An Integration of LeNet with Regularization Techniques for Electronic Nose in Air Contaminant Classification

Abstract

Accurate and rapid air contaminant classification is crucial for electronic nose (e-nose) systems in air quality monitoring applications. Several one-dimensional convolutional neural networks (1D-CNNs) have recently been proposed for the classification of air contaminants using e-nose systems. However, the lack of cross-model evaluation and the limited computational complexity analysis hinder consistent benchmarking among existing 1D-CNN architectures. Additionally, no recent studies have been conducted on integrating regularization techniques into 1D-CNNs in e-nose. Consequently, the effects of different 1D-CNN architectures, including the impact of integrating regularization techniques, have not been investigated. Thus, this study aims to evaluate three existing 1D-CNN architectures (i.e., LeNet, GasNet, and DenseNet) to propose an improved LeNet with regularization techniques (LeNet-R) for e-nose systems in classifying air contaminants. This study adapted the standard LeNet with three regularization techniques (i.e., batch normalization, dropout, and weight decay) to develop the proposed LeNet-R through a series of manual search experiments. Subsequently, LeNet-R was compared with three existing 1D-CNN models in terms of classification performance and computational complexity using a publicly accessible e-nose dataset. The results show that the proposed LeNet-R outperforms the other 1D-CNN models by achieving the highest average accuracy (i.e., 97.60%) and lowest average loss (i.e., 6.50%). Moreover, LeNet-R exhibited the shortest training time (i.e., 86.54 seconds), the shortest inference time (i.e., 1.91 seconds), the fewest total parameters (i.e., 11,644), and the smallest model size (i.e., 45.48 kB) among all the 1D-CNN models. Compared to the standard LeNet, the proposed LeNet-R improved the average accuracy by 1.35%, reduced total parameters and model size by 11%, shortened training time by 36.6%, and decreased inference time by 6.8%. These findings demonstrate that a simpler 1D-CNN integrated with regularization techniques can outperform more complex 1D-CNN models in classifying air contaminants for an e-nose system. This study is the first to show that integrating three regularization techniques into LeNet can improve accuracy and efficiency for e-nose-based air contaminant classification.

ABSTRAK: Pengelasan bahan pencemar udara yang tepat dan pantas adalah penting bagi sistem hidung elektronik (e-nose) pada aplikasi pemantauan kualiti udara. Kebelakangan ini, beberapa rangkaian neural konvolusi satu dimensi (1D-CNNs) telah dibina bagi tujuan klasifikasi bahan pencemar udara menggunakan sistem e-nose. Walau bagaimanapun, ketiadaan penilaian rentas model serta kekurangan kajian terhadap kerumitan pengiraan telah menyukarkan penanda aras konsisten pada model 1D-CNN sedia ada. Tambahan, tiada kajian terkini mengenai integrasi teknik regularisasi ke atas model 1D-CNN dalam bidang e-nose. Akibatnya, pelbagai senibina 1D-CNN, termasuk impak integrasi teknik regularisasi, belum dapat dikaji dengan sewajarnya. Oleh itu, kajian ini bertujuan menilai tiga senibina 1D-CNN sedia ada (iaitu LeNet, GasNet, dan DenseNet) dengan cadangan penambahbaikan model LeNet berintegrasikan teknik regularisasi (LeNet-R) untuk sistem e-nose dalam pengelasan bahan pencemar udara. Dalam kajian ini, model LeNet sedia ada, diubah suai dengan tiga teknik regularisasi (iaitu normalisasi kelompok, dropout, dan pereputan berat) bagi membangunkan LeNet-R yang dicadangkan melalui siri eksperimen secara carian manual. Seterusnya, LeNet-R dibandingkan dengan tiga model 1D-CNN sedia ada dari segi prestasi pengelasan serta kerumitan pengiraan menggunakan set data e-nose yang boleh diakses secara umum. Dapatan kajian menunjukkan bahawa LeNet-R mengatasi model 1D-CNN lain dengan mencapai ketepatan pengelasan tertinggi (i.e., 97.60%) dan purata ketidaktepatan terendah (i.e., 6.50%). Tambahan, malalui kaedah LeNet-R masa latihan adalah terpantas (i.e., 86.54 saat), masa inferens paling singkat (i.e., 1.91 saat), jumlah parameter paling sedikit (i.e., 11,644), serta saiz model paling kecil (i.e., 45.48 kB) berbanding model 1D-CNN yang lain. Berbanding LeNet biasa, ketepatan klasifikasi bagi LeNet-R meningkat sebanyak 1.35%, mengurangkan jumlah parameter dan saiz model sebanyak 11%, memendekkan masa latihan sebanyak 36.6%, dan menurunkan masa inferens sebanyak 6.8%. Dapatan menunjukkan bahawa model 1D-CNN yang lebih ringkas dengan integrasi bersama teknik regularisasi mampu mengatasi model 1D-CNN yang lebih kompleks dalam pengelasan bahan pencemar udara untuk sistem e-nose. Kajian ini adalah yang pertama menunjukkan bahawa penyepaduan tiga teknik regularisasi ke dalam LeNet dapat meningkatkan ketepatan dan kecekapan bagi pengelasan pencemar udara berasaskan sistem e-nose.