Abstract: To address the issues of model overfitting and training instability caused by noisy data in deep learning,this paper introduces,for the first time,a dynamic error compensation mechanism from control theory into code generation tasks.It proposes a code generation model named PID-Dynamic LSTM,based on a dynamically-tuned cross-entropy loss function(PID-CE Loss).Traditional cross-entropy loss is vulnerable to interference from anomalous samples under noisy conditions,leading to deviations in gradient updates and reduced convergence speed.To mitigate this,it integrates proportional(P),integral(I),and derivative(D) control terms to construct a dynamic error compensation mechanism.1)Proportional term preserves the immediate error response characteristic of cross-entropy.2)Integral term incorporates exponential moving average(EMA) differentialto capture long-term trends in loss variation,thereby correcting accumulated bias.3)Derivative term suppresses prediction fluctuations induced by noise by constraining the mean squared error(MSE) of probability distributions between adjacent training steps.Experimental results demonstrate that during 500 epochs of noisy training,the proposed method achieves an 96.28% validation accuracy on the test dataset(+3.42% improvement over baselines).Critically,it reduces the number of epochs required to first reach 80% accuracy by 31.7%(from 224 to 153 epochs).Furthermore,it reduces the overfitting gap by 6.4% and decreases loss fluctuation by 18.5%.Ablation experiment further verifies the key role and parameter characteristics of PID-CE in noise suppression.This method establishes a theoretically interpretable and engineering-friendly paradigm for noise-robust optimization,demonstrating significant application potential in noise-sensitive scenarios.

Key words: PID control, Deep learning, Cross-entropy loss, PID-CE Loss, PID-Dynamic LSTM, MCU driver code generation, TTA8