An Adaptive Deep Learning Method Combined With an Electronic Nose System for Quality Identification of Soybeans Storage Period

With the extension of storage period, the nutritional components of soybeans are lost, and the quality loss is severe, but the appearance difference is not obvious. Low-quality soybeans are often misrepresented as high-quality soybeans. In this work, an adaptive deep learning approach is proposed, integrating with an electronic nose (e-nose) system, to effectively identify the quality of soybeans with different storage periods. First, the PEN3 e-nose system is applied to obtain soybeans gas information under two different storage conditions. Second, a multispace self-attention mechanism (MSM) is proposed to selectively import features influencing the classification performance. A lightweight gas information classification network based on this attention mechanism is designed (MSM-Net). Finally, by conducting ablation experiments and comparing with state-of-the-art gas information classification methods, MSM-Net demonstrates superior classification results. Under storage temperature of 25 degrees C and relative humidity of 75% RH, an accuracy of 98.50%, a precision of 98.54%, and a recall of 98.48% are achieved. Under storage temperature of 25 degrees C and relative humidity of 45% RH, an accuracy of 96.50%, a precision of 96.62%, and a recall of 96.85% are achieved. The findings suggest that the integration of MSM-Net and the e-nose system offers an effective detection method for monitoring the quality of soybeans.