VOL. 6, NO. 4, APRIL 2022 7001004 Sensor signal processing Multidimensional Multiconvolution-Based Feature Extraction Approach for Drift Tolerant Robust Classifier for Gases/Odors Shiv Nath Chaudhri ∗ and Navin Singh Rajput ∗∗ Indian Institute of Technology (BHU) Varanasi, Varanasi, Uttar Pradesh 221005, India ∗ Student Member, IEEE ∗∗ Member, IEEE Manuscript received August 17, 2021; revised February 7, 2022; accepted February 20, 2022. Date of publication February 23, 2022; date of current version March 24, 2022. Abstract—Recently, convolutional neural networks (CNNs) have been used for the classification of gases/odors. These methods involve statistical algorithms for drift compensation of sensor characteristics inhibiting its application in real time. In this letter, we have proposed a hybrid CNN model called “drift tolerant robust classifier (DTRC),” which extracts multidimensional features from the raw sensor array responses and automatically compensates for any drift in the sensor response characteristics. The proposed DTRC comprises of 1-D, 2-D, and 3-D convolutional layers in a hybrid manner to compensate for the referred drift without any statistical algorithm. The efficacy of DTRC has been evaluated on a popular dataset and its published results, which comprise of ten batches of sensor characteristics exhibiting drift over a period of three years. Our proposed DTRC outperformed the referred results. In another experiment, DTRC outperformed other state-of-the-art methods. The proposed CNN architecture (DTRC) is a simpler, lightweight CNN with multidimensional multiconvolution end-to-end architecture, suitable for real-time applications. Index Terms—Sensor signal processing, 3-D convolutional neural networks (CNN), drift tolerant robust classifier (DTRC), gases/odors classification, hybrid CNN, sensor signals processing. I. INTRODUCTION An electronic nose (eNose) is an artificially mimicked system comprising of multiple gas sensor elements organized in the form of an array for classifying and quantifying gases/odors using uniquely generated signature patterns. This concept was pioneered in [1] by Persaud et al.. Now, eNose systems are being used in diverse ar- eas/industries as discussed in [2]. Broadly, an eNose captures sensor responses, preprocesses them, and classifies/quantifies the gas/odor by using pattern recognition (PR) techniques, as shown in Fig. 1. Traditionally, the classification of gases/odors is performed using various statistical pattern recognition techniques. An outstanding re- view of these techniques can be seen in [3]. However, statistical methods are more suitable for laboratory-level experiments and an- alytics and their application in real time is difficult. Lately, artificial neural network (ANN)-based PR techniques have been found more promising for gas sensor applications [4], [5]. Recently, nonmetal oxide (MOX)-based unsupervised sensor response cluster analysis has been presented in [6] and [7]. Unsupervised methods are applied on unlabeled data gathered from 16-element MOX-based gas sensor elements with limited success, while supervised methods applied on labeled dataset have been shown to achieve higher performance [8]. Further, ANN classifiers cannot achieve higher accuracy by directly operating on raw sensor responses without requiring additional man- ual intervention. Rajput et al. [4], [5] have proposed analysis space transformation approaches to achieve higher classification accuracies. Recently, convolutional neural networks (CNNs) have become popular to achieve higher performance in gas/odor classification. CNNs are Corresponding author: Navin Singh Rajput (e-mail: nsrajput.ece@iitbhu.ac.in). (Shiv Nath Chaudhri and Navin Singh Rajput contributed equally.) Associate Editor: Rolland Vida. Digital Object Identifier 10.1109/LSENS.2022.3153832 Fig. 1. Schematic block diagram of an electronic nose (eNose). natural extensions of ANNs, which contain multiple feature extraction layers. Feature extraction, selection, scaling, normalization, etc., are vari- ous data preprocessing techniques that contribute to attaining higher accuracies. Mishra et al. [9] have proposed a virtual sensor response generation technique called normalized difference sensor response technique (NDSRT), which makes the raw sensor responses exhibit significant intercluster distances between various classes of gas/odor samples; for more such techniques, see [4], [5], and [9]. CNNs consist of convolutional and pooling layers followed by fully connected multilayer perceptions that provide automatic fea- ture extraction capability to them. In convolutional layers, the highly distinguishable features are extracted using multidimensional kernels (1-D, 2-D, and 3-D). These serve a similar purpose to the traditional data preprocessing methods and virtual sensor response generation. Further, fully connected layers of neurons provide a high degree of learnability and signal processing capacity to the CNNs; thereby, these are best-suited candidates for developing high-performance eNose for real-time applications. Some interesting research on CNNs for eNose has recently been published [10]–[13]. Zhao et al. [10] have proposed a 1-D deep CNN 2475-1472 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See https://www.ieee.org/publications/rights/index.html for more information. Authorized licensed use limited to: Indian Institute Of Technology (Banaras Hindu University) Varanasi. Downloaded on July 17,2022 at 10:38:25 UTC from IEEE Xplore. Restrictions apply.