Recently, the emergence of low-cost sensors have allowed electronic noses to be considered for densifying the actual air pollution monitoring networks in urban areas. Electronic noses are affected by changes in environmental conditions and sensor drifts over time. Therefore, they need to be calibrated periodically and also individually because the characteristics of identical sensors are slightly different. For these reasons, the calibration process has become very expensive and time consuming. To cope with these drawbacks, calibration transfer between systems constitutes a satisfactory alternative. Among them, direct standardization shows good efficiency for calibration transfer. In this paper, we propose to improve this method by using kernel SPXY (sample set partitioning based on joint x-y distances) for data selection and support vector machine regression to match between electronic noses. The calibration transfer approach introduced in this paper was tested using two identical electronic noses dedicated to monitoring nitrogen dioxide. Experimental results show that our method gave the highest efficiency compared to classical direct standardization.

This paper deals with the prediction of wind turbines power output and proposes an approach to building a prediction model using the Artificial Neural Networks (ANN). The wind speed and output power measured on the site of Sendou, in Senegal, were used to identify the structure of the ANN. Spatiotemporal data on the climatic variables (wind speed, solar radiation, temperature, humidity, wind direction) collected on the same site were used to train the ANN. Data collected on three other sites (Goback, Keur Abdoul Ndoye and Sine Moussa Abdou), located on the northwest coast of Senegal, were used to validate the model and to analyze the influence of the spatial climatic variables on the performance of the model. Results showed the interest of considering climatic variables (wind speed, wind direction, solar radiation, temperature and humidity) as inputs to the ANN for wind turbines output power prediction. Further, this study showed that the prediction of the produced power depends strongly on the characteristics of the sites and the direction of the wind.

This paper investigates the long term drift phenomenon affecting electrochemical sensors used in real environmental conditions to monitor the nitrogen dioxide concentration [NO₂]. Electrochemical sensors are low-cost gas sensors able to detect pollutant gas at part per billion level and may be employed to enhance the air quality monitoring networks. However, they suffer from many forms of drift caused by climatic parameter variations, interfering gases and aging. Therefore, they require frequent, expensive and time-consuming calibrations, which constitute the main obstacle to the exploitation of these kinds of sensors. This paper proposes an empirical, linear and unsupervised drift correction model, allowing to extend the time between two successive full calibrations. First, a calibration model is established based on multiple linear regression. The influence of the air temperature and humidity is considered. Then, a correction model is proposed to solve the drift related to age issue. The slope and the intercept of the correction model compensate the change over time of the sensors’ sensitivity and baseline, respectively. The parameters of the correction model are identified using particle swarm optimization (PSO). Data considered in this work are continuously collected onsite close to a highway crossing Metz City (France) during a period of 6 months (July to December 2018) covering almost all the climatic conditions in this region. Experimental results show that the suggested correction model allows maintaining an adequate [NO₂] estimation accuracy for at least 3 consecutive months without needing any labeled data for the recalibration.

The present paper deals with gas concentration monitoring based on an electronic nose. The proposed approach investigates two regression methods for gas concentration estimation: the first is the most used in gas quantification with electronic nose and known as the partial least squares (PLS) and the second, known as the support vector machine (SVM) regression, is recently used by the electronic nose community. Data used in this work are collected using an E-nose device developed in our laboratory and responding to various concentrations of pine essential oil vapours. The comparison between the two regression methods studied in this paper is related to the accuracy, the universality as well as the number of samples needed for learning. The results are analyzed in order to select the more suitable prediction model for gas concentration estimation.