Low cost vector supercomputer
Based on high-end graphics cards with currently up to 240 SPs (streaming processors) vector supercomputers with a maximum of 960 SPs will be built for use with computer vision, real-time image processing applications and fast training of artificial neural networks (ANN). Open source development environments for computational intelligence and computer vision will be ported to the vector supercomputer.
Multi-computer multi-processor system
Multi-computer systems will be built on the basis of multi-core CPUs (central processing units) and high-speed Ethernet. Possible applications are real-time simulation of rapid dynamic processes, computer vision, training of large artificial neural networks (ANN with more than 1000 neurons) and the evolutionary optimization of the architecture of artificial neural networks for technical applications. The multi-computer multi-processor system with up to 32 CPUs is able to execute professional simulation software on standard multiprocessor capable operation systems.
Artificial Neural Network for the control of assembly-processes in aircraft construction
The behaviour of the material within the side shell of the fuselage section is virtualy modelled with artificial neural networks (ANN). Furthermore a neural-I-controller with the inverse nonlinear characteristic curve of the material behaviour of the side shell is developed to control the assembly process to control the geometry of the side shell at deviation from the requested position back into optimal assembly position.
Optical quality control for soot filters
A non-contact visual system is used for the quality inspection of filter particles. It scans the geometric characteristics of the opening of the filters. Measurements can be replicated on a sub pixel level through standard digital imaging and modern image processing technology. These programs are executed in C or C++ and have short run-times. Quick readings can be conducted with the existing infrastructure.
Energy management in motor vehicles
Vehicle manufacturers are receiving pressure from the EU to manufacture motor vehicles which produce less CO2 emissions. A cars electrical system without a Soft-Hybrid motor would need an average of 600 Watts. Fuel consumption and with it CO2 emission could be reduced by 5% if energy was fed into the electrical system from the energy used in braking, solar energy from solar panels on the roof (photovoltaic) and/or thermoelectric energy from a thermoelectric generator (which has been used in space flight for decades). A sophisticated energy management system based on computational intelligence and microelectronics can optimize energy consumption and reduce CO2 emissions effectively.
Using photogrammetry for measuring free form surfaces and outlines
A photogrammetric system uses two cameras to determine the height of any object with great detail on a value basis. The principles of the system are based on the stereopsis process of two human eyes. A structured surface is essential for an optimum correspondence analysis. Therefore a beamer lights up the structure providing an optimum measurement of faint objects. Combined with automation technologies like robotics, complex processes like lacquering a car can be automated.
Computational Intelligence Controller (CI-Controller), for nonlinear and time-variant control paths
Industrial processes are often non-linear and very complex regarding modelling. Robust rules with optimum quality control and detailed specifications can be too much for even popular procedures. A general approach was developed in this project for a self optimising neuro-adaptive PID industry controller. This controller uses evolutionary optimisation to adapt itself onto an unknown operating point the firm. It then uses parameter-logging (route variables and their associated controller parameters are stored in a database) and an artificial network (ANN) to very quickly recognize a non-linear process in the operating point, and regulate the changing process in accordance with the quality of control requirements. The artificial neural network is trained by the parameters from evolutionary optimisation. In a continuous operating time, tests show that the regulator behaves like an expert and enhances quality control. The specifications of the quality criterion and the adaptation come in time periods.
The evolutionary optimisation of the control process model runs parallel with real time control. The control process model is created from measurements of process values from the CI-controller.