Spectrum Occupancy Dataset

The dataset must capture Power Spectral Density (PSD) measurements with fine-grained timestamps. The sampling interval must be significantly shorter than the target Prediction Horizon to resolve the dynamics of primary user activity.

• Sub-millisecond resolution is required for bursty signals like radar.
• Consistent timestamping prevents temporal aliasing.
• Enables accurate modeling of inter-arrival times and channel holding times.

A robust dataset spans a wide, contiguous block of spectrum to capture adjacent channel interference and cross-band correlations. Narrowband datasets fail to train models that generalize across the electromagnetic environment.

• Covers target bands (e.g., sub-6 GHz, mmWave).
• Includes guard bands and unlicensed spectrum.
• Enables Spatiotemporal Forecasting across diverse propagation characteristics.

Spectrum usage is inherently location-dependent. A high-quality dataset aggregates measurements from multiple geographically distributed sensing nodes to capture spatial reuse patterns and propagation effects.

• Urban, suburban, and rural environments.
• Indoor vs. outdoor sensor placement.
• Essential for training Radio Environment Maps and federated models.

The binary state of a channel (IDLE or BUSY) must be derived from a reliable thresholding mechanism, typically using an Energy Detection algorithm with a known noise floor. Mislabeling introduces systematic bias.

• Requires calibrated Noise Floor Estimation.
• Validates against known Primary User transmission schedules.
• Prevents the model from learning phantom signals.

The dataset must span multiple weeks or months to capture the full range of cyclical human activity patterns. Short recordings fail to represent diurnal, weekly, and seasonal trends.

• Captures Seasonality Decomposition components.
• Includes weekdays, weekends, and holidays.
• Prevents Concept Drift caused by insufficient training on rare events.

Raw PSD data is insufficient without rich metadata describing the sensing context. Metadata enables proper normalization, filtering, and Domain Adaptation across different hardware configurations.

• Sensor specifications: antenna gain, noise figure, calibration date.
• Location coordinates and elevation.
• Timestamps in a standardized format (e.g., UTC Unix epoch).

A multi-dimensional data structure representing spectrum usage over time, frequency, and space. This tensor serves as the foundational input for spatiotemporal prediction models. Key dimensions include:

• Time axis: Sequential power measurements
• Frequency axis: Discrete frequency bins
• Spatial axis: Geolocated sensor coordinates

A recurrent neural network architecture specifically designed to capture long-range temporal dependencies in spectrum usage data. LSTMs overcome the vanishing gradient problem inherent in standard RNNs, making them highly effective for learning the complex periodic patterns found in human-driven spectrum activity, such as diurnal and weekly cycles.

The process of assigning a confidence score or prediction interval to a spectrum forecast. Rather than providing a single point estimate, this technique enables a cognitive radio to make risk-aware decisions. A radio might transmit in a slot predicted to be idle with 95% confidence, but defer if the uncertainty is too high to guarantee non-interference with a primary user.

The algorithmic monitoring of prediction errors and input data distributions to automatically identify when a deployed forecasting model has become stale. As wireless environments evolve due to new devices or changed human behavior, concept drift occurs. Drift detectors trigger automated retraining pipelines to maintain model accuracy in production systems.

A privacy-preserving distributed learning framework where multiple sensing nodes collaboratively train a shared forecasting model without exchanging raw spectrum data. Only encrypted model updates are shared, allowing defense agencies or competing commercial operators to build robust models on aggregated knowledge while maintaining strict data sovereignty and security.