Gated Recurrent Units Network Based on Adversarial Training for Multi-Step Fault Prediction of RF Circuits

RF (radio frequency) circuits have been widely used in radar systems, communication systems, cellular phones, etc. It is extremely important to ensure the proper operation of RF circuits and fault prediction is an important means to evaluate the reliable operation of RF circuits. Current fault prediction methods can usually predict the health state of a circuit at the next time step with high prediction accuracy. However, more often than not, circuit operators want to know if the circuit will operate properly in the next period, i.e., they need to know the health state of the circuit for the next multiple time steps. The manuscript proposes a GRU network based on adversarial training to implement RF circuit fault multi-step prediction, which can greatly reduce the prediction error accumulated in multi-step prediction using a recursive strategy. The core of the method is to use adversarial training to make the distribution of the time series generated by the multi-step prediction as close as possible to the data distribution of the series in the training set. Firstly, the GRU network is trained to achieve single-step prediction of RF circuit faults. Second, a recursive strategy is used to perform multi-step prediction. Finally, the GRU network is trained adversarially with a discriminator to determine whether the prediction series belongs to the training set; if it does not, the optimization of the GRU network continues; if it does, the prediction series of the network is shown to have been tuned and its distribution has been essentially identical to that of the training set. The method is validated in a low-noise amplifier circuit, experimental results show that when the prediction time step increases, the method greatly slows down the decreasing trend of the model performance. The RMSE and MAPE are higher than that of the traditional GRU model, and R2 is also the higher one, indicating that the method has the highest prediction accuracy and the best fit.