PaperlayerView exact revision on arXiv

Source-linked AI summary

Humans-as-a-sensor for buildings: Intensive longitudinal indoor comfort models

Prageeth Jayathissa, Matias Quintana, Mahmoud Abdelrahman, Clayton Miller

arXiv:2007.02014v2cs.HCstat.AP

TL;DR

Indoor comfort modelling must contend with many difficult-to-measure influences, while available models and sensors capture only part of that complexity. This paper tests smartwatch micro-EMA for collecting longitudinal occupant preferences and modelling comfort in field settings. The authors report that the resulting models can perform similarly to sensor measurements, while noting that prediction depends on available historical data.

  • Problem

    Comfort modelling involves many physiological, psychological, and environmental variables that are difficult to measure, and excluding factors can cause significant errors.

    Source: §1 · p. 1

  • Method

    The paper uses smartwatch micro-EMA to collect repeated subjective comfort preferences and combines longitudinal feedback with sensor data and preference tendencies for field-based prediction.

    Source: §1.1 · p. 2, §3.1 · p. 5, §4.1 · p. 7

  • Results

    The authors report that models for visual, aural, and thermal comfort can have performance similar to sensor measurements, and that smartwatch collection provided sufficient frequency for time-series models.

    Source: §5 · p. 11, §6 · p. 14

  • Takeaways & Limitations

    The models could supplement objective sensor data and could be used to detect building anomalies or inform subjective-driven building control.

    Source: §5 · p. 11, §6 · p. 14

    Prediction is constrained where historical data are absent or spaces are rarely used, and the study's Singapore setting had no seasons and little temperature variability.

    Source: §5.2 · pp. 13–14

Abstract

from arXiv · show

Evaluating and optimising human comfort within the built environment is challenging due to the large number of physiological, psychological and environmental variables that affect occupant comfort preference. Human perception could be helpful to capture these disparate phenomena and interpreting their impact; the challenge is collecting spatially and temporally diverse subjective feedback in a scalable way. This paper presents a methodology to collect intensive longitudinal subjective feedback of comfort-based preference using micro ecological momentary assessments on a smartwatch platform. An experiment with 30 occupants over two weeks produced 4,378 field-based surveys for thermal, noise, and acoustic preference. The occupants and the spaces in which they left feedback were then clustered according to these preference tendencies. These groups were used to create different feature sets with combinations of environmental and physiological variables, for use in a multi-class classification task. These classification models were trained on a feature set that was developed from time-series attributes, environmental and near-body sensors, heart rate, and the historical preferences of both the individual and the comfort group assigned. The most accurate model had multi-class classification F1 micro scores of 64%, 80% and 86% for thermal, light, and noise preference, respectively. The discussion outlines how these models can enhance comfort preference prediction when supplementing data from installed sensors. The approach presented prompts reflection on how the building analysis community evaluates, controls, and designs indoor environments through balancing the measurement of variables with strategically asking for occupant preferences in an intensive longitudinal way.

1. Introduction

Indoor comfort depends on many physiological, psychological, and environmental factors that are difficult to measure comprehensively. This paper tests whether smartwatch-based, intensive longitudinal preference feedback can complement sensor-based assessment and modelling.

1. Introduction

  • Comfort models face many influential variables that are difficult to measure, and omitting them can cause significant errors.Measurements may also be inaccurate or poorly calibrated, while Figure 1 excludes physical and mental ailments that add further variance.Source: §1 · p. 1
  • Existing thermal comfort models use selected personal and environmental variables, while real-building controls often rely on temperature, illuminance, and noise levels.The cited analysis reports that PMV is accurate 34% of the time.Source: §1 · p. 1

1.1. Can longitudinal human perception feedback supplement sensors?

  • Micro-EMA uses smartwatch prompts to collect momentary, right-here-right-now subjective feedback from one person over several weeks.The paper proposes using accumulated feedback to identify individual comfort tendencies and group people by how they perceive their environments.Source: §1.1 · p. 2
  • Grouped preference tendencies could help predict comfort and characterize spaces, supplementing installed-sensor data without requiring additional sensors for prediction.The proposed approach also considers short-term, commissioning, and long-term uses, as well as continuous system control and management.Source: §1.1 · p. 2

1.2. Paper overview

  • The paper presents a smartwatch-based preference collection and environmental measurement system, tests field-based preference models, and discusses integration, limitations, and future work.The overview identifies cozie as the watch-face used for collection.Source: §1.2 · pp. 2–3, Figure 2 · p. 3

2. Background and novelty

Prior indoor comfort approaches combine sensor measurements and occupant feedback, but both measurement and survey practices have limitations. The paper adapts intensive longitudinal sampling to buildings and proposes a practical smartwatch deployment with preference-group features and a cold-start variant.

2.1. Indoor environmental comfort variables and models

  • Sensor approaches face cost, calibration, placement, and spatial-interpolation challenges, while occupant surveys raise questions about what to ask and survey fatigue.Wearable and infrared sensors were reported to increase thermal sensation prediction accuracy by 3–4% in prior work.Source: §2.1 · p. 3

2.2. Ecological momentary assessments (EMA)

  • Ecological momentary assessment samples experience in the natural setting at the moment it occurs, repeatedly capturing temporal dynamics.Smartwatch micro-EMA has been reported as minimally disruptive and capable of eight-fold higher sampling frequency than smartphone use; a cited smartphone study collected only eight feedback points per occupant.Source: §2.2 · pp. 3–4

2.3. Similar work in intensive longitudinal data collection in the built environment

  • Prior built-environment studies used longitudinal sensing for urban experience, thermal comfort, and concentration, while related indoor comfort-feedback work used smartphone interfaces.Examples include thermal comfort data from 14 participants and smartphone feedback for activity-based workspace allocation and a campus sustainability tour.Source: §2.3 · p. 4

2.4. Novelty of proposed approach

  • Cozie is an open-source smartwatch clock-face for high-frequency micro-EMA, designed to make field deployment practical with a single device and brief surveys.The modelling approach converts longitudinal data into group comfort-preference features, tests a cold-start variant for people with limited or no preference history, and examines prediction without environmental sensors when enough occupant data exist.Source: §2.4 · p. 4

3. Methodology

The study collected frequent smartwatch-based comfort preferences, then combined them with location, environmental, physiological, and preference-history data. It clustered occupants and rooms by reported preferences and tested feature combinations in individual and grouped prediction tasks.

3.1. Tier 1: Smartwatch for micro-EMA

  • The smartwatch micro-assessments collected thermal, visual, and aural preferences, with participants responding about 5–15 times per day.The three-point preference scale captured how occupants wanted the environment to change.Source: §3.1 · p. 5

3.2. Tier 2: Indoor localisation

  • For the large SDE4 building, 100 Bluetooth beacons located occupants to one-meter precision and linked their feedback to building zones.Location data was used to geofence occupants and merge feedback in the cloud.Source: §3.2 · pp. 5–6

3.3. Tier 3: Preference data convergence with environmental sensors

  • Forty-five indoor and outdoor sensors measured temperature, humidity, noise, and illuminance, with at least one device in each building zone.Sensor readings were merged with preference feedback for comparison with environmental models.Source: §3.3 · p. 6

3.4. Tier 1b: Strap-mounted sensor kit

  • A strap-mounted temperature sensor logged data locally for transfer to the cloud after the experiment.The sensor was attached to the smartwatch using a custom 3D-printed case.Source: §3.4 · p. 6

3.5. Occupant and room preference clustering

  • Occupants and rooms were clustered by ratios of preference votes, with eight final clusters after merging groups because no one preferred louder conditions.The preference-history clustering used k-means with Euclidean distance and did not incorporate environmental or physiological context.Source: §3.5 · p. 6

3.6. Occupant comfort preference prediction

  • Individual models used each occupant’s data, while grouped models combined occupants’ training data; both used a time-ordered 60:40 train-test split.The time-based split used earlier votes for training and later votes for testing.Source: §3.6 · p. 6
  • Six feature-set combinations tested how additional data influenced prediction, using a random forest with fixed hyperparameters.The model used 1,000 trees, Gini splitting, and two minimum samples per split.Source: §3.6 · p. 6
  • Prediction features represented time cycles, heart rate, room and occupant preference histories, environmental measurements, and near-body temperature.Environmental readings were selected for spatial and temporal proximity to feedback.Source: §3.6 · pp. 6–7

4. Results

Grouped comfort models outperformed individual models across all three preference categories, while occupant and room clustering exposed distinct preference tendencies. The analysis also found substantial overlap in sensor measurements by preference and showed that peer data could support cold-start predictions.

4. Results

  • The two-week study collected 4,378 comfort-preference votes from 30 participants, with 1,474 localized to building environmental sensors for analysis.Figure 4 presents feedback sequentially by occupant across thermal, lighting, and acoustic preference categories.Source: §4 · p. 7, Figure 4 · p. 7

4.1. Grouping comfort preference tendencies

  • Occupant clustering captured differing preference tendencies and supplied the groups used in individual-versus-grouped prediction models.The segmentation was independent of environmental parameters; even among 30 occupants, comfort tendencies varied.Source: §4.1 · pp. 7, 9

4.2. Tagging the spatial context with preference feedback

  • Room-level clustering characterized which preferences occupants tended to report in different zones, providing facility managers an overview of their spaces.The room rows and cells represent the percentage of preference votes for each room.Source: Figure 5 · p. 8, §4.2 · p. 9

4.3. Correlation with indoor environmental quality variables

  • Sensor measurements overlapped substantially across preference classes, consistent with unmeasured variables and differing comfort tendencies limiting their descriptive power.Prefer-brighter votes were absent above 250 lux; near-body temperature and noise showed the most distinct differentiation.Source: §4.3 · p. 9, Figure 7 · p. 10

4.4. Predicting field-based indoor preference using intensive longitudinal data

  • Thermal prediction varied more across feature sets, while visual and aural models showed only small F1-score differences.The best thermal model used physiological, room, and preference-history features without environmental sensor data; aural preference had the highest F1 score and was binary because no prefer-louder responses were recorded.Source: §4.4 · pp. 9–10
  • Grouped comfort models performed better than individual models for all three preference categories.Participants with similar comfort preferences were clustered together, increasing the training data for each occupant type.Source: §4.4 · p. 10

4.5. Cold-start comfort preference prediction

  • Nine users for thermal and five for visual preference prediction matched the accuracy of a smartwatch-wearing user on average.This cold-start analysis used crowdsourced data from peers with similar comfort preferences.Source: §4.5 · pp. 10–11

4.6. Predicting continuous comfort preference without sensors

  • Time-based predictions showed different preference patterns across zones and times, but predictions during periods without data were often inaccurate.Square visual and aural peaks in the office from 22:00 to 07:00 reflected an absence of data for those hours.Source: §4.6 · p. 11

5. Discussion

The discussion presents intensive longitudinal preference feedback as a practical complement to building sensors, with potential uses in evaluation, space recommendations, and occupant-centered controls. Its predictions depend on historical data being available for the spaces and times being assessed.

5.1. Practical Application of Intensive Longitudinal Data in Industry

  • Preference time series could support occupant-centric controls that optimise for human comfort and energy, and reinterpret building conditions in terms of occupants’ recorded preferences.The authors describe using preference feedback to enrich reinforcement-learning control interactions by identifying motivations behind occupant control actions.Source: §5.1 · p. 11, §5.1.3 · p. 13

5.1.1. Post-occupancy evaluation, commissioning, and sensor calibration

  • For post-occupancy evaluation, two to four weeks of smartwatch feedback from non-transient occupants could supplement installed sensors and help identify comfort-related blind spots.The authors recommend automated indoor localisation to place feedback in spatial context without user intervention.Source: §5.1.1 · p. 11

5.1.2. Potential for spatial recommendation systems and impact on activity-based workspace design

  • In activity-based workspaces, matching occupant comfort tendencies to building zones could inform space recommendations and floor-plan decisions.The authors give the example of increasing cooling-oriented spaces if those zones are popular and at capacity.Source: §5.1.2 · pp. 11, 13
  • Excluding environmental sensor data produced the highest thermal-model F1 score, while visual and aural models showed minimal differences across feature sets.Figure 8 also compares grouped and individual models and the accuracy as more participants are added to training.Source: Figure 8 · p. 12

5.2. Prediction models are only as good as the training data

  • Time-series predictions can flag comfort anomalies, but missing historical data at particular hours or in rarely used spaces can make predictions inaccurate.Figure 9 shows errors during periods without data, including square-shaped office peaks.Source: Figure 9 · p. 13, §5.2 · p. 13

5.1.2. Potential for spatial recommendation systems and impact on activity-based workspace design

  • Six occupants, on average, sufficed to crowdsource predictions for the remaining 24 in this experiment, but greater preference variation would require more participants.The authors caution that this value is not generalisable across buildings.Source: §5.1.2 · p. 13

5.2. Prediction models are only as good as the training data

  • The study’s Singapore setting had minimal temperature variability; seasonal settings may require day-of-year features and up to a year of training data.The authors suggest combining sensor-based space characterisation with ongoing crowdsourced preference feedback as a direction for further work.Source: §5.2 · p. 14

6. Conclusions

The conclusion presents smartwatch-based micro-EMA as a way to collect intensive longitudinal subjective comfort data that can supplement environmental sensors and empirical comfort models. The resulting models may support anomaly detection, preference-driven controls, and recommendations for spaces aligned with occupant tendencies.

6. Conclusions

  • Smartwatch micro-EMA generated sufficiently large intensive longitudinal comfort-preference data to supplement objective environmental sensors and empirical comfort models.The implementation involved 30 occupants and revealed segmentation and variation in indoor comfort tendencies.Source: §6 · p. 14
  • The smartwatch sampling frequency enabled time-series models of indoor spaces that could detect building anomalies.Source: §6 · p. 14
  • The models could inform subjective-data-driven building control or recommend spaces that better match each occupant’s comfort tendencies.The conclusion identifies smartwatch feedback and occupant localisation as the preferred technological setup.Source: §6 · p. 14

Author contribution statement

The contribution statement assigns project, technical, analytical, and writing responsibilities across four contributors.

Author contribution statement

Funding

The research received support from Singapore’s Ministry of Education and the National University of Singapore.

Funding

  • The Singapore Ministry of Education and the National University of Singapore supported the research’s development and implementation.Source: Funding · p. 14
Loading 2007.02014v2…