Inference and learning algorithms for wireless iterative receivers

Abstract

With the development of science and technology, more and more users and internet nodes are connected to the wireless network, resulting in a more complex communication environment and the requirement to process a large amount of data in the network. In this case, machine intelligence technology has been widely used in wireless network processing big data by virtue of its powerful big data computing capability. The key to machine intelligence is to design inference and decoding algorithms for different scenarios, and both Compressive Sensing (CS) and Deep Learning (DL) are the same ideas of machine intelligence, which emphasizes the learning and construction of features. Hence, this thesis may focus on designing inference and decoding algorithms, which may employ CS and DL in Wireless Sensor Networks (WSNs) and Multiple-Input Multiple-Output (MIMO) schemes, and all of them involve iterative decoding. The implementation of iterative decoding in WSNs and the MIMO scheme is designed for providing better service in 5G and beyond wireless communication networks. In WSNs, they may suffer from the constrained energy problem, CS is an impressive technique for the WSNs to overcome this problem, and support identification is an important point in the CS technique. A novel support identification technique based on factor graphs and belief propagation is proposed for CS-aided WSNs, which reliably estimates the locations of non-zero entries in a sparse signal through an iterative process. This thesis shows that the proposed factor graph-based approach outperforms the state-of-the-art relative frequency-based approach and Orthogonal Matching Pursuit (OMP) algorithm. This thesis also demonstrates that the proposed support identification technique is capable of mitigating the signal reconstruction noise by as much as 4 dB upon pruning the sparse sensing matrix. Furthermore, by intrinsically amalgamating the relative frequency based and the proposed factor graph-based approach, this thesis conceives a hybrid support identification technique for reducing communication between the sensor nodes and the Fusion Center (FC), while maintaining high-accuracy support identification and simultaneously mitigating the noise contaminating signal reconstruction. Considering that WSNs suffer not only from constrained energy but also from unreliable channels, channel coding may be employed to protect the transmitted signals from unreliable channels. In this circumstance, the novel concept of joint CS and Low-Density Parity Check (LDPC) coding is conceived for Joint Source-Channel Coding (JSCC) in WSNs supporting a massive number of signals. More explicitly, this thesis demonstrates this concept for a specific scheme, which supports a massive number of signals simultaneously, using a small number of Sensor Nodes (SNs) based on the concept of CS. The compressed signals are LDPC coded in order to protect them from poor transmission channels. This thesis also proposes a new iterative joint source-channel decoding philosophy for exchanging soft extrinsic information, which combines CS decoding and LDPC decoding by merging their respective factor graphs. In this proposed iterative LDPC-CS decoding, an exhaustive Maximum A Posteriori (e-MAP) technique is employed in the CS decoding, which traverses all possible combinations of the connected signals. The BLock Error Rate (BLER) results show that the proposed iterative joint LDPC-CS decoding scheme attains about 1.5 dB gain at a BLER of 10−3 compared to a benchmarker, which employs separate CS and LDPC decoding. Naturally, this gain is achieved at the cost of approximately doubling the complexity of the proposed iterative joint LDPC-CS decoding scheme. The proposed e-MAP approach in the CS decoding of the JSCC scheme may suffer from high complexity according to the total number of combinations may be large. In order to solve the complexity problem of the e-MAP approach in the CS decoding, this thesis proposes two reduced complexity approaches based on the tree search, namely Soft-Input Soft-Output (SISO) Sphere Decoding (SD) approach and Hamming Distance (HD) approach. In the proposed SD and HD approaches, only the more likely combinations of source signals are tested for reducing the CS decoding complexity. More specifically, a tree search technique is used in the first step to finding the most likely combination of source signal values. Then, in the second step, the proposed SD continues the tree search to find a set of alternative hypotheses. This facilitates the generation of high-quality extrinsic information, which may be iteratively exchanged with the LDPC decoder. By contrast, in the HD approach, the second step obtains the alternative hypotheses within a certain HD of the most likely source signal combination. The proposed SD and HD approaches have been proven to reduce the complexity significantly while maintaining the performance of the e-MAP approach. In particular, this thesis shows that the e-MAP solution is about 56 times more complex than the SD approach and around 210 times more complex than the HD approach. Apart from the WSNs, the inference and decoding algorithms design is also important in the MIMO iterative receivers in order to improve decoding performance or reduce decoding complexity. In this case, this thesis proposes a DL-aided Logarithmic Likelihood Ratio (LLR) correction method, in order to improve the performance of Multiple-Input Multiple-Output (MIMO) receivers, where it is typical to adopt reduced-complexity algorithms for avoiding the excessive complexity of optimal full-search algorithms. These sub-optimal techniques typically express the probabilities of the detected bits using LLRs that often have values that are not consistent with their true reliability, either expressing too much confidence or not enough confidence in the value of the corresponding bits, leading to performance degradation. To circumvent this problem, a Deep Neural Network (DNN) is trained for detecting and correcting both over-confident and under-confident LLRs. It is demonstrated that the complexity of employing the DL-aided technique is relatively low compared to the popular reduced-complexity receiver detector techniques since it only depends on a small number of real-valued inputs. Furthermore, the proposed approach is applicable to a wide variety of iterative receivers as demonstrated in the context of an iterative detection and decoding-aided MIMO system, which uses a low-complexity Smart Ordering Candidate Adding (SOCA) scheme for MIMO detection and LDPC codes for channel coding. Extrinsic Information Transfer (EXIT) charts are adopted for quantifying the Mutual Information (MI) and show that the proposed DL method significantly improves the BLER. Explicitly, it is demonstrated that about 0.9 dB gain can be achieved at a BLER of 10−3 by employing the proposed DL-aided LLR correction method, at the modest cost of increasing the complexity by 16% compared to a benchmarker dispensing with LLR correction. Considering the complex communication environment of the 5G and beyond communication systems, the interference problem becomes vital. The demodulation reference signal in the 5G Multiple-Output Orthogonal Frequency-Division Multiplexing (MIMO-OFDM) wave has been designed to enable Minimum Mean-Square Error-Interference Rejection Combining (MMSEIRC) equalization, which becomes state-of-the-art. By contrast, in the 4G LTE system, the state-of-the-art non-linear turbo equalization techniques, where the nonlinearity afforded superior performance, compared to linear equalization. The reversion of the state-of-the-art from the non-linear technique in 4G to the linear technique in 5G represents an opportunity for nonlinear techniques to be introduced into the 5G systems, which preserves the IRC capability but additionally benefits from the performance enhancement of non-linear processing. A novel technique is proposed in the MIMO-OFDM system with interference, where a linear MMSE-IRC equalizer plus a non-linear equalizer, which makes the system non-linear is employed for interference mitigation and signal detection. The MMSE-IRC is a conventional linear algorithm, which is designed to not only enhance the transmitted signal but also attenuate the noise and the crucial interference as much as possible. This is in contrast to for example MMSE equalizers, which treat interference as noise and attempt only to attenuate this noise without the advantage of considering the particular nature of the interference. However, the MMSE-IRC may fail to detect the signals when all degrees of the freedom of the MIMO channel are occupied by the desired users, for example, when the transmitter adopts a high number of transmission layers. The linear MMSE-IRC equalizer is modified, which is compatible with the inputs of a non-linear equalizer. More specifically, the inputs need to be equivalent channel information, covariance matrix, and equivalent received signals. Having introduced this concept, it is demonstrated that it is for a particular novel example of the SOCA algorithm, which is a reduced complexity nonlinear detection algorithm and is particularly suited to practical implementation using parallel processing with low latency and suitable for hardware implementation. The proposed technique employs the MMSE-IRC algorithm to mitigate the interference and first estimate the signals, then uses the SOCA algorithm to further detect the signals and generate the soft information for the iterative detection and decoding scheme, in order to solve the problem that MMSE-IRC may fail to detect the valuable signals affected by the interference. We present BLER results, which show that the proposed scheme can always achieve superior performance to the conventional MMSE-IRC detector, at the cost of increasing the complexity. In some cases, our proposed scheme can obtain about 1.5 dB gain, at the cost of the complexity by 4 times. We demonstrate that the complexity of the SOCA detector can be reduced by adjusting its parameterization or being the cost of reducing the self-consistency of the soft information produced by the SOCA detector, which degrades the BLER performance. In order to mitigate this, we propose to use the DL technique to correct the soft information and make it become more self-consistent. Using this technique, we show that the MMSE-IRC detector in cooperating with DL can obtain about 3.25 dB gain at the cost of the complexity by 1.007 times, compared to the conventional MMSE-IRC detector.

More information

Identifiers

ORCID for Rob Maunder: orcid.org/0000-0002-7944-2615

ORCID for Soon Xin Ng: orcid.org/0000-0002-0930-7194

Catalogue record

Export record