Artificial intelligence: Is it a good fit for drying?

Not long ago we have believed that our century will celebrate the era of humanoid robots, who substitute humans on routine operations. The reality did not follow predictions of science fiction, but the concept of Artificial Intelligence (AI) is still a hot topic. AI is defined as intelligence exhibited by machines or more specifically, mimicking human reasoning for control of complex systems. Unfortunately (or maybe fortunately), human brain is not able to manage exponentially growing information. This paradox is formulated by Lotfi Zadeh^[1]: “As the complexity of a system increases, our ability to make precise yet significant statements about its behavior diminishes until a threshold is reached beyond which precision and significance become almost mutually exclusive characteristics”. Therefore AI applications, including medical diagnostics, electronic trading, education, communication, robotics and remote sensing, are becoming popular tools to support, but not substitute human thinking. Although we are far from building smart robots, significant progress in AI industrial applications, such as self-driving cars, navigation systems, android smartphones and more is already visible. The number of top journals, publishing AI research, such as artificial intelligence, pattern analysis and machine intelligence, knowledge-based systems, expert systems, machine learning, neural networks, fuzzy sets and systems, etc. keeps on growing. However, the number of AI applications in food industry and, particularly in drying, is very scarce. How we can explain this gap? Could we improve drying by simple feedback control and optimization? To what extent we really need “intelligent” drying systems? All of these questions require careful analysis and clarification.

Our review shows that most applications of AI in drying are dealing with artificial neural networks (ANN), fuzzy logic (FL), evolutionary algorithms (EA) or their combinations. While ANN was recognized as an excellent tool for the learning and modelling of nonlinear processes,^[2,3] their combination with FL in form of hybrid adaptive neuro-fuzzy inference system (ANFIS) showed promising results in optimization of different drying technologies.^[45678910] However, there is still the lack of commercially available AI systems for drying of agricultural products and biomaterials. It could be due to the difficulty in interpretation of AI language for the practical needs of drying community or maybe due to simpler alternatives, like PID control. Development of AI application is usually lengthy process, involving expert knowledge, setting control rules and in some cases fine-tuning of rules through supervised or unsupervised learning. All of these steps require specific knowledge in system analysis and soft computing. Despite multiple efforts to implement AI control in drying, they did not demonstrate critical advantages over conventional PID control. The other reason could be immature state of drying control strategies. Freire at al.^[11] identified three possible reasons for the limited number of applications in the field of drying control: (i) complex interactions among process variables; (ii) lack of fundamental mathematical models of drying process; (iii) lack of online “soft” sensors for key control variables. Obviously, insufficient knowledge about thermo-physical and chemical changes of biomaterials under drying slows down further development of AI applications.

Could we improve drying by simple feedback control and optimization? Excellent review by Dufour^[12] showed that majority of control applications in drying require clear understanding of process dynamics. Also, since drying is inherently transient process, the feedback control is more advantageous as compared to the feedforward or AI control. Moisture content, temperature or product quality are commonly used variables in feedback control and optimization. Statistical models derived from central composite design (CCD) or response surface methodology (RSM) are still successfully used for optimization of drying. We like to use these models, because they allow meaningful interpretation of factorial effects, which enrich our knowledge about relationships between input and output variables. So, if the knowledge about drying kinetics and quality degradation was available, we probably don’t need “intelligent” drying.

To what extent we really need “intelligent” drying systems? Drying is non-linear process with uncontrolled disturbances, coupling of key variables, which increase process uncertainty in terms of energy consumption and product quality. Here is the place for AI applications, which could manage highly nonlinear relationships. The promising trend in AI applications for drying is related to the development of “soft” or “smart” sensors of product quality.^[11,13,14] With the industrial focus on the product quality, the interest in the development of online sensors of product quality is increasing. One of the avenues is computer vision technology, suitable for online estimation of multiple product quality indicators, such as volume, moisture, texture, color, density and porosity during drying.^[13] Electronic nose and tongue, combined with ANN or FL estimators allow to control mechanical and biochemical transformations in food during drying.^[14] Apparently, in the close future, the combination of factorial, ANN, inferential and kinetic models will constitute the “intelligent” estimators of quality in drying.

How we can envision further development of AI as a part of smart drying systems? It follows that AI is able to minimize uncertainty of drying and ensure consistent product quality. We could predict that development of AI applications would be related, but not limited to:

1. Online estimation of the product quality

2. Hybrid drying technologies

3. Distributed drying systems (for example, cross-flow dryers).

4. Non-quantifiable control objectives (for example, product quality).

5. Significant disturbances or inherently coupled processes

6. Human-system interface

7. Converging of human expert knowledge into the set of rules (machine learning).

At this point it is difficult to predict the development of AI for drying applications. We already know that AI is extremely useful to develop “soft sensors” and for control of complex drying systems. We could expect that further development of hybrid drying technologies will substantially increase complexity of control strategies, which, in turn, will create demand on AI techniques. We envision that inferential models or rule-based control strategies will be able to follow optimum drying conditions with respect to reduced operating costs and improved product quality.