Data Augmentation

Applies a uniform random phase shift to the complex baseband I/Q signal. Since a power amplifier's nonlinear behavior is generally phase-invariant relative to the carrier, rotating the input constellation by an arbitrary angle θ creates a physically valid new training sample. This technique dramatically improves the model's robustness to carrier phase offsets and prevents the network from overfitting to a specific absolute phase alignment in the training data.

• Implementation: Multiply I/Q samples by e^(jθ) where θ ~ Uniform(0, 2π)
• Benefit: Teaches the network that distortion depends on envelope magnitude, not absolute phase
• Typical range: Full 360° rotation applied per training batch

Scales the input signal envelope by a small random factor to simulate varying average power levels and PAPR conditions. This augmentation forces the predistorter to generalize across different drive levels and prevents the model from learning a distortion profile tied to a single operating point.

• Implementation: Multiply signal magnitude by α where α ~ Uniform(0.8, 1.2)
• Benefit: Models PA behavior across the full gain compression curve
• Constraint: Must stay within the PA's linear and weakly nonlinear operating region to maintain physical validity

Adds controlled complex Gaussian noise to the I/Q samples to simulate measurement noise, thermal effects, and ADC quantization error. This regularization technique prevents the neural network from fitting to the noise floor of the specific measurement setup and improves generalization to real-world deployment conditions.

• Implementation: Add n ~ CN(0, σ²) where σ is calibrated to the target SNR
• Benefit: Acts as a form of implicit regularization, smoothing the learned distortion function
• Typical SNR range: 30-50 dB relative to signal power

Translates the signal's center frequency within the PA's operating bandwidth by applying a digital frequency shift. This exposes the neural network to the frequency-dependent memory effects of the power amplifier across its full band, ensuring the predistorter does not overfit to a single carrier frequency.

• Implementation: Multiply I/Q samples by e^(j2πΔf·t) where Δf spans the PA bandwidth
• Benefit: Captures variation in AM/PM distortion and memory depth across frequency
• Constraint: Total shifted bandwidth must remain within the PA's characterized range

Randomly shifts the temporal alignment of the input-output sequence pairs and crops to a fixed window. This augmentation teaches the network that the memory effect relationship is time-invariant and prevents the model from learning spurious correlations based on absolute sample index positioning.

• Implementation: Apply random integer delay τ to the sequence, then extract a fixed-length window
• Benefit: Enforces shift-invariance in the learned memory model
• Constraint: Delay must be less than the total sequence length minus the model's memory depth

Creates new training samples by linearly interpolating between two existing I/Q signal pairs and their corresponding PA outputs. This vicinal risk minimization technique forces the neural network to behave linearly between training points, producing smoother decision boundaries in the learned predistortion function.

• Implementation: x_new = λ·x_i + (1-λ)·x_j where λ ~ Beta(α, α)
• Benefit: Reduces adversarial vulnerability and improves generalization to unseen modulation schemes
• Typical parameter: α = 0.2 to 0.4 for stable interpolation