Online Training

Unlike offline training, online adaptation continuously recalculates predistorter coefficients during live signal transmission. The system captures a portion of the PA output through an observation receiver, compares it against the ideal reference, and updates model parameters in real-time. This closed-loop architecture enables the DPD to respond to thermal drift, load impedance changes, and component aging without interrupting service. The update rate must balance convergence speed against computational overhead, typically operating on a sample-by-sample or block-by-block basis.

Power amplifier behavior exhibits strong temperature dependence due to transistor junction heating and thermal memory effects. As the PA transitions from cold-start to thermal equilibrium—or experiences ambient temperature fluctuations—its AM-AM and AM-PM characteristics shift measurably. Online training directly addresses this by tracking the slow-varying thermal envelope and adjusting coefficients accordingly. This capability is critical for outdoor base station deployments where temperature swings of 40°C or more are common, preventing ACLR degradation during thermal transients.

Antenna impedance varies with environmental conditions, including ice accumulation, rain, and proximity to reflective surfaces. These variations alter the load presented to the power amplifier, shifting its optimal operating point and distortion profile. Online DPD systems monitor the voltage standing wave ratio (VSWR) and adapt the predistortion function to maintain linearity despite impedance fluctuations. This prevents spectral regrowth violations that would otherwise occur when a perfectly tuned offline model encounters real-world antenna mismatch.

Semiconductor devices degrade over operational lifetimes through mechanisms including hot carrier injection, gate oxide breakdown, and electromigration. These physical changes gradually alter the PA's non-linear transfer function. Online training provides a built-in aging compensation mechanism, continuously recalibrating the DPD model to track the amplifier's slowly evolving characteristics. This extends the effective service life of RF front-ends and maintains regulatory compliance without requiring manual recalibration or hardware replacement.

The choice of adaptation algorithm directly impacts convergence speed and steady-state error. Common approaches include:

• Least Mean Squares (LMS): Low complexity, suitable for slow-varying channels
• Recursive Least Squares (RLS): Faster convergence at higher computational cost
• Kalman filtering: Optimal for systems with known noise statistics
• Stochastic gradient descent: Applied when neural network DPD models require online fine-tuning The algorithm must balance tracking agility against coefficient stability to avoid introducing transient distortion during adaptation steps.

Online training depends critically on a high-fidelity observation receiver that captures the PA output for comparison. This path must exhibit significantly better linearity than the PA being corrected—typically 10-15 dB lower distortion—to avoid contaminating the error signal. Key specifications include wide dynamic range, phase coherence with the transmit path, and sufficient bandwidth to capture fifth-order intermodulation products. Any impairment in the observation path, such as I/Q imbalance or local oscillator phase noise, directly limits achievable linearization performance.