Algorithmic mitigation of sensor failure: is sensor replacement really necessary?

Abstract Data-driven models built on metal oxide (MOX) gas sensor arrays are broadly used to detect and discriminate a wide variety of chemical substances. Usually, a system composed of gas sensors and data processing algorithms is initially trained under controlled conditions with the aim to make accurate predictions of the new samples acquired. However, MOX gas sensors undergo two major impairments in the forms of sensor failures and drift that deteriorate the predictions’ capabilities of the previously calibrated models. While a variety of algorithms has been proposed to cope with the sensor drift problem, sensor failures have received much less attention despite their recurrent appearance after a certain time of operation. In this paper we propose a novel methodology based on multiple kernels to increase the robustness of e-nose systems against sensor failures. We built various multi-kernel models using multiple subsets of sensors and analyzed their performance when increasing the number of faulty sensors. Using an 8-sensor array module exposed to six different gases, we show that our proposed multi-kernel approach significantly increases the robustness of the system. In particular, we estimate that the percentage of multi-kernels free of faulty sensors has to be of at least 50% to maintain the performance of the classifier stable. The main conclusion drawn from this analysis is that instead of identifying which specific sensor fails, which is a nontrivial computational task and can be just temporal, the most convenient way would be to design a more robust system that requires minimal human intervention. Since the multi-kernel strategy copes with the sensor failure by itself, we claim that the lifetime of chemo-sensory systems, particularly their constituent chemical sensor devices, can be extended before the replacement of any particular chemical sensor is required.