An Online Test-Time Adaptive Drift Compensation Network for Electronic Nose

Sensor drift is a common problem in electronic nose (e-nose) systems. In recent years, research on this problem has also made significant progress, and various drift compensation methods have emerged. Although these methods have achieved impressive results in terms of recognition accuracy, they often fail to fully account for the drift situation in online environments. Therefore, an online test-time adaptive drift compensation network (OTTA-DriftNet) is proposed for e-nose in this article, which can adaptively adjust the prediction model online to address drift phenomena. OTTA-DriftNet consists of two steps: generating source model and online test-time adaptation. In the generating source model phase, the data are pre-processed, and the source model is trained using the distribution-aligned non-drifted source domain data. In the online test-time adaptation phase, the pre-trained source model is first used to initialize the teacher model in the network. Then, based on online test-time adaptation (OTTA), the student model gradually adapts to the changes in drifted data, ultimately achieving efficient online recognition of unlabeled drifted data. In this process, the parameters of the teacher model and the student model (prediction model) learn from each other using the knowledge distillation mechanism and are continuously optimized and updated. Due to the complexity of e-nose data and its temporal nature, attention mechanisms and gated recurrent unit (GRU) networks are used during model construction to enhance the model’s feature processing capability. The experimental results show that the proposed OTTA-DriftNet can effectively address the drift problem in the online situation on a public dataset and outperforms the existing methods in terms of recognition accuracy.