A Multibranch LSTM-Attention Ensemble Classification Network for Sensor Drift Compensation

Sensor drift has been one of the challenging problems in the development of electronic noses (E-noses), which can seriously affect the recognition accuracy of the sensor. In this article, a multibranch long short-term memory-attention ensemble classification network (MLAEC-Net) is proposed to address this problem. The framework simulates the working principle of the animal olfactory system and uses a multibranch structure to allow each branch to capture different features and attributes of the data. It combines the long short-term memory (LSTM) network and the attention mechanism to deal with long-term dependencies and extract key features. Furthermore, the framework improves the generalization ability and prediction accuracy of the overall model through the integration of classifiers. In addition, for E-nose systems with different numbers of sensors, it is only necessary to simply adjust the number of branches, providing excellent flexibility and generalizability to adapt to different sensor arrays and datasets. We validate the effectiveness of the MLAEC-Net framework by conducting experiments on two publicly available datasets. The experimental results show that the MLAEC-Net framework proposed in this article achieves significant advantages with better performance and robustness compared with other state-of-the-art approaches.