Intelligent Perception of Multiaroma Types Based on Machine Olfaction

Machine olfactory perception (MOP) is a new type of intelligent perception technology that senses and describes odors. The difference between MOP and machine olfaction is that MOP describes odors by using some semantic labels. One of the challenges in realizing MOP is an appropriate multilabel classifier from the high-dimensional sensor responses. Here, we propose a novel aroma-type (AT) oriented MOP method, which is an MOP method of multiple AT. Wherein, the method designed an intelligent perception of multiple ATs based on a convolutional neural network (IPMAT-CNN), in which an olfactory information-shared mechanism was introduced to enhance the feature extraction of the network. In addition, by analyzing the molecular correlation between multiple ATs, a matrix addition method was designed in IPMAT-CNN to strengthen the information on adjacent ATs. These designs effectively improved the accuracy of model detection. At last, IPMAT-CNN successfully sensed and recognized seven ATs of 38 monomer flavors. The effectiveness of IPMAT-CNN was evaluated by four baseline models [classifier chain based on random forest classifier (CC-RF), label powerset based on random forest classifier (LP-RF), multilabel decision tree (ML-DT), and multilabel
${k}$
-nearest neighbor (ML-kNN)]. The results showed that the IPMAT-CNN had the best performance with a 0.0636 hamming loss, a 93.65% accuracy, and a 0.8149 mean average precision (mAP) value for classifying the seven ATs. It reveals that the IPMAT-CNN had a decent prediction performance and can be used as an effective reference method for the ATs classification.