Stiefel-Attentive Transformer for Refrigerator Odor Detection

The detection and recognition of different odors in refrigerator food odor detection tasks prove to be complicated due to the strong correlation between odor features. Existing electronic nose technology faces challenges in distinguishing and identifying different odors. In this paper, we propose a mixed odor detection model for refrigerators that utilizes the multi-head attention mechanism of the Transformer and constrains it on the Stiefel manifold. Firstly, this constraint can better maintain the orthogonality and stability of attention weights, thus improving the reliability of attention computation. Secondly, for odor detection tasks, the Stiefel manifold can better capture the structural information in vector space, thus better capturing the correlation between odor features, ultimately improving the detection accuracy of the model. The proposed method exhibits superior performance in the refrigerator odor classification task, with an average accuracy of 90.6% and an average F1-score of 0.938. In addition, it demonstrates satisfactory performance in odor concentration prediction (MSE
$=$
0.009 and R2
$=$
0.983). These results indicate that our model outperforms conventional methods, such as decision trees (DT), support vector machines (SVM), multilayer perceptrons (MLP), and convolutional neural networks (CNN).
Note to Practitioners
—The motivation of this study is to address the problem of food spoilage detection in smart refrigerators. This research primarily focuses on identifying volatile organic compounds in mixed food, but it has potential applications in other electronic nose systems. Existing odor detection methods in refrigerators are mainly designed for specific foods, defining freshness levels in no more than three categories: fresh, slightly spoiled, and spoiled. These methods fail to accurately predict odor intensity, thereby hindering the reliable activation of disinfection devices. This paper proposes a novel approach that utilizes orthogonal constraints of the Stiefel manifold to reduce the degrees of freedom in the Transformer attention matrix. This constraint enables the model to better capture long-term odor sequence features. As a result, the proposed method can accurately predict five different freshness levels and continuous odor intensity values. The computational process of the Stiefel-Attentive Transformer is described in this paper, and visual comparisons of Transformer attention with and without the Stiefel manifold constraint are included. Experimental results demonstrate the effectiveness of the proposed method in addressing the issue of odor detection in smart refrigerators. Future research endeavors will concentrate on integrating user feedback into the framework of reinforcement learning, aiming to advance the precision of the electronic nose.