|제목(국문)||입자 크기 분포의 현장 모니터링을 위해 뉴럴 네트워크 알고리즘을 이용한 MEMS 기반 전기식 다단 임팩터|
|제목(영문)||Microfluidic electrical cascade impactor using neural-network-based inversion algorithm for on-site monitoring of particle size distribution|
Electrical particle sensors generally have problems associated with relatively poor resolution and assume geometric standard deviations when estimating the particle size distributions. A neural-network-based inversion algorithm was thus developed to improve the measurement accuracy with an alternative approach. The algorithm was applied to the microfluidic electrical cascade impactor, which was redesigned to enable monitoring in ambient atmospheric conditions. In the case of the our previous work, MEMS-based electrical cascade impactor, the number of impaction stages is only four, and the smallest cut-off diameter is 0.3 μm, because the minipump is difficult to operate under conditions with large pressure differences. Therefore, the measurement results of the particle size distribution from the electrical sensors often deviate from those of the reference instruments. Unlike the previous version, this impactor was redesigned for accurate monitoring in ambient atmospheric conditions; (i) a triode diffusion charger was adopted to minimize the particle loss and achieve robust charging in environments containing a mixture of various particle compositions, and (ii) the volumetric flow rate of the microfluidic electrical cascade impactor was increased and the number of stages was reduced to enhance measurement resolution. The components of this impactor were characterized using the singly charged aerosol reference (SCAR) to derive the theoretical particle size distribution using a conventional algorithm based on Tikhonov regularization. The neural network comprised an input layer (the set of raw data), an output layer (the parameters of the lognormal size distribution), and a hidden layer trained by the multilayer perceptron ensemble with the best performance in terms of validation. After training this ensemble, it is possible for the microfluidic electrical cascade impactor to derive the particle size distribution from only measured currents via the trained data. The measurement results of the developed algorithm were quantitatively compared to those of the reference instrument as well as the theoretical particle size distribution under ideal conditions. The measurement results of the neural-network-based algorithm were compared to the reference data measured by the scanning mobility particle sizer (SMPS) spectrometer as well as the Tikhonov regularization. The results showed that the neural-network-based inversion algorithm could be successfully applied to the microfluidic electrical cascade impactor to obtain higher measurement accuracy than the Tikhonov regularization. The predicted value of the neural-network-algorithm showed an error of less than 12% for all parameters in all cases compared to the SMPS data. These results indicate that the application of a neural-network-based algorithm is useful for improving the accuracy of the electrical particle sensors to enhance the temporal-spatial resolution of the particle size distribution.
|keyword||particle size distribution , neural network , aerosol , submicron particles|