Event Detection

The foundational mechanism that identifies significant magnitude changes in the aggregate power signal. The algorithm monitors the difference between consecutive steady-state windows.

• Primary metric: Change in mean active power (ΔP) exceeding a predefined threshold
• Dual-edge logic: Detects both positive edges (appliance ON) and negative edges (appliance OFF)
• Typical threshold: 10-50 watts for residential loads, tuned to balance sensitivity against false triggers from noise
• Example: A refrigerator compressor activating creates a distinct 150W positive step, while its defrost cycle ending produces a corresponding negative step

Captures the high-frequency burst of electrical noise and harmonic distortion that occurs during the brief moment an appliance switches state, typically lasting only milliseconds to a few AC cycles.

• Sampling requirement: High-frequency data at 1–100 kHz to capture transient morphology
• Key features: Turn-on current spike amplitude, transient duration, and spectral content
• Discriminative power: Transient shapes are often unique to specific appliance types, enabling identification without steady-state analysis
• Example: A vacuum cleaner motor exhibits a distinctive high-amplitude inrush current transient lasting 100–200ms, clearly distinguishable from an incandescent light bulb's instantaneous step

A statistical framework that evaluates whether an observed signal segment is better explained by a single steady state or by a state transition occurring within the window.

• Method: Generalized Likelihood Ratio (GLR) test comparing a null hypothesis (no event) against an alternative hypothesis (event at time t)
• Output: A detection statistic that peaks at the most likely change point
• Advantage: Provides a mathematically principled threshold based on desired false-alarm probability
• Example: A GLR detector can reliably identify a 60W load change even when superimposed on a slowly varying aggregate baseline with ±20W fluctuations

An iterative algorithm that accumulates deviations from a reference mean to detect subtle but sustained shifts that individual samples might miss.

• Mechanism: Maintains a running cumulative sum of residuals; triggers an event when the accumulated value exceeds a control limit
• Sensitivity tuning: The decision interval parameter (h) and reference value (k) control the trade-off between detection delay and false alarm rate
• Best for: Detecting small-magnitude events (5–20W) that would be lost in noise with simple thresholding
• Example: A CUSUM detector can identify a 15W LED lighting circuit activating even when the aggregate load fluctuates by ±30W due to other appliances

A signal processing technique that cross-correlates the aggregate signal with a known template of an appliance's characteristic turn-on or turn-off signature.

• Template library: Pre-extracted transient or steady-state step profiles for each known appliance
• Peak detection: Events are declared at time indices where the correlation coefficient exceeds a threshold
• Noise immunity: Optimal detection in additive white Gaussian noise when the template matches the true signature
• Example: Correlating the aggregate signal with a template of a washing machine motor start (gradual ramp over 2–3 seconds) distinguishes it from an instantaneous resistive load step

A non-parametric method that detects events by comparing the statistical dispersion of two adjacent sliding windows moving through the signal.

• Algorithm: Computes variance in a left window and right window; a significant ratio change indicates a transition
• Window size: Typically 5–30 seconds for low-frequency data, balancing responsiveness against noise suppression
• Advantage: Detects events even when the absolute power change is ambiguous, by identifying shifts in signal stability
• Example: A heat pump cycling on introduces not just a power increase but a change in variance due to compressor modulation, triggering detection even if the mean shift is gradual