Source-linked AI summary
Online Decorrelation of Humidity and Temperature in Chemical Sensors for Continuous Monitoring
Ramon Huerta, Thiago S. Mosqueiro, Jordi Fonollosa, Nikolai F Rulkov, Irene Rodriguez-Lujan
TL;DR
Humidity and temperature variations distort chemical-sensor readings, complicating continuous monitoring and gas discrimination. The paper models and subtracts these effects from MOX signals, reporting accurate sensor-response prediction and improved discrimination with filtered signals. The authors present the approach as suitable for online, continuous home monitoring, while noting unresolved highly ventilated scenarios.
Problem
Humidity cross-sensitivity and changing environmental conditions complicate volatile identification, while gas-dependent model parameters make correction challenging for complex mixtures.
Source: §1 · pp. 1–2Method
An online n-type semiconductor energy-band model estimates humidity- and temperature-related MOX response changes for subtraction, assuming small environmental changes and unchanged chemical content.
Source: §1 · p. 2, §4 · pp. 5–6Results
R^2 > 0.9 for all MOX sensors; in banana, wine, and baseline discrimination, including filtered data improved discrimination capability and stability.
Source: §1 · p. 2, §7 · p. 13Takeaways & Limitations
Simultaneous humidity and temperature recordings can help extract chemical signatures, and the proposed approximation is presented as an inexpensive option for online continuous home monitoring.
Source: §7 · p. 13The model assumes small humidity and temperature changes with unchanged chemical content; highly ventilated scenarios where environmental and chemical changes occur at similar rates remain for further study.
Source: §4 · p. 6, §5.2 · p. 9
Abstract
from arXiv · showhide
A method for online decorrelation of chemical sensor signals from the effects of environmental humidity and temperature variations is proposed. The goal is to improve the accuracy of electronic nose measurements for continuous monitoring by processing data from simultaneous readings of environmental humidity and temperature. The electronic nose setup built for this study included eight metal-oxide sensors, temperature and humidity sensors with a wireless communication link to external computer. This wireless electronic nose was used to monitor air for two years in the residence of one of the authors and it collected data continuously during 537 days with a sampling rate of 1 samples per second. To estimate the effects of variations in air humidity and temperature on the chemical sensors signals, we used a standard energy band model for an n-type metal-oxide (MOX) gas sensor. The main assumption of the model is that variations in sensor conductivity can be expressed as a nonlinear function of changes in the semiconductor energy bands in the presence of external humidity and temperature variations. Fitting this model to the collected data, we confirmed that the most statistically significant factors are humidity changes and correlated changes of temperature and humidity. This simple model achieves excellent accuracy with a coefficient of determination $R^2$ close to 1. To show how the humidity-temperature correction model works for gas discrimination, we constructed a model for online discrimination among banana, wine and baseline response. This shows that pattern recognition algorithms improve performance and reliability by including the filtered signal of the chemical sensors.
1 Introduction
Environmental humidity and temperature confound chemical-sensor measurements, while existing correction approaches face trade-offs in explanatory basis and applicability to mixed gases. This work proposes an online energy-band model to remove environmental effects and improve chemical-signature discrimination.
1 Introduction
- Humidity cross-sensitivity challenges identification and quantification of volatile compounds in uncontrolled monitoring settings.The introduction notes that humidity may itself drive a human-monitoring signal, complicating interpretation of chemical signatures.Source: §1 · pp. 1–2
- Preconditioning chambers control humidity but increase electronic-nose design costs for continuous environmental monitoring.Source: §1 · p. 2
- Power-law and fundamental models work under restricted conditions, while neural-network corrections lack an explanation of underlying processes and gas-dependent parameters complicate fundamental-model correction of mixtures.The prevailing power-law model is accurate when the gas is known and conditions are controlled, but becomes inaccurate as the environment changes.Source: §1 · p. 2
- The online method subtracts humidity- and temperature-driven changes from MOX responses, predicted all eight sensors with R2 > 0.9 on 537 days of home data, and significantly enhanced discrimination when decorrelated data were combined with raw time series.The discrimination evaluation used two distinct stimuli: wine and banana.Source: §1 · p. 2
2 Example of sensors correlation with humidity and temperature
A day-long home recording illustrates that environmental humidity and temperature vary alongside MOX sensor resistance. The authors use this example to motivate separating environmental effects from chemical variation.
2 Example of sensors correlation with humidity and temperature
- The plotted record comes from a regular working day in October 2014 at an author's residence, with humidity changes associated with household events.The passage identifies a morning shower, opening a bathroom window, and the family returning home as examples.Source: §2 · p. 3
- Environmental humidity and temperature changes are often correlated in the illustrated recording.Source: §2 · p. 3
- MOX resistance changes are strongly affected by humidity and temperature, although some sensor changes also reflect chemical variation.The example plots humidity, external temperature, and resistance from four sensors; the other four sensors are described as highly correlated with those shown.Source: Figure 1 · p. 3, §2 · p. 4
3 Design of the wireless electronic nose
The wireless electronic nose combines an eight-sensor MOX array with onboard processing and wireless communication. Its adjustable heaters and continuous power support long-term home monitoring.
3 Design of the wireless electronic nose
- The sensor array contains eight MOX gas sensors with six different sensitive surfaces, and each sensor's operating temperature can be adjusted independently.The board also integrates Sensirion SHT75 humidity and temperature sensors and is powered from an electrical outlet for continuous recording.Source: §3 · p. 4
- An MSP430F247 microcontroller collects sensor data at 100 Hz, controls heater temperature using PWM, and communicates bidirectionally with another device.Collection uses a 12-bit ADC; the board also provides serial communication via USB or a four-pin connector.Source: §3 · p. 4
- A separate wireless communication module links the microcontroller to the network through a WiFly radio module.The module supports wireless network communication and can accept UDP and TCP connections.Source: §3 · pp. 4–5
4 Online model for sensors response
The section develops a low-power energy-band model for estimating how humidity and external temperature changes affect MOX sensor resistance. It fits a three-parameter approximation under small environmental changes and unchanged chemical content.
4 Online model for sensors response
- The energy-band model expresses the resistance change before and after exposure as a nonlinear function of changes in semiconductor energy bands affected by humidity and external temperature.The model’s transduction is mediated by those energy-band changes.Source: §4 · p. 5
- The approximation assumes small changes in external humidity and temperature and unchanged chemical content during those environmental changes.The unchanged-content assumption matters because humidity can induce chemical-agent-dependent nonlinear energy changes.Source: §4 · p. 6
- Terms up to second order in humidity and temperature changes retain nonlinearities while keeping the model suitable for the microcontroller’s simple operations.The sensor temperature is controlled by its built-in heater and modulated by external temperature; the fitted model has three parameters, β1, β2, and β3, with β3/β1 = −µ/T.Source: §4 · pp. 5–6
- RMS error stayed below 3.0 and R2 stayed above 0.9; all fitted β parameters were significant at p-value below 10−10.The reported fit statistics apply across the sensors.Source: Table 2 · p. 7
5 Results
The humidity–temperature model explained sensor resistance variation with high accuracy, and subtracting its predictions filtered environmental drift. Its performance remained strong under forward testing but declined when sampling intervals exceeded two minutes.
5 Results
- R2 exceeded 90% for every sensor, and all model coefficients were statistically significant; fixed-heater sensors outperformed actively controlled sensors, with a nearly 10% R2 gap in the worst case.The fitted β1 and β3 coefficients had opposite signs across sensors, and β3/β1 ratios reflected their heater conditions.Source: §5 · p. 7
- Subtracting the predicted humidity- and temperature-driven response filtered sensor drift, but the resulting signal often had zero mean and partly lost inter-sensor relationships.Figure 3 illustrates the prediction-and-subtraction filter on sensor 1; the lost relationships matter for gas discrimination.Source: §5 · pp. 7–8, Figure 3 · p. 8
5.1 Parameter Stability
- Three-month training windows were chosen to maintain R2 above 0.9 across the year in forward tests, and the evaluation encompassed wireless and sensor-displacement failures.The accompanying time-series figure tracks sensor 1’s out-of-sample R2 and β parameters.Source: §5.1 · p. 8, Figure 4 · p. 9
5.2 Sampling rate
- Sampling periods beyond two minutes reduced filter R2 below 0.9, where the band-model approximations fail.Faster sampling may be needed with active heater control or rapidly changing environments; highly ventilated scenarios remain for future study.Source: §5.2 · p. 9
6 Impact on online discrimination of gas identity
The study tested whether humidity–temperature filtering could improve online discrimination among banana, wine, and baseline responses. Combining raw and filtered sensor signals significantly improved discrimination performance and consistency.
Data and task
- The approach aimed to predict stimulus presence from multivariate sensor responses rather than depend on particular banana ripeness or wine type.The authors note that banana signatures change with ripeness and wine signatures vary with alcohol content and grape origin.Source: §6 · pp. 9–10
- The experiment classified banana, wine, and baseline responses using 10-minute moving windows from 34 banana presentations, 36 wine presentations, and 30 baseline samples.The recordings spanned two months, with stimulus presentations lasting 10 minutes to one hour.Source: §6 · pp. 9–10
Classifier and evaluation
- The classification used a nonlinear Inhibitory Support Vector Machine and compared raw, humidity/temperature-augmented, filtered, and combined raw-plus-filtered feature sets.The evaluation held out presentations for testing and repeated the train/test procedure 50 times.Source: §6 · pp. 10–11
Results
- Combining raw and filtered signals (RS,FS) significantly improved online discrimination (Kolmogorov-Smirnov, p < 0.025); raw data alone reached 76% accuracy, while adding humidity and temperature did not improve it.Filtered signals alone underperformed raw signals, although their performance variance was lower.Source: §6 · p. 11
7 Conclusions
The energy-band model captures nonlinear humidity- and temperature-related changes in MOX sensor responses, with humidity changes and correlated humidity-temperature variations among the dominant terms. The authors report close-to-100% R2 across 537 days and improved discrimination when filtered sensor data are included.
7 Conclusions
- The model’s most dominant terms were humidity change, its quadratic term, and correlated humidity-temperature variations, and it provided robust corrections to environmental distortions.Source: §7 · pp. 12–13
- R2 was remarkably close to 100% when the model was fitted to all 537 days of sampling.Source: §7 · p. 13
- Reliable humidity-temperature filtering required sampling periods on the order of one minute, while a three-month training window kept R2 above 90% for all sensors throughout the year.Source: §7 · p. 13
- Including filtered sensor data improved the classifier’s discrimination capability and stability for banana, wine, and baseline activity.Source: §7 · p. 13
- The online decorrelation model was designed for simpler, low-energy microcontrollers, supporting cost-efficient continuous home monitoring with simultaneous humidity and temperature recordings.Source: §7 · pp. 13–14