Advertisement

Aerosol Science and Engineering

, Volume 2, Issue 4, pp 197–205 | Cite as

Predicting Fibrous Filter’s Efficiency by Two Methods: Artificial Neural Network (ANN) and Integration of Genetic Algorithm and Artificial Neural Network (GAINN)

  • Pooya Abdolghader
  • Fariborz Haghighat
  • Ali Bahloul
Original Paper

Abstract

In this study, we used both methods of ANN and GAINN for predicting the fibrous filter’s efficiency. In this regard, we collected the experimental penetration data for particles in the range of 10.7–191.1 nm. Experimental data were collected with different constant flow rates and from one type of N95 filtering facepiece respirator. A satisfactory number of data from experimental setup were exploited to build up a database. These methods are according to the back-propagation algorithm to map two components, namely, particle diameter and constant air flow rates into the corresponding penetration. The developed ANN and GAINN methods were capable of predicting precise values of penetration from experimental data. Also by comparing the results of these two methods, it is understandable that ANN method can predict the penetration data from examples of the experimental setup more efficiently than GAINN within an acceptable computational time.

Keywords

Nanoparticles Filtration Artificial neural networks Genetic algorithm HVAC filters 

1 Introduction

Air pollution has become a major problem in recent years. This matter is at highest importance in protecting workers against toxic inhaled nanoparticles. Besides, removing ultrafine particles (UFPs) in residential and light-commercial central HVAC systems highlights the significance of studies in this field. These concerns necessitate evaluating the efficiency of fibrous filters, used in ventilation systems (Stephens and Siegel 2013) or in filtering facepieces (Brochot et al. 2018; Mahdavi et al. 2015; Mostofi et al. 2011, 2012). As reported in several reviews, the filtration performance is evaluated based on two parameters: the pressure drop and filtration efficiency (Abdolghader et al. 2018; Givehchi and Tan 2014; Lee and Liu 1982; Mostofi et al. 2010; Wang and Otani 2013).

As an aerosol flow approaches the surface of a fiber, particle deposit on it is caused by several mechanisms such as gravitational settling, inertial impaction, interception, diffusion and electrostatic forces.

The first four are recognized as mechanical mechanisms. The gravitational settling is less significant for most particle sizes. Gravity sedimentation can be completely overlooked if size of the particle is smaller than 0.5 µm (Zhu et al. 2017). Figure 1 illustrates the particles size spectrum for each capture mechanism.
Fig. 1

Particle size spectrum for each capture mechanism in aerosol filtration (Zhu et al. 2017)

The filtration efficiency (η) is considered as a function of total single-fiber efficiencies \((E_{\sum } )\) and is given as (Hinds 1999):
$$\eta = 1 - \exp \left[ {\frac{{ - 4\alpha E_{\sum } L}}{{\pi d_{\text{f}} (1 - \alpha )}}} \right] ,$$
(1)
where α is considered as the packing density of the medium, L is the thickness of the medium and \(d_{\text{f}}\) is the fiber diameter. The equation was obtained theoretically by considering the interrelationship between total single-fiber efficiency and total efficiency of a medium. Single-fiber efficiency is the ratio of the particles settled down on a unit length of fiber surface. The total single-fiber efficiency is an approximation according to the assumption that all separate mechanisms are independent and all of them are smaller than unity (Brown 1993; Kasper et al. 1978; Lee and Liu 1982; Payet et al. 1992). In this regard, it will be assumed that total single-fiber efficiency \((E_{\sum } )\) is presented by the sum of the efficiencies for each of the individual mechanisms.
$$E_{\sum } \cong E_{\text{D}} + E_{\text{I}} + E_{\text{DR}} + E_{\text{R}} + E_{\text{P}} + E_{\text{C}} + E_{\text{IM}} ,$$
(2)
where \(E_{\text{D }}\) and \(E_{\text{R}}\) are due to diffusion and interception (Lee and Liu 1982); \(E_{\text{DR}}\) is due to interception of the particles undergoing diffusion; \(E_{\text{I}}\) is due to impaction (Hinds 1999); \(E_{\text{P}}\) is due to polarization force; \(E_{\text{C}}\) is due to Columbic force (Lathrache and Fissan 1987; Tennal et al. 1991); and \(E_{\text{IM}}\) is due to image force (Kanaoka et al. 1987).

Thus, it is obvious that theoretical studies are necessary to understand the performance of fibrous filters (Abdolghader et al. 2018; Givehchi and Tan 2014; Lee and Liu 1982). Also, it is important to carry out experiments to validate those theoretical studies. Meanwhile, predicting the experimental data using artificial neural networks (ANNs) is a very efficient approximation method. However, there are few studies to demonstrate that ANNs can be applied to predict the filtration efficiency of fibrous filters. Therefore, filling this gap is the prime motivation for the present study.

The objective of this work is to investigate the possibility of using two methods, artificial neural networks (ANNs), and the integration of genetic algorithm (GA) and neural network (GAINN) to predict the efficiency of the fibrous filters directly. At the next level, by comparing the results from the above-mentioned methods, we will indicate which method can be exploited to better estimate the penetration data from the examples of the experimental setup within an acceptable computational time.

2 Artificial Neural Network (ANN)

Artificial neural network is parallel information processing method which is capable of modeling complicated and nonlinear equations using a variety of input–output training data, obtained from accessible experiments or numerical results (Awad and Herzallah 2015; Kaushik et al. 2015; Su et al. 2014).

The most common neural network structure is the feed-forward model. If the signals or information are propagated in only one direction from input to output, the system is called to be feed-forward which has been broadly and favorably used in numerous engineering studies including building applications (Yang et al. 2005; Pala et al. 2008). A feed-forward multilayer neural network with back propagation can estimate any nonlinear continuous function with any desirable accuracy (Brown and Harris 1994; García-Pedrajas et al. 2003). In this study, this approved method of feed-forward has been exploited. Figure 2 illustrates a feed-forward double-layer neural network with back propagation.
Fig. 2

A feed-forward double-layer neural network with back propagation (Filletti et al. 2015)

\(x_{1}\) and \(x_{2}\) are input data which are particle diameter and constant air flow rates, respectively, and y is output value which is penetration.

Weights in neural networks are the most significant components for mapping input to the output. This is analogous to slope in linear regression, where a weight is multiplied to the input to add up to form the output. Figure 3 demonstrates the relationship between input and output in a typical neural network model.
Fig. 3

The relationship between input and output in a typical neural network model (Dongare et al. 2012)

If the inputs are \(x_{1}\), \(x_{2}\), … and \(x_{m}\), then the weights to be applied to them are denoted as \(w_{1}\), \(w_{2}\), … and \(w_{m}\). As a result, the y which is called output is given as:
$$y = f(Y - \theta ),$$
(3)
$${\text{where}}\quad Y = \mathop \sum \limits_{i = 1}^{m} w_{i} x_{i} .$$
(4)
f is a transfer function which generates scalar neuron output in accordance with weight, bias and neurons input. Among different types of transfer functions, the ‘S’-shaped log-sigmoid is commonly used in feed-forward multilayer neural networks with back propagation which is given as (Madaeni et al. 2010):
$$f(x) = \frac{1}{{1 + {\text{e}}^{ - x} }},$$
(5)
and \((w_{0} = - \;\theta )\) is the weight corresponding to the bias. A bias is an “additional” neuron added to each pre-output layer that has the value of 1. Bias is just appended to the start/end of the input and each hidden layer, and is not influenced by the values in the previous layer, which means that these neurons do not have any incoming connection. In other words, bias is like the intercept added in a linear equation. It is an extra component which is used to adjust the output along with the weighted sum of the inputs to the neuron (Krenker et al. 2011).
The network is trained by optimizing the weights for each node interconnection and bias terms until a satisfactory agreement is obtained among the output layer neurons and actual outputs. The mean squared error of the network (MSE) is given as:
$${\text{MSE}} = \frac{1}{2}\mathop \sum \limits_{k = 1}^{G} \mathop \sum \limits_{j = 1}^{m} [y_{{\left( {\text{ANN}} \right)j}} (k) - y_{{\left( {\text{Experiment}} \right)j}} (k)]^{2} .$$
(6)

In Eq. (6), m represents the number of output nodes, G denotes the number of training samples, while \(y_{{({\text{ANN}})j}} (k)\) is the expected output, and \(y_{{({\text{Experiment}})j}} (k)\) stands for the actual output (Parsian et al. 2017).

The network is trained by optimizing the weights for each node interconnection and bias terms until a satisfactory agreement is obtained among the output layer neurons and actual outputs.

The data are divided into two groups, a training data and a validating one. The validating data are used to estimate the accuracy of the model, whereas the training data are exploited to find optimal set of weights. In a case that the number of weights is bigger than the number of existing data, the error in-fitting the non-trained data reduces at first, but then increases as the network gets over-trained. In contrast, in a case that the number of weights is lower than the number of existing data, the over-fitting problem is not necessary.

3 Genetic Algorithm (GA)

Genetic algorithm is an optimization method which was established by Holland (1975) is based on Darwin’s theory of evolution. The GA has been applied in a wide range of studies from medicine (Lahanas et al. 2003) to engineering specially in building application (Huang and Lam 1997; Lu et al. 2005). Since the GA is a gradient-free method, it can deal with nonlinear functions to find global optimal solution without being restricted in local one (Magnier and Haghighat 2010). Additionally, the GA is able to provide a detailed optimization as well as finding optimal solution in an acceptable computational time (Sakamoto et al. 1999; Wetter and Wright 2004). A basic procedure for the GA approach is described in Fig. 4.
Fig. 4

A basic procedure for GA approach (Yengui et al. 2012)

3.1 GAINN

Integration of GA and artificial neural network (GAINN) is an interesting approach which is the combination of the GA and the neural network to speed up optimization time while exploiting the GA. The fundamental concept of GAINN is to take advantage of the swiftness of calculation provided by ANN as well as utilizing the optimization capability of the GA. A complete procedure for GAINN is demonstrated in Fig. 5 (Magnier and Haghighat 2010).
Fig. 5

A complete procedure for GAINN approach (Magnier and Haghighat 2010)

The simulation time and accuracy are the most important analytical points that should be mentioned in this study.

3.2 Simulation Time

Certainly, the runtime of the GAINN method is lengthier than ANN. The reason for this can relate to the structural characterization of these two methods. The neural network is a method of training based on gradient equations and derivatives and it knows the procedure to solve the problems. The GA has not been created to solve the problems, but it has been created to find the optimal answer. It obvious that the GA is a random algorithm which is searching in space; so, finding the answer by this method takes more time in comparison with the method of neural network.

3.3 Accuracy

According to the definition of the two methods in the previous section, the ANN will have a higher accuracy because the process of problem solving is done based on legitimate mathematical relationships. GA is a random searcher which is trying to find the answer by looking in the space of the states. The other thing about the GA is that this algorithm is the simplest algorithm in the evolutionary algorithm set which is no longer used in important and complex issues.

4 Experimental Procedure

In this study, the experimental data were obtained from one type of N95 filtering facepiece respirators. Figure 6 demonstrates the experimental setup used for measuring the penetration of N95 filtering facepiece respirators (FFRs) in terms of particle sizes between 10.7–191.1 nm and ten constant air flow rates. The selected constant airflow rates were: 42, 68, 85, 115, 135, 170, 210, 230, 270, 360 L/min. The experimental setup constituted of four main parts: an experimental chamber, a manikin head, a particle generation system and measurement devices. A six-jet collision nebulizer containing 0.1% v/v NaCl solution, was employed to generate 10–205.4 nm NaCl particles. In the next step, the flow was passed through the drying system (silica gel packs) to control the humidity of the chamber. A Kr-85 electrostatic neutralizer was used to globally neutralize charges carried by the generated particles to reach equilibrium charged particles. A set of measurement devices including an “electrostatic classifier (EC)” (Model 3080, TSI Inc.) containing a long DMA (differential mobility analyzer), and a “condensation particle counter “(CPC)” (Model 3775, TSI Inc.) were used to measure upstream and downstream concentrations distribution.
Fig. 6

Experimental setup configuration to measure the penetration of N95 filtering facepiece respirators (FFRs) with constant flows (Bahloul et al. 2014)

Finally, particle penetration percentage is defined as the ratio of downstream to upstream concentration which is given as:
$$P(\% ) = \left( {\frac{{C_{\text{d}} }}{{C_{\text{u}} }}} \right) \times 100,$$
(7)
C reflects the concentration, while the subscript d and u refer to downstream and upstream, respectively. Whereas efficiency of the filter is given (Bahloul et al. 2014):
$${\text{Efficiency }} = 1- P\;(\% ).$$
(8)

5 Results and Discussion

In this study the selected constant airflow rates were: 42, 68, 85, 115, 135, 170, 210, 230, 270, 360 L/min and also particles were in the range of 10.7–191.1 nm. The penetration range depends on the particle diameter and air flow rates. The output results of both ANN and GAINN are related to the implementation of a double-layer neural network (each layer of five neurons) and 200 populations for the GA. The GA is based on production of 2000 generations and it has stopped at 8000th repetition.

The results of the training of both ANN and GAINN demonstrated that they can reproduce the input/output relation of the training set data. Figures 7, 8, 9, 10, 11, 12, 13, 14, 15 and 16 present the estimated data for penetration calculated by both ANN and GAINN after the training period. The average MSE of all Figs. 7, 8, 9, 10, 11, 12, 13, 14, 15 and 16 for both ANN and GAINN are 0.23 and 1.19, respectively. In addition, the average R-squared value of all Figs. 7, 8, 9, 10, 11, 12, 13, 14, 15 and 16 for both ANN and GAINN are 9.8 and 0.9, respectively.
Fig. 7

Regression between estimated and experimental penetration data for the constant air flow rate of 42 L/min

Fig. 8

Regression between estimated and experimental penetration data for the constant air flow rate of 68 L/min

Fig. 9

Regression between estimated and experimental penetration data for the constant air flow rate of 85 L/min

Fig. 10

Regression between estimated and experimental penetration data for the constant air flow rate of 115 L/min

Fig. 11

Regression between estimated and experimental penetration data for the constant air flow rate of 135 L/min

Fig. 12

Regression between estimated and experimental penetration data for the constant air flow rate of 170 L/min

Fig. 13

Regression between estimated and experimental penetration data for the constant air flow rate of 210 L/min

Fig. 14

Regression between estimated and experimental penetration data for the constant air flow rate of 230 L/min

Fig. 15

Regression between estimated and experimental penetration data for the constant air flow rate of 270 L/min

Fig. 16

Regression between estimated and experimental penetration data for the constant air flow rate of 310 L/min

These results indicate that ANN method can predict the penetration data from examples of the experimental setup more efficiently than GAINN.

6 Conclusion and Future Studies

ANN and GAINN methods were trained with data derived from experimental setup to determine the fibrous filters efficiency. To create a database, a sufficient number of data were performed, from which the ANN parameters (weights and biases) were adjusted. Both methods were developed to map the two components, which are particle diameter and constant air flow rates into the corresponding penetration. The developed ANN and GAINN methods were capable of predicting correct values for penetration from experimental data which have not been used in the training. By comparing the results of two methods, it is obvious that ANN can predict the penetration data from examples of the experimental setup more efficiently than GAINN within an acceptable computational time.

This study is limited to predict the efficiency of filtering facepieces in the range of particles and airflow rates that were used in experimental setup under certain condition of the experiment, so in this study current prediction model with ANN can not be used to evaluate the efficiency of other filters under different experimental conditions because it has been trained and validated for certain experimental parameters.

From this study, it can be inferred that, the developed neural network method is an effective tool for prediction of fibrous filters efficiency. In this study, the N95 filtering facepiece respirator was chosen as a fibrous filter media and good results were obtained as a result, ANN method can be extrapolated to all general filters to evaluate their removal efficiencies. Future studies are still required to predict the filtration efficiency of fibrous filters especially in HVAC filters which are the dominant systems for removing ultrafine particles in buildings. Thus, exploiting the ANN can result in a good predication to evaluate the filtration performance in HVAC filters. Also, further studies are necessary to investigate the possibility of integrating neural network with other optimization algorithms; such as particle swarm optimization (PSO) to reach more precise prediction.

Notes

Acknowledgements

The authors would like to express their gratitude to the Concordia University for funding this work, and Ms. Farinaz Haghighat for her valuable comments and suggestions.

References

  1. Abdolghader P, Brochot C, Haghighat F, Bahloul A (2018) Airborne nanoparticles filtration performance of fibrous media: a review. Sci Technol Built Environ.  https://doi.org/10.1080/23744731.2018.1452454 CrossRefGoogle Scholar
  2. Awad WH, Herzallah R (2015) Driving licensing renewal policy using neural network-based probabilistic decision support system. Int J Comput Appl Technol 51(3):155–163CrossRefGoogle Scholar
  3. Bahloul A, Mahdavi A, Haghighat F, Ostiguy C (2014) Evaluation of N95 filtering facepiece respirator efficiency with cyclic and constant flows. J Occup Environ Hyg 11(8):499–508.  https://doi.org/10.1080/15459624.2013.877590 CrossRefGoogle Scholar
  4. Brochot C, Abdolghader P, Haghighat F, Bahloul A (2018) Filtration of nanoparticles applied in general ventilation. Sci Technol Built Environ.  https://doi.org/10.1080/23744731.2018.1500396 CrossRefGoogle Scholar
  5. Brown RC (1993) Air filtration: an integrated approach to the theory and applications of fibrous filters. Pergamon, New YorkGoogle Scholar
  6. Brown M, Harris CJ (1994) Neurofuzzy adaptive modelling and control. Prentice Hall, Englewood CliffsGoogle Scholar
  7. Dongare AD, Kharde RR, Kachare AD (2012) Introduction to artificial neural network. Int J Eng Innov Technol (IJEIT) 2(1):189–194Google Scholar
  8. Filletti ÉR, Da Silva JM, Ferreira VG (2015) Predicting of the fibrous filters efficiency for the removal particles from gas stream by artificial neural network. Adv Chem Eng Sci 5(03):317CrossRefGoogle Scholar
  9. García-Pedrajas N, Hervás-Martínez C, Muñoz-Pérez J (2003) COVNET: a cooperative coevolutionary model for evolving artificial neural networks. IEEE Trans Neural Netw 14(3):575–596CrossRefGoogle Scholar
  10. Givehchi R, Tan Z (2014) An overview of airborne nanoparticle filtration and thermal rebound theory. Aerosol Air Qual Res 14(1):46–63.  https://doi.org/10.4209/aaqr.2013.07.0239 CrossRefGoogle Scholar
  11. Hinds WC (1999) Aerosol technology: properties, behavior, and measurement of airborne particles. Wiley, New York. http://www.loc.gov/catdir/toc/onix02/98023683.html
  12. Holland JH (1975) Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence. University of Michigan Press, Ann ArborGoogle Scholar
  13. Huang W, Lam HN (1997) Using genetic algorithms to optimize controller parameters for HVAC systems. Energy Build 26(3):277–282CrossRefGoogle Scholar
  14. Kanaoka C, Emi H, Otani Y, Iiyama T (1987) Effect of charging state of particles on electret filtration. Aerosol Sci Technol 7(1):1–13.  https://doi.org/10.1080/02786828708959142 CrossRefGoogle Scholar
  15. Kasper G, Preining O, Matteson M (1978) Penetration of a multistage diffusion battery at various temperatures. J Aerosol Sci 9(4):331–338CrossRefGoogle Scholar
  16. Kaushik A, Soni A, Soni R (2015) A hybrid approach for software cost estimation using polynomial neural networks and intuitionistic fuzzy sets. Int J Comput Appl Technol 52(4):292–304CrossRefGoogle Scholar
  17. Krenker A, Bešter J, Kos A (2011) Introduction to the artificial neural networks. In: Suzuki K (ed) Artificial Neural Networks-Methodological Advances and Biomedical Applications. InTech, Rijeka, Carotia, pp 3–18Google Scholar
  18. Lahanas M, Schreibmann E, Milickovic N, Baltas D (2003) Intensity modulated beam radiation therapy dose optimization with multiobjective evolutionary algorithms. In: Fonseca CM, Fleming PJ, Zitzler E, Thiele L, Deb K (eds) Evolutionary multi-criterion optimization: second international conference, EMO 2003, Faro, Portugal, April 8–11, 2003. Proceedings. Springer, Berlin, pp 648–661.  https://doi.org/10.1007/3-540-36970-8_46 CrossRefGoogle Scholar
  19. Lathrache R, Fissan H (1987) Enhancement of particle deposition in filters due to electrostatic effects. Filtr Sep 24(6):418–422Google Scholar
  20. Lee K, Liu B (1982) Theoretical study of aerosol filtration by fibrous filters. Aerosol Sci Technol 1(2):147–161CrossRefGoogle Scholar
  21. Lu L, Cai W, Xie L, Li S, Soh YC (2005) HVAC system optimization—in-building section. Energy Build 37(1):11–22CrossRefGoogle Scholar
  22. Madaeni SS, Hasankiadeh NT, Kurdian AR, Rahimpour A (2010) Modeling and optimization of membrane fabrication using artificial neural network and genetic algorithm. Sep Purif Technol 76(1):33–43CrossRefGoogle Scholar
  23. Magnier L, Haghighat F (2010) Multiobjective optimization of building design using TRNSYS simulations, genetic algorithm, and Artificial Neural Network. Build Environ 45(3):739–746CrossRefGoogle Scholar
  24. Mahdavi A, Haghighat F, Bahloul A, Brochot C, Ostiguy C (2015) Particle loading time and humidity effects on the efficiency of an N95 filtering facepiece respirator model under constant and inhalation cyclic flows. Ann Occup Hyg 59(5):629–640.  https://doi.org/10.1093/annhyg/mev005 CrossRefGoogle Scholar
  25. Mostofi R, Wang B, Haghighat F, Bahloul A, Jaime L (2010) Performance of mechanical filters and respirators for capturing nanoparticles—Limitations and future direction. Ind health 48(3):296–304CrossRefGoogle Scholar
  26. Mostofi R, Bahloul A, Lara J, Wang B, Cloutier Y, Haghighat F (2011) Investigation of potential affecting factors on performance of N95 respirator. J Int Soc Respir Prot 28(1):26–39Google Scholar
  27. Mostofi R, Noël A, Haghighat F, Bahloul A, Lara J, Cloutier Y (2012) Impact of two particle measurement techniques on the determination of N95 class respirator filtration performance against ultrafine particles. J Hazard Mater.  https://doi.org/10.1016/j.jhazmat.2012.02.058 CrossRefGoogle Scholar
  28. Pala M, Caglar N, Elmas M, Cevik A, Saribiyik M (2008) Dynamic soil-structure interaction analysis of buildings by neural networks. Constr Build Mater 22(3):330–342CrossRefGoogle Scholar
  29. Parsian A, Ramezani M, Ghadimi N (2017) A hybrid neural network-gray wolf optimization algorithm for melanoma detection. Biomed Res 28(8):3408–3411Google Scholar
  30. Payet S, Boulaud D, Madelaine G, Renoux A (1992) Penetration and pressure drop of a HEPA filter during loading with submicron liquid particles. J Aerosol Sci 23(7):723–735CrossRefGoogle Scholar
  31. Sakamoto Y, Nagaiwa A, Kobayasi S, Shinozaki T (1999) An optimization method of district heating and cooling plant operation based on genetic algorithm. ASHRAE Trans 105:104Google Scholar
  32. Stephens B, Siegel JA (2013) Ultrafine particle removal by residential heating, ventilating, and air-conditioning filters. Indoor Air 23(6):488–497.  https://doi.org/10.1111/ina.12045 CrossRefGoogle Scholar
  33. Su H, Xie W, Zeng H (2014) Multiple response optimization based on the ANN theory of complex injection moulding process. Int J Comput Appl Technol 50(3–4):186–190CrossRefGoogle Scholar
  34. Tennal K, Mazumder M, Siag A, Reddy R (1991) Effect of loading with AIM oil aerosol on the collection efficiency of an electret filter. Part Sci Technol 9(1–2):19–29CrossRefGoogle Scholar
  35. Wang C, Otani Y (2013) Removal of nanoparticles from gas streams by fibrous filters: a review. Ind Eng Chem Res 52(1):5–17.  https://doi.org/10.1021/ie300574m CrossRefGoogle Scholar
  36. Wetter M, Wright J (2004) A comparison of deterministic and probabilistic optimization algorithms for nonsmooth simulation-based optimization. Build Environ 39(8):989–999CrossRefGoogle Scholar
  37. Yang J, Rivard H, Zmeureanu R (2005). Building energy prediction with adaptive artificial neural networks. In: Proceedings of the 9th international IBPSA conference, Montreal, Quebec, CanadaGoogle Scholar
  38. Yengui F, Labrak L, Frantz F, Daviot R, Abouchi N, O’Connor I (2012) A hybrid GA-SQP algorithm for analog circuits sizing. Circuits Syst 3(02):146CrossRefGoogle Scholar
  39. Zhu M, Han J, Wang F, Shao W, Xiong R, Zhang Q, Zhang F (2017) Electrospun nanofibers membranes for effective air filtration. Macromol Mater Eng 302(1):1600353CrossRefGoogle Scholar

Copyright information

© Institute of Earth Environment, Chinese Academy Sciences 2018

Authors and Affiliations

  1. 1.Department of Building, Civil and Environmental EngineeringConcordia UniversityMontrealCanada
  2. 2.Institut de recherche Robert-Sauvé en santé et en sécurité du travail (IRSST)MontrealCanada

Personalised recommendations