Model Extraction

Model extraction constructs a forward model that maps the PA's input to its output, capturing the nonlinear distortion and memory effects. This is distinct from identifying the inverse predistorter directly.

• The extracted model predicts the PA's output given any input signal
• Used in Direct Learning Architecture (DLA) to backpropagate the error
• Enables gradient computation without differentiating through the physical PA
• Typically a baseband equivalent model operating on complex I/Q samples

Model extraction can be performed offline using pre-recorded datasets or online during live operation, each with distinct trade-offs.

• Offline extraction: Uses laboratory measurements with dedicated test signals for high accuracy
• Online extraction: Adapts the model in real-time to track temperature drift and aging
• Offline methods allow computationally intensive algorithms like batch least squares
• Online methods require recursive algorithms such as Recursive Least Squares (RLS) or LMS

The quality of the extracted model depends critically on the persistent excitation of the training signal. The input must sufficiently exercise the PA's nonlinear operating range.

• Signals must span the full amplitude and bandwidth of expected operation
• Modern communication waveforms like OFDM provide natural excitation
• Insufficient excitation leads to ill-conditioned correlation matrices
• Dedicated training sequences may be injected during idle periods for background calibration

The choice of model structure determines the trade-off between fidelity and complexity. Common structures include memory polynomial, Volterra series, and neural network architectures.

• Memory Polynomial: Captures nonlinearity and memory with manageable complexity
• Generalized Memory Polynomial (GMP): Adds cross-terms for improved accuracy
• Volterra Series: The most general but computationally prohibitive for high orders
• Neural Networks: Emerging approach for capturing complex nonlinear behaviors

Once the model structure is chosen, the coefficients are estimated by minimizing a cost function—typically the mean squared error between the model's predicted output and the measured PA output.

• Least Squares (LS): Batch solution for offline extraction with optimal accuracy
• Recursive Least Squares (RLS): Online algorithm with fast convergence and a forgetting factor for tracking
• LMS and NLMS: Low-complexity stochastic gradient methods suitable for hardware implementation
• QR Decomposition: Provides superior numerical stability for ill-conditioned problems

Accurate model extraction requires precise time alignment between the transmitted input and the observed feedback signal. Even sub-sample misalignment degrades model fidelity.

• Loop delay through the transmit chain and feedback receiver must be estimated
• Integer-sample alignment via cross-correlation peak detection
• Fractional delay filters correct sub-sample misalignment using interpolation
• Misalignment introduces dispersion that the model incorrectly attributes to memory effects

The engineering discipline of constructing mathematical models of dynamic systems from observed input-output data. In DPD, system identification is applied to characterize the forward PA behavior—capturing both nonlinear distortion and memory effects. The identified model serves as the computational surrogate through which the error gradient is backpropagated during predistorter training.

• Encompasses model structure selection and parameter estimation
• Requires persistent excitation of the input signal for accurate identification
• Forms the theoretical bridge between raw measurements and usable behavioral models

A dedicated observation receiver chain that down-converts and digitizes a coupled sample of the PA output. The fidelity of this receiver—its bandwidth, dynamic range, and linearity—directly bounds the accuracy of model extraction. Any distortion introduced by the feedback path becomes inseparable from the PA's own nonlinearity in the extracted model.

• Must capture at least 3-5x the signal bandwidth to observe spectral regrowth
• Requires precise IQ imbalance compensation to avoid corrupting the extracted model
• Time alignment with the reference signal is a prerequisite for valid extraction

The process of precisely synchronizing the transmitted reference signal with the observed feedback signal in the digital domain. Even sub-sample misalignment introduces phase dispersion that manifests as artificial memory effects in the extracted model. Accurate alignment is the gatekeeper step—without it, all subsequent model extraction is invalid.

• Integer-sample alignment via cross-correlation peak detection
• Sub-sample refinement using fractional delay filters or Farrow structures
• Loop delay must be estimated and compensated before coefficient estimation begins

A predefined nonlinear transformation applied to the input signal to construct the model's output. In model extraction, the choice of basis functions—such as memory polynomial terms, Volterra kernels, or orthogonal functions—defines the hypothesis space within which the PA behavior is approximated. The basis set must be rich enough to capture the amplifier's nonlinear order and memory depth without introducing ill-conditioning.

• Memory polynomial: |x(n-m)|^k · x(n-m) terms
• Orthogonal bases improve numerical conditioning during least-squares estimation
• Basis selection trades off model fidelity against computational complexity

A numerical state where the correlation matrix formed during model extraction has a high condition number, making coefficient estimates hypersensitive to small perturbations in measurement data. Ill-conditioning arises when basis functions are highly correlated—common with dense polynomial expansions. Mitigation strategies include regularization parameters, orthogonal basis functions, and QR decomposition for numerically stable solutions.

• Condition number quantifies sensitivity: κ >> 1 indicates ill-conditioning
• Leads to coefficient drift and unreliable predistorter performance
• Particularly problematic in fixed-point FPGA implementations