AirSense: an intelligent home-based sensing system for indoor air quality analytics

Biyi Fang, Qiumin Xu, Taiwoo Park, Mi ZhangUbiComp 2016 · DOI

TL;DR

This paper introduces AirSense, an intelligent home-based system that goes beyond simply measuring indoor air quality (IAQ). It automatically detects pollution events, identifies their sources (like cooking or smoking), forecasts future IAQ levels, and provides actionable advice to users. This aims to empower individuals to better understand and manage their home environment, addressing the common problem of IAQ being overlooked despite its significant health impacts.

What Problem Does This Paper Address?

Indoor air quality (IAQ) is a critical but often neglected aspect of our health. People spend about 90% of their time indoors, and indoor air can be significantly more polluted than outdoor air. This pollution can lead to serious health issues, including respiratory infections, chronic lung diseases, and even cancer. The problem is compounded because many indoor pollutants are colorless and odorless, making them undetectable by humans. Furthermore, the immediate health effects can be subtle and easily mistaken for common illnesses. Existing IAQ monitoring systems primarily focus on measurement and visualization, leaving users feeling powerless due to a lack of information about pollution sources and how to mitigate them. This paper addresses this gap by developing a system that provides comprehensive IAQ analytics, including source identification and actionable recommendations.

Methods

AirSense employs a multi-stage approach for IAQ analysis:

  • Pollution Event Detection: The system uses a normalized standard deviation (NSTD) based scheme. It monitors sensor data (PM2.5, humidity, VOCs) within a sliding window. A sharp increase in sensor readings, indicated by a high NSTD value, signals the beginning of a pollution event.
  • Pollution Source Identification: Once an event is detected, AirSense extracts a set of 18 features from sensor data. These features capture unique patterns in PM2.5, humidity, and VOCs levels, including rates of change, magnitudes, standard deviations, and cross-sensor relationships. A linear kernel-based Support Vector Machine (SVM) is then used to classify the pollution source based on these features.
  • IAQ Forecast: To estimate the seriousness of pollution and personal exposure, AirSense uses a non-parametric regression model. This model predicts future IAQ sensor values by finding the nearest historical trials of similar pollution events and using their patterns to forecast the current event. This approach is designed to handle the high variability within different pollution events.

The system architecture comprises an IAQ sensing platform (with temperature, humidity, PM2.5, and VOC sensors), a cloud server for data storage and analytics, and a smartphone application for user interaction.

Datasets

The paper describes experiments conducted in two main settings:

  • Controlled Settings: Experiments were conducted in two homes for ten weeks, and in three homes for nine weeks in real-world settings.
  • Real-World Deployment Study: This involved three additional families over two phases (six weeks without the app, three weeks with the app). Participants were instructed to perform specific pollution-generating activities (cooking, smoking, spraying pesticide) and to log ground truth data.

The paper does not introduce a new benchmark dataset but rather utilizes data collected from these experimental deployments for evaluation.

Key Insights and Ideas

  • Beyond Measurement: The core insight is that simply measuring IAQ isn't enough. Users need to know what is causing the pollution and how to fix it to feel empowered and motivated to act.
  • Sensor Fusion for Source ID: Combining data from multiple sensors (PM2.5, humidity, VOCs) and analyzing their temporal patterns is crucial for accurately identifying the source of pollution events.
  • Non-Parametric Forecasting for Variability: Recognizing that pollution events have high within-class variance, a non-parametric approach to forecasting is more robust than traditional parametric models.
  • Timely Notifications Drive Behavior: The system's ability to provide rapid detection and actionable suggestions can act as a motivational trigger, encouraging users to adopt healthier IAQ practices.
  • Visualizations Enhance Awareness: Presenting IAQ data through intuitive visualizations (graphs, pie charts) significantly increases user awareness and understanding of their home environment.

Key Results

  • Pollution Event Detection: AirSense accurately detects pollution events with an average accuracy of 95.8% in controlled settings. In real-world deployments, the average true positive and true negative rates were 99.0% and 99.5% across three families.
  • Pollution Source Identification: The system achieves high accuracy in identifying pollution sources, with an average of 94.2% for Family 1 and 97.3% for Family 2 in controlled settings. In real-world settings, average accuracies were 87.0%, 90.7%, and 92.2% across three families. The system can identify sources within five minutes of an event.
  • IAQ Forecast: The IAQ forecast algorithm achieves an average NRMSD of 6.8% for PM2.5 and 8.1% for VOCs in controlled settings, demonstrating accurate prediction of future IAQ levels.
  • User Impact: Real-world deployment showed that AirSense significantly increased users' awareness of IAQ and their perceived competence in managing it, leading to a reduction in polluted air time.

Limitations and Future Work

  • Limited Pollution Sources: Currently, AirSense is designed to detect cooking, smoking, and spraying pesticide. Future work aims to expand this to include more sources like wood stoves or outdoor air pollution.
  • Detection of Mild Pollution: The current event detection algorithm relies on sharp increases in sensor data, which might miss milder pollution sources that cause slow, gradual changes.
  • IAQ Forecast Model: The forecast algorithm currently relies on historical data from similar events. Future improvements could incorporate external factors like weather and outdoor air quality reports.
  • Deployment Location Impact: While the study showed minor impact of deployment location on accuracy and latency, further investigation into optimal placement could be beneficial.
  • Subcategory Performance: While generally accurate, the system showed slightly lower identification accuracies for events involving multiple pollution sources simultaneously.

Summary updated

BibTeX
@inproceedings{fang2016airsense,
  author = {Biyi Fang and Qiumin Xu and Taiwoo Park and Mi Zhang},
  title = {{AirSense: an intelligent home-based sensing system for indoor air quality analytics}},
  booktitle = {UbiComp},
  year = {2016},
  doi = {10.1145/2971648.2971720},
}
Figure