1. Introduction
Machine olfaction is an advanced technology that captures odorous materials and identifies them by distinguishing the differences in response patterns. Usually, electronic noses (e-noses) are used, which consist of an array of gas sensors and intelligent identification algorithms mimicking biological noses, to ‘smell’ and ‘sense’ odors [
1,
2].
Gas sensors typically detect gases by measuring the change in electrical conductivity. Sensitivity, selectivity, response time, and recovery time are the major specifications to evaluate the performance of a gas sensor [
3]. There are different types of gas sensors: catalytic combustion, electrochemical, thermal-conductive, infrared absorption, paramagnetic, solid electrolyte, and metal oxide semiconductor sensors [
3]. In recent years, paper-based sensors, which are a new type of gas sensor fabricated by cellulose paper, have the characteristics of flexibility, tailorability, being low-cost, lightweight, and environmentally friendly [
4]. The response of a gas sensor detecting an odor is a synthetic process since the sensor may be sensitive to a group of different molecules, which is usually called ‘cross-sensitivity’. Cross-sensitivity is a characteristic of gas sensors that arises because of poor selectivity [
5]. It is an issue when measuring the gas concentration using a single gas sensor. However, it can be utilized as a feature to identify odors when an array of gas sensors is used. Response patterns of sensor signals are different from various odors. It is difficult to interpret sensing responses due to the synthetically non-linear sensing process of gas sensors. Most gas sensors are fabricated for detecting industrial gases or volatile organic chemicals (VOCs).
Developments in odor identifications have progressed in recent years and have been applied to specific fields. However, such methods ignore the essence of odors. An odor is usually composed of a group of odorous compounds. We human beings sniff the odorous mixture, discriminate, and identify the odor if people are trained to recognize the odor. We have difficulties describing an unknown odor without prior knowledge. Instead, we describe it by using some semantic words. Accordingly, is there a method to describe the odor space so that odor can be recorded and encoded in some general forms?
It is a challenge to determine the dimensionality of the olfactory perceptual space because there are still a lot of efforts required in the investigation of the mechanism of olfactory perceptions. Physiological studies had identified that the human olfactory system consists of around 400 odorant receptor types [
6]. An odor activates some of these odorant receptor types to generate a specific pattern so that humans can discriminate against it. The number of odorant receptor types sets the upper bound on the dimensionality of the perceptual space. There is no dedicated vocabulary to describe odors in major languages. Instead, words about objects, for example, flowers and animals, or emotions such as pleasantness are applied to describe olfactory perceptions. J. E. Amoore claimed that odors were divided into seven groups which were regarded as primary odors [
7]. Markus Meister suggested that olfactory perceptual space may contain around 20 dimensions or less [
8], and Yaara and Noam reviewed that humans are good at odor detection and discrimination, but are poor at odor identification and naming [
9]. Semantic descriptors profiled from a list of defined verbal words are rated by human sniffers. Up to now, there is no universal list of odor semantic descriptors yet.
Currently, there is no odor space to describe the variety of odors in nature. Some studies revealed a significant relationship between odor molecular structure information and olfactory perceptions [
10]. Functional groups and hydrocarbon structural features were considered to be factors influencing olfactory perceptions. A hypothesis demonstrated that odorants possessing the same functional groups activate the same glomerular modules [
11] which generate similar perceptual patterns so that humans identify them as the same type of odor. Recent studies revealed that 3D structure information of odorous molecules has a more noticeable impact on olfactory perceptions [
12]. Considering the complexity of molecular structure information, the mapping to odor space may be non-linear.
Several studies investigated the map between odor responses and odorous perceptual labels. T. Nakamoto designed an odor sensing system that consisted of a mass spectrum and large-scale neural networks to predict odor perceptual information [
13]. R. Haddad et al. investigated the relationship between odor pleasantness and e-nose sensing responses by modeling a feed-forward back-propagation neural network [
14]. D. Wu et al. designed a convolutional neural network for predicting odor pleasantness [
15]. These models used in predicting odor perceptual descriptors perform decently in some particular datasets. However, machine percepts and describes odors using distributed representation is still a challenge for us.
It is worth establishing some forms of odor space to describe a sufficiently complete group of odors in nature. An odor space should be some form of numerical values with definite dimensionality. The odor space should be a linear space for convenient interpretation because of the non-linear map. Those semantic olfactory descriptors are only some points in the quantization, just as the color “red” is quantified to (255, 0, 0) in RGB color space. The importance of such odor space is a quantization form so that odors can be converted to information for data storage or transmission. The odors can be reproduced by blending some similar odorants to generate the odor.
Machine olfactions have been applied widely to many fields in recent years. Some linear methods such as principal component analysis (PCA), linear discriminant analysis (LDA), support vector machines (SVM), etc., were used in the analysis of odor discrimination [
16]. PCA is an unsupervised method ignoring discriminant information, which is a popular method for dimensionality reduction [
17]. LDA is a supervised method for classification by finding decision surfaces and calculating the signed orthogonal distance of data points. It has been used in the identification of Chinese herbal medicines [
18]. SVM is a kind of regularization in which the aim is to find the maximum margin between classes. K. Brudzewski applied SVM as the classification tool for identifying tobacco [
19]. Classifiers using linear methods can be transferred to convex problems which have the advantages of mathematical interpretability. Non-linear methods such as artificial neural networks (ANN) were also introduced in machine olfactions. In recent years, deep learning methods have been dramatically developing and widely used in various fields such as computer visions, speech processing, automatic driving, etc. They also have been introduced in machine olfaction for odor identification [
15,
20].
There are several advantages that mean that machine olfaction technology applies to many fields. Firstly, it is a non-destructive technique to detect volatiles released from the surface of objects [
21]. Secondly, e-nose is usually portable, which is convenient to detect odors anywhere and anytime [
22]. Thirdly, e-noses have the capacity of extending out olfactory perception scopes since gas sensors are capable of detecting those chemicals which humans are unable to smell and sense [
23]. Furthermore, e-noses can be used in some unpleasant environments [
24,
25].
Linear methods for classifications usually require highly correlated features and high calibration costs, which limit the number of training data [
16]. Non-linear methods have difficulties in interpretation. Nonetheless, non-linear methods, especially deep learning methods, have a higher capacity of odor identification. In this paper, we borrowed the idea from an auto-encoder and proposed a novel deep learning algorithm for odor identification—Odor Labeling Convolutional Encoder–Decoder (OLCE). OLCE consists of an encoder and a decoder, where the encoder output is constrained to odor labels. OLCE has a decoder structure which offers some clues on how the model learns features. In the following paragraphs, we will first describe the experimental setups and the modeling of OLCE. After that, the performance of the model, comparison with other methods, and an overview of decoded response results will be illustrated. Furthermore, the perspective of machine olfaction will be discussed.
4. Discussion
Using an e-nose to identify odor is a process of detecting and discriminating those ingredients that gas sensors are sensitive to. It is different from other measuring instruments such as GC–MS that have the capacity of identifying ingredients of an odor. An E-nose with an array of gas sensors and a suitable identification algorithm mimics human olfaction to identify odors, which can be applied to many fields where fast detections are required because it has advantages of portability, easy-to-design, and low-cost. Hence, a reliable algorithm to discriminate various response patterns is necessary.
OLCE has an elegant and symmetrical structure using a convolutional neural network, which makes it easy to build the model. The experimental results show that OLCE performs decently in odor identification of Chinese herbal medicines according to several performance indexes. It may also suggest that OLCE can be used in other odor identifications. The OLCE encoder encodes sensor responses to odor labels using a convolutional neural network. The OLCE decoder reproduces sensor responses using a convolutional neural network with a symmetrical structure. The reproduced responses on the decoder side reveal some clues on which features OLCE focuses on. The one-hot encoding labels in the representation layer, the intermediate layer, make the classification more robust than categorical encoding because of the mutual exclusivity of the encoding bits.
OLCE is a multi-class classifier that uses one-hot encoding codes to output the identification results. Multi-class classifiers are suitable to be used in the scenario where the identification category is mutually exclusive. The other type is the multi-label classifier that an instance may belong to more than one class. It is interesting to consider that the one-hot encoding labels in the representation layer of OLCE can be replaced by binary encoding labels so that the model can be used as a multi-label classifier.