TDACNN: Target-domain-free domain adaptation convolutional neural network for drift compensation in gas sensors

Sensor drift is a long-existing unpredictable problem that deteriorates the performance of gaseous substance recognition, calling for an antidrift domain adaptation algorithm. However, prerequisite traditional methods to achieve fine results have data from both nondrift distributions (source domain) and (target alignment, which usually unrealistic unachievable in real-life scenarios. To compensate this, this paper, deep learning based on target-domain-free convolutional neural network (TDACNN) proposed. The main concept CNNs extract not only domain-specific features samples but also domain-invariant underlying source target domains. Making full use these various levels embedding can lead comprehensive utilization different characteristics, thus achieving compensation by extracted intermediate between two In TDACNN, flexible multibranch backbone with multiclassifier structure proposed under guidance bionics, utilizes multiple comprehensively without involving during training. A classifier ensemble method maximum mean discrepancy (MMD) evaluate all classifiers jointly credibility pseudolabel. optimize training, additive angular margin softmax loss parameter dynamic adjustment utilized. Experiments datasets settings demonstrate superiority TDACNN compared several state-of-the-art methods.