TLDR: The i-Mask is a novel intelligent mask that uses integrated sensors to capture exhaled breath patterns (temperature and humidity variations) for human activity recognition (HAR). It processes this data through noise filtering, time-series decomposition, and machine learning models. The system achieved over 95% accuracy in classifying activities like running, walking, sitting, and sleeping, with the kNN classifier performing best at 96.4%. This non-invasive, real-time approach offers significant potential for healthcare and fitness applications.
A new intelligent mask, dubbed i-Mask, has been developed to recognize human activities by analyzing patterns in exhaled breath. This innovative approach offers a non-invasive and real-time method for monitoring well-being, with potential applications in healthcare and fitness.
Traditional methods for human activity recognition (HAR) often rely on wearable sensors, vision-based systems, or environmental sensors. However, these methods can come with drawbacks like discomfort, privacy concerns, or complex setup requirements. The i-Mask project, developed by researchers from the Indian Institute of Technology Patna and Dhirubhai Ambani Institute of Information and Communication Technology (DA-IICT), addresses these limitations by focusing on the rich physiological signals contained within our breath patterns.
How i-Mask Works
The core of the i-Mask system is a custom-developed mask equipped with integrated sensors. These sensors primarily capture variations in temperature and humidity from exhaled breath. The researchers found that different human activities – such as running, walking, sitting, and sleeping – produce distinct patterns in these breath parameters. For instance, more active states like running and walking show higher humidity and lower average temperatures due to increased ventilation, while stationary activities like sitting and sleeping exhibit higher average temperatures and more stable humidity levels.
The mask prototype includes an AHT10 sensor for precise temperature and humidity measurements, and an MQ135 sensor to detect changes in air composition, providing additional contextual data. A D1 Mini V2 NodeMCU microcontroller with built-in Wi-Fi handles data acquisition and wireless transmission to a connected smartphone. The AHT10 sensor is strategically placed inside the mask near the nose and mouth, while the MQ135 and NodeMCU are mounted externally to avoid direct airflow interference.
Data Collection and Analysis
To train the system, data was collected from twenty volunteers aged 20 to 30, who performed the four key activities while wearing the mask. Data was recorded at a one-second sampling rate over 30-minute sessions, with ambient conditions also noted. This comprehensive dataset allowed for the examination of both short-term and long-term breath pattern variations.
Before analysis, the raw data undergoes several crucial preprocessing steps. This includes noise and vibration filtering to remove high-frequency noise and highlight respiratory cycles. Techniques like wavelet transforms and Hilbert transforms are used to identify significant peaks in temperature and humidity signals, which correspond to inhalation and exhalation events. The data is also scaled and synchronized to ensure consistency, and outliers (abrupt spikes or drops) are detected and removed using activity-specific thresholds, with missing values filled by linear interpolation.
Further, the time-series data is decomposed into trend, seasonality, and residual components using the Seasonal-Trend Decomposition using Loess (STL) method. This helps in isolating long-term shifts, regular breathing cycle variations, and random noise, making the underlying breath patterns clearer. Correlation analysis also revealed interesting relationships, such as running temperature being moderately correlated with sitting and walking humidity, and negatively correlated with sitting and sleeping temperatures.
Also Read:
- SePA: An AI Agent for Proactive and Personalized Health Coaching
- Advancing Marine Bioacoustic Monitoring with Robust AI and Ecologically Sound Evaluation
Activity Recognition with Machine Learning
After preprocessing and feature extraction, machine learning models are trained to classify activities based on the processed breath patterns. The researchers evaluated four different models: Decision Tree (DT), k-Nearest Neighbors (kNN), Random Forest (RF), and Support Vector Machine (SVM).
The experimental results showed promising accuracy, with the kNN classifier achieving the highest performance at 96.4% accuracy. Decision Tree and Random Forest models also performed well, with accuracies of 95.7% and 94.8% respectively, while SVM had the lowest accuracy at 74.4%. The kNN model’s success is attributed to its ability to group similar activity states in the feature space, effectively distinguishing between different activities based on their unique breath signatures. Even closely related activities like walking and running showed minimal misclassification.
This research highlights the significant potential of breath-based physiological signals for human activity recognition. The i-Mask system represents a step forward in non-invasive, real-time monitoring for various applications, from personal fitness tracking to health monitoring. Future work aims to integrate deep learning models for even greater accuracy and robustness, expand the study with more diverse datasets, and explore the potential for disease prediction through breath analysis. You can read the full research paper here.
