The results of PSYCHO are summarized in the next following list of software developments and technical know-how.
A set of new identification and control schemes based on the former schemes.
Algorithms for traning the neural networks, including genetic algorithms and stochastic methods.
Comparisons between classical and advanced neural controllers.
Software implementation and test of neural controllers for specific industrial applications defined by the endorsers.
Specifications for an industrial research project to be carried out by the endorsers.
The PSYCHO project has developed, generalized and extrapolated a wider set of non-linear identification and control schemes than initially expected, based on the following algorithms:
Identification algorithms: MLP, Modified Elman Networks, RBF (Radial Basis Functions), Genetic Algorithms, NGST, Reinforcement Learning, FasArt (Fuzzy Adaptive System ART-based), Genetic Algorithms.
Control algorithms: NMPC (Non-linear Predictive Control), MRNC (Model Reference Non-linear Control), Reinforcement learning, NGST, FasArt, FasBack, Genetic Algorithms.
The models developed are highly oriented towards industrial applications. All these identification and control models have been tested with data from the proposed theoretical testbed plants, with simulated data from the testbed plants (including a binary NEREFCO petroleum distillation column), real data of penicillin fermentation from ANTIBIOTICOS and data based on other biochemical models.
The models have been generalized and extrapolated for an important number of industrial applications. Other industrial and theoretical testbeds have been incorporated by the consortium with the aim of demonstrating the applicability of these models in a broader number of industrial testbeds than initially expected, like SHELL Crude Distillation Unit, Catalytic Reformer Debutaniser Column, 3 theoretical systems, Baker's yeast fed-fatch fermentation, Waste water neutralization plant, mobile robots, Mobile robot, Traffic control in ATM Networks, Batch Chemical Reactor Temperature Control.
The proposed research is directed at developing advanced neural-network-based schemes for identification and contro, with increased stability and robustness. They will be highly oriented to industrial applications, and especially for chemical and biochemical reactors. The technical objectives of the project are:
-Neural networks capable of implicit identification of complex non linear processes.
-Adaptive neural controllers which, trained in laboratory using experimental reactors, will be capable of generalizing their skills in the scaling-up process to real plants.
The scientific objectives of the project are:
- New neural schemes based on Elman and Jordan models,ART(Adaptive Resonance Theory),NGST(Neural Group Selection Theory)and use of Value Repertoires.
- Efficient algorithms for tracing the neural networks, including genetic algorithms and stochastic search methods.
- Software implementation and test of neural controllers for specific industrial applications.
- Specifications for a subsequent industrial research project to be carried out by the endorsers
The achievement of positive results in the present project will imply more reliable operations, reduced scrap, energy consumption and savings in materials. It will produce a considerable improvement in the quality, safety and environmental impact of the existing processes by reducing the production time, increasing yields and reducing byproducts(waste). Moreover it will imply a reduction of the time and efforts necessary in the scaling-up from laboratory to production and a reduction of chemical or biological hazards.
Funding SchemeCSC - Cost-sharing contracts
6200 AL Maastricht
CF2 1XH Cardiff