Electrical and Electronic Engineering - Theses
Permanent URI for this collection
Search Results
1 - 1 of 1
-
ItemAdvanced Neural Network-Based Equalization for Short-Reach Direct Detection Systems( 2022)Driven by the exponential growth of Internet traffic mostly from cloud and mobile services in recent years, there is an increasing demand for high-speed low-cost optical communication systems in short-reach applications such as data center interconnects. Compared with coherent detection, direct detection optical links are well-suited for such applications due to its low cost and simple structure. However, the intensity-only direct detection, or the simple square-law detection of optical field, produces a nonlinear channel when mixed with chromatic dispersion. Moreover, to meet the low-cost target, bandwidth-limited transceivers and cheap lasers such as directly modulated lasers (DML) are preferred which possess non-ideal frequency response and the chirp impairments. The mixed linear and nonlinear impairments could strongly degrade bit error rate (BER) performance and limit the system achievable capacity. As such, efficient nonlinear equalization techniques are of vital importance to guarantee a desired system BER performance. With the rapid development of machine learning technologies, various neural network (NN)-based equalizers have been proposed recently as the underlying digital signal processing (DSP) tools to effectively deal with the system impairments. NN-based equalizers attract a lot of attention since they usually outperform traditional methods such as feedforward equalization (FFE), decision feedback equalization (DFE) and the Volterra series-based equalization in BER performance, which enables higher data-rate signal transmission. Besides system BER performance, another important concern lies in the computational complexity (CC) of the receiver. For NN-based equalization, the CC concern lies in both the training and equalization processes. The training process of NNs usually requires a large number of training symbols and epochs, and when the link scenario is changed, the performance of the old NNs will degrade and the NNs may need to be retrained to fit for the new scenario, which is rather computationally inefficient. As for the equalization process, the number of multiplications per equalized symbol can only be around a few tens considering real-time DSP implementation. The increase of CC would lead to higher latency and larger power consumption of the receiver. Based on these important facts, it is highly desirable to reduce CC, whether in NN training or the equalization process. In this thesis, the performance and CC of NN-based equalizers are the main concerns for short-reach direct detection links. The CC of four commonly-used NN-based equalizers, i.e., feedforward NN (FNN), radial basis function NN (RBF-NN), auto-regressive recurrent NN (AR-RNN), and layer recurrent NN (L-RNN), are theoretically derived and their BER performance are compared in numerical simulation of a pulse amplitude modulation (PAM)4 direct detection optical link. FNN-based equalizers are found to have the lowest CC while the AR-RNN-based equalizers exhibit the best BER performance. Guidelines are provided on a proper NN selection considering the tradeoff between BER and CC. This thesis then focuses on performance-enhanced NN-based equalizers, and some advanced NN designs are investigated. A novel cascade FNN/RNN is developed and demonstrated in a 100-Gb/s PAM4 transmission experiment, which shows its superior performance over traditional approaches with limited additional CC involved. Different equalization schemes are then compared with the aim to jointly equalize both linear and nonlinear impairments. Besides cascade NNs, inserting an FFE/DFE block after NNs could also slightly improve system performance. Moreover, a thorough analysis and discussion are made on the BER and CC impact of all the possible additional connections which can be added onto a 2-layer FNN. Among all the possible connections, the cascade and recurrent ones are found most useful in determining equalization performance. Lastly, this thesis discusses several approaches to reduce CC for NN-based equalization. For the NN training part, transfer learning (TL) is applied in short reach applications, and the number of training symbols and epochs for equalization in the target system can be greatly decreased with the help of the trained NNs from source systems (systems different from target one but somewhat related). Instead of trained from scratch, the NNs are trained with preserved information gained from the source system, which enables an expeditious transition from one optical link to the other. For the NN equalization part, the idea of multi-symbol prediction is proposed for different types of NNs, which enables effective weight sharing and lower the number of multiplications required per received symbol. For FNN and cascade FNN, the CC reduction can be achieved without the loss of BER, while for RNN, multi-symbol equalization sacrifices slight BER performance. Nevertheless, if the BER is not strictly required, the CC reduction of RNN is still significant. Moreover, pruning technique is employed in NN-based equalization for short-reach links to reduce CC. With the aid of weight pruning, a sparsely-connected cascade RNN is demonstrated for equalization in a 100-Gb/s PAM4 link, while upholding the BER achieved by the fully-connected version. Pruning is also combined with multi-symbol equalization schemes, which further reduce the overall CC.