基于双向气体特征分类网络的电子鼻小米产地识别方法

Electronic Nose-based Millet Origin Identification via Bidirectional Gas Feature Classification Network

摘要:

开展小米产地精准溯源识别,对保护地理标志农产品、规范市场秩序具有重要现实意义,而现有检测方法普遍存在操作繁琐、耗时长、有损样本且识别精度不足的问题。电子鼻凭借快速无损、气味指纹表征能力强的优势,为小米产地识别提供了可靠检测手段。为实现基于电子鼻系统的小米产地识别,本文结合气体数据特点设计了一种双向气体特征分类网络（Bidirectional gas feature classification network, BGFC-Net）。电子鼻系统检测了6个不同产地的小米气体信息。由于电子鼻获得的气体检测数据具有交叉敏感性和时间序列属性,且高维非线性特性使现有数据分析方法难以挖掘影响分类性能的重要特征,提出双向气体特征计算模块（Bidirectional gas feature calculation module, BGFCM）,分别从传感器和时间序列双方向协同计算影响气体分类性能的深层特征,并结合交叉自注意力实现双向特征融合。基于BGFCM构建轻量化的BGFC-Net,用于实现小米产地识别。消融试验验证了BGFCM设计的有效性。在与先进气体分类方法对比结果中,BGFC-Net获得了最优分类准确率为97.40%、最优F1分数为97.59%,且在深度学习方法中,计算复杂度呈明显优势。

Abstract:

Accurate traceability and identification of millet origins are of great practical significance for protecting geographical indication agricultural products and standardizing market order. However, existing detection methods generally suffer from cumbersome operation, time consumption, sample damage and insufficient identification accuracy. With the advantages of rapid non-destructive detection and strong odor fingerprint characterization capability, electronic nose provided a reliable detection method for millet origin identification. To realize millet origin identification, a bidirectional gas feature classification network (BGFC-Net) was designed according to the characteristics of gas data. The gas information of millet from six different origins was detected by an electronic nose system. The gas detection data acquired by electronic nose had cross-sensitivity and time-series attributes, and their high-dimensional nonlinearity made it difficult for existing data analysis methods to mine key features affecting classification performance. Therefore, a bidirectional gas feature calculation module (BGFCM) was proposed, which collaboratively extracted deep features affecting gas information classification performance from both sensor and time-series dimensions, and integrated bidirectional features through cross self-attention. Based on BGFCM, a lightweight BGFC-Net was constructed to identify the origin of millet. Ablation experiments verified the effectiveness of the designed BGFCM. Comparative results with state-of-the-art gas information classification methods showed that BGFC-Net achieved the optimal classification accuracy of 97.40% and the optimal F1-score of 97.59%, and also exhibited obvious superiority in computational complexity compared with state-of-the-art deep learning methods.