IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO.

2, APRIL 2022 169

Gas Recognition in E-Nose System: A Review

Hong Chen , Senior Member, IEEE, Dexuan Huo , and Jilin Zhang , Member, IEEE

Abstract—Gas recognition is essential in an electronic nose (E- fine-grained E-nose including sensor arrays, readout interface
nose) system, which is responsible for recognizing multivariate circuits, gas recognition circuits. Compared with the work in
responses obtained by gas sensors in various applications. Over [24], the framework of E-nose in [25] was more effective so that
the past decades, classical gas recognition approaches such as
principal component analysis (PCA) have been widely applied in it has been adopted extensively afterward.
E-nose systems. In recent years, artificial neural network (ANN) Although the better E-nose framework is proposed, it is still
has revolutionized the field of E-nose, especially spiking neural unable to meet the requirement of the market, such as portability
network (SNN). In this paper, we investigate recent gas recognition and real-time detection ability. As pointed out in [24], E-nose has
methods for E-nose, and compare and analyze them in terms of evolved from being large in size, expensive, and power-hungry
algorithms and hardware implementations. We find each classical
gas recognition method has a relatively fixed framework and a few instruments to portable, low cost and power devices with low
parameters, which makes it easy to be designed and perform well power consumption. In order to design such E-nose systems,
with limited gas samples, but weak in multi-gas recognition under researchers usually focus on their three main parts: sensor arrays,
noise. While ANN-based methods obtain better recognition accu- readout interface circuits, and gas recognition circuits.
racy with flexible architectures and lots of parameters. However, The sensor array is the combination of several different gas
some ANNs are too complex to be implemented in portable E-nose
systems, such as deep convolutional neural networks (CNNs). In sensors with a microfabrication process (MEMS). Each sensor
contrast, SNN-based gas recognition methods achieve satisfying in the array behaves like a receptor responding to different
accuracy and recognize more types of gases, and could be im- odors to varying degrees. The interface readout circuits convert
plemented with energy-efficient hardware, which makes them a sensor signals into electrical signals [26]. Generally, the sensor
promising candidate in multi-gas identification.
array with MEMS process and the readout interface circuit with
Index Terms—Gas recognition, electronic nose, artificial neural complementary-metal-oxide semiconductor (CMOS) process
network, spiking neural network, energy-efficient hardware. can be integrated into an E-nose chip [27], which is discussed
in [28]–[30]. After processing signals transmitted by interface
readout circuits, the gas recognition circuits output final recog-
I. INTRODUCTION nition results with various gas recognition algorithms.
INCE Persaud, K et al. [1] first proposed an electronic In the 2000s, researchers focused on the classical gas recog-
S nose (E-nose) based on biological olfaction in 1982, the
E-nose system [2] has been developed for nearly 40 years. As
nition algorithms, like K-Nearest neighbor (KNN), support
vector machine (SVM), decision tree (DT), naive Bayesian
reported in [3]–[5], E-nose has been remarkably successful in model (NBM), principal component analysis (PCA), and linear
many applications including public health [6]–[8], food safety discriminant analysis (LDA). Each classical gas recognition
[9]–[15], environment monitoring [16], [17], agriculture [18], method has a relatively fixed framework and a few parameters,
industry [19], air quality testing [20]–[22], military [23] and so which is easy-design and widely used, but it is difficult for them
on. to recognize multiple odors accurately under noise. Moreover,
A distinct historical landmark of E-nose was the work pre- they are often implemented by general microprocessors (such as
sented by Gardner, J.W et al. [24] in 1994, which was known MPU [30]–[33]), which consume too much power to be suitable
for modularizing the E-nose system into two models: sensors for a portable E-nose system.
and signal processing circuits. Seven years later, Tang K. T With the development of artificial intelligence in recent years,
and Goodman R. M [25] extended the study and proposed a gas recognition circuits based on artificial neural networks
(ANNs) is proved more effective [34]–[36] with flexible archi-
tectures. However, a large amount of gas sensor data is required
Manuscript received December 7, 2021; revised February 24, 2022 and April
7, 2022; accepted April 7, 2022. Date of publication April 12, 2022; date when training an ANN. Besides, some ANNs are too complex
of current version May 23, 2022. This work was supported in part by the to be implemented in the portable E-nose system. Different
National Science and Technology Major Project from Minister of Science and from the classical gas recognition algorithm, ANN-based gas
Technology, China, under Grant 2018AAA0103100, in part by the National
Natural Science Foundation of China under Grants 92164110 and U19B2041, recognition algorithm is often designed with an application-
in part by the Beijing Engineering Research Center under Grant BG0149. This specific integrated circuit (ASIC) [37] or Field Programmable
paper was recommended by Associate Editor K. Tang. (Corresponding author: Gate Array (FPGA) ([38]–[41]).
Hong Chen.)
The authors are with the School of Integrated Circuits, Tsinghua Spiking neural network (SNN) is proposed as a biologically
University, Beijing 100084, China (e-mail: [email protected]; realistic algorithm with lower computing cost and is arguably
[email protected]; [email protected]). the only viable option at the neuronal description level [42]. To-
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TBCAS.2022.3166530. wards linking biologically plausible learning methods and better
Digital Object Identifier 10.1109/TBCAS.2022.3166530 recognition capability, a number of SNN-based gas recognition
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
170 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

algorithms have been developed [43]–[51]. In addition, SNNs whether it is less dense than air, then formaldehyde by toxicity,
have the power-saving ability in hardware implementation like and the remaining one is carbon dioxide. As a result, a decision
[52], [53]. Therefore, with high accuracy and good hardware tree model is built for gas recognition. The work presented by
friendliness, SNN-based gas recognition algorithms have been a Schroeder et al. [62] in 2019 used a decision tree to distinguish
promising solution recently. Furthermore, they can be deployed cheese, wine, and edible oil samples with an accuracy rate of
in neuromorphic chips (such as TrueNorth [54], Loihi [55]) in 91%.
recent works. 3) SVM: Support vector machine (SVM) is a supervised-
In this paper, we review recent studies on gas recognition al- learning algorithm based on statistical learning theory. SVM
gorithms and their hardware circuits of E-nose systems with the is a binary classification model to find the largest interval in the
focus on: (1) Classical gas recognition methods; (2) Gas recogni- data feature space [63]. In 2017, Qiu, S et al. [64] realized the
tion approaches using artificial neural networks; (3) SNN-based gas recognition of food additives through an SVM classifier. In
gas recognition methods; (4) Comprehensive comparison of addition, some modified SVMs, such as the least-squares support
the performance of classical, ANN-based, and SNN-based gas vector machine (LSSVM) [65] were applied in the determination
recognition approaches. We expect that this paper can offer of gas mixture components [66]. Furthermore, researchers often
researchers an overview of recent reports in gas recognition, adopted two or more algorithms to obtain better accuracy, such
and help designers develop efficient E-nose chips, and finally as PCA+SVM [67], CNN+SVM [68]. The study presented
foster a cross-fertilization in future experimental and theoretical in 2018 used the PCA+SVM method [67] to distinguish the
work. ripeness of bananas with the highest accuracy of 97.14%.
4) KNN: K-nearest neighbor (KNN) is a classic instance-
II. CLASSICAL GAS RECOGNITION METHODS based and supervised learning algorithm. It is a nonparametric
statistical method widely used in classification and regression
In the past few decades, many classical gas recognition meth-
[70]. W. Harsono, R. et al. [71] used KNN algorithm to recognize
ods have been competent in E-nose [56]. In the section, we
the original Arabica civet coffee with 97.77% accuracy. The
describe these algorithms and their hardware implementations
combination of LDA/PCA+KNN was proposed by C. Wu et al.
and give a comprehensive comparison of them.
[69] to recognize drug cigarettes with an accuracy rate of 95.92%
in 2019. Besides, Mirzaee-Ghaleh, E et al. [72] developed a
A. Classical Gas Recognition Algorithms fuzzy K-nearest neighbors (F-KNN) algorithm to identify fresh-
The important principle of classical gas recognition algo- chilled and frozen-thawed chicken meat with 95.83% accuracy.
rithms is the feature discovery hierarchy where increasingly 5) NBM: Naive Bayesian model is a probabilistic model with
discriminative, abstract, and invariant features are acquired in the topology of a directed acyclic graph (DAG). It is suitable for
the process of features extraction for gas data. Generally, five expressing and analyzing uncertain and probabilistic events [73].
algorithms are widely used in gas recognition tasks as follows. Some works concerned for gas recognition have been reported.
1) PCA & LDA: Principal component analysis (PCA) is For instance, the research presented by D. R. Wijaya et al. [74]
an unsupervised algorithm for reducing the dimensionality of obtained 75% accuracy in fresh beef and pork odor detections,
such datasets, increasing interpretability but at the same time and the work by B. Grodniyomchai et al. [75] reached 100%
minimizing information loss [57]. While linear discriminant accuracy in alcohol products.
analysis (LDA) is a supervised-learning dimensionality reduc- Besides gas recognition, gas concentration estimation is an-
tion technique. The main principles of PCA and LDA are to other key task in E-nose systems. For example, in chemical,
project high-dimensional data to low-dimensional space. After metallurgy, and other industries, the safety level can be evaluated
projection, it is expected that the projection points of similar data by the concentration of flammable, explosive, and toxic gases.
categories are as close as possible, and the distance between However, due to sensor-drifting [76], sensor-poisoning, and
the category centers of different data categories is as large as challenges from the external environment, much uncertainty
possible [58]. These two methods have been successfully used exists in gas sensors.
in gas recognition to reduce the data dimensionality, and then the Hence, the algorithms for gas concentration estimation are
dimensionality-reduced data is classified through other methods strongly related to the sensor response [27]. If the sensor re-
[59]. sponse is proportional to the gas concentration, the gas con-
2) DT: Decision tree (DT) is a tree structure model with centration can be estimated by least squares, linear regression,
data attributes. It is often adopted to solve classification and etc. For instance, researchers adopt the weighted constrained
regression problems [60]. The most important feature of DT least squares (WCLS) method based on the features of sensor
is its capability to break down a complex decision-making data to estimate the component concentration [76]. On the other
process into a collection of simpler decisions, thus providing hand, if the sensor response is proportional to the square of the
a solution that is often easier to interpret [61]. In theory, we gas concentration, polynomial regression can be used, such as
can use the properties (such as toxicity, flammability, density, C = ax2 + b [50], in which gas concentration estimation can be
etc.) of different gases to decompose the complex multi-gas regarded as a regression problem. Unlike classification, regres-
identification task into several simple tasks. For example, for the sion is used to predict continuous real values. Therefore, root-
three-gas identification task of carbon dioxide, formaldehyde, mean-squared error (RMSE) and mean absolute error (MAE)
and hydrogen, hydrogen can be filtered out on the basis of are often adopted to analyze the error between the predicted
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 171

value and the true value in the E-nose system to evaluate the
performance of gas concentration estimation [50], [77].

B. Hardware Implementation
To our knowledge, the first hardware implementation of
the E-nose system was introduced in 1994 by Gardner,
J.W et al.[24], which consisted of a simple sensor array, a signal
preprocessing circuit, and a gas recognition circuit with the back-
propagation neural network. Five years later, another E-nose
prototype system was proposed by B. C. Muñoz et al.[78]. Like
Fig. 1. A two-chip solution for E-nose system, comprising a front-end sensor
[24], this system was composed of a 32-channel sensor array, array and a back-end low-power signal-processing chip in 2014 [82].
an acquisition card with a 22-bit ADC and microprocessor, and
a computer as a gas recognition circuit. Unfortunately, with the
relatively backward chip manufacturing process before 2000,
the E-nose was a large-size instrument.
An electronic olfaction system on a chip was introduced in
2001 by Tang, K.-T et al.[25]. The system integrated interface
readout circuits and gas recognition circuits with Euclidean
distances. It could distinguish eight odors with recording sensor
data features. However, without an effective data bus between
different circuits and lack of advanced energy-efficient design,
the chip had low response speed and much power dissipation.
Besides gas recognition circuits, some works focused on gas
concentration estimation. A micro chemical sensor system chip
along with three sensors was designed to estimate two different
concentrations of ethanol and toluene by C.Hagleitner et al. [79]
in 2002. This chip was realized with digital and analog hybrid
circuits. Unfortunately, no gas recognition circuits were realized
in this work. Another gas sensor system chip was put forward in
2003 by Covington, J.A et al. [80]. Compared with the previous Fig. 2. Target application scenario and top-level system view of the nose-on-
work in [79], this work fabricated a seventy-sensor array and an a-chip in 2014 [32].
integrated drive, gain and baseline removal circuitry. It reached
0.00012%/ppm and 0.00644%/ppm sensitivity to ethanol and
toluene respectively. this work was a two-chip E-nose system, comprising eight
Before 2006, no fully integrated E-nose chips including on-chip sensors and a signal-processing chip including a readout
sensor arrays, readout interface circuits, and gas recognition interface circuit and a gas recognition circuit. A microprocessor
circuits were reported. From 2006, researchers began to find an was adopted to realize the KNN algorithm as a gas recognition
energy-efficient and high-integration solution with less power circuit. The overall system had 100% classification accuracy in
dissipation and longer battery lifetime. testing five alcoholic products. Besides, the signal-processing
In 2006 Tang, K.-T et al. [81] designed a chip distinguishing chip, fabricated in 0.18µm process, consumed only 1.05 mW,
8 different odors with 3 polymer sensors using 1.5µm process. lower than 7.6 mW in [81] and 2.81 mW in [31].
An on-chip analog memory was adopted to replace the digital A fully-integrated E-nose system-on-chip (SoC) was reported
counterpart. As a result, by storing data in analog form, analog in the work of Tang, K.-T. et al. [32] in 2014. Different from
to digital (A/D) and digital to analog (D/A) converters, which [82], as shown in Fig. 2, this fully-integrated chip consisted
cost most of the power, were avoided. The power consumption of eight on-chip sensors, a readout interface circuit, and a gas
of the chip in standby and full load modes was 1.3 mW and recognition circuit. The gas recognition circuit integrated two
7.6 mW respectively. In 2011, they developed a lower-power cores: (1) a digital continuous restricted Boltzmann machine
chip for 100% classification of three odors with 8 sensors [31]. (CRBM) core loaded with a probabilistic neural network for
The power consumption of the gas recognition circuit fabricated data dimension reduction; (2) a RISC core as a KNN classifier
in 0.18µm process was 2.81 mW with 1.8V supply. Compared to output recognition results. The chip consumed 1.27 mW at
with their previous work in [81], this work made much progress 0.5V and provided a solution for the rapid diagnostic strategy
in saving power. to ventilator-associated pneumonia with 95.73% accuracy using
In an effort to further reduce power consumption, Shih-Wen 0.18µm process.
Chiu et al. [82] showed a power-efficient and high-integration Compared with the previous work with 1.05 mW in [82],
solution for E-nose system in 2014. As depicted in Fig. 1, this chip consumed 0.22 mW more power but achieved the
172 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

TABLE I
PERFORMANCE COMPARISON OF E-NOSE CHIP FOR CLASSICAL GAS RECOGNITION

∗
not fully-integrated NA: Not Available NR: Not Reported DC: Digital Controller LTA: Loser-Take-All CB: Carbon-Black MWNTs: Multi-walled Carbon Nanotubes AFE:
Analog Front-end CRBM: Convolution Restricted Boltzmann Machine KNN: K-Nearest Neighbor CFI: Clustering Fisher Index EC: environmental compensation FBAR: Film
Bulk Acoustic Resonator VAP: ventilator-associated pneumonia

whole system integration and better computation capability by C. Limitations of Classical Gas Recognition in E-nose
incorporating two processing cores and a series of buses for fast
According to our survey, the classical gas recognition algo-
data transmission. Moreover, the E-nose chip supported on-chip
rithms in E-nose achieve good accuracy for small samples with
learning and common interface protocols for connections of
limited gas sensors [62], [67]–[69], [71], [72]. One of the reasons
other devices conveniently. With the similar framework in [32],
is their relatively good generalization. Most of the classical
the authors presented another E-nose SoC consuming 1.68 mW
approaches are non-linear methods such as SVM. The non-linear
at 0.5V for chronic obstructive pulmonary disease screening
relationship between data and features is easy to be found with
with 92% accuracy in 2016 [33].
SVM when the sample size is small. However, in some cases,
Tang, K.-T et al. presented a following-up study [30] of an
the classical algorithms aren’t the best. The studies in [34], [35]
E-nose SoC with on-chip learning in 2018. Compared with
show that on the same dataset, the classical methods (such as
their previous work with 8 sensors [32], this chip integrated
SVM) have lower accuracy than ANNs-based gas recognition
36 on-chip sensors including temperature and humidity sensors
approaches.
acquiring environmental information. The chip was fabricated
From our investigation, we find that few of the proposed
with 180nm CMOS process and consumed 2.6 mW at 1V. Be-
gas recognition algorithms could be implemented on chips,
sides, together with CRBM+KNN for classification, a modified
which was due to two main limitations of the algorithms: hard-
PCA-component correction was adopted as a compensation
ware friendliness and anti-noise capability. Lots of classical gas
algorithm according to humidity and temperature. To verify
recognition algorithms have built-in functions requiring matrix
the performance of the chip, they collected 70 gas sensor data
multiplications and a large amount of sensor data, which makes
including methanol and ethanol, and finally obtained 93% clas-
them difficult to be realized in low-power chips. From Table I, it
sification accuracy.
is observed that KNN is implemented in many E-noses because
According to the analysis above, we compare the perfor-
it is more effective and hardware-friendly than other algorithms.
mance of E-nose using classical gas recognition circuits as
Another limitation is the poor anti-noise capability. Gener-
shown in Table I, from which we find that recent studies
ally, the gas sensor is sensitive to the external environment, so
focused on the integration of multiple on-chip sensors [30]–
the sensor data is often inaccurate [27]. In a relatively stable
[33], [82]. Additionally, the more complex gas recognition
environment, like laboratory, good performance of sensors and
algorithms are adopted, the higher accuracy is achieved [30],
E-nose system could be obtained. However, the performance of
[32], [82]. Moreover, environment compensation circuits are
the same E-nose system with the same sensors will be greatly
important for interface readout circuits [26], [28] and gas
affected in the wild. Therefore, it is necessary to find anti-
recognition circuits [30], which help to alleviate the influence
noise and hardware-friendly gas recognition algorithms with
of temperature and humidity on the chip results to a great
the ability to recognize more types of gases for the E-nose
extent.
system.
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 173

III. GAS RECOGNITION APPROACHES USING ARTIFICIAL performance using the leave-1-out cross validation method for
NEURAL NETWORKS gas recognition, and the test accuracy reached over 90%.
For complex gas recognition tasks, ANN has become a good
choice. In comparison with classical gas recognition algorithms, 4) SNN
ANN achieves higher accuracy by adjusting its network layers, As the third generation of the neural network, spiking neural
number of neurons in each layer, built-in functions and parame- network (SNN) is originally inspired by the human brain and its
ters of neurons, and so on. Generally, the more gas data samples, communication scheme where neurons are used for information
the better performance will be obtained. transformation via discrete action potentials (spikes) through
adaptive synapses. A spike is generated when the running sum of
1) BPNN changes in the membrane potential, which can result from presy-
naptic stimulation, crosses a threshold [42]. Compared with
Back propagation neural network (BPNN) is a basic ANN
other ANNs, SNN costs with less power dissipation as sparse
with a forward structure which maps a set of input vectors into
spikes are adopted. Many works [50], [51], [92] have verified
output vectors. The product of the input vector and the initial
the performance of SNN-based algorithms for gas recognition
weight is activated by an activation function and then is passed
and their hardware implementations, which will be discussed in
to the neurons of the next layer until to the output neurons.
Section IV.
After that, the output value and the real value are processed by
the loss function, and the result is used to update the weight
of each neuron in each layer through the chain rule reversely. 5) Summary and Comparison
Through iterations, the local optimal solution of this network In designing an E-nose system, designers often choose an
for classification or regression is found [83]. In 2009, Yu, H appropriate gas recognition algorithm according to the features
et al. [34] developed a PCA+BPNN classifier to identify the and size of sensor data. Among the classical and ANN-based
quality level of green tea with an accuracy rate of 88%. More gas recognition algorithms discussed above, some (such as DT)
applications of BPNN for gas recognition are reported in recent are suitable for high-dimensional gas data containing various
years, such as the smell detection for the freshness of chicken distinctive chemical characteristics, and others (such as SVM)
with 94% accuracy [35] in 2016, and the odor identification [36] are good for sensor data with large fluctuations. Therefore, we
of aspergillus species and quantity with 96.4% accuracy in 2019. summarize and compare the properties of these algorithms in
Table II.
2) CNN From Table II, we find that ANNs(such as BPNN, CNN,
RBFNN) have advantages in terms of recognition accuracy and
Convolutional neural network (CNN) is a well-known deep
the number of identified gas types. Generally, with enough
learning architecture inspired by the natural visual perception
data for training, ANNs have often higher accuracy and are
mechanism of the living creatures [94]. To our knowledge, CNN
optimized by increasing the size of the network or changing
was first applied in gas recognition in the work of Pai Peng et al.
built-in parameters. Unfortunately, it’s difficult to obtain enough
[84] in 2018. They presented a CNN with up to 38 layers to test 4
data to train the networks in the wild.
different odors with 95.2% accuracy in a range of concentrations.
As said above, many classical gas recognition algorithms are
Another work [85] was introduced by Xiaofang Pan et al. in
effective with small samples from limited gas sensors. One main
2019 about a novel hybrid approach which combined convo-
reason is their good generalization of the classical methods. Such
lutional and recurrent neural networks (RNNs) to recognize 4
methods often have only a few parameters and limited feature
different gases with the highest accuracy of 98.28%. In 2020,
space. Therefore, for learning on small sample sizes with limited
Wang, Y et al. [86] put forward an optimized CNN, and a special
dimensions of input data, over-fitting seldom happens with the
striped 1D kernel was used in the convolutional and pooling
methods. The non-linear relationship between data and features
layers. In the work, various types of Chinese herbal medicine
is easy to be found when the sample size is small.
were classified with 87.56% accuracy. In 2021, D. Ma et al. [87]
On the other hand, from the studies in [36], [84], [85], [87], it is
proposed a deep conventional neural network with 15 layers
revealed that ANN-based algorithms perform better in gas recog-
through a sensor array response map to recognize 10 different
nition. One of the important reasons is that when the data size
VOCs (volatile organic compounds) with 92% accuracy.
is big enough, the classical algorithms will have performance
bottlenecks with limited parameters and solid frameworks, but
3) RBFNN
the performance (such as accuracy) of ANN-based methods can
Radial basis function neural network (RBFNN) is a feed- be improved further.
forward neural network which is easy to be designed with good From [64], [69], [71], [90], [91], we find researchers often
generalization, strong tolerance to input noise, and online learn- choose and implement the best algorithm through lots of exper-
ing ability [88]. It has been proven that RBFNN can approximate iments, or organically combining two or more algorithms for
any continuous nonlinear function with arbitrary accuracy. In given gas recognition scenarios [67], [68], [93]. Compared with
the 2018 study of Thazin, Y.et al [90], RBFNN was adopted ANN, many classical algorithms (such as KNN and SVM) are
to recognize the acidity level of coffee with 95% accuracy. considered as “lazy” learning methods as they have relatively
Another study [91] used RBFNN to extract features and tested fixed frameworks and a few parameters. Recent studies [64],
174 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

TABLE II
COMPARISON OF ALGORITHMS FOR GAS RECOGNITION

[67]–[69], [71], [72], [93] show that the classical gas recognition Biological neurons interact with each other by transmitting
algorithms are feasible in the classification of limited types spikes through adjustable synapses. Hebbian theory [97] de-
of gases. According to the analysis of the literatures, we find scribes the basic principle of synaptic plasticity, that is, the
that it is appropriate to distinguish less than five gases. In continuous and repeated stimulation from presynaptic neurons to
contrast, some ANNs (such as BPNN and RBFNN) achieve postsynaptic neurons leads to the increase of synaptic transmis-
higher accuracy in gas recognition of more than five types of sion efficiency. And spike timing dependent plasticity (STDP) is
odors. However, some works [71], [35], [36], [90], [91] show a temporally asymmetric form of Hebbian learning induced by
that as the number of identified gases increases, the amount of tight temporal correlations between the spikes of pre- and post-
required sensor data increases as well. synaptic neurons [98].
From the perspective of hardware implementations, the As defined in [99], spike-timing dependent plasticity (STDP)
synapse weights of ANNs need to be updated in each iteration. is a form of synaptic plasticity sensitive to the precise timing
Generally, a considerable number of iterations are needed in of spike firing relative to impinging presynaptic spike times.
ANNs. As a result, much calculation during the training stage It relies on local information driven by backpropagation of
is required, which costs much time and power in hardware action potential (BPAP) through the dendrites of the postsy-
implementation. Unfortunately, few ANN-based gas recognition naptic neuron. Once gas information is encoded in the form of
circuits in E-nose systems are reported, most are about the spikes adjusted by STDP, higher brain regions can process this
realization of gas recognition with a field-programmable gate information for odor recognition [100], [101].
array (FPGA) [38]–[41]. In [102], an insect-olfaction SNN for gas classification was
introduced by Schmuker et al. in 2007. In the work, a three-step
IV. SNN-BASED GAS RECOGNITION METHODS approach mimicked neurocomputational olfactory systems was
discussed for high-precision prediction of the biological activity
The computational power of bio-inspired systems has at-
of pharmaceutical active compounds. The highlight of this work
tracted increasing attention from the research community [103].
was the introduction of virtual receptors as modeling response
In essence, as the third generation neural networks [95], SNNs
patterns to code the gas stimulus into a self-organizing map,
can be considered the most biologically plausible neuron models
which expressed spiking intensity response to degrees of gas
to capture informational dynamics observed among real biolog-
stimulus.
ical neurons, and to represent and integrate several information
In addition, the authors presented their insect olfaction sparse
dimensions [96]. Moreover, through modeling neuromorphic
neuromorphic SNN for multivariate data with 94% accuracy
olfaction, many encouraging researches about SNNs have been
in 2014 [45] with the learning engines of STDP and global
emerging in the field of gas recognition.
weakly-supervised spike reward-dependent plasticity (SRDP).
Subsequently, they expanded the size of the network to over
A. Gas Recognition Algorithms Based on Biology-Inspired 6000 neurons and 6 million synapses in 2016 [92]. Dense
SNN plastic connectivity was added between the projection neuron
In biology, olfactory receptor neurons (ORN) are responsible (PN) and association neuron (AN), and a simple reinforcement-
for transmitting gas information in the form of spike sequences. style learning rule was applied while the training dataset was
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 175

Fig. 3. Neuromorphic self-organized mapping and spiking classifier model in

2019 [104].

Fig. 4. Encoding a sample sensor response into a spike pattern in 2015 [106].
presented. Additional winners-take-all (WTA) inhibition was
applied between the AN layer to create a final clear decision.
The hardware of this SNN was deployed in GPU, and the SNN
was able to recognize over 20 types of gases with an accuracy membrane potential curve of the output neurons was considered
of 92%. as classification features transmitted to back-end classifiers.
In [104], another STDP-based SNN model with lateral inhi- The best combination (SNN+SVM) with a total correct rate of
bition applied within layers to perform unsupervised clustering 93% was chosen in testing all seven odors. In this method, the
was put forward in 2019. Authors adopted time-invariant and STDP-based SNN was used for data dimension reduction, which
rated spiking coding mapped into a feature space with a specified makes a great contribution to back-end classifiers to classify
resolution. Fig. 3 illustrates the network, in which layers are multiple odors.
implemented as single populations with subpopulations demar- The biology-inspired networks were further optimized in
cated by connectivity. An additional population of Poisson was [108], the dynamic characteristics of OB models were studied in
added to implement supervised learning by exciting or inhibiting their 2017 work in [109], in which the influence of various inhibi-
designated output subpopulations depending on the class of the tion mechanisms on recognition performance was demonstrated,
presented input. including lateral inhibition, reciprocal inhibition, and both. This
Besides the SNN model, spike coding methods of sensor data work supported strongly the necessity of lateral inhibition in
have influence on the accuracy of gas recognition, which is SNN for gas recognition. In 2018, a mature scheme to convert
discussed in [42], [95]. In accordance with the principles of unsupervised SNN into supervised D-SNN with backpropaga-
biological neurons, spike coding methods commonly include tion was presented in [110], [111]. In theory, with the teaching
rated coding and temporal coding. Various spike coding methods signal, the D-SNN should have better recognition performance
were proposed by Yamani, J.A. in [105], in which a temporal than the STDP-based SNN. Unfortunately, no report of D-SNN
spike latency coding with the time interval between spikes by used for gas recognition in E-nose is found.
each neuron was described, and a more compact glomerular A research on different neuron behaviors in the AL model
latency coding method with a higher recognition performance was raised much interest in [112] in 2018. The authors divided
than the original spike latency coding was discussed. As shown the neurons in AL into two categories: specialists (neurons that
in Fig. 4, similar work in [106] was presented by Sankho react to a few stimuli) and generalists (neurons that respond to
et al., in which a novel spike coding method was designed using a lot of stimuli). Moreover, the relationship between the ratio
multiple Gaussian receptive fields superior over the temporal of the number of these two classes of neurons in SNN and the
e-nose responses. The SNN with the spike coding method had the stimulus complexity was discussed in detail. This work brought
highest accuracy of 94.68% for odor classification of orthodox inspiration for complicated gas recognition tasks, that is, two
black tea. or more types of neuron models of SNN could improve the
However, in practice, many SNNs based on STDP learning performance of the networks.
cannot achieve satisfying precision with unsupervised learn- To reduce calculation costs, an SNN with few-shot learning
ing algorithms which lack effective learning directions. Lateral was proposed by T. A. Cleland et al. in 2019 in [50]. It was
inhibition is an effective solution as proved by the works in inspired by the external plexiform layer (EPL) in OB. Similar
[92]. On the other hand, additional classifiers with supervised to AL in [92], each sensor output was transmitted into a specific
learning could be added to improve recognition performance. excitatory and inhibitory neuron, thus forming many groups like
For example, Ya-Qi Jing et al. proposed an SNN together glomerulus for each sensor data. Furthermore, they developed
with LDA/SVM/BPNN respectively to recognize seven types precedence coding as input spikes of SNN for each sensor
of Chinese liquor odors in 2016 [108]. The variance of the channel. More uniquely, they proposed a novel lateral inhibition
176 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

of designing SNN algorithms for specific gas recognition tasks,

such as Brian [116].

B. Hardware Implementation of SNNs for Gas Recognition

As mentioned above, Koickal, T.J et al. [123] first proposed
the SNN neuromorphic olfaction chip in 2006, which is a mixed-
signal circuit with an on-chip sensor array, an interface circuit,
and a gas recognition circuit. A two-layer SNN including 12
neurons and 81 synapses was realized by analog circuits. The
learning rules of global STDP and local lateral inhibition in the
Fig. 5. Schematic model of EPL network circuitry in 2019 [50]. output layer were used to train the spike weights of synapses.
Another hardware prototype was proposed in 2010 by Beyeler
et al. [44]. Different from that in [123], the neuromorphic VLSI
method used in the data preprocessing stage and matched for the chip included 2048 neurons and 6144 synapses realized with
glomerular layer as illustrated in Fig. 5. This network supported digital circuits. The works in [44], [123] could be considered as
few-shot learning with an average accuracy of 94.93% for the early SNN olfaction chips and laid a solid foundation for other
six odors classification. future works. However, the accuracy and power consumption of
Based on the work in [50], a state-of-the-art work for gas these two works were not given and no follow-up design of the
recognition was put forward in 2020 in [51]. Imam Nabil and system was published.
Cleland T.A presented a neural network algorithm supporting In 2010, H. S. Abdel-Aty-Zohdy et al. developed a neuromor-
one-shot learning and identification of odorant samples. The phic olfaction SNN chip consisting of 32000 synapses and 255
network reached 92% accuracy for 10 kinds of gases recognition neurons for gas recognition, which supported on-chip learning
under 40% noise. Besides STDP and lateral inhibition, an addi- and identified gases under noise [124]. However, SRAM mem-
tional inhibition layer was adopted to restrain excitatory neurons ory and sensors array were not integrated into the chip, and no
in other different glomeruli. Rather than Hamming Distance test results and power consumption were reported.
used in previous work [50], the Jaccard similarity coefficient With spike sparsity, the power-efficient SNN chips were ap-
was designed as the classification standard in this work. The propriate for portable E-nose devices, which was proved by the
work has made a great breakthrough in gas recognition, in which 2011 work of Tang, K.T et al. [49]. In the work, an SNN chip for
distributed, event-driven computations and online incremental gas recognition was introduced, which encoded the sensors array
one-shot learning were adopted to identify various kinds of gases data into 2D spike sequences, and identified gas by matching the
accurately. In addition, the work was implemented on the Intel reference gas stored in the 2D spatiotemporal feature library. The
Loihi [55] neuromorphic platform. chip comprised two parts: a self-made 4 × 4 SnO2 gas sensors
Other researchers focus on linking biologically plausible array and SNN circuits for odor identification. It consumed
learning SNNs with brain-inspired simulators. Vanarse. A et al. 6.6 mW at the voltage ranging from 0.15V to 2.7V with 0.35µm
reported their works in 2019 about a new SNN encoder docked process. Classification performances were evaluated with three
with BrainChip’s Akida development environment (ADE) with gases and a 94.9% correct detection rate was achieved.
92.5% accuracy in [113]. In 2020, they explored spike coding An improved work about SNN olfaction circuits was proposed
methods to identify odors based on the 3D SNN of NewCube in 2012 by Hsieh. and Tang. K.T in [125], in which an energy-
[114] in [115]. Their works prove that SNN modeling on deep saving SNN chip mimicking the OB model was described.
layer networks of the brain has great potential for gas recogni- Compared with that in [49], the chip area and power consump-
tion. tion were further optimized through sub-threshold oscillation
As mentioned above, through modeling OB/AL, various bi- and onset-latency representation. This chip was fabricated with
ological shallow SNNs using STDP and lateral inhibition have 0.18µm process and consumed 3.6µW without sensors array.
been developed. Here, we review the SNNs for odor identifi- The classification performance was evaluated with three kinds
cation in recent 20 years and compare them in Table III, from of gases data sampled by commercial e-nose sensors (Cyra-
which we find that many approaches were adopted to improve nose329) with an 87.59% correct detection rate.
the recognition performance, such as modeling according to In 2017, Hsieh, H.Y et al. [126] proposed an on-chip learning
insect olfaction [102], various mechanisms of lateral inhibition analog circuit for probabilistic spike neural network (PSNN)
[50], [51], [92], [109], spike coding for specific odor identifi- with a gas recognition accuracy of 92.3% for three fruity odors.
cation tasks [106], adding extra classification algorithms [50], Fabricated in 0.18µm CMOS process, this chip consumed less
[51], [108], self-regulated neuron threshold through its spiking than 10µW at 1V supply under the condition that all of the
rate [50], spike reward-dependent plasticity [104], [102], using transistors were operated in subthreshold region. Even though
multiple neuron models [112], and glomerulus-like design [50], no on-chip sensors or SRAM were integrated with the chip, such
[51], [92], [104]. Besides, it is reported that using brain-like the low-power approach is promising in designing wearable,
simulators to classify gas data may be another approach instead portable and implantable E-nose devices.
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 177

TABLE III
OVERVIEW OF SNN ALGORITHMS FOR GAS RECOGNITION

NR: Not reported MCC: Matthew’s Correlation Coefficient ORN: Olfactory Receptor Neuron PN: Projection Neuron LN: Local-inhibitory Neuron
EN: Exhibitory Neuron IN: Inhibitory Neuron VOC: Volatile Organic Compounds AN: Association Neuron M-P: McCulloch-Pitts Model SRM: Spike Response Model
IF: Intergrate-and-Fire Model LIF: Leaky-Intergrate-and-Fire Model MC: Mitral Cell CC: Cortical Cell GC: Granule Cell EN: Exhibitory Neuron AERO: Address Event
Representation for Olfaction SF: Step Forward H-R: Hindmarsh–Rose model

In the same year of 2017, P. Huang et al. put forward another reason is the implementation of Euclidean vector normalizers as
analog chip simulating the front end of gas sensing and realized a calibration method was used to mitigate the current mismatch
a trainable feature extraction algorithm in [52]. The authors between branches. These allowed the transistors to be small in
designed a 6-channel frontend and simulated gas sensor response size, thereby reducing current consumption. The third reason is
for a specific gas with current inputs. Each channel consisted that the algorithms had direct implementation in subthreshold
of two logarithmic converters, a voltage-to-current converter, analog circuits with low power supply, achieving high energy
and an integrate-and-fire (IF) circuit. Besides, they implemented efficiency.
an on-line learning circuit (including a vector normalizer, and In comparison with other works, the SNN-based algorithm
sample-and-hold (S/H) circuits), which was able to adapt itself in [52] extracted one composite feature of all analyses by
to the input statistics and capture some of the parameter shifts performing the gradient descent algorithm during training and
due to variations and mismatches. transformed the sensor responses into concentration-invariant
The circuits in [52] were fabricated using the 65nm pro- spike patterns, with subthreshold analog circuits the gas sensing
cess, consumed only 519nW and 463nW at 0.6V and 0.65V frontend had much lower power consumption.
respectively, lower than that of other analog chips [125], [126]. As mentioned above, some researchers implement their SNN-
There are three main reasons. The first one is that the spike time based gas recognition algorithms on neuromorphic hardware
encoding scheme was adopted in the algorithm. It was robust to such as Loihi [42] and Spikey [107]. The work in [45] imple-
mismatches between channels, as it would appear as a constant mented a configurable chip with digital-analog mixed signals
shift in the output delay vector for all analytes. The second based on Spikey [107] as a general neural network simulator,
178 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

TABLE IV
COMPARISON OF SNNS HARDWARE IMPLEMENTATION

NA: Not Available NR: Not Reported

which had the analog circuits of neurons and synapses and the reported in [51], neuromorphic systems are custom integrated
digital transmission of the action potential. The chip was com- circuits that model biological neural computations, typically
prised of a three-layer SNN and a naive Bayes classifier, which with orders of magnitude greater speed and energy efficiency
was considered as a hardware approach for olfaction-inspired than general-purpose computers. Meanwhile, asynchronous de-
processing and classification of gas data. sign methods have long been seen as the appropriate tool for
From the works discussed above, we think that SNN-based prototyping SNNs [55]. The authors in [51] combined the spike
circuits are often power-efficient for E-nose systems. Neverthe- sparsity of SNN with the event-driven feature of asynchronous
less, as the gas sensor is sensitive to the environment, it’s still circuits to achieve high-precision fast gas recognition and power
hard for the E-nose system to be applied in practice. In order saving, which is of importance in gas recognition.
to overcome the challenge, N. Imam et al. proposed anti-noise So far there are fewer reports about SNN hardware im-
gas recognition SNN algorithms and hardware implementation plementation including ASIC design and neuromorphic chip
as presented in [117] in 2012. For the algorithms, few-shot deployed compared with conventional gas recognition circuits.
even one-shot learning SNN for gas detection under noise has Table IV depicts the comparison of the properties of classical,
been developed as described in [50], [51]. With respect to the ANN-based and SNN-based gas recognition described above.
hardware, a scalable digital SNN chip to simulate the OB model A fully-integrated, more on-chip sensors, more power-efficient
was proposed in [117] in 2012, which could remove similar odor SNN ASIC is a promising solution to E-nose. On the other hand,
representations without knowing the statistical distribution of SNNs can be deployed in programmable neuromorphic chips
analyte characteristics in advance. like FPGA, which is another hotspot in the field.
In 2020, the SNN for gas recognition in [51] was deployed in
neuromorphic systems like Loihi. The network included 72 sen-
C. Comparison of SNN-based Gas Recognition and Other
sor channels, each column of the model was configured within
Methods
one neuromorphic core of Loihi, thereby a total of 72 cores
were used. One inference cycle of the 72-core network required As depicted in Table III, the performance of the SNN-based
2.75ms and consumed 0.43mJ. This work provided a useful gas recognition algorithm includes the number of recognized
testbed to evaluate hypotheses on olfactory SNNs. Besides, as gas types and accuracy, which is closely related to the spiking
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 179

TABLE V
COMPARISON FOR CLASSICAL, ANN-BASED AND SNN-BASED GAS RECOGNITION

coding, neuron model, learning method, neural network scale, affected by external surroundings. This is of great interest for gas
and so on. According to [51], [92], [106], [108], [109], it is recognition.
revealed that three or four-layer SNN are enough to obtain As said in [136], SNNs usually perform worse than ANNs in
satisfying performance in some gas recognition tasks. To fur- terms of accuracy. However, SNNs have advantages over most
ther improve the performance, researchers often adopt lateral ANNs in robustness. In the study [137], the authors compare the
inhibition mechanisms. Furthermore, adding a teaching signal attack of SNNs and ANNs and reveal the robustness of SNNs
helps to improve the training of SNN, and better gas recognition against adversarial attacks. Besides, the authors conduct exten-
results can be obtained as a result [45], [92], [104]. sive experiments on various datasets and show a 99%+ attack
In regard to the hardware implementation of SNN-based gas success rate in most cases, which is the best result on SNN attack.
recognition, as shown in Table IV, an increasing number of The results indicate the better ability of anti-noise of SNNs
neurons and synapses are integrated into the chips, the power over most ANNs. Therefore, the SNN-based gas recognition
consumption doesn’t increase as fast as the size of the SNNs. algorithms are more robust with the gas data easily affected by
Besides, we find some researchers design analog neurons and surroundings like temperature and humidity.
synapses to realize gas recognition [49], [52], [123], [125]– Besides, SNNs have better online learning ability. Most ANNs
[127]. The reason is that analog circuits are more biomimetic work on non-local information that restricts their scalability
to model biological signals. However, digital circuits seem when they are implemented in hardware. Consequently, the
more suitable to realize SNN-based olfaction chips [44], [117], methods cannot perform online and incremental fast and energy
[124]. First, digital hardware can integrate more neurons and efficient learning, similar to the learning observed in humans
synapses, which makes it possible to obtain higher accuracy and and animals [138]. In comparison, SNNs based on biological
recognize more types of gas. Moreover, asynchronous digital plausible learning methods like STDP perform well in online
circuits exchange signals with a handshake mechanism, which learning, which is meaningful for gas recognition. As well
naturally matches the event-driven and sparsity characteristics known, one fundamentally ill-posed problem that needs to be
of SNNs. Some SNN-based gas recognition algorithms have addressed urgently is sensor drift, in which the sensitivity and
been deployed in asynchronous neuromorphic Loihi [51], [55] selectivity profiles of gas sensors gradually change over time of
in recent works. Moreover, SNN-based gas recognition circuits use or disuse [50]. In [50], the SNN exhibit high classification
are often energy-efficient like [52]. From [52], we concluded performance after few-shot learning without catastrophic forget-
that one of the key reasons for low-power consumption is due ting, which is a challenge for most ANNs [128]. Some ANNs like
to the reduction of current consumption caused by small size of extreme learning machine (ELM) also offer a feasible solution
transistors. And an important reason for the viability of small through designing highly introspective or specific corrective
size transistors is that the spike time encoding scheme of SNN algorithms. But they require algorithmically guided selection
is robust to mismatches between different input channels. of relevant samples across batches and/or the utilization of test
The comparison of classical, ANN-based, SNN-based gas data as unlabeled data for additional training [129]–[130].
recognition is shown in Table V, from which we find that Moreover, SNNs have better hardware friendliness. Com-
SNN-based gas recognition has good accuracy and anti-noise bined with bio-plausible local learning rules, this makes it easier
ability. In [51], it is reported that the SNN with one-shot to build low-power hardware for SNNs [42]. The study in [136]
learning achieves over 90% accuracy in 10 kinds of gases analyses the cost of hardware implementation for ANNs and
under more than 40% noise. These encouraging results help SNNs, and the results show that the memory cost of SNNs is
to improve the gas recognition accuracy, which is always comparable with that of ANNs, but SNNs have lower compute
180 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

Fig. 6. Major contributions in gas recognition for E-nose.

cost. In the study of [138], the authors find that SNNs are compat- the low-power and high-integration E-nose system is an im-
ible for implementation in hardware as they can be more energy portant research topic. In 2014, the fully-integrated E-nose was
efficient compared to classical neural networks that work based implemented using classical and ANN-based gas recognition
on continuous values. Additionally, the nature of information algorithms [32]. Since 2014, SNN-based gas recognition works
communication in SNNs, which is based on spike activities, is have been emerging [92], [106], [108], [109], [112]. Besides,
close to the binary processing in a hardware platform. SNN-based gas recognition achieves precise recognition over
However, SNNs with STDP cannot always achieve satisfy- 20 kinds of gases with 92% accuracy [92] and over 10 kinds of
ing accuracy. With the local unsupervised learning rule, the gases under 40% noise with 92% accuracy [50], [51].
algorithm has no prior information, which controls the training
process, allowing the algorithm to obtain the best results without
violating it. Besides, with its encoding methods (transforming V. CONCLUSION
data into spikes), some key features of the data may be lost, In this paper, we review gas recognition approaches in E-nose
resulting in poor accuracy. In order to obtain better accuracy, including classical, ANN-based, and SNN-based algorithms and
researchers adopt the SNN+ANN approach in gas recognition their hardware implementations. From our investigation, we find
and realize 97.14% and 100% accuracy in [108] and [109] that classical gas recognition algorithms are easy-design and
respectively. Alternatively, ANN to SNN conversion (S-ANN) perform well in the case of limited gas samples with a few
techniques also improve the accuracy, which are able to obtain parameters and relatively fixed structures, but weak in multi-
competitive results with ANNs in terms of accuracy and com- gas recognition under noise. They are often used to recognize
putational power [131]–[133]. less than 5 kinds of gases. While ANN-based gas recognition
In addition, design complexity is another significant barrier to algorithms are able to achieve satisfying accuracy with a number
the development of SNNs. The complexity is mainly due to the of gas data. Instead of MPU with the classical methods, ANNs
sensitivity to parameter values, a large number of parameters to are often designed in form of ASIC.
be optimized, and the lack of a coherent framework for many Besides, current researches prove shallow SNNs can recog-
different models and coding schemes [96]. Moreover, using local nize over 10 kinds of gases under over 40% noise successfully,
learning methods, such as STDP which is bio-plausible, to train showing that SNNs have considerable potential for anti-noise.
large-scale networks remains a deep challenge [42]. Besides, online learning capability of SNN provides a novel
In order to intuitively illustrate the overall development of solution against sensor aging and drift. Similar to the shallow
gas recognition for E-nose, we summarize major contributions in ANNs, SNNs are often implemented in low power ASIC for
terms of algorithms and hardware in the recent 20 years in Fig. 6. gas recognition. Moreover, SNNs can be deployed in neuromor-
From Fig. 6, it is observed that in the 2000s researchers proposed phic platforms like Loihi. This brings great convenience to the
feasible solutions for classical, ANN-based, and SNN-based gas hardware implementation of SNN for gas recognition.
recognition [25], [123] and reported a few novel SNN gas recog- However, SNNs for gas recognition have challenges in terms
nition algorithms [102] and theories [44]. In the last decade, of accuracy and design complexity. SNNs with STDP cannot
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 181

always achieve satisfying accuracy. In order to obtain accuracy, [18] R. Jansen, J. W. Hofstee, H. Bouwmeester, and E. V. Henten, “Automated
the SNN+ANN and ANN to SNN conversion approaches are signal processing applied to volatile-based inspection of greenhouse
crops,” Sensors, vol. 10, no. 8, pp. 7122–7133, 2010.
designed and there have been some studies on the former with [19] M. Morvan, T. Talou, A. Gaset, and J. F. Beziau, “Electronic-
success in gas recognition. In addition, sensitive to parameter nose systems for control quality applications in automotive indus-
values, a large number of parameters that need to be optimized, try,” Sensors Actuators B: Chem., vol. 69, no. 3, pp. 384–388, Oct.
2000.
and the lack of a coherent framework for the many different [20] J. W. Gardner, P. Bartlett, G. Dodd, and H. Shurmer, “Pattern recognition
models and coding schemes make SNNs complicated in design in the warwick electronic nose,” in Proc. 8th Int. Congr. Eur. Chemore-
In future, the development trend of gas recognition circuits ception Res. Organisation, 1988, p. 9.
[21] A.-C. Romain, D. Godefroid, and J. Nicolas, “Monitoring the exhaust
built with classical, ANN-based, and SNN-based gas recogni- air of a compost pile with an e-nose and comparison with GC–MS data,”
tion is in sync with the current interest in enabling ubiquitous Sensors Actuators B: Chem., vol. 106, no. 1, pp. 317–324, 2005.
intelligence and power-efficient design. It is now the time for [22] S. Zampolli et al., “An electronic nose based on solid state sensor
arrays for low-cost indoor air quality monitoring applications,” Sensors
the interchange of ideas, with multidisciplinary efforts spanning Actuators B: Chem., vol. 101, no. 1-2, pp. 39–46, 2004.
circuits and algorithms to synthesize energy-efficient and intel- [23] R. E. Baby, M. Cabezas, and E. W. De Reca, “Electronic nose: A useful
ligent E-nose systems. tool for monitoring environmental contamination,” Sensors Actuators B:
Chem., vol. 69, no. 3, pp. 214–218, 2000.
[24] L. Cheng, Q.-H. Meng, A. J. Lilienthal, and P.-F. Qi, “Development of
compact electronic noses: A review,” Meas. Sci. Technol., vol. 32, no. 6,
2021, Art. no. 062002.
REFERENCES [25] K.-T. Tang and R. M. Goodman, “Electronic olfaction system on a chip,”
in Proc. SCI2001/ISAS2001 Int. Conf. Inf. Syst., Anal. Synth., 2001,
[1] K. Persaud and G. Dodd, “Analysis of discrimination mechanisms in vol. 15, p. 534.
the mammalian olfactory system using a model nose,” Nature, vol. 299, [26] V. Petrescu, J. Pettine, D. M. Karabacak, M. Vandecasteele, M. C.
no. 5881, pp. 352–355, Sep. 1982. Calama, and C. Van Hoof, “Power-efficient readout circuit for minia-
[2] C. A. de March, S. Ryu, G. Sicard, C. Moon, and J. Golebiowski, turized electronic nose,” in Proc. IEEE Int. Solid-State Circuits Conf.,
“Structure–odour relationships reviewed in the postgenomic era,” 2012, pp. 318–320.
Flavour Fragrance J., vol. 30, no. 5, pp. 342–361, 2015. [27] K. Arshak, E. Moore, G. Lyons, J. Harris, and S. Clifford, “A review
[3] J. Tan and J. Xu, “Applications of electronic nose (e-nose) and electronic of gas sensors employed in electronic nose applications,” Sensor Rev.,
tongue (e-tongue) in food quality-related properties determination: A vol. 24, no. 2, pp. 181–198, 2004.
review,” Artif. Intell. Agriculture, vol. 4, pp. 104–115, Jan. 2020. [28] S. H. Lee et al., “20.3 A 4.0× 3.7× 1.0 mm 3-MEMS CMOS integrated E-
[4] A. Sanaeifar, H. ZakiDizaji, A. Jafari, and M. d. l. Guardia, “Early Nose with embedded 4× gas sensors, a temperature sensor and a relative
detection of contamination and defect in foodstuffs by electronic nose: humidity sensor,” in Proc. IEEE Int. Solid-State Circuits Conf., 2020,
A review,” TrAC Trends Anal. Chem., vol. 97, pp. 257–271, Dec. 2017. pp. 316–318.
[5] H. Shi, M. Zhang, and B. Adhikari, “Advances of electronic nose and its [29] I. Nikolaou, H. Hallil, V. Conédéra, G. Deligeorgis, C. Dejous, and D.
application in fresh foods: A review,” Crit. Rev. Food Sci. Nutr., vol. 58, Rebiere, “Inkjet-printed graphene oxide thin layers on love wave devices
no. 16, pp. 2700–2710, Nov. 2018. for humidity and vapor detection,” IEEE Sensors J., vol. 16, no. 21,
[6] D. Harwood, “Something in the air [electronic nose],” IEE Rev., vol. 47, pp. 7620–7627, Nov. 2016.
no. 1, pp. 10–14, 2001. [30] T.-I. Chou et al., “A 1-V 2.6- mW environmental compensated fully
[7] L. Zhang et al., “Classification of multiple indoor air contaminants by an integrated Nose-on-a-Chip,” IEEE Trans. Circuits Syst. II: Exp. Briefs,
electronic nose and a hybrid support vector machine,” Sensors Actuators vol. 65, no. 10, pp. 1365–1369, Oct. 2018.
B: Chem., vol. 174, pp. 114–125, Nov. 2012. [31] K.-T. Tang, S.-W. Chiu, M.-F. Chang, C.-C. Hsieh, and J.-M. Shyu, “A
[8] S. Dragonieri et al., “An electronic nose distinguishes exhaled breath low-power electronic nose signal-processing chip for a portable artificial
of patients with malignant pleural mesothelioma from controls,” Lung olfaction system,” IEEE Trans. Biomed. Circuits Syst., vol. 5, no. 4,
Cancer, vol. 75, no. 3, pp. 326–331, Mar. 2012. pp. 380–390, Aug. 2011.
[9] A. D. Wilson and M. Baietto, “Advances in electronic-nose technolo- [32] K.-T. Tang et al., “24.5 A 0.5 v 1.27 mW nose-on-a-chip for rapid diag-
gies developed for biomedical applications,” Sensors, vol. 11, no. 1, nosis of ventilator-associated Pneumonia,” in Proc. IEEE Int. Solid-State
pp. 1105–1176, 2011. Circuits Conf. Dig. Tech. Papers, 2014, pp. 420–421.
[10] C. Di Natale, A. Macagnano, E. Martinelli, R. Paolesse, E. Proietti, and A. [33] T.-I. Chou et al., “Design of a 0.5 v 1.68 mW nose-on-a-chip for rapid
D’Amico, “The evaluation of quality of post-harvest oranges and apples screen of chronic obstructive pulmonary disease,” in Proc. IEEE Biomed.
by means of an electronic nose,” Sensors Actuators B: Chem., vol. 78, Circuits Syst. Conf. (BioCAS), 2016, pp. 592–595.
no. 1, pp. 26–31, Aug. 2001. [34] H. Yu, J. Wang, H. Xiao, and M. Liu, “Quality grade identifica-
[11] D. Guo, D. Zhang, N. Li, L. Zhang, and J. Yang, “A novel breath analysis tion of green tea using the eigenvalues of PCA based on the E-nose
system based on electronic olfaction,” IEEE Trans. Biomed. Eng., vol. 57, signals,” Sensors Actuators B: Chem., vol. 140, no. 2, pp. 378–382,
no. 11, pp. 2753–2763, Nov. 2010. 2009.
[12] I. Concina, M. Falasconi, and V. Sberveglieri, “Electronic noses as [35] K. Timsorn, T. Thoopboochagorn, N. Lertwattanasakul, and C. Wong-
flexible tools to assess food quality and safety: Should we trust them?,” choosuk, “Evaluation of bacterial population on chicken meats using a
IEEE Sensors J., vol. 12, no. 11, pp. 3232–3237, Nov. 2012. briefcase electronic nose,” Biosyst. Eng., vol. 151, pp. 116–125, 2016.
[13] S. Ampuero and J. O. Bosset, “The electronic nose applied to dairy [36] S. Gu, J. Wang, and Y. Wang, “Early discrimination and growth tracking
products: A review,” Sensors Actuators B: Chem., vol. 94, no. 1, pp. 1–12, of aspergillus spp. contamination in rice kernels using electronic nose,”
Aug. 2003. Food Chem., vol. 292, pp. 325–335, 2019.
[14] A. Berna, “Metal oxide sensors for electronic noses and their application [37] M. Al-Nsour and H. S. Abdel-Aty-Zohdy, “MOS fully analog reinforce-
to food analysis,” Sensors, vol. 10, no. 4, pp. 3882–3910, 2010. ment neural network chip,” in Proc. IEEE Int. Symp. Circuits Syst. (Cat.
[15] M. M. Macías, A. G. Manso, C. J. G. Orellana, H. M. G. Velasco, R. No.01CH37196), 2001, vol. 2, pp. 237–240.
G. Caballero, and J. C. P. Chamizo, “Acetic acid detection threshold in [38] X. Zhai, A. A. S. Ali, A. Amira, and F. Bensaali, “MLP neural network
synthetic wine samples of a portable electronic nose,” Sensors, vol. 13, based gas classification system on zynq SoC,” in IEEE Access, vol. 4,
no. 1, pp. 208–220, 2013. pp. 8138–8146, 2016.
[16] D. Yao, “A gas sensing system for indoor air quality control and polluted [39] T. Jia et al., “Mixed natural gas online recognition device based on a
environmental monitoring,” in Proc. 9th IEEE Conf. Nanotechnol. (IEEE- neural network algorithm implemented by a FPGA,” Sensors (Basel),
NANO), 2009, pp. 806–811. vol. 19, no. 9, 2019, Art. no. 2090.
[17] C. Arnold, M. Harms, and J. Goschnick, “Air quality monitoring and [40] F. Benrekia, M. Attari, A. Bermak, and K. Belhout, “FPGA implemen-
fire detection with the Karlsruhe electronic micronose KAMINA,” IEEE tation of a neural network classifier for gas sensor array applications,” in
Sensors J., vol. 2, no. 3, pp. 179–188, Jan. 2002. Proc. 6th Int. Multi-Conf. Syst., Signals Devices, 2009, pp. 1–6.
182 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

[41] F. Benrekia, M. Attari, and M. Bouhedda, “Gas sensors characteriza- [66] J. Smulko, R. Ionescu, C. Granqvist, and L. Kish, “Determination of
tion and multilayer perceptron (MLP) hardware implementation for gas gas mixture components using fluctuation enhanced sensing and the
identification using a field programmable gate array (FPGA),” Sensors LS-SVM regression algorithm,” Metrol. Meas. Syst., vol. 22, no. 3,
(Basel), vol. 13, no. 3, pp. 2967–2985, 2013. pp. 341–350,2015.
[42] A. Tavanaei, M. Ghodrati, S. R. Kheradpisheh, T. Masquelier, and [67] L.-Y. Chen, C.-C. Wu, T.-I. Chou, S.-W. Chiu, and K.-T. Tang, “De-
A. Maida, “Deep learning in spiking neural networks,” Neural Netw., velopment of a dual MOS electronic nose/camera system for im-
vol. 111, pp. 47–63, 2019. proving fruit ripeness classification,” Sensors, vol. 18, no. 10, 2018,
[43] J. Carter, J. Rego, D. Schwartz, V. Bhandawat, and E. Kim, “Learning Art. no. 3256.
spiking neural network models of drosophila olfaction,” in Proc. Int. [68] Y. Shi, F. Gong, M. Wang, J. Liu, Y. Wu, and H. Men, “A deep fea-
Conf. Neuromorphic Syst., 2020, pp. 1–5. ture mining method of electronic nose sensor data for identifying beer
[44] M. Beyeler, F. Stefanini, H. Proske, G. Galizia, and E. Chicca, “Exploring olfactory information,” J. Food Eng., vol. 263, pp. 437–445, 2019.
olfactory sensory networks: Simulations and hardware emulation,” in [69] W. Jia, G. Liang, H. Tian, J. Sun, and C. Wan, “Electronic nose-based
Proc. Biomed. Circuits Syst. Conf. (BioCAS), 2010, pp. 270–273. technique for rapid detection and recognition of moldy apples,” Sensors,
[45] M. Schmuker, T. Pfeil, and M. P. Nawrot, “A neuromorphic network for vol. 19, no. 7, 2019, Art. no. 1526.
generic multivariate data classification,” Proc. Nat. Acad. Sci., vol. 111, [70] S. Zhang, X. Li, M. Zong, X. Zhu, and R. Wang, “Efficient kNN
no. 6, pp. 2081–2086, 2014. classification with different numbers of nearest neighbors,” IEEE Trans.
[46] A. Vanarse, A. Osseiran, and A. Rassau, “An investigation into spike- neural Netw. Learn. Syst., vol. 29, no. 5, pp. 1774–1785, May 2018.
based neuromorphic approaches for artificial olfactory systems,” Sensors, [71] W. Harsono, R. Sarno, and S. I. Sabilla, “Recognition of original arabica
vol. 17, no. 11, 2017, Art. no. 2591. civet coffee based on odor using electronic nose and machine learning,”
[47] T. Pfeil et al., “Six networks on a universal neuromorphic computing in Proc. Int. Seminar Application for Technol. Inf. Commun. (iSemantic),
substrate,” Front. Neurosci., vol. 7, p. 11, 2013. 2020, pp. 333–339.
[48] K. T. Ng, F. Boussaid, and A. Bermak, “A frequency-based signature [72] E. Mirzaee-Ghaleh, A. Taheri-Garavand, F. Ayari, and J. Lozano, “Iden-
gas identification circuit for SnO2 gas sensors,” in Proc. IEEE Int. Symp. tification of fresh-chilled and frozen-thawed chicken meat and estimation
Circuits Syst.: Nano-Bio Circuit Fabrics Syst., 2010, Art. no. 2275. of their shelf life using an E-nose machine coupled fuzzy KNN,” Food
[49] K. T. Ng, F. Boussaid, and A. Bermak, “A CMOS single-chip gas recog- Anal. Methods, vol. 13, no. 3, pp. 678–689, 2020.
nition circuit for metal oxide gas sensor arrays,” IEEE Trans. Circuits [73] X. Wu et al., “Top 10 algorithms in data mining,” Knowl. Inf. Syst., vol. 14,
Syst. I: Reg. Papers, vol. 58, no. 7, pp. 1569–1580, Jul. 2011. no. 1, pp. 1–37, 2008.
[50] A. Borthakur and T. A. Cleland, “A spike time-dependent online learning [74] D. R. Wijaya, R. Sarno, and A. F. Daiva, “Electronic nose for classifying
algorithm derived from biological olfaction,” Front. Neurosci., vol. 13, beef and pork using naïve Bayes,” in Proc. Int. Seminar Sensors, Instrum.,
p. 656, 2019. Meas. Metrol., 2017, pp. 104–108.
[51] N. Imam and T. A. Cleland, “Rapid online learning and robust recall [75] B. Grodniyomchai, K. Chalapat, K. Jitkajornwanich, and S. Jaiyen, “A
in a neuromorphic olfactory circuit,” Nature Mach. Intell., vol. 2, no. 3, deep learning model for odor classification using deep neural network,”
pp. 181–191, 2020. in Proc. 5th Int. Conf. Eng., Appl. Sci. Technol., 2019, pp. 1–4.
[52] P.-C. Huang and J. M. Rabaey, “A bio-inspired analog gas sensing [76] A. Vergara, S. Vembu, T. Ayhan, M. A. Ryan, M. L. Homer, and R. Huerta,
front end,” IEEE Trans. Circuits Syst. I: Reg. Papers, vol. 64, no. 9, “Chemical gas sensor drift compensation using classifier ensembles,”
pp. 2611–2623, Sep. 2017. Sensors Actuators B: Chem., vol. 166, pp. 320–329, 2012.
[53] J. Park, J. Lee, and D. Jeon, “7.6 A 65nm 236.5nJ/classification neu- [77] K.-T. Tang, Y.-S. Lin, and J.-M. Shyu, “A local weighted nearest neighbor
romorphic processor with 7.5% energy overhead on-chip learning using algorithm and a weighted and constrained least-squared method for mixed
direct spike-only feedback,” in Proc. IEEE Int. Solid- State Circuits Conf., odor analysis by electronic nose systems,” Sensors, vol. 10, no. 11,
2019, pp. 140–142. pp. 10467–10483, 2010.
[54] F. Akopyan et al., “Truenorth: Design and tool flow of a 65 mw 1 mil- [78] B. C. Muñoz, G. Steinthal, and S. Sunshine, “Conductive polymer-carbon
lion neuron programmable neurosynaptic chip,” IEEE Trans. Comput.- black composites-based sensor arrays for use in an electronic nose,”
aided Des. Integr. circuits Syst., vol. 34, no. 10, pp. 1537–1557, Oct. Sensor Rev., vol. 19, no. 4, pp. 300–305,1999.
2015. [79] C. Hagleitner, D. Lange, N. Kerness, A. Hierlemann, O. Brand, and H.
[55] M. Davies et al., “Loihi: A neuromorphic manycore processor with Baltes, “A gas detection system on a single CMOS chip comprising
on-chip learning,” IEEE Micro, vol. 38, no. 1, pp. 82–99, Jan./Feb. capacitive, calorimetric, and mass-sensitive microsensors,” in Proc. IEEE
2018. Int. Solid-State Circuits Conf.. Dig. Tech. Papers (Cat. No. 02CH37315),
[56] S. Marco and A. Gutierrez-Galvez, “Signal and data processing for 2002, vol. 1, pp. 430–479.
machine olfaction and chemical sensing: A review,” IEEE Sensors J., [80] J. Covington, S. Tan, J. Gardner, A. Hamilton, T. Koickal, and T. Pearce,
vol. 12, no. 11, pp. 3189–3214, Nov. 2012. “Combined smart chemFET/resistive sensor array,” in Proc. Sensors,
[57] I. T. Jolliffe and J. Cadima, “Principal component analysis: A review and IEEE, 2003, vol. 2, pp. 1120–1123.
recent developments,” Philos. Trans. Roy. Soc. A: Math., Phys. Eng. Sci., [81] K.-T. Tang and R. M. Goodman, “Towards a wearable electronic nose
vol. 374, 2016, Art. no. 2065. chip,” in Proc. IEEE Custom Integr. Circuits Conf., 2006, pp. 273–276.
[58] T. L. Hayes and C. Kanan, “Lifelong machine learning with deep stream- [82] S.-W. Chiu, H.-C. Wu, T.-I. Chou, H. Chen, and K.-T. Tang, “A miniature
ing linear discriminant analysis,” in Proc. IEEE/CVF Conf. Comput. Vis. electronic nose system based on an MWNT–polymer microsensor array
Pattern Recognit. Workshops, 2020, pp. 220–221. and a low-power signal-processing chip,” Anal. Bioanalytical Chem.,
[59] D. Du, J. Wang, B. Wang, L. Zhu, and X. Hong, “Ripeness prediction vol. 406, no. 16, pp. 3985–3994, 2014.
of postharvest kiwifruit using a MOS E-Nose combined with chemomet- [83] L. Zhang and F. Tian, “Performance study of multilayer perceptrons in
rics,” Sensors, vol. 19, no. 2, pp. 419, 2019. a low-cost electronic nose,” IEEE Trans. Instrum. Meas., vol. 63, no. 7,
[60] J. R. Quinlan, “Induction of decision trees,” Mach. Learn., vol. 1, no. 1, pp. 1670–1679, Jul. 2014.
pp. 81–106, 1986. [84] P. Peng, X. Zhao, X. Pan, and W. Ye, “Gas classification using deep
[61] S. R. Safavian and D. Landgrebe, “A survey of decision tree classi- convolutional neural networks,” Sensors, vol. 18, no. 1, pp. 157,2018.
fier methodology,” IEEE Trans. Syst., Man, Cybern., vol. 21, no. 3, [85] X. Pan, H. Zhang, W. Ye, A. Bermak, and X. Zhao, “A fast and
pp. 660–674, May/Jun. 1991. robust gas recognition algorithm based on hybrid convolutional and
[62] V. Schroeder et al., “Chemiresistive sensor array and machine learning recurrent neural network,” IEEE Access, vol. 7, pp. 100954–100963,
classification of food,” ACS Sensors, vol. 4, no. 8, pp. 2101–2108, 2019. 2019.
[63] C.-C. Chang and C.-J. Lin, “LIBSVM: A library for support vector ma- [86] Y. Wang et al., “An optimized deep convolutional neural network for
chines,” ACM Trans. Intell. Syst. Technol. (TIST), vol. 2, no. 3, pp. 1–27, dendrobium classification based on electronic nose,” Sensors Actuators
2011. A: Phys., vol. 307, 2020, Art. no. 111874.
[64] S. Qiu and J. Wang, “The prediction of food additives in the fruit juice [87] D. Ma, J. Gao, Z. Zhang, and H. Zhao, “Gas recognition method based on
based on electronic nose with chemometrics,” Food Chem., vol. 230, the deep learning model of sensor array response map,” Sensors Actuators
pp. 208–214, 2017. B: Chem., vol. 330, 2021, Art. no. 129349.
[65] A. Uçar and R. Özalp, “Efficient android electronic nose design for [88] H. Yu, T. Xie, S. Paszczynski, and B. M. Wilamowski, “Advantages
recognition and perception of fruit odors using kernel extreme learning of radial basis function networks for dynamic system design,” in IEEE
machines,” Chemometrics Intell. Lab. Syst., vol. 166, pp. 69–80, 2017. Trans. Ind. Electron., vol. 58, no. 12, pp. 5438–5450, Dec. 2011.
CHEN et al.: GAS RECOGNITION IN E-NOSE SYSTEM: A REVIEW 183

[89] W. W. Ng, D. S. Yeung, M. Firth, E. C. Tsang, and X.-Z. Wang, [113] A. Vanarse, A. Osseiran, A. Rassau, and P. van der Made, “A
“Feature selection using localized generalization error for supervised hardware-deployable neuromorphic solution for encoding and clas-
classification problems using RBFNN,” Pattern Recognit., vol. 41, no. 12, sification of electronic nose data,” Sensors, vol. 19, no. 22, 2019,
pp. 3706–3719, 2008. Art. no. 4831.
[90] Y. Thazin, T. Pobkrut, and T. Kerdcharoen, “Prediction of acid- [114] N. K. Kasabov, “NeuCube: A spiking neural network architecture for
ity levels of fresh roasted coffees using e-nose and artificial neu- mapping, learning and understanding of spatio-temporal brain data,”
ral network,” in Proc. 10th Int. Conf. Knowl. Smart Technol., 2018, Neural Netw., vol. 52, pp. 62–76, 2014.
pp. 210–215. [115] A. Vanarse, J. I. Espinosa-Ramos, A. Osseiran, A. Rassau, and N.
[91] S. Balasubramanian, S. Panigrahi, C. Logue, H. Gu, and M. Marchello, Kasabov, “Application of a brain-inspired spiking neural network ar-
“Neural networks-integrated metal oxide-based artificial olfactory sys- chitecture to odor data classification,” Sensors, vol. 20, no. 10, 2020,
tem for meat spoilage identification,” J. Food Eng., vol. 91, no. 1, Art. no. 2756.
pp. 91–98, 2009. [116] D. Goodman and R. Brette, “Brian: A simulator for spiking neural
[92] A. Diamond, M. Schmuker, A. Berna, S. Trowell, and T. Nowotny, networks in python,” Front. Neuroinform., vol. 2, p. 5, 2008.
“Classifying continuous, real-time e-nose sensor data using a bio-inspired [117] N. Imam et al., “Implementation of olfactory bulb glomerular-layer
spiking network modelled on the insect olfactory system,” Bioinspiration computations in a digital neurosynaptic core,” Front. Neurosci., vol. 6,
Biomimetics, vol. 11, no. 2, 2016, Art. no. 026002. p. 83, 2012.
[93] C.-C. Wu, S.-W. Chiu, and K.-T. Tang, “An electronic nose system for [118] R. A. Fisher, “The use of multiple measurements in taxonomic problems,”
rapid detection of ketamine smoke,” IEEE Sensors Lett., vol. 3, no. 8, Ann. Eugenics, vol. 7, no. 2, pp. 179–188, 1936.
2019, Art. no. 6001604. [119] T. Nowotny, A. Z. Berna, R. Binions, and S. Trowell, “Optimal feature
[94] J. Gu et al., “Recent advances in convolutional neural networks,” Pattern selection for classifying a large set of chemicals using metal oxide
Recognit., vol. 77, pp. 354–377, 2018. sensors,” Sensors Actuators B: Chem., vol. 187, pp. 471–480, 2013.
[95] W. Maass, “Networks of spiking neurons: The third generation of neural [120] I. Rodriguez-Lujan, J. Fonollosa, A. Vergara, M. Homer, and R. Huerta,
network models,” Neural Netw., vol. 10, no. 9, pp. 1659–1671, 1997. “On the calibration of sensor arrays for pattern recognition using the min-
[96] J. L. Lobo, J. Del Ser, A. Bifet, and N. Kasabov, “Spiking neural imal number of experiments,” Chemometrics Intell. Lab. Syst., vol. 130,
networks and online learning: An overview and perspectives,” Neural pp. 123–134, 2014.
Netw., vol. 121, pp. 88–100, 2020. [121] CSIRO, A. Berna and T. Stephen, Electronic Nose (Fox) Recording of 20
[97] C. Keysers and D. I. Perrett, “Demystifying social cognition: A hebbian Chemicals. Canberra, Australia: CSIRO Data Collection, 2015.
perspective,” Trends Cogn. Sci., vol. 8, no. 11, pp. 501–507, 2004. [122] A. Vergara and J. Fonollosa, J. Mahiques, M. Trincavelli, N. Rulkov, and
[98] J. Sjöström and W. Gerstner, “Spike-timing dependent plasticity,” Spike- R. Huerta, “On the performance of gas sensor arrays in open sampling
timing dependent Plast., vol. 5, no. 2, Art. no. 1362, 2010. systems using inhibitory support vector machines,” Sens. Actuators B:
[99] H. Paugam-Moisy and S. Bohte, “Computing with spiking neuron net- Chem., vol. 185, pp. 462–277, 2013.
works,” in Handbook of Natural Computing, G. Rozenberg, T. Bäck, and [123] T. J. Koickal, A. Hamilton, T. C. Pearce, S.-L. Tan, J. A. Covington, and
J. N. Kok, eds. Berlin, Heidelberg: Springer, 2012, pp. 335–376. J. W. Gardner, “Analog VLSI design of an adaptive neuromorphic chip
[100] K. Narusuye, F. Kawai, and E.-. Miyachi, “Spike encoding of olfactory for olfactory systems,” in Proc. IEEE Int. Symp. Circuits Syst., 2006,
receptor cells,” Neurosci. Res., vol. 46, no. 4, pp. 407–413, 2003. pp. 4547–4550.
[101] C. Di Natale, “An investigation on the role of spike latency in an artificial [124] H. S. Abdel-Aty-Zohdy, J. N. Allen, and R. L. Ewing, “Spiking neural
olfactory system,” Front. Neuroengineering, vol. 4, p. 16, 2011. network e-nose classifier chip,” in Proc. IEEE Nat. Aerosp. Electron.
[102] M. Schmuker and G. Schneider, “Processing and classification of chem- Conf., 2010, pp. 374–378.
ical data inspired by insect olfaction,” Proc. Nat. Acad. Sci., vol. 104, [125] H.-Y. Hsieh and K.-T. Tang, “VLSI implementation of a bio-inspired
no. 51, pp. 20285–20289, 2007. olfactory spiking neural network,” IEEE Trans. neural Netw. Learn. Syst.,
[103] N. K. Kasabov, “Deep learning of multisensory streaming data for pre- vol. 23, no. 7, pp. 1065–1073, 2012.
dictive modelling with applications in finance, ecology, transport and en- [126] H.-Y. Hsieh, P.-Y. Li, and K.-T. Tang, “An analog probabilistic spiking
vironment,” in Time-Space, Spiking Neural Networks and Brain-Inspired neural network with on-chip learning,” in Proc. Int. Conf. Neural Inf.
Artificial Intelligence, N. K. Kasabov, ed. Berlin, Heidelberg: Springer, Process., 2017, Springer, pp. 777–785.
2019, pp. 619–658. [127] H.-Y. Hsieh and K.-T. Tang, “A spiking neural network chip for odor
[104] A. Diamond, M. Schmuker, and T. Nowotny, “An unsupervised data classification,” in Proc. IEEE Asia Pacific Conf. Circuits Syst., 2012,
neuromorphic clustering algorithm,” Biol. Cybern., vol. 113, no. 4, IEEE, pp. 88–91.
pp. 423–437, 2019. [128] G. I. Parisi et al., “Continual lifelong learning with neural networks: A
[105] J. Al Yamani, F. Boussaid, A. Bermak, and D. Martinez, “Glomerular review,” Neural Netw., vol. 113, pp. 54–71, 2019.
latency coding in artificial olfaction,” Front. Neuroengineering, vol. 4, [129] L. Zhang and D. Zhang, “Domain adaptation extreme learning machines
p. 18, 2012. for drift compensation in E-Nose systems,” in IEEE Trans. Instrum.
[106] S. T. Sarkar et al., “Towards biological plausibility of electronic noses: Meas., vol. 64, no. 7, pp. 1790–1801, Jul. 2015.
A spiking neural network based approach for tea odour classification,” [130] Z. Ma, G. Luo, K. Qin, N. Wang, and W. Niu, “Online sensor drift
Neural Netw., vol. 71, pp. 142–149, 2015. compensation for E-Nose systems using domain adaptation and extreme
[107] D. Brüderle, E. Müller, A. P. Davison, E. Muller, J. Schemmel, and K. learning machine,” Sensors, vol. 18, no. 3, pp. 742, Mar. 2018.
Meier, “Establishing a novel modeling tool: A python-based interface [131] J. H. Lee, T. Delbruck, and M. Pfeiffer, “Training deep spiking neural
for a neuromorphic hardware system,” Front. Neuroinform., vol. 3, p. 17, networks using backpropagation,” Front Neurosci., vol. 10, pp. 508, 2016.
2009. [132] S. B. Shrestha and G. Orchard, “SLAYER: Spike layer error reassignment
[108] Y.-Q. Jing, Q.-H. Meng, P.-F. Qi, M.-L. Cao, M. Zeng, and S.-G. Ma, in time,” in Proc. 32nd Int. Adv. Neural Inf. Process. Syst., vol. 31, New
“A bioinspired neural network for data processing in an electronic York, Curran Associates, 2018, pp. 1419–1428.
nose,” IEEE Trans. Instrum. Meas., vol. 65, no. 10, pp. 2369–2380, [133] P. Arena, M. Calí, L. Patané, A. Portera, and A. G. Spinosa, “A CNN-
2016. based neuromorphic model for classification and decision control,” Non-
[109] Y.-J. Liu, Q.-H. Meng, P.-F. Qi, B. Sun, and X.-S. Zhu, “Using spike-based linear Dyn., vol. 95, no. 3, pp. 1999–2017, 2018.
bio-inspired olfactory model for data processing in electronic noses,” [134] K. G. Fahlbusch et al., in Ullmann’s Encyclopedia of Industrial Chem-
IEEE Sensors J., vol. 18, no. 2, pp. 692–702, Jan. 2018. istry, Weinheim,Germany, Wiley-VCH Verlag GmbH and Co.: 2003,
[110] F. Zenke and S. Ganguli, “Superspike: Supervised learning in multi- ISBN 9783527306732.
layer spiking neural networks,” Neural Computation, vol. 30, no. 6, [135] J. Wang, C. Tang, and H. Chen, “An adaptable continuous restricted boltz-
pp. 1514–1541, 2018. mann machine in VLSI for fusing the sensory data of an electronic nose,”
[111] P. U. Diehl, D. Neil, J. Binas, M. Cook, S. -C. Liu, and M. Pfeiffer, in IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 4, pp. 961–974,
“Fast-classifying, high-accuracy spiking deep networks through weight Apr. 2017.
and threshold balancing,” in Proc. Int. Joint Conf. Neural Netw., 2015, [136] L. Deng et al., “Rethinking the performance comparison between SNNS
pp. 1–8. and ANNS,” Neural Netw., vol. 121, pp. 294–307, 2020.
[112] A. Montero, R. Huerta, and F. B. Rodriguez, “Stimulus space complexity [137] L. Liang et al., “Exploring adversarial attack in spiking neural networks
determines the ratio of specialist and generalist neurons during pattern with spike-compatible gradient,” IEEE Trans. Neural Netw. Learn. Syst.,
recognition,” J. Franklin Inst., vol. 355, no. 5, pp. 2951–2977, 2018. to be published, doi: 10.1109/TNNLS.2021.3106961.
184 IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 16, NO. 2, APRIL 2022

[138] A. Taherkhani, A. Belatreche, Y. Li, G. Cosma, L. P. Maguire, and T. Dexuan Huo received the B.S. degree in microelec-
M. McGinnity, “A review of learning in biologically plausible spiking tronics science and engineering from Jilin Univer-
neural networks,” Neural Netw., vol. 122, pp. 253–272, 2020. sity, Changchun, China, in 2017, the M.S. degree
in integrated circuit engineering from the Univer-
sity of Electronic Science and Technology of China,
Chengdu, China, in 2020. He is currently working
toward the Ph.D. degree in electronic science and
technology with Tsinghua University, Beijing, China.
His Ph.D. project, is focused on the design of digital
Hong Chen (Senior Member, IEEE) received the asynchronous spiking neural network accelerator for
Ph.D. degree from the Department of Electronic En-
gas recognition. His research interests include spiking
gineering, Tsinghua University, Beijing, China, in
neural network, machine olfaction, and digital asynchronous circuit design.
2005. From 2005 to 2007, she was with the Institute
of Microelectronics of Tsinghua University (IMETU)
as a Postdoctoral Fellow. Since 2007, she has been
with IMETU. She is currently an Associate Profes-
sor. In 2006 and 2016, she was a Visiting Scholar
with Medical Center, Nebraska University, Lincoln, Jilin Zhang (Member, IEEE) received the B.S. de-
NE, USA, and the Department of Electronics and gree in microelectronic engineering from Lanzhou
Computer Engineering, Georgia Tech, Atlanta, GA, University, Lanzhou, China, in 2019. He is currently
USA, respectively. She has authored or coauthored 110 journal and conference working toward the master degree in microelectronics
papers and three books and chapters. She also holds 25 granted Chinese and with Tsinghua University, Beijing, China. His re-
American patents. Her research interests include monitoring-system design for search interests include digital asynchronous circuit
TKR/THR surgery, low-power digital integrated-circuit design, asynchronous design and spiking neural network.
circuit design, asynchronous SNN accelerators, and low-power mixed-signal
SoC design. She has been a Technologies Program Committee Member of
ASYNC and ISCAS since 2014 and 2018, respectively. She was the recipient of
the Best Demo Award at ISCAS2013 and Best Paper Nominee at ASYNC2021.