Principal Component Analysis (PCA)

PCA projects high-dimensional I/Q samples or power spectral density features onto a set of orthogonal basis vectors called principal components. The first component captures the direction of maximum variance in the data, with each subsequent component capturing the next highest variance while remaining orthogonal to all previous components. This creates a compact representation where normal spectrum behavior is encoded in the first k components, while noise and anomalies reside in the discarded dimensions.

Anomalies are detected by measuring the reconstruction error—the difference between the original signal and its projection back from the reduced principal component space. Normal signals reconstruct accurately because their structure is captured by the dominant components. Anomalous signals, such as rogue emitters or jamming waveforms, have significant energy in the residual subspace, producing a high squared prediction error (SPE) or Q-statistic that serves as the anomaly score.

In addition to the residual-based Q-statistic, PCA enables anomaly detection through Hotelling's T², which measures the Mahalanobis distance of a sample's projection within the principal component subspace. While the Q-statistic detects anomalies that violate the correlation structure of normal data, T² identifies samples that are extreme within the normal operating region. Combining both metrics provides comprehensive coverage for detecting different types of spectrum anomalies.

The mathematical foundation of PCA is the eigendecomposition of the covariance matrix of the spectrum data. The eigenvectors become the principal component directions, and the corresponding eigenvalues quantify the variance explained by each component. In RF applications, this decomposition reveals the underlying correlation structure between frequency bins or time samples, enabling the separation of structured signal content from unstructured noise and interference.

For real-time spectrum monitoring, incremental PCA and recursive PCA algorithms update the principal components as new data arrives without recomputing the full eigendecomposition. These adaptive variants track slow environmental changes—such as diurnal spectrum usage patterns—while maintaining sensitivity to sudden anomalies. Moving window PCA applies the decomposition to a sliding window of recent observations, balancing adaptability with computational efficiency.

Effective PCA-based anomaly detection requires careful preprocessing of spectrum data:

• Mean centering: Subtract the mean of each feature to center the data at the origin
• Variance scaling: Normalize features to unit variance to prevent high-power frequency bins from dominating the decomposition
• Complex-valued extensions: For raw I/Q data, complex PCA preserves phase relationships by operating on the complex covariance matrix
• Windowing: Apply appropriate time-frequency windows to capture transient anomalies

The quantitative difference between the original high-dimensional input and its projection back from the lower-dimensional PCA subspace. In spectrum anomaly detection, the residual error is used as an anomaly score.

• Calculation: Typically the Euclidean distance (L2 norm) between the original vector and the reconstructed vector.
• Mechanism: PCA learns the subspace of normal data variance. Anomalous signals contain energy in directions orthogonal to the principal components, resulting in a high reconstruction error.
• Application: A threshold on the reconstruction error separates normal background noise from rogue emitters or interference.

A multivariate distance metric that measures how many standard deviations a point is from the mean of a distribution, accounting for the covariance structure of the data.

• Relation to PCA: In the PCA-whitened space, the Mahalanobis distance simplifies to the Euclidean distance from the origin.
• Advantage: Unlike Euclidean distance on raw features, it is scale-invariant and corrects for correlations between variables.
• Use Case: Identifying signal vectors that are statistically distant from the normal cluster, even if they fall within the range of individual feature dimensions.

A non-linear generalization of the PCA concept. A neural network is trained to reconstruct normal signal data through a bottleneck layer.

• Comparison to PCA: An autoencoder with linear activation functions is mathematically equivalent to PCA. Non-linear activations allow it to learn complex, non-linear manifolds of normality.
• Architecture: An encoder compresses the input I/Q data into a latent code, and a decoder reconstructs the original signal.
• Scoring: The reconstruction error between the input and output serves as the anomaly score, similar to the PCA residual.

Deep Support Vector Data Description is a neural one-class classification method that learns to map normal data into a minimal hypersphere in a learned feature space.

• Core Idea: The network is trained to minimize the volume of a hypersphere that encloses the feature representations of normal data.
• Anomaly Scoring: The distance from the center of the hypersphere to the feature representation of a new sample is the anomaly score.
• Contrast with PCA: While PCA finds a linear subspace, Deep SVDD finds a non-linear, compact boundary around the normal class, offering a tighter envelope for complex signal distributions.

The task of identifying inputs that differ fundamentally from the training data distribution, crucial for detecting novel signal types in open-world spectrum environments.

• PCA Role: PCA can serve as a baseline OOD detector by flagging samples with high reconstruction error or low likelihood in the principal component subspace.
• Semantic Shift: OOD detection must distinguish between covariate shift (e.g., a known signal at a different SNR) and semantic shift (a completely new modulation scheme).
• Importance: Essential for cognitive radios operating in uncharacterized electromagnetic environments where unknown emitters are the norm.

The identification of changes in the underlying statistical properties of spectrum data over time, which can indicate a new emitter or a change in the RF environment.

• PCA Application: A sudden shift in the principal component directions or a persistent increase in the residual subspace variance can signal concept drift.
• Monitoring: By tracking the eigenvalues and eigenvectors of the covariance matrix over sliding windows, one can detect when the learned model of normality becomes stale.
• Adaptation: Triggers model retraining or alerts an operator to investigate the new spectral activity.