Research Overview

Research Background: What Problem Did This Study Address?

Individual thermal comfort varies from person to person. Even in the same indoor environment, two occupants may have different thermal sensations and different preferences for cooling or heating. This creates a challenge for smart buildings: if building controls rely only on general comfort standards or room-level temperature settings, they may fail to reflect individual needs.

Recent studies have used wearable devices, IoT sensors, and machine learning to build personalized thermal comfort models. However, collecting enough high-quality individual data remains difficult. Conventional data collection often asks participants to submit thermal comfort surveys at fixed time intervals, such as every 15 or 30 minutes. This can be intrusive, tiring, and inefficient. It may also collect many repetitive data points that do not help the model learn more effectively.

This study addressed the inefficiency of individualized thermal comfort data collection. The researchers proposed a closed-loop, human-in-the-loop system that combines environmental sensing, wearable physiological sensing, smartphone-based surveys, web-server data management, alert-based feedback requests, and online threshold tuning.

UbiBot WS1 was used to collect the occupant’s surrounding indoor environmental data. These data were combined with physiological data from an E4 wristband and thermal preference feedback from a smartphone application. The goal was to determine whether online threshold tuning could reduce the number of required survey data points while maintaining or improving the predictive performance of individualized thermal preference models.

The Specific Role of UbiBot in the Study

In this study, UbiBot WS1 was used as the portable indoor environmental sensing device in the individualized data collection system. The paper does not claim that UbiBot alone predicted thermal preference or proved the closed-loop model. Instead, UbiBot collected real microenvironmental data around each occupant, and these data became one part of the multimodal dataset used for alert activation, feature analysis, threshold tuning, and model training.

The system consisted of three main data collection components: UbiBot WS1 for environmental sensing, an E4 wristband for physiological sensing, and a smartphone application for thermal comfort and thermal preference survey responses. A web server received and stored time-stamped data from the sensors and survey application through wireless communication.

UbiBot WS1 was placed near the participant in their own living environment. The paper’s experimental setup figure shows the sensor placed close to the occupant during daily indoor activity. The two subjects collected data in their own living settings rather than in a controlled climate chamber. They were allowed to adjust air-conditioning setpoints, turn fans on or off, or open windows for natural ventilation. Under these real-life conditions, UbiBot measured the microenvironment around the subject rather than only a fixed room-level condition.

The UbiBot WS1 collected three environmental variables: air temperature, relative humidity, and light level. The paper reports the device accuracy as ±0.3 °C for temperature, ±3% for relative humidity, and ±2% for light level. Its IoT platform was used to support data sharing and synchronization with the research team’s web server.

The UbiBot data were used in several ways.

First, environmental data were used as candidate predictor variables for individualized thermal preference models. Air temperature, relative humidity, and light level were combined with clothing insulation, wrist skin temperature, electrodermal activity, heart rate, and acceleration.

Second, UbiBot data supported alert-based survey triggering. The web server monitored changes in predictor variables and sent survey notifications through the smartphone app when a variable’s variation exceeded its predefined threshold. For environmental variables, the thresholds covered air temperature, relative humidity, and light level.

Third, UbiBot data were involved in the online threshold-tuning process. During Phase II, feature importance analysis identified which variables mattered more for each subject’s model. If an important environmental feature, such as air temperature or humidity, appeared too infrequently in alert activations, its threshold could be lowered to trigger more relevant data collection in the next mini-batch.

Therefore, UbiBot’s role was microenvironmental sensing and data support. It provided real indoor environmental measurements that helped the system decide when to request feedback, how to tune thresholds, and how to train individualized thermal preference models.

Research Methods and Data Collection Approach

The study used a two-phase experimental design with two subjects: one female subject and one male subject. Both participants were generally healthy adults in their early thirties. The experiments were approved by the Institutional Review Board of Pennsylvania State University.

The data collection system included UbiBot WS1, an E4 wristband, a smartphone app, and a web server. UbiBot WS1 measured air temperature, relative humidity, and light level. The E4 wristband measured physiological and motion-related signals, including heart rate, wrist skin electrical conductance, wrist skin temperature, and acceleration. The smartphone app collected subjective feedback on thermal sensation and thermal preference.

Thermal sensation was collected on a five-point scale: Cold, Cool, Neutral, Warm, and Hot. Thermal preference was collected on a three-point scale: Want warmer, No change, and Want cooler. Clothing information was also collected and converted into clothing insulation values based on ASHRAE 55.

In Phase I, the researchers developed baseline individualized models using fixed alert thresholds. Data collection occurred in May and June 2021. The stopping criteria required at least 300 data points per subject, at least 14 days of measurement, and overall prediction accuracy of at least 90%. Phase I collected 302 data points for each subject and lasted about three weeks for Subject A and four weeks for Subject B.

In Phase II, the same subjects participated again after one month. The researchers introduced online threshold tuning and mini-batch modeling. Data were analyzed in batches of 32 points. After each batch, the system checked model performance and feature importance. The threshold values for important features were then adjusted to improve data collection efficiency. Phase II collected 160 data points for each subject and lasted about two weeks for each subject.

For modeling, the researchers tested multiple classification algorithms, including linear discriminant analysis, support vector machine, k-nearest neighbors, random forest methods, classification trees, naïve Bayes, and XGBoost. The final comparison used ranger models, an efficient implementation of random forest. Model performance was evaluated using overall accuracy, cross-validation accuracy, balanced accuracy, and AUC.

Key Research Findings

The baseline Phase I models reached high overall accuracy but required more data. With 302 data points per subject, the ranger model achieved overall accuracy of 94.14% for Subject A and 94.96% for Subject B. However, the cross-validation accuracy was lower: 79.14% for Subject A and 69.04% for Subject B.

The alert activations in Phase I were mainly triggered by electrodermal activity, light level, heart rate, and skin temperature. UbiBot environmental variables, especially light level and air temperature, contributed to the candidate predictor set and alert logic.

In Phase II, the online tuning process reduced the number of required data points. The optimized mini-batch-based models required approximately 96 data points for both subjects to achieve similar overall accuracy to the baseline model. This represented a 12.8% data reduction for Subject A and a 36% data reduction for Subject B when comparing the amount of data needed to reach similar overall accuracy.

When cross-validation accuracy was considered, Phase II also required less data than Phase I. The reductions were approximately 20.0% for Subject A and 27.3% for Subject B.

For the same input data quantity of 160 data points, the optimized Phase II models achieved slightly higher overall accuracy than the Phase I baseline models, with approximately 1% to 2% improvement. More importantly, cross-validation accuracy improved by 5.5% for Subject A and 25% for Subject B.

Feature importance analysis showed that electrodermal activity, skin temperature, and ambient air temperature were among the most important variables for both subjects. Light level also ranked as an important feature, although its relationship with thermal preference differed by subject.

The study concluded that online threshold tuning based on feature importance can reduce user burden, improve data collection efficiency, and support more adaptive individualized thermal preference modeling.

What This Means for Individualized Thermal Comfort Monitoring

This study shows that personalized thermal comfort research can move beyond fixed-interval surveys and passive data logging. Instead of repeatedly asking occupants for feedback at arbitrary times, the system can monitor physiological and environmental changes and request feedback only when meaningful variation occurs.

For smart buildings, this is important because occupant-centered control depends on reliable individual comfort data. A system that requires fewer surveys is less disruptive and more suitable for long-term use. If individual preference models can be updated continuously, building controls may eventually respond more precisely to occupants’ comfort needs.

For indoor environmental monitoring, the study highlights the value of combining environmental and physiological data. UbiBot WS1 measured the occupant’s surrounding air temperature, relative humidity, and light level, while the wearable device captured body responses. This combination gave the model a more complete view of the human-environment interaction.

For healthcare facilities, senior living, and patient rooms, the approach may be especially valuable. The paper notes that such a framework could support long-term monitoring of indoor comfort and potentially alert caregivers when comfort-related conditions require attention.

Application Value of UbiBot Devices

UbiBot WS1 demonstrated value as a portable microenvironment sensing device in this human-in-the-loop thermal comfort study.

First, it provided real indoor environmental data around each occupant. Instead of relying only on thermostat settings or general room temperature, UbiBot measured the microenvironment close to the participant.

Second, it supported multimodal data fusion. UbiBot’s air temperature, humidity, and light data were combined with wearable physiological signals, clothing information, and subjective survey responses.

Third, UbiBot data supported alert-based feedback collection. The web server used variations in environmental variables to help determine when a smartphone survey should be triggered.

Fourth, the environmental data contributed to online threshold tuning. In Phase II, important variables could receive lower thresholds so that the system collected more useful feedback when those features changed.

Fifth, the device’s IoT data-sharing capability allowed synchronization with a web server, making it suitable for automated data collection and long-term monitoring workflows.

Overall, UbiBot’s value in this study was not standalone comfort prediction. Its value was providing reliable, time-stamped indoor microenvironment data that helped connect occupant physiology, subjective thermal preference, online model tuning, and potential smart building control.

Extended Application Scenarios

The sensing approach used in this study can be extended to several related scenarios:

1. Smart building comfort control
   Used to collect occupant-level environmental data for personalized HVAC control.
2. Healthcare facility comfort monitoring
   Used to monitor indoor comfort conditions around patients, residents, or older adults.
3. Office thermal preference modeling
   Used to build individual or group comfort models in open-plan offices and flexible workspaces.
4. Long-term indoor environment monitoring
   Used to record temperature, humidity, and light conditions in occupied indoor spaces.
5. Human-building interaction research
   Used to study how occupants perceive and respond to changing indoor environments.
6. Wearable and environmental data fusion
   Used together with wristbands, smart watches, or biosensors to model comfort, stress, or activity.
7. Energy-saving occupant-centric control
   Used to support HVAC strategies that balance occupant comfort and energy efficiency.
8. Adaptive survey systems
   Used to trigger feedback requests only when meaningful environmental or physiological changes occur.

FAQ

1. Which UbiBot product was used in the study?

The study used the portable UbiBot WS1 sensor.

2. What data did UbiBot collect?

UbiBot WS1 collected indoor air temperature, relative humidity, and light level around the occupant.

3. Where was UbiBot deployed?

It was placed near participants in their own living environments during the experiments, measuring the microenvironment around each subject.

4. What was the sampling frequency?

The paper does not explicitly state the raw UbiBot logging interval. For alert activation, the system compared the average value of one four-minute interval with the average value of the next four-minute interval.

5. How long did the research last?

Phase I was conducted in May and June 2021 and lasted about three weeks for Subject A and four weeks for Subject B. Phase II was conducted one month later and lasted about two weeks for each subject.

6. How were UbiBot data used?

The data were used as environmental predictor variables, as inputs for alert-based survey triggering, and as part of the online threshold-tuning process for individualized thermal preference modeling.

7. Did UbiBot predict thermal comfort by itself?

No. UbiBot collected environmental data. The research team combined UbiBot data with E4 wristband physiological signals, clothing information, and smartphone survey feedback to train prediction models.

8. What other devices were used?

The study used an E4 wristband for physiological sensing, a smartphone application for survey feedback, and a web server for data storage, alert monitoring, and threshold tuning.

9. What was the main result?

The Phase II online tuning framework reduced the required data points by 12.8% for Subject A and 36% for Subject B while achieving similar overall accuracy. With the same data quantity, cross-validation accuracy improved by 5.5% and 25%, respectively.

10. Why is this research important?

It shows how environmental IoT sensing, wearable physiological data, and adaptive feedback collection can make individualized thermal comfort modeling more efficient and less disruptive for occupants.