Ricardo Rendall

A Comparison Study of Single‐Scale and Multiscale Approaches for Data‐Driven and Model‐Based Online Denoising

Signal denoising is a pervasive operation in most online applications, such as engineering process control and online optimization, strongly affecting the outcome of these higher-level tasks and impacting the overall variability exhibited by processes and products. Therefore, it plays a fundamental role in improving process capability, which is, however, often overlooked. In this work, we compare the performance of different types of currently available online denoising filters using a variety of test signals that represent the diversity of situations likely to be found in practice, properly corrupted with additive noise of varying magnitudes. Both single-scale/multiscale, data-driven/model-based and time domain/frequency domain, online filtering approaches were contemplated, in what is, to the best of the authors knowledge, the more extensive comparison study conducted on online denoising (or filtering) methodologies. A new class of multiscale denoising algorithms is also considered in this study, based on the online wavelet multiresolution decomposition. In this context, we propose and test a new formulation, called the online multiscale hybrid Kalman filter. After proper tuning, the methods are tested and their performances compared. As a result of the comparison study, clear guidelines are provided for practitioners on the use of online denoising methodologies, which allow for a better management of the impact of the propagation of unstructured components of variability in the final outcome of the processes. Copyright

Profile-driven Features for Offline Quality Prediction in Batch Processes

Abstract The present paper addresses the task of quality prediction in batch processes, where measurements from process variables are used to predict one or more quality variables of interest. The majority of current methods for batch quality prediction are based on complete time profiles for all variables, requiring synchronization before batch-wise unfolding (BWU). Synchronization is complex to implement and requires trained personnel whereas BWU leads to a matrix with thousands of pseudo-variables, increasing the potential for model overfitting. In this context, the development and validation of reliable data-driven predictive models is challenging and time consuming. On the other hand, low complexity approaches for batch processes remain vastly unexplored and only a few examples are available in the literature. Therefore, in this work we present a new methodology called profile-driven features (PdF) for offline quality prediction. PdF presents low modelling and implementation complexity, is able to cope with the dynamics presented by batch process variables and generate useful features for building predictive models. In order to test the proposed method, datasets from two simulated batch processes were obtained and partial least squares models were developed to predict end-of-batch quality parameters. Upon comparison with the benchmark method based on BWU and other feature-oriented approaches, PdF presented similar or superior prediction performances under independent testing conditions, despite its lower complexity.

An extended comparison study of large scale datadriven prediction methods based on variable selection, latent variables, penalized regression and machine learning

Abstract Regression methods are pervasive in most data-driven predictive studies performed with industrial and laboratorial data. They provide the means to obtain reliable estimates of output variables (related to product quality or other properties) based on a set predictor variables which are usually easier to measure and less expensive. In this paper, a large scale comparison framework is developed in order to assess the performance of a rich variety of regression methods, compare them and provide guidelines for choosing a suitable regression method in a given application scenario. Regression methods were grouped in four classes: variable selection, latent variables, penalized regression and ensemble methods. The framework was applied to three case studies: two based on simulated data and one with real data from a wine age prediction study. Improved results were obtained when the models prior assumptions, regarding sparsity and collinearity, matched the data generating mechanism.

Managing Uncertainty Information for Improved Data-Driven Modelling

Abstract Obtaining uncertainty information and using it in an optimized way in the context of the various PSE tasks, is still a challenge in the modern Chemical Processing Industry (CPI). Among the variables most affected by measurement uncertainty, one typically finds the process outputs. Examples include stream concentrations (main product and by-products) and measurements of quality properties (mechanical, chemical, etc.). With the increasing flexibility of processing units, these quantities can easily span different orders of magnitude and present rather different uncertainties associated with their measurements. Therefore, heteroscedasticity in the process outputs (Y’s) is a prevalent feature in CPI. In this article we address two related and complementary aspects of this problem with practical relevancy: i) how to estimate measurement uncertainty in heteroscedastic contexts? ii) how to take advantage of the availability of measurement uncertainty information for optimal model development?