Oliver Ploder,
"Machine Learning Aided Self-Interference Cancellation in 4G/5G Mobile Transceivers"
, 4-2024
Original Titel:
Machine Learning Aided Self-Interference Cancellation in 4G/5G Mobile Transceivers
Sprache des Titels:
Englisch
Original Kurzfassung:
The limited transmitter-to-receiver stop-band isolation of the duplexer in long term evolution (LTE), LTE advanced (LTE-A) and 5G New Radio (5G NR) frequency division duplex (FDD) transceivers causes a leakage signal from the transmitter (Tx) into the receiver (Rx). These leakage signals can implicate a multitude of self-interference (SI) signals in the receiver path(s) diminishing the receivers? sensitivity. These effects include second-order intermodulation distortions (IMD2), happening in combination of a Tx leakage signal with a coupling of the downconversion mixer?s RF- and local oscillator ports. In the case of LTE-A carrier aggregation (CA), so called modulated spur interferences can occur as a result of a coupling between the local oscillator lines, creating harmonics on the chip which can lead to a downconversion of the leakage signal to the Rx baseband. The last effect being studied for FDD transceivers are disturbing Tx harmonics, which might result because of a non-ideal Tx power amplifier (PA) coupled with downlink CA. Lastly, in-band full duplex SI, as a result of a non-ideal PA coupled with the usage of the same frequency band for both, Rx and Tx, is studied and possible solutions are presented and compared. This last SI effect cannot occur in LTE(-A) or 5G transceivers, as the standard does not include in-band full duplex communications, but is interesting from a researcher?s perspective and for potential future communication systems.
Technological advances of the recent decade made neural networks (NNs) and machine learning (ML) a feasible approach for solving complex problems that are even able to surpass human accuracy in certain areas. While significant research has been conducted on the topic of SI cancellation (SIC), most proposed solutions rely on traditional adaptive filter based approaches and very little work has been conducted on the usability of NNs for such problems. Therefore, this work aims to investigate to which extent NNs can replace and/or aid current, traditional approaches for SIC in mobile transceivers. Besides neural networks, different other machine learning concepts (e.g., random forests, tensors) are evaluated for their suitability for SIC. Further, special focus has been laid
on the complexity of solutions in terms of hardware needed to properly cancel the SI signal. This is especially true for the NN based architectures presented, where this work proposes a novel online learning (i.e., sample-adaptive) approach allowing NNs to be used for SIC, which has not been done before. Further, the field of traditional SIC is extended by combining fully-digital and mixed-signal algorithms into a new hybrid approach. Finally, ML concepts are used to enhance traditional SIC architectures by eliminating the need for tuning of the algorithms, while overcoming the performance of standard architectures.
Forschungs-