A data-driven process control strategy aligned with quality by design using a local linear modelling method and fault detection & diagnosis for twin-screw granulation

Abstract

The commercial implementation of continuous granulation requires an intelligent control framework compliant with Quality by Design (QbD) principles. To satisfy regulatory requirements this ‘control system’ cannot simply vary multiple operational variables simultaneously to regain quality attributes but should recognize the source of a disturbance and make the appropriate correction. A Fault Detection and Diagnosis (FDD) method is proposed as a novel element for such a framework, automating identification of the root cause within a non-linear design space. The disclosed method is considered a first step to realizing a QbD control system by presently assuming only one disturbance can occur at a time to highlight the value of this new approach, not producing a ready-to-use product.

The design space for twin-screw granulation was modelled using linear Partial Least Squares (PLS) with data collected using the Prediction Reliability Enhancing Parameter (PREP) method, which navigates non-linear data, producing a more comprehensive dataset. Particle size distribution (PSD) was the primary output used for assessing the granulation process. To address the inherently complex design space, local models were dynamically generated around the operation/disturbance point to reduce the error of fit for the FDD algorithm. Fault detection involved the verification of a disturbance and perturbation of a verified input whereas the fault diagnosis phase employed an optimization framework comparing observed and predicted PSDs associated with the deviations. This design helps overcome the interdependent behavior of process inputs and enables systematic isolation of the correct root cause. The algorithm is evaluated in this study by three case studies.

Introduction

An intelligent control framework as envisioned under Quality by Design (QbD) by the International Council for Harmonisation (ICH) is capable of relating critical quality attributes (CQA) to critical process parameters (CPP) within the design space of a process. Aspects of this framework are discussed in the ICH guidelines Q8, Q9 and Q10 (Williams, 2009) that profess quality derived by process knowledge is superior to reliance on product testing (though the latter is still required). Realistically, such a control system is meant to maintain ideal operations and in the face of a deviation, identify the root cause(s) of the fault and then enact knowledge-based interventions to correct the system. For pharmaceutical manufacturing, intelligence infers a mechanistic awareness of the process to ensure the corrective actions taken are explainable as to how they returned the system to its validated operating state rather than simply varying multiple parameters simultaneously to mask an underlying problem. Moreover, intelligent control of a continuous process should involve real-time assessments of quality and the mentioned corrective actions, when necessary, are made rapidly so that minimal waste occurs. With twin-screw granulation beginning to be implemented commercially in drug production, research on developing responsive and accurate predictive models for its design space and using those models in an intelligent control strategy for this continuous process needs to be accelerated.

The design space for twin-screw granulation describes all of the relationships between process parameters (e.g. operations and formulation variables) and product attributes. A well-defined design space model captures process behaviour and can be used for optimization or in the case of the present study, as a framework for a process control strategy. Evaluating this design space can be done experimentally using Design of Experiments (DOE) methodologies and statistical tools (Fonteyne et al., 2014, Heo and Lee, 2018, Kosanovich et al., 1994, Pauli et al., 2018, Sierra-Vega et al., 2024, Willecke et al., 2018), or numerically by discrete particle simulations (Wang et al., 2025, Zheng et al., 2022) or population balance modelling (Barrasso et al., 2013). These approaches at present provide a partial representation of the process within specific operating regions; their challenges in modeling twin-screw granulation are well recognized: experiments are costly to run, limiting the number and levels of parameters to be explored; discrete particle simulations lack the proper contact-force mechanics to closely resemble the real process and generate results slower than doing experiments; and population balance models, while still being developed, will always be constrained to the experimental data used in their validation. This means that all of these approaches should be considered equally well suited to defining the design space for twin-screw granulation. When considering multiple step operations (like a twin-screw granulator and fluid bed dryer), another approach has also been considered called Model Driven Design which uses dynamic flowsheet modelling to link physics-based and PBM models together to mapping a design space (Wang et al., 2022). Recent approaches have also utilized machine learning based methods for the purpose of modeling and control of the wet granulation process (Vega-Zambrano et al., 2025, Alharby et al., 2024, Shirazian et al., 2017).

A review authored by many of the largest global pharmaceutical manufacturers covered the process, process analytical technologies (PAT), and general control strategies for automation of twin-screw wet granulation (TSWG) (Dahlgren et al., 2019), noting the major challenges in implementing a control framework will be associated with oversight of multiple unit operations making up a production line since they will have differing response times to start-up conditions or disturbances. To date, the literature on control strategies for TSWG has focused either on the extruder (including its feeders and pumps) (Pereira et al., 2019) or the extruder with downstream units like a fluid bed dryer (Celikovic et al., 2023, Jelsch et al., 2023, Pauli et al., 2018, Silva et al., 2017, Singh et al., 2014). A classic feedback control system for TSWG was demonstrated by Singh et al. (Singh et al., 2014), which oversaw a twin-screw extruder, feeders and pumps as well as a fluid bed dryer, mill and tablet press, with inline feedback quality descriptors based on API concentration and granule size. In their case, a CPP was pre-assigned to control each CQA being monitored in order to prevent multiple variables being adjusted each time corrective action was needed. This is an example of a multiloop single-input single-output (SISO) control strategy, whereas most of the controllers in the literature appear to adopt a multivariable multi-input multi-output (MIMO) strategy where multiple CPP can be adjusted simultaneously. Multivariable MIMO strategies for controlling TSWG generally fall under the definition of model predictive control (MPC) strategies and are based on multivariate statistics (Vega-Zambrano et al., 2025). MPC strategies reportedly perform much better for a continuous granulation process compared to other traditional approaches like linear quadratic regulator (LQR) and proportional-integral-derivative (PID) due to their knowledge of the process design space and their predictive capabilities (Chindrus et al., 2023). In a study by Silva et al (Silva et al., 2017), they explored the identification of seven different disturbances to sensors in the ConsiGma-25 system (GEA) which, was trained/validated on five normal operating conditions to establish a reduced design space for its 35 monitored process variables using principal component analysis (PCA). Hotelling’s T2 values and Q residuals (Kourti and MacGregor, 1995, MacGregor and Kourti, 1995) were used by Silva et al to measure variations in the trial samples and identify how much each monitored variable contributed to a detected disturbance. Describing particle size distribution (PSD) as an immediate CQA affecting product quality, one group has developed a data-driven MPC while highlighting the non-linearity of the design space for TSWG to which they addressed by using a local-linear model tree (LoLiMoT) (Celikovic et al., 2023). The LoLiMoT method uses pre-determined weighting coefficients to tell the MPC how to partition areas of the design space for its decision making. In a subsequent paper, the authors (Celikovic et al., 2024) demonstrate the control system’s ability to detect faults in target specifications for API and moisture content. Their fault detection algorithm, which was designed to alert operators to process changes, used a disturbance estimation determined by sensors measuring the output composition. It is assumed in their control system that disturbances will be distinctive to each fault, which for API content or liquid content may be true but generally that approach seems to ignore the recognized non-linearity of the process. Despite some concerns over these assumptions, we consider the work of Celikovic to be the most advanced to date in compliancy with the principles of QbD.

Reported MPC-based approaches depend on multivariate regression models to correct deviations, however, they offer no diagnosis of the root cause of a disturbance and may ultimately mask the fault(s) (undesired product quality) by adjusting multiple process parameters simultaneously to regain CQA targets. We propose that the missing element in these current studies on QbD-compliant control strategies for TSWG using an MPC is Fault Detection and Diagnosis (FDD), which is automation with the primary responsibility of seeking out the root cause of a fault without pre-determination by relying on neighbor information in the design space. We are focused on process faults in the present study, which are caused by changes in process parameters rather than input variances related to material properties (for which at least one study has started to examine such faults (Puig and Simani, 2021)). FDD includes a means to determine that a fault is present (detection) followed by performing a diagnosis, which is in reference to isolating (type, location and time of detection) and identifying (size and time-invariant behavior) the fault. Our data-driven FDD algorithm uses a statistical model of the design space to monitor and predict output variables using process parameters, and hence its accuracy is dependent on the quality of the model (Heo and Lee, 2018, Puig and Simani, 2021). Alternatively, a FDD algorithm can be based on inverse modeling (de Peralta Menendez and Andino, 1998) where the process model is used to predict the input conditions responsible for the measured quality variable. However, this method is highly sensitive to model accuracy and can lead to non-unique or unstable solutions. Dunia et al. (Dunia et al., 1996) developed a data-driven FDD method where a Principal Component Analysis (PCA) model based on historical sensor data was used to detect faulty sensors. By their approach, each sensor was temporarily considered missing, and the corresponding measurement was reconstructed using correlations with other sensors identified by the PCA model. A follow-up study by Dunia and Qin (Dunia and Joe Qin, 1998) developed a PCA-based method for sensor and process fault detection and reconstruction, introducing a Fault Subspace for each potential fault to the system, representing the expected deviation of measurements when the specified fault occurs. When a fault was detected, a new measurement vector would be constructed by searching through the different Fault Subspaces and minimizing the distance to the principal-component subspace.

To advance approaches being considered in QbD for twin-screw granulation, this study outlines and tests the first known automated FDD strategy as part of a MPC to restrain the feasible corrective measures taken upon detection of a disturbance. As a first study on this approach, disturbances caused by only one fault were considered. Included in this work, a novel method for handling non-linearities detected in the design space of a granulation process will be included that avoids the necessity of pre-determined weighting coefficients (LoLiMoT method) by generating target-specific local Partial Least Squares (PLS) models based on neighboring PSD data in the PCA score space. These local models and controlled perturbation are integrated into the proposed FDD framework enabling adaptation to local process behavior. Three case studies will be used to examine the strengths and weakness of these novel additions to control strategy theory for twin-screw granulation.

Download the full article as PDF here A data-driven process control strategy aligned with quality by design using a local linear modelling method and fault detection & diagnosis for twin-screw granulation

or continue reading here

Materials

A single model formulation was studied consisting of 39% α-lactose monohydrate (Flowlac®100; Meggle Pharma, Germany), 20% microcrystalline cellulose (MCC, Avicel PH102; International Flavors & Fragrances Inc., Midland, MI, USA), 40% hydroxypropyl methylcellulose (HPMC, METHOCEL ™ K4M; International Flavors & Fragrances Inc.) and, 1% acetaminophen (Sigma-Aldrich, St. Louis, MO, USA); the acetaminophen was included in all samples produced but only used as a tracer during initial screen studies prior to the work reported in this paper. The binder solution was made with distilled water and varying concentrations (2–6 wt%) of hydroxypropyl methylcellulose (METHOCEL™ E3PLV; International Flavors & Fragrances Inc.). All ingredients were oven-dried at 50 °C for at least five hours prior to trials.

Kavitha Sivanathan, Prashant Mhaskar, Michael R. Thompson, A data-driven process control strategy aligned with quality by design using a local linear modelling method and fault detection & diagnosis for twin-screw granulation, International Journal of Pharmaceutics, Volume 700, 2026, 127082, ISSN 0378-5173, https://doi.org/10.1016/j.ijpharm.2026.127082.

Read also our introduction article on Quality by Design (QbD) here: