Model Extraction

Offline extraction uses a complete, pre-recorded dataset of synchronized input and output waveforms to solve for model parameters in batch processing.

• Batch Least Squares: Computes the optimal coefficient vector in a single step by minimizing the sum of squared errors over the entire dataset.
• Data Requirements: Requires high-quality, time-aligned measurements captured with a vector signal analyzer and a representative excitation signal.
• Advantage: Provides the globally optimal solution for the given data and is ideal for initial model characterization in a lab environment.
• Limitation: Cannot adapt to changes in amplifier behavior due to temperature drift or aging without re-extraction.

Online extraction updates model parameters iteratively during live operation, allowing the model to track time-varying amplifier characteristics.

• Recursive Least Squares (RLS): An adaptive algorithm that updates coefficients with each new sample, weighting recent data more heavily to track dynamic changes.
• Least Mean Squares (LMS): A simpler, gradient-based adaptive algorithm that updates coefficients based on the instantaneous error, suitable for hardware-efficient implementation.
• Application: Essential for closed-loop Digital Predistortion (DPD) systems where the power amplifier's nonlinear profile shifts due to thermal memory effects or supply voltage variations.

The fidelity of an extracted model is highly dependent on the statistical properties of the signal used to stimulate the power amplifier during measurement.

• Persistent Excitation: The input signal must have sufficient bandwidth and amplitude variation to excite all nonlinear modes and memory depths of the amplifier.
• Peak-to-Average Power Ratio (PAPR): The test signal's PAPR must match the target communication standard (e.g., OFDM) to accurately capture compression characteristics.
• Avoiding Extrapolation: A model extracted with a low-power signal will fail unpredictably when predicting behavior at higher drive levels not present in the training data.

A critical failure mode in model extraction is overfitting, where the model memorizes the specific noise and artifacts of the training dataset rather than learning the true underlying amplifier physics.

• Cross-Validation: Partitioning measured data into distinct training and validation sets. The model is extracted on the training set, and its true performance is evaluated on the unseen validation set.
• Regularization: Adding a penalty term (e.g., ridge regression or LASSO) to the cost function during extraction to constrain coefficient magnitudes, which improves numerical stability and prevents overfitting.
• Validation Metrics: Normalized Mean Square Error (NMSE) on the validation set is the primary metric; a low training NMSE with a high validation NMSE indicates overfitting.

The mathematical robustness of the extraction process is determined by the condition number of the data matrix formed from the input signal.

• Ill-Conditioning: When basis functions (e.g., polynomial terms) are highly correlated, the data matrix becomes near-singular, leading to wildly oscillating coefficient values and poor model fidelity.
• Orthogonalization: Techniques like using orthogonal polynomials (e.g., Chebyshev or Legendre polynomials) instead of standard power series can dramatically improve numerical stability.
• Coefficient Sparsity: Post-extraction, near-zero coefficients can be pruned to reduce model complexity without significant accuracy loss, which is critical for real-time implementation on FPGAs.

The model extraction architecture defines the signal flow and error calculation used to identify the parameters of the predistorter or amplifier model.

• Direct Learning Architecture: The predistorter model is extracted by directly minimizing the error between the desired linear output and the actual output of the cascaded predistorter and power amplifier.
• Indirect Learning Architecture: The power amplifier's post-distorter is extracted first by swapping the input and output roles. The predistorter is then assumed to be an exact copy of this post-distorter, simplifying the extraction problem.
• Trade-off: The indirect architecture is simpler to implement but assumes the PA model is invertible, which is not strictly true for all nonlinear systems.