Gaussian process based modeling and experimental design for sensor calibration in drifting environments
Introduction
Chemical sensors have been widely used in indoor and outdoor environment monitoring, vehicle exhaust measurement, human breath detection, etc [1], [2], [3]. It has been long recognized that the responses of chemical sensors, especially chemiresistors, are affected by the drift of environmental factors such as temperature and humidity [4], [5], [6], [7]. To reduce detection errors and false alarm, it is important to accurately calibrate a sensor in a drifting environment, which primarily motivated this work. The environmental factors are denoted as the vector x, and the task of sensor calibration is to establish the functional dependence of the sensor response r upon the analyte concentration c as well as x.
Quantifying the c − x − r relationship is challenging due to two main reasons: First, the variables (c, x) may affect the response r in a nonlinear fashion and also interact nonlinearly with each other. The underlying mechanism is complicated [8], [9], [10], [11] and difficult to be adequately captured by traditional regression analysis [6]. Second, to estimate a calibration model of high dimension, an extremely large sample size is typically required by the classic design of experiments (DOE) [6], [12]. Thus, there is a need to develop new modeling and DOE methods for the efficient calibration of sensors subject to environmental drift.
While focusing on calibrating sensors with environmental drift, this work falls into the research efforts to calibrate sensors with general drifting behaviors, which can be classified into two categories [13], [14]: external (i.e. environmental) and internal drifts. The latter is caused by the physical and/or chemical changes of the sensor itself, and examples of such changes include re-organization of the sensing materials and irreversible interaction with analytes. When calibrating drifting sensors, most of the literature used a reference-based linear compensation or linear regression to quantify the drifting effects [15], [16], [17], [18]. Recognizing the possible nonlinear nature of sensor drifts, powerful nonlinear models have also been employed, such as neural network [6], [19], kernel ridge regression [20] and nonlinear supporting vector machine [21]. However, in this stream of nonlinear modeling work, no effort was ever made to quantify the uncertainty of the target estimates (e.g., the analyte concentration estimated by the calibration model from an observed sensor response). This is at least partly due to the difficulties in deriving valid statistical inference (i.e., quantifying model uncertainty) based on those models [22], [23]. It is known that statistical inference lays the basis for optimum DOE: Experiments are designed to minimize the uncertainty on the model estimates of interest [24], [25], [26]. Thus, optimum DOE is a research issue that has barely been touched in the nonlinear model-based sensor calibration.
In light of the discussions above, our objective is to develop a statistical procedure, which leads to a calibration model of the highest quality by using the least experimental effort. In this work, the calibration model assumes the form of a Gaussian process (GP), which is highly flexible and able to capture practically any continuous functional relationships. GP is chosen over other powerful nonlinear models because of its statistical inference capability [27], which allows for uncertainty quantification and provides the necessary basis for optimum DOE. For sensor calibration, the inference issues are further complicated by the coexistence of forward modeling and inverse estimation (as will become clearer in Section 2.1), and hence a GP-based bootstrap resampling method is developed in this work. The DOE is performed in a batch sequential manner to circumvent the dilemma that the optimum DOE depends on the true c − x − r relationship, which however, is unknown at the stage of designing experiments [28], [25], [24]. A learning process is allowed in such a sequential procedure: For the design of a new batch of experiments to be performed, all the information derived from the experimental data already collected is utilized to search for the optimum DOE of that new batch; and the DOE optimization seeks to minimize the calibration model uncertainty with a given batch size.
The remainder of the paper is organized as follows: Section 2 presents the formulation of the calibration model, which takes the form of a GP. The GP-based model fitting and statistical inference issues are discussed in Section 3. The batch sequential procedure for sensor calibration is described in Section 4. Section 5 is devoted to an empirical study to evaluate the effectiveness and efficiency of the calibration procedure. A brief summary is given in Section 6.
Section snippets
Calibration model
For a sensor exposed to drifting environments, its calibration model needs to functionally relate the sensor response r to the target analyte concentration c as well as the environmental factors x. For notational convenience, all the exposure-condition factors are denoted as the vector of d dimension, with d being a positive integer. The sensor response can be generally written as
where F(w) quantifies the expected sensor response E[r(w)] as a function of w.
Experimental data for sensor calibration
To calibrate a sensor, experimental data has to be collected at a range of exposure conditions. The calibration sample data can be represented as
In (7), wi denotes the ith design point (exposure condition at which experiments are performed) out of a total of I distinct design points; rj(wi) denotes the observed response from the jth replication at wi, and n is the number of replications at each design point. The sample average at wi can then be calculated as
The batch sequential procedure
There are two questions that a general DOE typically addresses: (i) At what design points should samples be allocated? (ii) At each design point, how many replications should be assigned? For the physical and/or chemical experiments involved in sensor calibration, it is a common practice to have a predetermined and fixed number of replications assigned to a design point, each time that exposure condition is selected; this is to ensure the reliability of the response measurements which are
Empirical results
The GP-based procedure was applied to calibrate a chemiresistor sensor, whose response can be substantially drifted by the fluctuations in humidity and temperature. The effectiveness of the calibration procedure is illustrated, and its efficiency over the traditional DOE method is demonstrated.
Summary
To efficiently calibrate sensors in drifting environments, a Gaussian process (GP)-based batch sequential procedure is developed. A GP form is employed for the calibration model, and used to capture the possibly comprehensive and nonlinear relationship between the sensor's response and its exposure condition including the target analyte concentrations as well as other environmental factors. Based on GP modeling, a bootstrap resampling method is developed to quantify the uncertainty of the
Acknowledgments
Research reported in this publication was partially supported by the National Science Foundation under Award Number CMMI-1068131 and by the National Institute Of Neurological Disorders And Stroke of the National Institutes of Health (NIH) under Award Number R15NS087515. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the NSF or NIH. We thank Drs. Hossein-Babaei and Ghafarinia for providing
Zongyu Geng is a Ph.D candidate in the Industrial and Management Systems Engineering Department at West Virginia University. His research work has been focused on statistics and chemometrics.
References (55)
- et al.
Human activity monitoring using gas sensor arrays
Sens. Actuators B: Chem.
(2014) - et al.
Optimum design of sensor arrays via simulation-based multivariate calibration
Sens. Actuators B: Chem.
(2011) - et al.
Combustion control and sensors: a review
Prog. Energy Combust. Sci.
(2002) - et al.
Thick-film gas sensors
Sens. Actuators B
(1995) - et al.
Metal oxide sensors
Sens. Actuators B
(1996) - et al.
Compensation for the drift-like terms caused by environmental fluctuations in the responses of chemoresistive gas sensors
Sens. Actuators B
(2010) - et al.
Interactions of tin oxide with O2, H2O and H2
Surf. Sci.
(1979) Managing dynamic thermal exchanges in commercial semiconducting gas sensors
Sens. Actuators B
(1996)- et al.
A calibration method for handling the temporal drift of solid state gas-sensors
Anal. Chim. Acta
(2000) - et al.
Long term stability of metal oxide-based gas sensors for e-nose environmental applications: an overview
Sens. Actuators B
(2010)
Drift compensation of gas sensor array data by orthogonal signal correction
Chemomet. Intell. Lab. Syst.
Drift compensation of gas sensor array data by common principal component analysis
Sens. Actuators B: Chem.
Aging fingerprint characterization of beer using electronic nose
Sens. Actuators B: Chem.
TiO2-based sensor arrays modeled with nonlinear regression analysis for simultaneously determining CO and O2 concentrations at high temperatures
Sens. Actuators B
Chemical gas sensor drift compensation using classifier ensembles
Sens. Actuators B: Chem.
Model selection in neural networks: some difficulties
Eur. J. Oper. Res.
High temperature sensor array for simultaneous determination of O2, CO, and CO2 with kernel ridge regression data analysis
Sens. Actuators B
Modeling carbon black/polymer composite sensors
Sensors and Actuators B
A bootstrapping-based statistical procedure for multivariate calibration of sensor arrays
Sens. Actuators B
Batch sequential designs for computer experiments
J. Stat. Plan. Infer.
Customized sequential designs for random simulation experiments: Kriging metamodeling and bootstrapping
Eur. J. Oper. Res.
Design of experiments and data-fitting techniques applied to calibration of high-frequency electromagnetic field probes
Measurement
On the calibration of sensor arrays for pattern recognition using the minimal number of experiments
Chemomet. Intell. Lab. Syst.
Minimax and maximin distance designs
J. Stat. Plan. Infer.
Demonstration of fast and accurate discrimination and quantification of chemically similar species utilizing a single cross-selective chemiresistor
Anal. Chem.
The effects of CO water vapor and surface temperature on the conductivity of a SnO2 gas sensor
J. Electron. Mater.
Understanding the fundamental principles of metal oxide based gas sensors; the example of CO sensing with SnO2 sensors in the presence of humidity
J. Phys.: Condens. Matter
Cited by (25)
Current issues and perspectives in nanosensors-based artificial olfactory systems for breath diagnostics and environmental exposure monitoring
2024, TrAC - Trends in Analytical ChemistryRelative humidity control during shiitake mushroom (Lentinus edodes) hot air drying based on appearance quality
2022, Journal of Food EngineeringCitation Excerpt :When different processes are similar, we can model all processes simultaneously and take the advantage of the common aspects to improve predictive performance (Luo et al., 2018). The GPR model has been applied to many engineering applications such as photovoltaic power forecasting (Geng et al., 2015), atmospheric temperature and humidity prediction (Zhou et al., 2018) and injection molding processes Luo et al. (2018). However, it is seldom used to predict changes of material quality in drying process.
Various uncertainties self-correction method for the supervisory control of a hybrid cooling system in data centers
2021, Journal of Building EngineeringCitation Excerpt :Virtual in-situ self-correction of uncertain parameters refers to the correction of parameter errors/uncertainties by means of models and other available measurement parameters. It has been proposed in different fields due to the excellent performance of reducing the uncertainty of parameters, such as computer science [21,22], chemical [23], optical [24], etc. However, HVAC systems in buildings are not mass produced or well instrumented, but usually unique [25].
In-situ sensor calibration in an operational air-handling unit coupling autoencoder and Bayesian inference
2020, Energy and BuildingsCitation Excerpt :These calibrated datasets can be widely used for (1) monitoring/controlling systems, and (2) training/testing various data-driven applications, thereby improving their performance; in this regard, in-situ calibration potentially offer greater advantages when compared with SFDD. In various fields, in-situ calibration methods have been developed; for example, computer science [19,20], chemistry [21–23], optical engineering [24,25], etc. [26,27]. These methods provide good calibration performance when judged on the basis of dense deployment of sensors, reference sensor readings, known conditions, and use of additional equations; they effectively eliminate unknown parameters for a known (better) calibration environment.
Mixed-effects Gaussian process modeling approach with application in injection molding processes
2018, Journal of Process ControlExtended virtual in-situ calibration method in building systems using Bayesian inference
2017, Automation in ConstructionCitation Excerpt :Several modeling-based methods have been proposed in different research fields to convert these to determined problems, such as an on-line calibration [17], a collaborative calibration [9], a blind calibration [18], and a self-calibration [19]. In chemistry, various calibration methods [20–23] use more than one reference sensor for benchmarks, which can be considered as a known calibration environment. Literature in this regard mainly comes from fields where sensor redundancy, high quality sensors, or known relationships between sensors do exist.
Zongyu Geng is a Ph.D candidate in the Industrial and Management Systems Engineering Department at West Virginia University. His research work has been focused on statistics and chemometrics.
Feng Yang is currently an associate professor in the Industrial and Management Systems Engineering Department at West Virginia University. She received her Ph.D. degree in Industrial Engineering and Management Sciences from Northwestern University in 2006. Her research interests include stochastic simulation and metamodeling, design of experiments, and applied statistics.
Xi Chen is an assistant professor in the Department of Industrial and Systems Engineering at Virginia Tech. Her research interests include stochastic modeling and simulation, applied probability and statistics, computer experiment design and analysis, and simulation optimization.
Nianqiang (Nick) Wu received a Ph.D degree in materials science and engineering in 1997. He worked at Keck Interdisciplinary Surface Science Center in Northwestern University from 2001 to 2005. Currently he is a professor of materials science in Mechanical Engineering and Aerospace Engineering, West Virginia University. His research interest lies in low-dimensional nanomaterials, chemical sensors and biosensors, photocatalysts, photoelectrochemical cells and solar cells.