No abstract

The drastically increasing availability of low-cost sensors for environmental monitoring has fostered a large interest in the literature. One particular challenge for such devices is the fast degradation over time of the quality of their data. Therefore, the instruments require frequent calibrations. Traditionally, this operation is carried out on each sensor in dedicated laboratories. This is not economically sustainable for dense networks of low-cost sensors. An alternative that has been investigated is in situ calibration: exploiting the properties of the sensor network, the instruments are calibrated while staying in the field and preferably without any physical intervention. The literature indicates there is wide variety of in situ calibration strategies depending on the type of sensor network deployed. However, there is a lack for a systematic benchmark of calibration algorithms. In this paper, we propose the first framework for the simulation of sensor networks enabling a systematic comparison of in situ calibration strategies with reproducibility, and scalability. We showcase it on a primary test case applied to several calibration strategies for blind and static sensor networks. The performances of calibration are shown to be tightly related to the deployment of the network itself, the parameters of the algorithm and the metrics used to evaluate the results. We study the impact of the main modelling choices and adjustments of parameters in our framework and highlight their influence on the results of the calibration algorithms. We also show how our framework can be used as a tool for the design of a network of low-cost sensors.

No abstract

In recent years, low-cost sensors have raised strong interest for environmental monitoring applications. These instruments often suffer from degraded data quality. Notably, they are prone to drift. It can be mitigated with costly periodic calibrations. To reduce this cost, in situ calibration strategies have emerged, enabling the recalibration of instruments while leaving them in the field. However, they rarely identify which instruments actually need a calibration because of drift, so that in situ calibration may instead degrade performances. Therefore, a novel drift detection algorithm is presented in this work, exploiting the concept of rendez-vous between measuring instruments. Its originality lies mainly in the comparisons of values determining the state of the instruments, for which the quality of the measurement results is taken into account. It defines the concept of compatibility between measurement results. A case study is developed, showing an accuracy of 88% for correct detection of drifting instruments. The results of the diagnosis algorithm are then combined with calibration approaches. Results show a significant improvement of the measurement results. Notably, an increase of 15% of the coefficient of determination of the linear regression between their true values and the measured values is observed with the correction and the error on the slope and on the intercept respectively is reduced by 50% and 60% at least.

Monitoring Activities of Daily Living (ADL) has become a major occupation to respond to the aging population and prevent frailty. To do this, the scientific community is using Machine Learning (ML) techniques to learn the lifestyle habits of people at home. The most-used formalism to represent the behaviour of the inhabitant is the Hidden Markov Model (HMM) or Probabilistic Finite Automata (PFA), where events streams are considered. A common decomposition to design ADL using a mathematical model is Activities–Actions–Events (AAE). In this paper, we propose mathematical criteria to evaluate a priori the performance of these instrumentations for the goals of ADL recognition. We also present a case study to illustrate the use of these criteria.