Long-term drift compensation algorithms based on the kernel-orthogonal signal correction in electronic nose systems

In order to compensate the drift of the performance of gas sensors in recognition, the kernel-orthogonal signal correction algorithm (K-OSC) is proposed. In the K-OSC, the feature space is mapped into a higher dimension space by using the kernel principal component analysis (KPCA) at first. Then the orthogonal signal correction (OSC) is used to remove undesired components which do not correlate to the label from the feature vector. The feature vector processed by the K-OSC can improve the accuracy of pattern recognition tools, such as support vector machines (SVMs) or deep neural networks. The experimental results of K-OSC demonstrate that the K-OSC has a better performance than other methods considered over a longer interval time between training samples and test samples.