The food industry needs tools to improve the efficiency of their production processes by minimizing waste, detecting timely potential process issues, as well as reducing the efforts and workforce devoted to laboratory analysis while, at the same time, maintaining high-quality standards of products. This can be achieved by developing on-line monitoring systems and models. The present work presents a feasibility study toward establishing the on-line monitoring of a pesto sauce production process by means of NIR spectroscopy and chemometric tools. The spectra of an intermediate product were acquired on-line and continuously by a NIR probe installed directly on the process line. Principal Component Analysis (PCA) was used both to perform an exploratory data analysis and to build Multivariate Statistical Process Control (MSPC) charts. Moreover, Partial Least Squares (PLS) regression was employed to compute real time prediction models for two different pesto quality parameters, namely, consistency and total lipids content. PCA highlighted some differences related to the origin of basil plants, the main pesto ingredient, such as plant age and supplier. MSPC charts were able to detect production stops/restarts. Finally, it was possible to obtain a rough estimation of the quality of some properties in the early production stage through PLS.

Petrochemical companies aim at assessing final product quality in real time, in order to rapidly deal with possible plant faults and to reduce chemical wastes and staff effort resulting from the many laboratory analyses performed every day. In order to answer these needs, the main purpose of the current work is to explore the feasibility of multiblock regression methods to build real-time monitoring models for the prediction of two quality properties of Acrylonitrile-Butadiene-Styrene (ABS) by fusing near-infrared (NIR) and process sensors data. Data come from a production plant, which operates continuously, and where four NIR probes are installed on-line, in addition to standard process sensors. Multiblock-PLS (MB-PLS) and Response-Oriented Sequential Alternation (ROSA) methods were here utilized to assess which of such sensors and plant areas were the most relevant for the quality parameters prediction. Several prediction models were constructed exploiting measurements provided by sensors active at different ABS production process stages. Both methods provided good prediction performances and permitted identification of the most relevant data blocks for the quality parameters’ prediction. Moreover, models built without considering recordings from the final stage of the process yielded prediction errors comparable to those involving all available data blocks. Thus, in principle, allowing final ABS quality to be estimated in real-time before the end of the process itself.

Partial Least Squares regression (PLS) is a widely used tool for predictive modelling, particularly when dealing with multivariate datasets with dependent variables exhibiting strong collinearities. However, when relationships between variables are non-linear or atypical data points have to be coped with, PLS calibration models may face challenges. In recent years, different variants of the original PLS algorithm have been proposed to overcome these limitations. On the one hand, several robust regression methods that down-weigh outlying observations during the model training phase like RoBoost-PLS have been developed to reduce the detrimental effect of outliers on the performance of PLS. On the other hand, local modelling approaches, like K-Nearest-Neighbours-Locally-Weighted-PLS (KNN-LW-PLS), have been designed to handle non-linearities by fitting for each new incoming sample a separate linear calibration model considering only its nearest-neighbours. Unfortunately, none of these strategies can address the two aforementioned problems simultaneously. This paper introduces a novel approach named Locally-Weighted-RoBoost-PLS (LW-RoBoost-PLS), that combines the strengths of both local and robust modelling methodologies in order to deal with non-linearities while mitigating at the same time the influence of outliers. The performance of LW-RoBoost-PLS was evaluated on simulated and real industrial data (with this latter resulting from a continuous Acrylonitrile-Butadiene-Styrene ABS production process conducted at Versalis S.p.A.), both characterised by the simultaneous presence of outliers and non-linear relationships among measured variables. In the two case-studies investigated here, LW-RoBoost-PLS outperformed RoBoost-PLS and KNN-LW-PLS, achieving considerable reductions in the prediction error and prediction bias, which demonstrates that this technique permits to effectively overcome the limitations of the other approaches. This paper describes a novel multivariate calibration approach named LW-RoBoost-PLS, which provides a solution for predictive modelling in scenarios where outliers and non-linearities co-exist. LW-RoBoost-PLS simultaneously handles non-linearities and outliers by combining local and robust modelling strategies, leading to improved prediction accuracy and reduced bias.