Optimization of a MLP network structure for a real-time PLC application

The neural networks find many applications today in different kinds of real-time working systems. To obtain short execution times and high recognition accuracy in real-time decision-making systems becomes a question of first importance. Therefore, the requirements to the recognition stage in such systems in reference to reduce the reaction time grow up. In the proposed research a new method for optimization of a MLP network structure for a real-time programmable logic controllers (PLC) application is presented. The optimization is accomplished in two steps. First the DCT coefficients are calculated over radial profiles of the objects which form a vector in the frequency parametrical space. This vector describes the corresponding 2D object and is applied as Initial Input Set to the MLP neural network structure. The size of each input for MLP vector is reduced applying modified coefficient of variations (MCV) to evaluate the outlier values. Second the reduced input set is divided and grouped into a number of small MLPs based on analysis of the degree of correlation between the inputs. The trained MLPs are downloaded in a Siemens PLC S7-300 for on-line real-time work in a parallel recognition mode. The proposed optimization is tested for four different 2D objects captured by a CCD matrix camera. The achieved results are represented and analyzed.