Improvement Of Output Characteristic Of Humidity Sensors Of Lanthanum Ferrite/polymer Composite By Neural Network And Its FPGA Realization

Humidity sensors were made of nanometer material of lanthanum ferrite/polymer quaternary acrylic resin. The characteristics of the humidity sensor were measured such as sensitivity, humidity hysteresis, capacitance, impedance, response and recover properties, temperature influence, etc. The measurement voltage was IV with a frequency range of 10Hz ~ 100KHz. The measurement temperature was 5 to 30 oC. The humidity sensors possess high sensitivity: the resistance changes 4 orders in the relative humidity (RH) ranges from 11%RH to 98%RH. The maximum humidity hysteresis is about 4%RH. The linearities of resistance vs RH of the sensor are good when the testing frequency are 100Hz ~ 10KHz. The response time is 90s. And the recover time is 100sTemperature affects the applications of humidity sensors. The temperature coefficient of the humidity sensor was about 2.5~3%RH/℃. It can bring error in measurement. In this paper, we use artificial neural network (ANN) to improve the temperature affection of the humidity sensor. Firstly, we analyzed a BP ANN by Matlab. The different hide layer nerve cells, different transfer functions, and different learning algorithm were compared. It could be convergence when an ANN included 5 hide layer nerve cells, the transfer functions was an atanh function and the slearning algorithm was gradient with momentum term. Secondly, we discussed the precision of the hardware. The input of the net was 8 bits signed digits. The value of weight was 16 bits, and the value of threshold was 18bit. We simulated this ANN by Matlab, and proved that the system was correction. Finally, the BP ANN was realized by FPGA. The software of function and sequence were validated by useing Modelsim. The software was synthesized by Leonard 2002a and P&R was realized by using QuartusII. Then, we validated this system on ACEXlk EPlk50208-3 of Altera. The working frequency of 12MHz was set. A hardware system of off-line ANN can be gained. The precision for measement was about 3%RH. The system can resolve the problem come from temperature influence.