Pyramid Scattering Convolutional Network for Odor Classification With Small Sample Size

Deep learning models can autonomously learn intricate feature representations directly from raw data, thereby alleviating the burden of feature engineering and achieving notable performance enhancements across various tasks. However, due to the limited sample size of gas sensor array datasets, deep learning models are susceptible to issues such as feature learning and overfitting in odor classification tasks. This article proposed a pyramid scattering convolutional network (PSCN) framework based on wavelet scattering transform (WST) to address the above issues. PSCN first constructs time-series pyramids by downsampling the original sensor response signals, and then employs a wavelet filter bank that does not require numerous samples to learn to extract multiscale features. Subsequently, PSCN trains subnetworks for classification based on multiscale features, simultaneously utilizes the extreme learning machine (ELM) algorithm to optimize subnetwork parameters, and then uses the logistic regression model to fuse the outputs of all subnetworks. Finally, the efficacy of PSCN is verified on gas sensor array datasets with small sample size (SSS), accompanied by comparative experiments with traditional methods and ablation experiments. The experimental results demonstrate that PSCN extracts more effective features and exhibits superior classification accuracy compared to traditional methods.