Search

The operation of industrial processes to achieve economic and safety objectives significantly depends on the use of advanced fault diagnosis and process control strategies. To design model-based fault diagnosis and control frameworks, developing sufficiently accurate models is an essential task. With the recent advance in computational power and data storage technologies, there have been growing interests in the use of machine learning techniques for process modeling and control. The performance of these techniques, however, depends on the use of a sufficient amount of high-quality data. The presence of uninformative and redundant data can hinder the model's ability to accurately predict dynamic behaviors. In this situation, the over-fitting also remains a challenging problem. As a result, it is necessary to use appropriate pre-processing tools, data mining techniques, and first-principles knowledge (if applicable) to achieve a reliable model. Motivated by the above considerations, this thesis focuses on the problem of hybrid and machine learning based modeling, fault detection, and control when collected data are not sufficient or there exist correlations between data samples. The first part of this thesis addresses the problem of system identification for heating, ventilation, and air conditioning (HVAC) systems when an insufficient amount of data is available. To this end, a hybrid machine learning based approach is developed where a pre-trained recurrent neural network (RNN) model (trained on a large amount of data from a representative zone) is leveraged to build a model for the zone in question. In the next phase, first-principles knowledge is integrated with data to develop a fault detection mechanism for HVAC systems using principal component analysis (PCA). The superior performance of the proposed approach, over the individual first-principles and data-driven based methods is shown. Finally, the problem of handling correlated data for the RNN-based model predictive control (MPC) implementations is addressed. To this end, PCA and autoencoder (AE)-based strategies are used to recognize the correlations that exist between data samples in both the input and output spaces. The constrained RNN-based MPC is then formulated by adding a PCA-based squared prediction error (SPE) constraint. PCA and AE-based optimization problems are also defined to calculate the achievable set-points. The efficacy of the proposed approaches over the nominal (standard) RNN-based MPC is demonstrated using different set-point tracking scenarios for a chemical reactor example. Thesis Doctor of Philosophy (PhD)

This thesis focuses on data-driven modeling and model predictive control for a monoclonal antibody process. The process uses live cell cultures such as Chinese Hamster Ovary (CHO) cells and thus needs special consideration. With the trend in the industry to move towards perfusion processes which are continuous allowing better productivity, advanced process control would play a vital role. Model Predictive Control (MPC) requires a suitable model to optimally control the process. First principles models for such live cell processes are complex and unsuitable for direct use in MPC. In this work, we focus on data-driven modeling given its suitability for use in control. The data-driven model is eventually also incorporated with some first principles knowledge for better performance and robustness. Data-driven modelling requires some input perturbation and data generation methods such as Pseudo-Random Binary Sequence (PRBS) inputs. By themselves, these methods are unsuitable for live cells as they can shock the system. To account for this, a suitable, intensified design of experiments (DOE) approach is used to perturb the data frequently enough to build a model of reasonable accuracy without significantly impacting live cells. This method is general enough to be used to identify appropriate input perturbation and data generation for many bioprocesses, and for our particular process of study it identified an input perturbation frequency of once per three days. The data-driven model is used in a model predictive control (MPC) scheme which respects biological constraints and can meet desired objectives. Further improvement in model robustness are made possible by incorporating first principles knowledge into the data-driven model. The methods are demonstrated on an advanced simulator developed by Sartorius with significant improvement over current industry standard. Thesis Master of Applied Science (MASc) Biopharma processes involving live cells are utilised to produce several critical products such as monoclonal antibodies which are state-of-the-art cancer therapeutics. Improving productivity for these would require advanced process control methods which in turn would require good models. This thesis focuses on introducing a method to build control relevant data driven models by using input perturbation and data generation suitable for live cell systems. The data driven model is utilised in a model predictive control scheme which meets control objectives while respecting biological constraints with better performance than industrial standards. Further improvement is made by incorporating first principles knowledge into the data driven model through a novel implementation of constrained subspace identification. The approaches are showcased on an advanced simulator test bed created by Sartorius.

Third submission of my master thesis, first one didn't finish for some reason, the second I think I forgot to include my name. This thesis focuses on the creation of new parallel hybrid model designs for prediction and control in batch and fed-batch reactors within Model Predictive Control (MPC) frameworks. In the hybrid model, the first principle (FP) explains the dynamics and the residual Subspace Identification (SID) model explains the error between the FP and the process. Modifications to the structure of the hybrid model are motivated by limitations of MPC frameworks. MPCs need accurate models to explain the system dynamics to make informed control decisions, and mechanistic models can be difficult to implement due to challenges of solving the optimization problem in real time. Two tools are demonstrated to help solve these problems. The first tool, Residual First Principle 0 Hybrid (RFP0H) model, helps to deal with the intractability of a mechanistic model in a hybrid modelling framework. The input for the FP model is kept constant and the SID predicts the error between the first principle and the process. Allowing for the desired output to be subtracted by the predicted FP to create a desired error value. Thus, MPC control only needs to be solved using the linear SID model in a linear or quadratic framework. Making a potentially intractable problem, tractable in MPC. This is demonstrated using a simulated fed-batch crystallization process. The second tool, Scaling Factor First Principle 0 Hybrid (SFFP0H) model, modifies the hybrid model structure to multiple the sub-models’ outputs together. The SID data driven model predicts a factor to scale the FP output for the process prediction. The results demonstrate that the SFFP0H model has increased predictive ability and has smaller variability in control compared to the RFP0H model. Helping to solve the problem of needing accurate models within an MPC formulation. This is demonstrated by using a laboratory scale batch polymerization process. Thesis Master of Applied Science (MASc)

The focus of this thesis is on building models by utilizing process information: from data, from our knowledge of physics, or both. The closer the model approximates reality, the better is the expected performance in forecasting, soft-sensing, process monitoring, optimization and advanced process control. In the domain of batch and continuous manufacturing, quality models can help in ensuring tightly controlled product quality, having safe and reliable operating conditions and reducing production/operation costs. To this end, first a parallel grey box model was built which makes use of a mechanistic model, and a subspace identification model for modeling a batch poly methyl methacrylate (PMMA) polymerisation process. The efficacy of such a parallel hybrid model in the context of a control problem was illustrated thereafter for reducing the volume of fines. Real-time implementation of models in many cases demand the model to be tractable and simple enough, and thus the parallel hybrid model was next adapted to have a linear representation, and then used for control computations. While the parallel hybrid modelling strategy shows great advantages in many applications, there can be other avenues of using fundamental process knowledge in conjunction with historical data. In one such approach, a unique way of adding mechanistic knowledge to improve the estimation ability of PLS models was proposed. The predictor matrix of PLS was augmented with additional trajectory information coming strategically from a mechanistic model. This augmented model was used as a soft-sensor to estimate batch end quality for a seeded batch crystallizer process. In a collaborative work with an industrial partner focusing on estimating important variables of a hydroprocessing unit, an operational data based input-output model was chosen as the right fit in the absence of available mechanistic knowledge. The usefulness of linear dynamic modeling tools for such applications was demonstrated. Thesis Doctor of Philosophy (PhD)

Wastewater treatment plants (WWTPs) face increasingly stringent effluent quality constraints as a result of rising environmental concerns. Efficient operation of the secondary clarification process is essential to be able to meet these strict regulations. Treatment plants can benefit greatly from making better use of available resources through improved automation and implementing more process systems engineering techniques to enhance plant performance. As such, the primary objective of this research is to utilize data-driven modeling techniques to obtain a representative model of a simplified secondary clarification unit in a WWTP. First, a deterministic subspace-based identification approach is used to estimate a linear state-space model of the secondary clarification process that can accurately predict process dynamics, with the ultimate objective of motivating the use of the subspace model in a model predictive control (MPC) framework for closed-loop control of the clarification process. To this end, a low-order subspace model which relates a set of typical measured outputs from a secondary clarifier to a set of typical inputs is identified and subsequently validated on simulated data obtained via Hydromantis's WWTP simulation software, GPS-X. Results illustrate that the subspace model is able to approximate the nonlinear process behaviour well and can effectively predict the dynamic output trajectory for various candidate input profiles, thus establishing its candidacy for use in MPC. Subsequently, a framework for forecasting the occurrence of sludge bulking--and consequently clarification failure--based on an engineered interaction variable that aims to capture the relationship between key input variables is proposed. Partial least squares discriminant analysis (PLS-DA) is used to discriminate between process conditions associated with clarification failure versus effective clarification. Preliminary results show that PLS-DA models augmented with the interaction variable demonstrate improved predictions and higher classification accuracy. Thesis Master of Applied Science (MASc)