TL;DR
PresenceSense (PS) is a "zero-training" semi-supervised algorithm that detects individual office occupancy by analyzing plug-load power consumption. By leveraging existing power monitoring infrastructure and an iterative self-labeling process, it achieves high accuracy (up to 93%) without requiring intrusive sensors or the labor-intensive manual labeling typically needed to train occupancy models.
What Problem Does This Paper Address?
Building managers want to optimize HVAC and lighting to save energy, but they need to know where people actually are. Traditional occupancy sensing faces a "Goldilocks" problem:
- Passive Infrared (PIR): Cheap but coarse; they often fail to detect a person sitting still at a desk.
- Cameras/Acoustic Sensors: Highly accurate but intrusive, raising significant privacy concerns.
- The Training Bottleneck: Most high-performing algorithms require "ground truth" labels—weeks of manual logs where users record every time they arrive or leave—to train the model. This is tedious, error-prone, and doesn't scale.
The authors tackle this by asking: Can we use the power data already being collected by "smart plugs" to infer presence without ever asking the user for a training label?
Methods
The core of the paper is the PresenceSense (PS) algorithm, which treats occupancy detection as a semi-supervised learning problem.
1. Feature Engineering
Instead of just looking at raw wattage, the authors identify three distinct "views" of power data:
- Power Level: The baseline demand (e.g., a laptop charging).
- Edge Effects: Sudden spikes or drops in power that signal a state change (turning a monitor on or off).
- Rippling Effects: High-frequency fluctuations that occur when a device is actively being used. They use metrics like Standard Deviation (SD) and Mean Absolute Difference (MAD) to capture this "electronic signature" of activity.
2. The "Zero-Training" Algorithm
The PS algorithm uses a variation of Co-training. It starts with a very rough "prior"—an educated guess of a typical work schedule (e.g., "user is present from 9 AM to 8 PM").
- It trains multiple classifiers on different "views" of the unlabeled data using this rough prior.
- The classifiers then "vote" on the labels for the next iteration.
- The algorithm iteratively refines these labels, effectively teaching itself the user's actual patterns.
3. Early Stopping
To prevent the model from eventually "learning" noise or random errors, the authors derive a theoretical early stopping rule based on the Probably Approximately Correct (PAC) learning framework. This ensures the algorithm stops once it reaches a point of diminishing returns in accuracy.
Datasets
The study was conducted in a graduate student office testbed at UC Berkeley (CREST).
- Primary Data: Individual plug-load data (desktops, monitors, laptops, lamps) collected via ACme sensors at 1-second resolution.
- Ground Truth: 4 users provided 5-minute resolution logs of their actual presence over several weeks.
- Comparison Sensors: To benchmark PS, the authors also deployed ultrasonic sensors (measuring distance to the user), acceleration sensors (attached to chairs), and WiFi access point logs (tracking phone proximity).
Key Insights and Ideas
- Rippling is the "Aha!" Moment: Simple power thresholds fail because many devices (like desktops) stay on when a user leaves. However, the fluctuations in power (rippling) are much more tightly correlated with active human presence than the absolute power level.
- Robustness to Bad Priors: The algorithm is surprisingly resilient. Even if the initial "expert guess" of a schedule is inaccurate, the iterative voting process allows the model to converge toward the true occupancy pattern.
- The Power of Multi-View Learning: By splitting features into different "views" (level vs. ripple), the algorithm can resolve ambiguities. If the power level is high but there is no ripple, the model can correctly infer the user has left their computer on and departed.
Key Results
- Competitive Accuracy: PresenceSense achieved an overall accuracy of 87% to 93% across different users.
- Outperforming Supervised Models: Remarkably, PS (which uses no training labels) outperformed traditional threshold-based models that were optimized using the full training set.
- Comparison to Hardware: While dedicated hardware like ultrasonic sensors performed slightly better (93.7% accuracy), PS provides a "software-only" alternative that requires zero additional infrastructure if power monitoring is already in place.
- WiFi and Accelerometers Lagged: WiFi proximity was found to be a poor indicator of desk presence (47% accuracy for presence), as phones often connect to APs even when the user is in a different room.
Limitations and Future Work
- Device Dependency: The system relies on the user actually using their plug-loads. If a user sits at their desk reading a paper book without turning on a lamp or computer, the system will likely mark them as absent.
- Granularity: The current model is designed for individual cubicles/desks. Scaling this to shared equipment or open-plan offices without individual metering remains a challenge.
- Future Directions: The authors suggest integrating this with "social games" to motivate energy savings or using the algorithm on other non-intrusive data sources like indoor positioning systems.
Summary updated