Feature Ensemble Learning for Sensor Array Data Classification Under Low-Concentration Gas

Gas sensor array (GSA) data usually has high-dimensional features and a small sample size. When a classifier is directly used for GSA data classification, it is prone to overfitting and has a high time cost. The traditional solution is to perform feature dimensionality reduction before classification. However, selecting a suitable dimensionality reduction method is time-consuming and laborious, and some features useful for classification may be lost after dimensionality reduction, especially for the weak sensor response data to low-concentration gases. In this article, we proposed a feature ensemble-based extreme learning machine framework (FE-ELM) for GSA data classification. In FE-ELM, downsampling is first performed on the time series of each sensor, and then the downsampled features of different sensors are combined to obtain fused feature subsets. Next, a base ELM is trained independently on each fused feature subset with all training samples by solving the least-squares problem. The final FE-ELM predictions for input samples are obtained by voting the prediction results of all base ELMs. Compared with traditional methods, the proposed method solves the overfitting problem and can be directly used for GSA data classification without prior feature dimension reduction. Furthermore, the ensemble of all base classifiers with little loss of original features enables the proposed FE-ELM to have a more efficient and robust classification performance. Experimental results on data from both homemade GSA under low-concentration gases (ppb) and publicly available confirm that the proposed FE-ELM exceeds traditional methods and extends the detection limit of the sensor array.

Automated summary

In this article, a feature ensemble-based extreme learning machine framework (FE-ELM) is proposed for GSA data classification. The method performs downsampling on the time series of each sensor and combines the downsampled features to obtain fused feature subsets. Each fused feature subset is trained independently using a base ELM, and the final predictions are obtained by voting the results of all base ELMs. Experimental results show that the proposed FE-ELM surpasses traditional methods and extends the detection limit of the sensor array.