Sensors selection for gas mixtures analysis systems

The paper addresses the problem of sensors selection providing effective operation of the sensor system e.g. for the multi-component gas mixtures analysis. The neural network sensitivity analysis approach to the problem was investigated. In the initial phase several neural network structures were created using the experimental data. In further steps the sensitivities of the neural network outputs for the inputs were calculated using derivative calculus. Eventually the least significant inputs of the neural networks were found, pointing to the redundant sensor in the experimental setup. The sensor was removed and the whole cycle repeated leading to further reductions in the sensor array. It is shown that the performance of the system with reduced array is similar or better than the original one. Sensitivity method, derivating from the neural networks theory may be easily adapted to other fields where the problem of distinguishing between necessary and redundant information occurs.