Computational Intelligence Techniques for Bioprocess Modelling, Supervision and Control

This is an introductory chapter that presents a general review of some Computational Intelligence (CI) techniques used today, both in the biotechnology industry and in academic research. Various applications in bioprocess-related tasks are presented and discussed. The aim of putting forth a surveying view of the main tendencies in this field is to provide a broad panorama of the research in the intersection between the two areas, to highlight the popularity of a few CI techniques in Bioprocess applications and to discuss the potential benefits that other not so explored CI techniques could offer.

Industrial bioprocesses present a very difficult challenge to control engineers. Problems associated with the nature of the organisms in the process and difficulties related to obtaining accurate information regarding the progression of the process make controlling and monitoring particularly challenging. The lack of suitable and robust on-line sensors for key variables such as biomass or product concentration has been considered as a serious obstruction for the implementation of control and optimization of bioprocesses. Considering biomass concentration alone, there are typically two methods available to measure this value - direct or indirect methods. To measure the biomass directly, several techniques have been applied: optical density measurements, capacity measurements, high performance liquid chromatography (HPLC), nuclear magnetic resonance (NMR), laser cytometry or biosensors. In addition to the high costs associated with these measuring devices, their reliability can be poor when applied to large-scale systems. It is still the case that most industrial bioprocess control policies are based upon the use of infrequent off-line assay information for process operator supervision. The low sampling frequency associated with such measurements and the inevitable delays in taking samples and performing laboratory tests inevitably compromises the quality of control that is possible using such measurements. As a result of this an alternative approach, that of indirect measurement has attracted a great deal of attention over the last 20 years or so. Indirect measurements of biomass are mathematical algorithms that can produce estimates of unmeasured biomass concentration using the continuously measured variables such as temperature, dissolved oxygen, pH and off-gas concentration. The method of estimating the quality related variables from measurements of secondary variables is referred to as ‘Inferential Estimation’ and these mathematical estimators are usually referred to as ‘Software Sensors’. Improved control of the process can be achieved by measuring and setting up a feedback control system using these secondary variables. Such control strategies are referred to as ‘Inferential Controllers’. Software sensors usually rely on a model to describe the process, thus different techniques have been proposed for on-line inferential estimation in bioprocesses just as different models exist. Among these applications, the majority have been based upon mechanistic, artificial neural network (ANN) or other empirical models. In this chapter, some of the important and recent research conducted on Software Sensors is reviewed and the associated techniques are introduced with examples and case studies.

Nowadays, bioprocesses play a key role in the production of high-added value products in the pharmaceutical industry and measurements of the main component concentrations are of great importance for monitoring cell cultures. Although some hardware sensors are readily available, they often have several drawbacks, including purchase and maintenance costs, sample destruction, discrete-time measurements (instead of continuous ones), processing delay, calibration, sterilization, disturbances in the hydrodynamic conditions inside the bioreactor, etc. It is therefore of interest to use software sensors which reconstruct on-line some component concentrations in continuous time. Software sensors are based on the theory of state estimation. In this chapter, some state estimation techniques are reviewed, and two important situations are distinguished: (a) some component concentrations can be measured and a dynamic model of the bioprocess can be established and (b) only basic operating signals, such as pH, base addition, stirrer speed, feed rates, can be measured on-line and it is difficult (or even impossible) to build a mechanistic model linking these variables. In the latter case, a neural network approach appears particularly suitable, and is largely illustrated in this chapter by real-life experimental applications.

This chapter is focused on developing more efficient computational schemes for modeling and control of chemical and biochemical process systems. In the first part of the chapter a theoretical framework for estimation of general process kinetic rates based on Artificial Neural Network (ANN) models is introduced. Two scenarios are considered: i) Partly known (measured) process states and completely known kinetic parameters; ii) Partly known process states and kinetic parameters. The two scenarios are successfully tested for two benchmark problems, estimation of the precipitation rate of calcium phosphate and estimation of sugar crystallization growth rate. In the second part of the chapter the developed ANN-based models are integrated into the structure of a nonlinear model predictive control (MPC). The proposed ANN-MPC control scheme is a promising framework when the process is strongly nonlinear and input-output data is the only process information available.

Wastewater treatment systems (WWTS) are based on complex, dynamic, and highly nonlinear processes. Depending on the design and the specific application, these systems can achieve biological nitrogen and phosphorus removal, besides removal of organic carbon substances. Also, depending on the type/quantity of effluent to be treated, different configurations can be proposed being the most common the aerobic, anoxic and anaerobic schemes working in continuous or sequential modes. In common they have the fact that they deal with communities of different microorganisms which are more or less sensitive to external and/or internal variations of the process conditions. This is a real problem having in mind that usually, influents to be treated are highly inconsistent in flow and concentration being the changes most of the times completely unpredictable. In this way, the knowledge and experience obtained from operational difficulties of one wastewater treatment plant cannot be easily generalized to another.

Recent increased regulation over discharge of nutrients to receiving waterways, associated with operational difficulties of wastewater treatment plants, resulted in an increased need for tools to evaluate the organic matter and nutrient-removal capabilities of wastewater treatment processes. However, the description of its behavior requires complex models that involve a very large number of state variables, parameters and biochemical phenomena that have to be accurately identified and quantified. When deterministic models as the activated sludge model (ASM) and anaerobic digestion model (ADM), fail in predicting the WWT process, alternative modeling methodologies usually known as black-box models, may complement and support the knowledge about the wastewater treatment process and operation. Black-box models are entirely identified based on input–output data without reflecting physical, biological or chemical process knowledge in the model structure. What we purpose in this chapter is to identify and detail the black-box models, also known as artificial intelligent (AI) techniques that are being used for WWT monitoring and control. Particularly focused will be the Multivariate Statistical Methods (MVS), Knowledge Based Systems (KBS), Fuzzy Logic (FL)and Artificial Neural Networks (ANN) as they already proved its potential in different real applications.

The combination of multiobjective genetic algorithms with wastewater treatment plant (WWTP) models provides an efficient framework for the evaluation, optimisation and comparison of WWTP control laws. This chapter presents a methodology developed for this efficient combination. Existing models and simulation software are used. They are combined with NSGA-II, a multiobjective genetic algorithm capable of finding the best tradeoffs (Pareto front) among multiple opposed objectives. Long term evaluations of the optimized solutions are proposed to check their robustness. An application of the methodology on the Benchmark Simulation Model 1 is presented and illustrates the benefits of the methodology.

The oxygen mass transfer coefficient (K_La) is of paramount importance in conducting aerobic fermentation. K_La also serves to compare the efficiency of bioreactors and their mixing devices as well as being an important scale-up factor. In submerged fermentations, four methods are available to estimate the overall oxygen mass transfer coefficient (K_La): the dynamic method, the stationary method based on a previous determination of the oxygen uptake rate (Q_O2X), the gaseous oxygen balance and the carbon dioxide balance. Each method provides a distinct estimation of the value of K_La. Data reconciliation can be used to obtain the most probable value of K_La by minimizing an objective function that includes measurement terms and oxygen conservation models, each being weighted according to their level of confidence. Another alternative, for a more rapid determination of K_La, is using a neural network which has been previously trained to predict K_La from the series of oxygen conservation models. Results obtained with this new approach show that K_La can be predicted rapidly and gives values that are equivalent to those obtained with the conventional data reconciliation algorithm.

Batch and Fed-Batch cultivation processes are used extensively in many industries where a major issue today is to reduce the production losses due to sensitivity to disturbances occurring between batches and within batches. In order to ensure consistent product quality by eliminating the influence of process disturbances it is very important to consider implementation of monitoring and control and thereby significantly improve the economic impact for these industries. A data driven modeling methodology is described for batch and fed batch processes which is based upon data obtained from operating processes. The chapter illustrates how additional production experiments may be designed to improve model quality for control. The chapter also describes how the developed models may be used for process monitoring, for ensuring process reproducibility through control and for optimizing process performance by enforcing learning from previous batch runs through Learning Model Predictive Control (L-MPC).

This Chapter describes two approaches to ensuring the production quality of batch biotechnological processes. The first makes use of the multivariate statistical data analysis and multivariate statistical process control (MSPC) or better termed multivariate statistical process performance monitoring (MSPM). An industrial application is described to the interrogation of data from a reaction vessel producing an active pharmaceutical ingredient (API) which enabled the realization of a better understanding of the factors causing the onset of an impurity formation to be obtained as well demonstrating the power of multivariate statistical data analysis techniques to provide an enhanced understanding of the process. In the second application, a simulation study of batch-to-batch iterative learning control strategy is presented where the batch control actions for the next batch are adjusted using the information obtained from current and previous batches. The control policy updating is calculated using a model linearized around a reference batch. In order to cope with process variations and disturbances, the reference batch can be taken as the immediate previous batch. After each batch, the newly obtained process operation data is added to the historical data base and an updated linearized model is re-identified. Since the control actions during different stages of a batch are usually correlated, it is proposed here that the linearized model can be identified from partial least square regression.

In this chapter we describe a software tool for modelling fermentation processes, the FerMoANN, which allows researchers in biology and biotechnology areas to access the potential of Artificial Neural Networks (ANNs) for this task. The FerMoANN is tested and validated using two fermentation processes, an Escherichia coli recombinant protein production and the production of a secreted protein with Saccharomyces cerevisiae in fed-batch reactors. The application to these two case studies, tested for different configurations of feedforward ANNs, illustrate the usefulness of these structures, when trained according to a supervised learning paradigm.