Sensors and Actuators B 140 (2009) 143–148
Contents lists available at ScienceDirect
Sensors and Actuators B: Chemical
journal homepage: www.elsevier.com/locate/snb
Classification of data from electronic nose using relevance vector machines
Xiaodong Wang
a,∗
, Meiying Ye
b
, C.J. Duanmu
a
a
Department of Electronic Engineering, Zhejiang Normal University, Jinhua 321004, PR China
b
Department of Physics, Zhejiang Normal University, Jinhua 321004, PR China
article info
Article history:
Received 3 December 20 08
Received in revised form 11 March 2009
Accepted 21 April 2009
Available online 3 May 2009
Keywords:
Electronic nose
Gas sensors
Relevance vector machines
Support vector machines
Principal component analysis
Classification
abstract
An approach to classify the data from an electronic nose is investigated in this paper. This approach is
based on the method of the relevance vector machines (RVM). The electronic nose data are first converted
into principal components using the principal component analysis (PCA) method and then directly sent
as inputs to a RVM classifier. The performance of the developed approach is validated by cross-validation
procedure. Some experiments have been performed using different combinations of original coffee data,
including the test data from a multi-class classification problem as well as the data from some two-
class classification problems having different kinds of hardness. Experimental results show that the RVM
method is an effective technique for the classification of electronic nose data. Compared with the support
vector machine (SVM) method, the RVM method can provide similar classification accuracy with dra-
matically fewer kernel functions. In addition, another advantage of RVM method is its fewer parameter
settings, in which case only one kernel parameter is needed.
© 2009 Elsevier B.V. All rights reserved.
1. Introduction
An electronic nose, which imitates the perceptional mecha-
nisms of biological olfactory, is defined as an instrument that
consist of an array of chemical sensors (usually gas sensors) and
a pattern recognition system. Electronic noses have been used
to discern the categories of odors mainly in relation to the food
industry, environmental, agricultural, medical areas, and automo-
tive industry. In particular, an electronic nose can be a better
alternative to conventional methods for the continuous real-time
monitoring and controlling of food quality during production
process.
So far, many pattern recognition techniques based on artifi-
cial intelligence have been explored for the analysis of sensors’
array data, e.g. k-nearest neighbor (k-NN), artificial neural network
(ANN), and support vector machine (SVM). As a new class of the
learning algorithm, the SVM can be widely used for classification
and regression. Also, many practical applications in the sensor areas
have been carried out by means of SVM method [1–23]. Some of
them for electronic nose data analysis have been demonstrated, e.g.
in Refs. [2–16]. Despite the fact that SVM classifier provides success-
ful results, a number of significant and practical disadvantages are
identified as follows [24]:
∗
Corresponding author. Tel.: +86 579 82298904; fax: +86 579 82298188.
E-mail address: wxd@zjnu.cn (X. Wang).
•
Although relatively sparse, the number of support vectors (SVs)
typically grows linearly with the size of the training set, and there-
fore, SVM makes unnecessarily liberal use of basis functions.
•
Predictions are not probabilistic, and therefore, SVM is not suit-
able for classification tasks in which posterior probabilities of
class membership are necessary.
•
It is required to estimate the regularizing parameter C, which gen-
erally entails a cross-validation procedure, which can be a waste
of data as well as computation.
•
The kernel function must satisfy Mercer’s condition; hence, it
must be a continuous symmetric kernel of a positive integral
operator.
Some of these problems of SVM can be efficiently alleviated by
relevance vector machine (RVM), a recently developed machine
learning technique, which was originally introduced by Tipping
[25]. As a Bayesian treatment of the sparse learning problem, the
RVM has a comparable generalization performance, yet it also yields
a probabilistic output, as well as circumventing other limitations of
SVM, such as the need for Mercer kernels and the definition of the
regularizing parameter C. Most importantly, the fact that the RVM
requires dramatically fewer kernel functions can lead to significant
reduction in the computational complexity of the decision function,
thereby making the RVM more suitable for real-time applications
[26–29].
Like the SVM, the RVM may be used for regression and clas-
sification. Nowadays, RVM has been applied to some practical
applications and demonstrated that such an approach could
0925-4005/$ – see front matter © 2009 Elsevier B.V. All rights reserved.
doi:10.1016/j.snb.2009.04.030