Study Of Dynamic Reconfigurable Neural Network And Method Of Gas Recognition

Volatile organic compounds (VOC), which is mainly from building materials, indoor decoration materials, life and office supplies, outdoor industrial waste gas, and so on, have attracted increasing attention since it influences human health. So, preparation of gas sensor that can detect various VOC and study of mixing gas recognition are meaningful work.Because of cross-sensitivity of semiconductor gas sensors, it is impossible for a single gas sensor to recognize mixed gas exactly. Gas sensor array, composing by various doped SnO₂ gas sensors, is used for recognizing the mixed gas, which is composed of formaldehyde, ethanol, toluene and acetone. The concentration rang of each gas from 1 ppm～5 ppm.Artificial Neural Network (ANN) is one of the methods which are widely used in mixed gas recognition. BP, RBF and ANFIS neural networks are studied on application of mixed gas recognition and measurement in this paper. The prediction errors are: BP: 3.16%, RBF: 2.70%, ANFIS: 1.96% respectively. It shows that, in prediction accuracy: ANFIS>RBF>BP. PCA and ICA have been proved to be helpful on raising prediction accuracy.Realized by using programmable logical device (such as CPLD and FPGA), the neural network can implement its parallel computing function, accelerate the computation speed and meet real-time requirements. BP and RBF neural network are built and simulated in Xillinx Vertex II Pro development board separately in this paper, and the result shows that, the hardware implementation of the ANN is 2 orders of magnitude faster than the software in the operating speed of the neural network.Considering some problems existing in practical application of ANN, for example, the number of the neurons is difficult to be decided, the dynamic reconfigurable ANN (DRANN) is proposed in this paper. DRANN is based on the structure of the Network-on-chip (NoC) which has some characters, such as good scalability and parallel communication. DRANN not only can adjust the number of nodes in input, hidden and output layer real-time and dynamically, but also can realize the distribution neural network which is based on varied mathematical models, and make it possible for parallel communication between the distribution neural networks.