Sensor failure mitigation based on multiple kernels

Electronic-nose systems are trained to build computational algorithms that provide predictions based on the multivariate response of a chemical detection platform when measuring an unknown sample. However, the inevitable appearance of sensor failures after a certain period of time represents a major impairment that deteriorates the system accuracy based on previous calibrated models. In this work, we present a general formulation based on multiple kernels to maximize the robustness of sensor arrays against sensor failures. The method consists of building an engine containing a combination of subsets of sensor models (that we call kernels) to maximize the accuracy in the predictions of an electronic nose upon the appearance of faulty sensors. Using a MOX sensor array - one of the most common choices to detect and discriminate chemical analytes in a wide variety of applications - exposed to six pure gaseous substances, we explore the benefits of the proposed methodology to maintain the accuracy of the classifier for a longer period of time. We also determine that the percentage of multi-kernels free of faulty sensors has to be of at least 50 % to keep the robustness of the classifier. The 50% rule of kernels with sensor failures can be considered as a general guide for building calibration algorithms for sensor arrays of any kind.