Gas Sensor Drift Compensation Using Semi-Supervised Ensemble Classifiers with Multi-Level Features and Center Loss.
Journal:
ACS sensors
PMID:
40198237
Abstract
The drift compensation of gas sensors is a significant and challenging issue in the field of electronic noses (E-nose). Compensating sensor drift has a great benefit in improving the performance of E-nose systems. However, conventional methods often perform poorly due to complex data relationships before and after drifting, or require label information for both nondrift (source data) and drift data (target data) to enhance performance, which is hard to achieve and even unrealistic. In this study, we propose a semisupervised domain adaptive convolutional neural network (CNN) based on ensemble classifiers of multilevel features, pretraining, and center loss to tackle the drift problem. The main idea is to make full use of multilevel features extracted from the network and apply Hilbert space's maximum mean discrepancy (MMD) to evaluate the domain similarity of the features at different levels. Then the corresponding MMD is used as a weight to achieve the weighted fusion of predictions in the classifier ensemble module, so as to obtain a more reliable result. Furthermore, to optimize training, MMD is used as a loss for pretraining to help feature extractors learn more robust and common features in two domains. Center loss is also applied to achieve more focused learning for features of the same class. The results on two data sets demonstrate the effectiveness of our method. The average classification accuracies under different settings reach 76.06% (long-drift) and 82.07% (short-drift), respectively, and the average score reaches 0.804 in the regression task, which has significant improvements compared with several conventional methods. Our work provides an effective and reliable method at the algorithm level to solve the drift compensation problem of gas sensors.