of scrutiny required for the blending application of interest.