High-Frequency NILM

Captures the high-frequency electromagnetic interference generated during an appliance's startup or state change. Unlike steady-state power draw, these transient spikes last microseconds to milliseconds and are unique to the physical switching mechanism.\n\n- Turn-on transients: The inrush current profile of a motor starting\n- Commutation noise: Electrical arcs from brushed motors in drills or vacuum cleaners\n- Switch-mode power supply oscillations: High-frequency ringing from modern electronics\n\nThese signatures are sampled at 1 kHz to 1 MHz to resolve the full waveform morphology.

Decomposes the aggregate current waveform into its fundamental frequency and integer multiples using Fast Fourier Transform (FFT) analysis. Non-linear loads draw current in non-sinusoidal patterns, generating distinct harmonic fingerprints.\n\n- Triplen harmonics (3rd, 9th, 15th): Strongly indicative of electronic loads with rectifier front-ends\n- Even harmonics: Often signify asymmetrical waveforms from half-wave rectifiers or faulty equipment\n- Phase angle of harmonics: The relative timing of harmonic components provides a secondary discriminative feature\n\nHarmonic analysis is particularly effective for distinguishing variable frequency drives, LED lighting, and consumer electronics that have similar real power consumption.

Plots the instantaneous voltage against current over one complete AC cycle (typically 1/50 or 1/60 of a second) to create a closed V-I trajectory loop. The shape, area, and curvature of this loop form a highly discriminative appliance fingerprint.\n\n- Loop area: Proportional to reactive power consumption\n- Asymmetry: Indicates non-linear behavior or harmonic distortion\n- Direction of traversal: Clockwise or counter-clockwise rotation reveals the nature of the reactive load\n- Self-intersections: Characteristic of phase-controlled devices like dimmer switches\n\nThis method is robust against voltage fluctuations because the trajectory normalizes current against the instantaneous driving voltage.

Digitizes the raw current waveform at sampling rates from 10 kHz to 1 MHz using precision analog-to-digital converters (ADCs). This preserves the fine-grained morphology of the current draw, revealing subtle features lost in RMS or power-averaged measurements.\n\n- Crest factor: The ratio of peak to RMS current, high for devices with sharp current pulses\n- Phase-controlled waveforms: The characteristic chopped sinusoid of a TRIAC-based dimmer\n- Pulse-width modulation ripple: High-frequency current ripple from motor drives and inverters\n\nThis raw waveform data serves as the input to convolutional neural networks that learn hierarchical feature representations directly from the time-series signal.

Exploits the conducted electromagnetic noise that appliances inject back onto the power line. Every device with a switch-mode power supply or motor generates characteristic EMI patterns in the 10 kHz to 30 MHz range.\n\n- Common-mode noise: Propagates on both line and neutral conductors relative to ground\n- Differential-mode noise: Propagates between line and neutral conductors\n- Switching frequency: The fundamental operating frequency of a power supply's oscillator\n- Frequency modulation patterns: Spread-spectrum clocking creates identifiable modulation signatures\n\nEMI-based disaggregation can identify devices even when their real power consumption is negligible or masked by larger loads, making it uniquely powerful for detecting modern electronic appliances.

Identifies appliance state changes at a temporal resolution finer than a single AC cycle (less than 16.67 ms at 60 Hz). This requires sampling at several kilohertz or higher to capture the precise moment of switching.\n\n- Zero-crossing synchronization: Events are timestamped relative to the voltage zero-crossing for phase consistency\n- Edge detection algorithms: Identify the exact sample index where current deviates from the predicted steady-state waveform\n- Overlapping event separation: Distinguishes two appliances switching within the same cycle using high-frequency ringing analysis\n\nSub-cycle resolution enables the disaggregation of simultaneously switching loads that would appear as a single event in low-frequency data.

A core high-frequency feature extraction method that plots the normalized voltage against the current waveform over a single AC cycle. The resulting V-I trajectory forms a unique shape fingerprint that captures the non-linear reactance and harmonic distortion of a specific appliance. Key characteristics:

• Resistive loads (heaters) produce straight lines
• Inductive loads (motors) generate elliptical loops
• Non-linear loads (electronics) create complex, asymmetric shapes Clustering algorithms like Dynamic Time Warping (DTW) then group similar trajectories to classify unknown devices.

The study of the brief, high-magnitude electrical disturbances that occur when an appliance changes state. Unlike steady-state signatures, transient events last only microseconds to milliseconds and are only visible at high sampling rates. Captured phenomena include:

• Inrush current spikes from motor startups
• Switching arcs and contact bounce noise
• High-frequency ringing from power supply capacitors These transient patterns are often more distinctive than steady-state consumption, enabling the differentiation of appliances with similar power ratings.

The process of decomposing a periodic current waveform into its fundamental frequency (50/60 Hz) and higher-order harmonic components using the Fast Fourier Transform (FFT). Non-linear loads inject significant energy at odd harmonics (3rd, 5th, 7th), creating a spectral fingerprint. Analysis dimensions:

• Magnitude and phase angle of each harmonic order
• Total Harmonic Distortion (THD) percentage
• Inter-harmonic content for variable-speed drives Harmonic signatures are additive in the aggregate signal, requiring source separation techniques to attribute specific harmonic contributions to individual appliances.

Specialized sensing infrastructure required to capture electrical signals in the kHz to MHz range. Unlike standard smart meters, high-frequency NILM demands dedicated hardware with specific capabilities:

• Current transformers (CTs) or Rogowski coils with wide bandwidth
• Analog-to-digital converters (ADCs) sampling at 1-100 kHz minimum
• Voltage dividers for simultaneous voltage waveform capture
• Edge compute modules (NVIDIA Jetson, Raspberry Pi) for on-device preprocessing
• GPS or NTP time synchronization for phase-angle accuracy This hardware is typically installed at the main electrical panel, representing a higher upfront cost than low-frequency approaches but enabling far richer feature extraction.

A specialized high-frequency technique that analyzes the conducted electromagnetic noise appliances inject onto the power line in the 100 kHz to 30 MHz range. Every switch-mode power supply and motor generates unique EMI signatures based on its internal oscillator frequencies and switching harmonics. Key advantages:

• EMI signatures propagate throughout the entire electrical network from a single sensing point
• Can identify devices in different rooms or circuits without sub-metering
• Highly stable signatures that don't vary with mechanical load This approach treats the building's electrical wiring as a giant antenna, capturing a rich signal that low-frequency methods completely miss.

The algorithmic identification of state transitions in streaming high-frequency data. Unlike low-frequency step-change detection, high-frequency event detectors must process massive data volumes in real-time. Common approaches:

• Cumulative Sum (CUSUM) algorithms adapted for high-dimensional feature streams
• Wavelet transform analysis to localize events in both time and frequency
• Edge detection operators applied to the raw waveform envelope
• Matched filtering using known transient templates False positives from noisy loads and false negatives from overlapping events are critical challenges requiring sophisticated thresholding and multi-resolution analysis.