Blind Source Separation Disaggregation

BSS is fundamentally the cocktail party problem applied to energy. Just as you isolate a single voice in a noisy room, BSS algorithms separate overlapping appliance signals from a single-point measurement.

• No prior information: The algorithm has no labeled data about which appliances exist.
• Statistical independence: Assumes source signals are statistically independent.
• Mixing matrix: The aggregate signal is a linear combination of unknown sources.

ICA is the core mathematical engine for BSS disaggregation. It searches for a linear transformation that maximizes the non-Gaussianity of the separated signals.

• Central Limit Theorem: A mixture of independent signals is always more Gaussian than the sources.
• Kurtosis maximization: ICA finds projection directions that maximize the peakedness of the distribution.
• Whitening: A preprocessing step that decorrelates the input data to simplify the separation problem.

SOBI exploits temporal structure rather than higher-order statistics. It uses time-delayed covariance matrices to separate sources with distinct spectral content.

• Joint diagonalization: Simultaneously diagonalizes multiple covariance matrices at different time lags.
• Colored signals: Excels when sources have unique autocorrelation structures, like a refrigerator compressor vs. a resistive heater.
• Robustness: Less sensitive to outliers than kurtosis-based ICA methods.

NMF decomposes the aggregate power spectrogram into a parts-based representation, enforcing the physical constraint that power consumption cannot be negative.

• Additive decomposition: V ≈ WH, where W contains spectral bases and H contains activations.
• Sparsity constraints: Encourages each basis to represent a single appliance's signature.
• Multiplicative updates: Iterative optimization that preserves non-negativity throughout training.

In practice, the number of appliances far exceeds the number of measurement channels. Underdetermined BSS tackles the case where there are more sources than sensors.

• Sparse component analysis: Assumes sources are inactive most of the time, creating separable patterns in the time-frequency domain.
• Single-channel BSS: The extreme case with one aggregate meter and dozens of appliances.
• Dictionary learning: Builds a library of appliance signatures from the aggregate data itself.

Unlike supervised NILM which requires labeled appliance data, BSS is unsupervised. This is both its strength and limitation.

• Zero training data: No need for sub-metered appliance recordings.
• Source labeling problem: BSS separates signals but cannot name them—a refrigerator and freezer may be separated but not identified.
• Hybrid approaches: Modern systems often use BSS for discovery and supervised models for classification.

The foundational statistical technique for blind source separation that recovers independent source signals from mixed observations. ICA operates on the principle that appliance loads are statistically independent and have non-Gaussian distributions.

• Maximizes statistical independence between recovered sources
• Requires at least as many sensors as sources in classical formulation
• Effective for separating steady-state appliance signatures with distinct probability distributions
• Often combined with FastICA or Infomax algorithms for computational efficiency

A dictionary learning approach that represents aggregate power signals as a sparse linear combination of appliance-specific basis functions. Each appliance contributes a unique activation pattern that is active only intermittently.

• Learns an overcomplete dictionary of appliance signatures from training data
• Enforces sparsity constraints to ensure only a few appliances are active simultaneously
• Excels at separating loads with transient spikes and short-duration events
• Uses L1-regularization to promote sparse coefficient vectors during decomposition

A decomposition method that factorizes the aggregate power matrix into non-negative basis vectors and activation coefficients. This constraint naturally aligns with the physical reality that appliances only consume positive power.

• Enforces physical plausibility through non-negativity constraints
• Decomposes aggregate signal into appliance power signatures and their time-varying activations
• Particularly effective for low-frequency smart meter data with additive load structures
• Iteratively updates factors using multiplicative update rules to minimize reconstruction error

The specific blind source separation problem where only one aggregate measurement is available, making it an underdetermined system. This is the standard scenario in NILM where a single smart meter monitors the entire building.

• More challenging than multi-sensor ICA due to fewer observations than sources
• Requires strong prior knowledge or learned models of appliance behavior
• Solved using deep learning architectures like sequence-to-sequence models
• Relies on temporal structure and spectral differences rather than spatial diversity

The classic metaphor for blind source separation: isolating individual voices from a room full of overlapping conversations using only a few microphones. In energy disaggregation, the aggregate power signal is the noisy room recording.

• Each appliance represents a distinct 'speaker' contributing to the mixed signal
• The challenge is to 'unmix' without knowing how many appliances are active
• Human auditory system solves this through spatial hearing and spectral cues
• Energy disaggregation must solve the same problem using only electrical waveform characteristics

A hybrid approach that incorporates partial prior knowledge about appliance characteristics or operational constraints to improve separation accuracy. Unlike fully blind methods, it leverages available metadata.

• Uses known appliance power ratings as constraints on recovered signals
• Incorporates time-of-day priors reflecting typical usage patterns
• Reduces solution ambiguity compared to completely blind approaches
• Balances the flexibility of blind methods with the accuracy of supervised techniques