Chemical Oxygen Demand (COD) On-Line Monitoring Network Node And Pattern Recognition Algorithm Of Sensor Array

Chemical oxygen demand (COD) is the most important index for evaluating the degree of water pollution. However, the traditional detecting methods are tedious and time-consuming, exhibit weak oxidation ability for some organics difficult to degrade and prone to secondary pollution. Some new detecting methods such as ozone oxidation method significantly improve efficiency, but are heavily dependent on the laboratory environment and thus not suitable for real-time field detection. Therefore, this paper developed a set of intelligent and convenient COD on-line monitoring network node system with stable performance and rapid response, supporting wireless communication function, can carry out multi-point real-time COD monitoring in the wild.Firstly, we analysed the scheme of photocatalytic degradation of organics with TiO₂ nanotube array electrode, proposed the general design scheme of COD on-line monitoring network node. Then to carry out from two aspects of hardware and software design: in hardware design, emphatically introduced the design of constant potential control circuit with solution ohmic-drop compensation function, differential measurement circuit and ZigBee wireless communication module, built a complete hardware system and took performance test for constant potential control. Software system design was divided into two aspects of lower and upper position machine. In the development of lower position machine, we emphatically introduced the algorithms of filtering procession, final value judgment, curve fitting and numerical integration. In the development of upper position machine we introduced the function of measurement curve display based on Labview virtual instrument. The system preliminarily developed had been tested for the actual COD detection experiments.Finally, using Learning Vector Quantization (LVQ) algorithms for clustering, then the clustering vectors were used as hidden layer neurons of probabilistic neural network (PNN) to implement LVQ-PNN recognition algorithms. Besides, measures such as normalization of training samples and clustering vectors, reasonable setting the number of initial clustering vectors, removing dead neurons and improving the kernel width were taken to further optimize the algorithms. The results of simulation experiments demonstrated that the improved LVQ-PNN algorithms overcome the shortcomings of PNN as occupying huge memory space and slow response speed, was particularly suitable for on-line detection.