The limitations of the bioprocess control systems do not concern only the measurements or models, but at the same time much valuable human knowledge is only available in a qualitative heuristic form.

Hence, it has been found that the knowledge-based control structures using the human decisional factor (i. e. a subjectively element) offer sometimes better results. Moreover, the computer performances are developed in the detriment of the general knowledge concerning life phenomena and do not promote advanced comprehension upon the metabolic routes of bioprocesses. Consequently, the intelligent techniques (i. e. neural nets, fuzzy structures, genetic algorithms or expert systems) are capable of simulating human expert-like reasoning and decision making, dealing with uncertainties and imprecise information [24].

As the human perception about the bioprocess is commonly altered by the psychological factors, the intelligent control systems founded (only) on the human subjective knowledge is less valuable than the control systems who utilize the objective information fitted by a conceptual model. Hence, the literature recommends the intelligent control techniques utilization only if the control structure based on quantitative models fails.

Frequently, different process parameters are controlled in order to follow predefined transitory trajectories. Such control strategies can be designed by a trial-and-error approach in combination with operator’s experience and statistical analysis of historic data.

a) One method for automatic bioprocesses control using AI is based on expert systems (ES) that reproduce the human operator’ rules of action. The literature presents several examples how to transfer the knowledge from operators into knowledge-rule bases [38]. An ES conceptual architecture is presented in Figure 4.

The most used ES systems in bioprocess control are applied for supervisory control, or process monitoring and diagnosis.

Moreover, the ES logic is used to translate human language into a mathematical description. The parameters tuning is then regulated by phase detection based on if…then rules, conditional statements representing heuristic reasoning in which if expresses the condition to be applied and then-the action to be done. Of course, at the same time, it is not possible to calculate optimal parameter’ values with this method. For example [39] an ES was developed in order to supervise a conventional control system applied to fed-batch baker’s yeast cultivation and to surmount its limitation. Expert system BIOGENES can execute standard process control tasks, but also advanced control tasks: process data classification;

qualitative process state identification (metabolic state, process phase, substrate feeding); supervisory control through corrective actions.

SUPERVISOR

One of the main limits in developing ES is the knowledge acquisition process due to: (1) the linguistic rules formulation by human experts, i. e. the analytical description of their actions during the bioprocess (manual) control; (2) the loss of information during the transfer between the bioprocess expert and the IT specialist; (3) the subjectivity of human expert regarding own decision / rules.

b) Another AI technique, the fuzzy approach is based on fuzzy sets and fuzzy reasoning. In actual AI systems, fuzzy rules are often applied together with different types of models / parameter / state estimators. These fuzzy rules can be regarded as problem specific basis function system [40]. Any variable can be a fuzzy variable, particularly recommended when it is not possible to define its value in a given situation. One define fuzzy sets in the form of membership functions (between 0 and 1) in order to express what is likely to be considered as degree / level for a certain characteristic (high, medium, low). Relationships between fuzzy variables can be formulated with fuzzy logic operators (and, or, not) and processed by fuzzy logic. Fuzzy rules reflect the rules of thumb used in everyday practice and can be processed as if…then expressions. With a set of fuzzy rules, considered as universal process approximates, the behavior of a system can be described quite accurately. There are many applications: (a) hierarchical fuzzy models within the framework of orthonormal basis functions41 (Laguerre and Kautz functions); (b) several important use of fuzzy control in the Japanese bioindustry by the companies Ajinomoto, Sankyo or Nippon Roche [42]; (c) the control of the a-amylase fed batch bioprocess with the recombinant E. coli to maintain glucose and ethanol at low concentrations with 2 fuzzy controllers for feed rate control: feed forward and feedback [43] (see Figure 5 below):

c) One can use artificial neural networks (ANN) to get predictions about biosystem behavior. The traditionally used format of ANN is the feed forward. Given a set of process measurements, the output of ANN can be estimated parameters or process variables. The weights applied to the process measurements as inputs are determined through the "training process" of the ANN [44]. To train the ANN it is to get complete process information, corresponding to the NN inputs and outputs, from the data gathered in a set of fermentation runs. This set defines "an experimental space" and the ANN will predict outputs accurately only within this range and not beyond it.

Various applications were studied: (a) biomass and recombinant protein concentration estimation via feed forward NN for a fed batch bioprocess with a recombinant E. coli [45, 46]; (b) two types of NN (input/output and continuous externally recurrent) can control the batch and fed batch piruvat production from glucose and acetate with a recombinant strain of E. coli [47]; (c) two NN also to control the submerged bioprocess of Monascus anka fungus cultivation (the temperature and the dissolved oxygen are the inputs and the controlled outputs are the glucoamylase activity and the concentration of the red pigment [48]; (d) NN based soft sensor for online biomass estimation in fed bioprocess for polyhydroxibutirate production [49]; (e) media formulation optimization with genetic algorithm evaluated by ANN [50].

d) Because all types of information must be used in order to improve the bioprocess control: mathematical / deterministic models, heuristic knowledge, rule-based reasoning, a new control structure is developed in the last years — i. e. hybrid control system (HCS). HCS acts on both parts of bioprocess control: conventional control systems (i. e. based on a priori model) merge with AI techniques, in a complementary way: if a priori (mathematical) model exists, it will be preferred; else the linguistic rules (i. e. expert systems / fuzzy techniques) will be used.

This is an Intelligent Control Structure (ICS) based on Hybrid Control Techniques (HCT). The most widely used hybrid structure combines balance equations with ANN: (a) balance equations for substrate and cell concentrations coupled with ANN for growth rate model in case of bakers ‘yeast fed batch cultivation [51]; (b) ANN is responsible for modeling the unknown kinetics in applications with the yeast Saccharomyces cerevisiae or activated sludge urban wastewaters bio treatment [52]; (c) batch bioprocess of animal cells [53].