LSTM and Transfer Learning Based Underwater Acoustics Broadband Power Amplifier Behavior Modeling

Sonar transmitter identification is an important way to identify the target. The key here is the recognition of the internal power amplifiers, which depends on the accuracy of the power amplifier behavioral model. Typically, power amplifier models are suitable only for some specific types of signals. A new method of power amplifier behavior modeling is proposed in this paper to improve the adaptability of power amplifier models, that integrates long and short-term memory networks (LSTM) with transfer learning. The trained parameters from a source domain are used as the initial value of the target domain network. While the first two layers of the network are fixed, only the target domain data is used to train the following layers. This approach ensures that the model trained under a specific signal becomes independent of the signal type, and could learn the power amplifier characteristics. Experimental results show that, comparing to conventional modeling methods, the proposed method improves the power amplifier model robustness to diverse signal types while maintaining higher accuracy.