Advanced Data-Driven Approaches for Intelligent Olfaction

1 Introduction

Data science and data analytics play a crucial role in the development of intelligent systems (Drobot, 2020). These disciplines help enhance the performance of conventional systems (Bag et al., 2011) while reducing the overall system design costs (Chaudhri, Rajput, Alsamhi et al, 2022). In today's scenario, the effectiveness of advanced data-driven approaches in developing intelligent systems has been well-proven (Chaudhri, Rajput, Alsamhi et al, 2022; Chaudhri, Rajput, & Mishra, 2022). Additionally, the performance of intelligent systems heavily relies on the presentation of input data. The term “Intelligent Systems” encompasses almost every field of science, engineering, and technology, enabling the creation of smart applications (Ghaffari et al., 2010; Cole, Covington and Gardener, 2011; Gholam Hosseini et al., 2007; Choden et al., 2017). This chapter specifically focuses on artificial olfaction systems (AOS), which rely on electronic gas/odor sensors (Zhang, Wang, Chen et al, 2021). Traditional AOS implementations incorporate conventional pattern recognition techniques with handcrafted feature extraction and selection methods for data processing (Moore et al., 1993). Data processing in such AOS systems relies on statistical and probabilistic methods (Estakhroueiyeh and Rashedi, 2015; Sanaeifar et al., 2014; Sari et al., 2021). However, in artificial neural network (ANN) based AOS systems, advanced data processing utilizes a one-dimensional (1D) feature extraction process (Sari et al., 2021; Mishra, Rajput, and Han, 2018; Ying et al., 2015). In contrast, a 2D feature extraction process provides enhanced and information-rich features that further improve the system's performance (Chaudhri & Rajput, 2022). Hence, advanced data-driven approaches require the transformation of 1D features into 2D features.

The incorporation of several layers of convolutional neural networks (CNN), such as convolutional and pooling layers, involves the use of 2D features, leading to improved performance compared to previous intelligent olfaction systems. CNNs gained popularity due to their remarkable performance on image datasets (LeCun et al., 1998). Images can consist of single-band (e.g., Binary and Grayscale Images) or multi-band (e.g., RGB/False-Color-Composite, Multispectral, Hyperspectral Images) data (LeCun et al., 1998; Chaudhri, Rajput, and Singh, 2020; Ghamisi, Benedicktsson, and Ulfarsson, 2013; Chaudhri et al., 2021). Apart from processing two-dimensional (2D) or three-dimensional (3D) images, one-dimensional (1D) datasets, such as time-series datasets and gas sensor responses, are processed using fully-connected layers-based MLP-ANNs (Mishra & Rajput, 2018). However, ANNs have several disadvantages compared to CNNs, such as allowing only complex fully connected layers and lacking weight-sharing capabilities. This limitation has motivated researchers to harness the power of CNNs for non-imaging datasets. Consequently, the first challenge lies in using CNNs on gas sensor array response datasets. Another challenge involves generalizing the applicability of CNNs independent of the modality of gas sensor array responses, including transient and steady-state responses. If CNNs can be effectively applied to gas sensor responses, they can be utilized to tackle various data-driven problems related to AOS.