神經(jīng)模糊定量分類器分析厚膜氧化錫氣體傳感器陣列對酒精以及含酒精飲料的響應(yīng)-外文翻譯（英文）

1、IEEE SENSORS JOURNAL, VOL. 10, NO. 9, SEPTEMBER 2010 1461A Neuro-Fuzzy Classifier-Cum-Quantifier for Analysis of Alcohols and Alcoholic Beverages Using Responses of Thick-Film Tin Oxide Gas Sensor ArrayRavi Kumar, R. R.

2、Das, V. N. Mishra, and R. DwivediAbstract—A novel neuro-fuzzy classifier-cum-quantifier is presented. The proposed classifier retrieves both qualitative and quantitative information simultaneously from the steady-state r

3、esponses of thick-film tin oxide gas sensor array when it was exposed to seven different kinds of alcohols and alcoholic bever- ages. The individual concentration bands were represented in the output feature space by fuz

4、zy subsethood measure. The qualitative and quantitative classifications were done by training an artificial neural network (ANN) with backpropagation algorithm. Each output neuron of the network represented one out of th

5、e seven alcohols and alcoholic beverage classes and was trained to fire at the fuzzy subsethood value of the particular concentration band of a particular alcohol or alcoholic beverage whose sample was presented to the n

6、etwork. The proposed network gave satisfactory performance and simultaneous qualitative and quantitative classi- fication of the alcohols and alcoholic beverages was obtained using a single neural network.Index Terms—Alg

7、orithm, electronic nose, fuzzy subsethood, in- telligent gas sensor, neural networks.I. INTRODUCTION SMELL processing in humans is a very complex task which involves processing of many different categories of infor- mati

8、on which are both qualitative and quantitative in nature. Electronic nose (e-nose) technology strives to mimic the human system of smell processing and intelligent gas sensors (IGSs) are the foundation pillars of e-nose

9、technology. Pattern Recog- nition (PR) techniques play a key role in IGS technology by improving the selectivity of poorly selective sensors. Artificial neural networks have emerged as one of the most sought after PR tec

10、hnique owing to their massively parallel nature and the ability to mimic human response to pattern recognition. Varying degrees of success have been achieved by applying ANN tech- niques for gas/odor discrimination [1]–[

11、5]. In the course of de- velopment, there has been a gradual shift of the interest of re- searchers from crisp algorithms to those which employ fuzzy logic to generate class information, since, unambiguous classi- ficati

12、on criteria do not hold well in real-world problems.Manuscript received February 23, 2010; accepted February 24, 2010. Date of publication June 07, 2010; date of current version July 14, 2010. The associate editor coordi

13、nating the review of this paper and approving it for publication was Prof. Evgeny Katz. The authors are with the Department of Electronics Engineering, Institute of Technology, Banaras Hindu University, Varanasi, Uttar P

14、radesh, India (e-mail: roybhu_royravs@yahoo.co.in; roybhu83@gmail.com; rrdas.ece@itbhu.ac.in; vnmishra.ece@itbhu.ac.in; rdwivedi.ece@itbhu.ac.in). Color versions of one or more of the figures in this paper are available

15、online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSEN.2010.2045369Both fuzzy logic and ANN being biologically inspired, their amalgamation has been attempted by researchers in the form of neuro-fuz

16、zy systems [6], where neural networks have been used to implement fuzzy logic. Neuro-fuzzy systems have also been used successfully for function-approximation tasks [7], and also to tune membership functions in fuzzy sys

17、tems [8]. In this paper, we report a neural network trained on an advanced fuzzy mea- sure viz. fuzzy subsethood applied to retrieve the qualitative and quantitative information from the steady-state responses of thick-f

18、ilm tin oxide gas sensor array fabricated at our laboratory and exposed to seven different types of alcohols and alcoholic beverages. The proposed network functions both as a classifier and a quantifier. The quantitative

19、 information is represented in the form of fuzzy subsethood values in the output feature space [8]. The network when presented with an input vector results in the firing of the neuron corresponding to that particular cla

20、ss of odor to which the input vector belongs. The rest of the neurons remain deactivated and thus the class information is achieved. Our approach differs from earlier reported techniques in many ways. We have used an adv

21、anced fuzzy measure instead of com- monly used fuzzy membership functions to represent fuzzy in- formation in the output feature space. Furthermore, the fuzzy in- formation has been incorporated in the training target se

22、t, while the neural network only tunes the weight vector, thus making the computation simpler and eliminating the need for linguistic fuzzy rule formulation as is done in conventional fuzzy sys- tems. The proposed networ

23、k overcomes the limits posed by sat- urating tendency of the sensor response at higher concentrations and performs fairly good quantitative classification along with near-perfect qualitative classification of seven diffe

24、rent alcohols and alcoholic beverages. The hardware implementation for this approach is expected to be simpler than that required by the con- ventional fuzzy computations. The present paper is organized into following si

25、x sections. Section II describes experimental set up and Sections III and IV present the theories of fuzzy-sub- sethood and backpropagation algorithm, respectively. Problem formulation is described in Section V and Secti

26、ons VI and VII are on discussion of results and conclusion, respectively.II. EXPERIMENTAL [9]In the present study, already reported responses of four sen- sors of an integrated gas sensor array fabricated at our labora-

27、tory have been used. Integrated gas sensor array comprises of four tin oxide thick-film sensors. Three different dopants viz. ZnO, , and NiO were used, respectively, with tin oxide to result in three different types of g

28、as sensors, whereas the fourth was tin oxide sensor without any doping. Sensors were1530-437X/$26.00 © 2010 IEEEKUMAR et al.: A NEURO-FUZZY CLASSIFIER-CUM-QUANTIFIER FOR ANALYSIS OF ALCOHOLS AND ALCOHOLIC BEVERAGES

29、1463Fig. 3. (e) Steady-state response of the sensor array upon exposure to Rum-1. (f) Steady-state response of the sensor array upon exposure to Rum-2. (g) Steady-state response of the sensor array upon exposure to Ethan

30、ol.and doped with , NiO, and ZnO sensors with concentrations of four different types of whiskies two different types of rums and ethanol. The concentration of vapors is indicated in number of drops of alcoholic beverages

31、 in liquid phase, which after being injected into the test chamber get vaporized due to integrated heater associated with sensor array mounted inside the test chamber [9]. In the text in subsequent sections, different va

32、pors of alcohols and alcoholic beverages will be mentioned as “gas.”III. FUZZY SUBSETHOOD [10]Fuzzy set theory is nothing but a generalization of the con- ventional crisp set theory. It measures the degree to which an ev

33、ent occurs [11]. Each element of a fuzzy set has a degree of membership assigned to it in accordance with a member-ship function. The most commonly used membership functions in the literature being triangular and trapezo

34、idal membership functions. Let be a nonfuzzy set. The subsets of are called bit vectors or bivalent messages. If, then , , and the subset is represented as . The 1s and 0s indicate the presence or absence of the th eleme

35、ntin the subset. Each non fuzzy subset can be defined as one of the two-valued membership functions : . The power set of is the set of all of ’s subsets. There arepossible messages defined on (in ). In the example, there

36、 are possible messages. In contrast, fuzzy subsets ofare referred to as fit vectors or fit messages. Each subset ofcan be defined as one of the continuum-many continuous-valued membership functions : . Fuzzy sets can als

37、o be represented geometrically and this representation gives us more insight into the intricacies of fuzzy sets and operations related to them [10]. According to this representation, the fuzzy power set, which is the set

38、 of all fuzzy subsets of , is visualized as a unit hypercube and a fuzzy set is any point in the cube . Vertices of the cube define nonfuzzy or crisp sets which are a subset of . Thus, crisp sets are nothing but special

39、cases of the fuzzy sets. Fig. 1(a) depicts the geometrical representation of fuzzy sets. The sets on vertices are nonfuzzy sets and long diagonals connect nonfuzzy set complements. A fuzzy set with fit values is represen

40、ted inside the unit square consisting of all possible fuzzy subsets of two elements. The midpoint of the unit square shown in Fig. 4(a) is the point of maximum fuzziness. Thus, the proposition for fuzziness can be summar

41、ized as follows:(1)(2)The terms and are termed as overlap and un- derlap, respectively. The positions of a fuzzy set along with its complement, overlap set and underlap set are shown in Fig. 4(b). With the increasing fuz

42、ziness of , all the four points shrink towards the midpoint of the fuzzy square, which is the point of maximum fuzziness. The size or cardinality of a fuzzy set is given by , as follows:(3)where is the membership value o

43、f the th element of-valued fuzzy set . Fuzzy subsethood measures the degree of belongingness of a fuzzy set to its superset and is denoted by(4)A fuzzy set can be a subset of another fuzzy set iffThe fuzzy-subsethood the