Using Artificial Neural Network for Error Reduction in a Nondispersive Thermopile Device

The outputs of electronic devices are sensitive to the noise (thermal noise, background noise, etc.) and the change of operating conditions such as temperature or power supply voltage. As a result, the output data may have errors. Theoretically, if these changes are monitored, they can be used to support to correct the errors. The device using three nondispersive thermopiles to detect Fusarium encounters the same problem. The three outputs come from broadband, lambda(1), and lambda(2) thermopiles. The broadband thermopile works in 1 mu m to 20 mu mrange; lambda(1) and lambda(2)thermopileswork in 6.09 mu m +/- 0.02 mu m and 9.49 mu m +/- 0.22 mu m respectively. One temperature sensor and two voltage-monitoringmoduleswere installed tomonitor the operating conditions of the device. The information from the changes and the background noise of the device are used to train an artificial neural network. The training data are collected under unstable operating conditions. After the training, the trained neural network is used to fix errors in the output data. From the experiments results, the best error ratios of the raw and corrected data, Eraw/ Ecorrected, are 23.6, 13.8, and 18.5, respectively. The results were achieved by applying the external training and forcing method. Through the promising results, the technique of using support inputs and artificial neural network to correct data can be applied in any device which encounters similar problem. This method can help to improve accuracy and reliability of the sensor systems.