Using particle size distributions to identify indoor emissions: a machine learning method for source recognition

Variation in emitted particle sizes between sources offers potential for source identification. Seven thousand two hundred particle size distributions collected in a single-family home were categorized into four classes: (1) smoking, (2) cooking, (3) household cleaning, and (4) background (no source). We created a multi-class classifier (neural network) to predict the presence of these three source types and background air. Number, surface, and volume concentration profiles (each with 16 particle size bins from 0.3 to 10 µm) were used as input features. Surface or volume concentration profiles were better predictors than number concentration profiles; accuracy was highest when concatenating the number, surface, and volume profiles (48 features). The model achieved ~ 98% accuracy in predicting the four classes when using a 10-unit hidden layer for the 48 input features (compared to ~ 66% for multinomial logistic regression). Smoking has the highest F-score among the four classes (0.999). We applied the model to time series of size-resolved concentrations for three successive indoor particle emissions (pan frying food, sweeping the floor, and smoldering marijuana). Time-varying probabilities of the four classes were visualized. We examined the effects of source, space, ventilation, and temperature/relative humidity on source identification. Using training examples from the single-family home, the model reasonably predicted smoking in another residence under different indoor settings (accuracies > 96%). This study applied, for the first time, the neural network method to identify patterns across different particle size distribution types (number, surface, and volume). It demonstrates the potential of using machine learning for source recognition of transient indoor aerosol emissions (e.g., identifying smoking in a rental property).