Design of Machine Learning Based Optimization Algorithms for Electrical Balance Duplexers

Electrical Balance Duplexers (EBD) have demonstrated remarkable capabilities in achieving high TX-to-RX isolation within Frequency-Division Duplex (FDD) applications, rivaling conventional Surface Acoustic Wave (SAW) filters commonly used in cell phones. Unlike SAW filters, EBDs offer the advantage of achieving high TX-to-RX isolation over a wider frequency range, making them more adaptable for modern communication standards. However, the inherent challenge lies in maintaining optimal isolation levels, as EBD performance is sensitive to the impedance mismatch between the antenna and the balance network. The need for rapid readjustment of the balance network to counteract impedance variations at the antenna is crucial for sustaining high TX-to-RX isolation. Traditional optimization algorithms employed for tuning EBDs often exhibit extended convergence times and susceptibility to local minima, hindering their effectiveness in dynamic environments. This work addresses these challenges by proposing innovative tuning strategies that leverage the capabilities of machine learning techniques. The thesis begins with an explanation of the fundamental principles of EBDs and a novel EBD concept, the Phase Gradient Supported Electrical Balance Duplexer (PBD), which overcomes the physical limitations of the EBD. The balancing mechanisms are derived and the effects on the tuning of the balance network are discussed. Different system modeling techniques are explained and a machine learning based model, which is the basis of the proposed tuning concept, is presented. The dissertation examines tuning concepts in detail and focuses on machine-learning based approaches, offering a thorough grasp of their capabilities and possible use cases. To validate the theoretical groundwork, a balance network is assembled and analyzed extensively to evaluate its Smith chart coverage and the effectiveness of antenna mismatch compensation. The measurement data obtained from the balance network serves as the training data for the development of a machine learning model. Utilizing neural networks, the inverse function of the balance network is learned to serve as an inverse model of the balance network. The presented methods achieve significant initial isolation (average isolation >40 dB) and more than 55 dB in less than 5 iterations of retuning for LTE band 8 (10 MHz bandwidth, 45 MHz duplex spacing). The performance of the algorithms is limited only by Smith chart coverage of the balance network and training data accuracy, both of which can be optimized offline. In realistic simulations with a dynamic antenna impedance, the proposed concepts maintain high isolation by adjusting the balance network to the antenna impedance. The results were compared with conventional balanced network optimization algorithms and demonstrate superior performance with less overshooting, higher average and worst-case isolation.