Machine learning (ML) and tensor-based methods have been of significant interest for the scientific community for the last few decades. In a previous work we presented a novel tensor-based system identification framework to ease the computational burden of tensor-only architectures while still being able to achieve exceptionally good performance. However, the derived approach only allows to process real-valued problems and is therefore not directly applicable on a wide range of signal processing and communications problems, which often deal with complex-valued systems. In this work we therefore derive two new architectures to allow the processing of complex-valued signals, and show that these extensions are able to surpass the trivial, complex-valued extension of the original architecture in terms of performance, while only requiring a slight overhead in computational resources to allow for complex-valued operations.

Function approximation from input and output data is one of the most investigated problems in signal processing. This problem has been tackled with various signal processing and machine learning methods. Although tensors have a rich history upon numerous disciplines, tensor-based estimation has recently become of particular interest in system identification. In this paper we focus on the problem of adaptive nonlinear system identification solved with interpolated tensor methods. We introduce three novel approaches where we combine the existing tensor-based estimation techniques with multidimensional linear interpolation. To keep the reduced complexity, we stick to the concept where the algorithms employ a Wiener or Hammerstein structure and the tensors are combined with the well-known LMS algorithm. The update of the tensor is based on a stochastic gradient decent concept. Moreover, an appropriate step size normalization for the update of the tensors and the LMS supports the convergence. Finally, in several experiments we show that the proposed algorithms almost always clearly outperform the state-of-the-art methods with lower or comparable complexity.

The limited transmitter-to-receiver stop-band isolation of the duplexers in long term evolution (LTE) and 5G/NR frequency division duplex transceivers induces leakage signals from the transmitter(s) (Tx) into the receiver(s) (Rx). These leakage signals are the root cause of a multitude of self-interference (SI) problems in the receiver path(s) diminishing a receiver's sensitivity. Traditionally, these effects are counteracted by the use of various different SI cancellation (SIC) architectures which typically solely target one specific problem. In this paper, we propose two novel neural networks based architectures that can handle a variety of different SI effects without the need for a different architecture for each effect. We additionally show the suitability of the proposed architecture on SI effects occurring in in-band full duplex transceivers. Further, we introduce two novel low-cost training algorithms to enable online adaptation (as opposed to offline training currently proposed in literature). The combination of these two concepts is shown to not only beat existing algorithms in their cancellation performance, but also to provide sufficiently low computational complexity allowing on-chip implementations.

Many researchers and practitioners make heavy use of the least mean squares (LMS) algorithm as an efficient adaptive filter suitable for a multitude of problems. Despite being versatile and efficient, a drawback of this algorithm is that the adaptation rate, i.e. step-size, has to be chosen very carefully in order to get the desired result (optimum compromise between fast adaptation and low steady state error). This choice was simplified by the invention of the normalised LMS, which bounds the step-size and guarantees convergence. However, the optimum choice of the normalisation becomes non-trivial if the system to be approximated is part of a bigger, non-trivial model, e.g. cascaded filters or linear paths followed by nonlinearities. Such cases usually require approximations or worst-case estimates in order to yield a normalised update algorithm, which might result in sub-optimal performance. To counteract this problem, a new class of LMS algorithms which automatically choose their own normalisation terms, the so-called self normalising LMS, is introduced. The simulations show that this new algorithm not only outperforms state-of-the-art solutions in terms of steady state performance in a cascaded filter scenario but also converges just as fast as all other considered algorithms.

State-of-the-art radio frequency transceivers for mobile communication devices suffer from transmitter-to-receiver (Tx-Rx) leakage in frequency division duplex operation, which, in combination with further non-idealities in the analog front-end, may lead to diverse self-interference (SI) effects. Digital as well as mixed-signal architectures have been proposed for self-interference cancellation. In this work we present a digitally intensive mixed-signal approach, where a low-cost auxiliary receiver senses the leaked Tx-signal. Firstly, the sensed signal is used to adaptively estimate the leakage channel, whereas in a second step a cleaned version of the leaked Tx-signal is reconstructed digitally. This reconstructed Tx-leakage signal is then used as input for a low complex adaptive interference cancellation unit to suppress modulated spurs or intermodulation distortions. We show that this approach allows to significantly relax the analog auxiliary receiver specifications, while different to conventional all-digital solutions being able to deal with multiple different types of SI with minimal configuration overhead.

Tensor-based estimation has been of particular interest of the scientific community for several years now. While showing promising results on system estimation and other tasks, one big downside is the tremendous amount of computational power and memory required – especially during training – to achieve satisfactory performance. We present a novel framework for different classes of nonlinear systems, that allows to significantly reduce the complexity by introducing a least-mean-squares block before, after, or between tensors to reduce the necessary dimensions and rank required to model a given system. Our simulations show promising results that outperform traditional tensor models, and achieve equal performance to comparable algorithms for all problems considered while requiring significantly less operations per time step than either of the state-of-the-art architectures.

Modern radio frequency transceivers for wireless communication standards operating in frequency division duplex suffer from an unwanted Tx leakage signal. In combination with carrier aggregation this might harm the receiver in form of a so-called modulated spur interference. Different digital interference cancellation techniques have been proposed but suffer from a slow adaptation rate due to the high correlation of the involved Long Term Evolution (LTE) signals. This work derives and analyzes a new form of the Least-Mean-Squares (LMS) algorithm that incorporates knowledge of the signal statistics. Simulations show that the presented algorithm can effectively improve the cancellation and adaptation performance for real world interference scenarios.

In frequency division duplex transceivers, the non-ideal analog duplexer has only a limited stop-band attenuation, and therefore a part of the transmit signal leaks into the receive path. Although operating on a different frequency band, non-ideal effects in the receive path cause different kinds of self-interferences, which can have a higher power level than the actual wanted receive signal. A possible way to tackle this problem are adaptive filters. These approaches are mostly model based, and for each kind of interference a different algorithm is needed. Kernel adaptive filtering offers the possibility to deal with different sorts of interferences with the same algorithm. In this work, we investigate the capabilities of kernel adaptive filtering to cancel especially the second-order intermodulation distortion (IMD2) and the transmitter (Tx)-harmonics interference.

The reciprocal function, 1/x, is important for many real-time algorithms. It is used in a large variety of algorithms from areas ranging from iterative estimation to machine learning. Many of these algorithms are iterative in nature and require the online computation of the reciprocal. Such an iterative structure often prevents effective use of pipelining for implementation of the reciprocal. For this reason, a reciprocal algorithm requiring only a low amount of clock cycles is desired. Many real-time algorithms, often being of approximate nature, can tolerate the use of only an approximate solution of the reciprocal. For this reason, we present a low complexity non-iterative approximation of the reciprocal function. This approximation can be calculated using only combinatorial logic. We present synthesis results showing that the proposed approach can be implemented with low area requirements at high clock frequencies. We analytically describe the error of the approximation and show that by optimizing a constant value used in the approximation, different variants with different error behaviors can be obtained. We furthermore present performance results of application examples that, when using our proposed method, show only negligible performance degradation compared to when using the exact reciprocal function, demonstrating the versatility of our proposed approach.

In frequency division duplex transceivers, a part of the transmit signal leaks into the receive path due to the non-ideality of the analog duplexer. Although operating on a different frequency band, non-ideal effects in the receive path lead to self-interferences with potentially higher power level than the wanted receive signal. One option to tackle this problem is the use of adaptive filtering algorithms for interference cancellation. In this work, we discuss support vector machines (SVMs) as an alternative approach. Different to the existing methods, the proposed concept does not need a model of the type of interference. We investigate the cancellation performance of SVMs compared to a recently published nonlinear adaptive filter for a specific type of interference called the second-order intermodulation distortion. It turns out that SVMs clearly outperform the adaptive filtering approach even for the severe case of narrow allocated transmit signals and high transmit power levels.

The limited transmitter-to-receiver stop-band isolation of the duplexers in long term evolution (LTE) frequency division duplex transceivers induces leakage signals from the transmitter(s) (Tx) into the receiver(s) (Rx). These leakage signals are the root cause of a multitude of self-interference (SI) problems in the receiver path(s) diminishing a receiver's sensitivity. This work proposes a novel architecture combating the second-order intermodulation distortion (IMD2), arising from the Tx leakage signal in combination with a coupling between the RF- and local oscillator (LO)-ports of the Rx IQ-mixer. In contrast to traditional adaptive filter solutions, the presented work relies on a neural network based approach for estimating the transmitter induced IMD2 SI signal used to cancel the interference in the receiver. The proposed architecture outperforms existing work based on least mean squares (LMS), recursive least squares (RLS) and Volterra kernel algorithms while maintaining comparable complexity.

The limited transmitter-to-receiver stop-band isolation of the duplexers in long term evolution and 5G frequency division duplex transceivers induces leakage signals from the transmitter(s) (Tx) into the receiver(s) (Rx). These leakage signals are the root cause of a multitude of self-interference (SI) problems in the receiver path(s) diminishing a receiver's sensitivity. This work deals with second-order intermodulation distortion, arising from the Tx leakage signal in combination with a coupling between the RF- and local oscillator-ports of the Rx IQ-mixer. We propose a novel adaptive architecture, utilizing neural networks, to cancel this type of interference. In contrast to traditional adaptive filter solutions, the proposed architecture can be used even if there is no model of the system available, making it robust against modeling noise and flexible in terms of interferences that it is able to cancel. The proposed architecture outperforms existing work based on least mean squares (LMS) algorithms and converges as fast as recursive least squares algorithms while maintaining comparably low complexity as the LMS approach.

A machine type communication (MTC) system for wireless transmission of sampled sensor data between two devices is considered. In order to reduce the latency between collecting the sample at the transmitting device and making it accessible at the receiving device, a novel cross-layer design is proposed and appropriate channel coding is discussed. The novel scheme deviates from traditional packet oriented communication in that the sampling is triggered by the physical layer of the transceiver once the packet transmission has started, instead of commanding a packet transmission at the application layer after the sample is available at the CPU. Thereby, delays from data buffering and packet generation are eliminated. Furthermore, in order to maintain guaranteed transmission latencies, appropriate channel codes are selected and their error correction performance is assessed in industrial environments.