paper survey ——deep learning or machine learing and optical communication

machine learning 或者说deep learning已经被广泛应用于各种领域，之前本人也发表了几篇ML或者DL跟VLC相结合的论文。本博文主要是对16年后ML或DL跟optical communication结合的相关的论文的调研。仅供本人学习记录用

 

 

 

 

Modulation Format Recognition and OSNR Estimation Using CNN-Based Deep Learning

An intelligent eye-diagram analyzer is proposed to implement both modulation format recognition (MFR，调制格式识别) and optical signal-to-noise rate (OSNR，光信噪比) estimation by using a convolution neural network (CNN)-based deep learning technique.

Recently, techniques from MACHINE learning algorithms have been equally well used in optical communications to promote the development of intelligent systems. The published works mainly focused on optical performance monitoring (OPM) and nonlinear impairments compensation using different algorithms from MLA family, including expectation maximum (EM), random forest, back-propagation artificial neural network (BP-ANN), as well as our recently proposed k-nearest neighbors (KNN) and support vector machine (SVM).

The previous MFR and OSNR estimation schemes were mainly based on the techniques of amplitude histograms, delay-tap scatter plots, Q-factor, or two-tap sampling , but they could not process raw data directly and had to extract the corresponding features artificially （人为的）.To realize real intellectualization and automation without manual intervention, we adopt convolutional neural network (CNN) as deep learning algorithm and select eye diagram as processing objective. With the ability of feature extraction and self-learning, CNN can process eye diagram in its raw form (namely the pixel values of an image) without knowing other eye-diagram parameters nor original bit information.

这篇应该就是属于optical performance monitoring（光学性能检测）通过训练一个CNN网络来对眼图进行识别分析。

 

Deep Learning for Interference Cancellation in Non-orthogonal Signal Based Optical Communication Systems

Interference Cancellation（干扰抵消）

关于SEFDM可以参考博客《学习笔记之——高频谱效率频分复用（SEFDM）》

A recognized waveform, termed spectrally efficient frequency division multiplexing (SEFDM，频谱有效频分复用), which was a technique initially proposed for wireless systems, has been extensively studied in 60 GHz millimeter wave communications, optical access network design and long haul optical fiber transmission. Experimental demonstrations have shown the advantages of SEFDM in its bandwidth saving, data rate improvement, power efficiency improvement and transmission distance extension compared to conventional orthogonal communication techniques. However, the achieved success of SEFDM is at the cost of complex signal processing for the mitigation （缓解） of the self-created inter carrier interference (ICI). Thus, a low complexity interference cancellation approach is in urgent need

Recently, deep learning has been applied in optical communication systems to compensate for linear and non-linear distortions in orthogonal frequency division multiplexing (OFDM) signals. The multiple processing layers of deep neural networks (DNN) can simplify signal processing models and can efficiently solve un-deterministic problems（不确定性问题）.

Furthermore, future 5G systems require ultra-low latency communications. Thus, the signal processing at both transmitter and receiver should be correspondingly as simple as possible. In realistic communication systems, due to complex channel conditions, complex channel estimation and equalization algorithms have to be used at the receiver leading to increased receiver complexity and extra signal processing power consumption. In addition, due to the existence of ICI in some signals, variable interference cancellation algorithms have to be used to remove interference. Other signal distortions such as frequency offset, timing synchronization, phase offset （相位偏移） and mismatched local oscillators （本振子） would introduce extra interference. All the effects have to be compensated at the receiver with the assistance of complex signal processing.

The non-orthogonal signal waveform SEFDM is shown in figure 1

It is clearly seen that the SEFDM signal waveform saves bandwidth relative to the typical OFDM. The higher bandwidth saving the higher ICI. The mathematical principle of the discrete representation of an SEFDM signal symbol is expressed as

It is mathematically shown that the second term in (2) is the interference component thathas to be removed.

Basically, the detection algorithm is a trade-off between performance and complexity. On one hand, the better performance indicates higher detector complexity. On the other hand, lower complexity results in worse performance. The novelty of this work is to use deep learning to learn the complex interference via backpropagation mechanism and mitigate its impact. The efficiency of the removal (去除) is dependent on the signal waveform characteristics and neural networks architectures.

Two signal detection methods are studied and compared. The first one is the hard decision detectorand the second one is the neural networks based detector, with the architecture shown in the greyblocks.

Neural Network Design for Interference Mitigation

Depending on different interference mitigation principles, the design of the neural network is different. In this section, we designed five neural networks for the non-orthogonal signal and each one has specific interference mitigation scheme. Five different neural network topologies are illustrated in Fig. 3.

The first one, termed no connection-neural network (N-NN), has the simplest network connections where each symbol associated with one sub-carrier forms an independent network termed sub-nets. There are multiple sub-nets making up the entire network. The number of sub-nets is determined by the number of sub-carriers in SEFDM signals. The reason for this architecture is that sub-carriers within OFDM are orthogonally packed and no cross interference exists. The complexity is proportional to the number of sub-carriers. In terms of SEFDM, since sub-carriers are non-orthogonally packed, interference will be introduced between adjacent sub-carriers. Thus, this network architecture would be no longer accurate.

The second topology, termed partial (部分) connection-neural network (P-NN), is an evolved (进化的) version of the N-NN. In such evolved network, two symbols on adjacent input cells are connected in one subnet and processed jointly. Therefore, the decision of one output cell is jointly determined by two adjacent input cells. This scenario assumes that the interference comes only from one neighbouring input cell. This will include the adjacent interference factor and would derive a more accurate network model. The parallel architecture still exists within P-NN but each sub-net will jointly process two input cells and two output cells.

The third topology, termed hybrid (混合)1-neural network (H1-NN), divides the input cells into two groups. The two edge cells form independent sub-net similar to the one in N-NN. The middle input cells take into account interference from adjacent cells and make up interleaved connections, hence the “hybrid” in the name. Since the interference suffered on middle input cells is considered in the network, a more accurate model than the P-NN can be derived

The forth topology, termed hybrid2-neural network (H2-NN), optimizing the connections for the edge input cells. As a matter of fact, the two edge input cells are interfered from their adjacent input cells as well. This crucial information has to be included in the network. The way to emulate this impact is to connect the edge input cell to its neighbouring cell. With a more accurate network connection, a more accurate model can be derived. H2-NN can efficiently emulate the interference suffered on each sub-carrier of SEFDM signals. Therefore, the H2-NN outperforms the H1-NN.

The last topology, termed full connection-neural network (F-NN), effects a fully connected mesh network. This network connects all the input cells and considers interference from both adjacent cells and non-adjacent cells. The decision of one symbol has to be determined by all the input symbols. This fully connected network architecture explores the interference completely within SEFDM signals and will show the best performance among the five investigated network architectures.

总的来说，就是单链接、两链接、全链接（相对于隐含层）、及他们的排列组合

结果对比

 

An Optical Communication’s Perspective （观点） on Machine Learning and Its Applications

ML applications in optical communications and networking are also gaining more attention, particularly in the areas of nonlinear transmission systems, optical performance monitoring（光学性能监测）, and cross-layer network optimizations for software-defined networks. however，the extent （程度）to which ML techniques can benefit optical communications and networking is not clear and this is partly due to an insufficient understanding of the nature of ML concepts.This paper aims to describe the mathematical foundations of basic ML techniques from communication theory and signal processing perspectives, which in turn will shed light on the types of problems in optical communications and networking that naturally warrant ML use.

这篇论文是一篇很不错的论文。将ML理论与communication理论相结合。作一个理论分析。非常适合入门者查阅。后面会针对这篇论文写一篇中文的博客。现在先看看这篇论文给我们带来什么新的东西

MLis an interdisciplinary fieldwhich shares common threads with the fields of statistics, optimization, information theory, and game theory. Most ML algorithms perform one of the following two types of pattern recognition tasks as shown in Fig. 1. In the
first type, the algorithm tries to find some functional description of given data with the aim of predicting values for new inputs, i.e., regression problem. The second type attempts to find suitable decision boundaries to distinguish different data classes, i.e., classification problem, which is more commonly referred to as clustering problem in ML literature. ML techniques are well known for performing exceptionally well in scenarios in which it is too hard to explicitly describe the problem’s underlying （构成…的基础） physics and mathematics. 

Optical communication researchers are no strangers to regressions and classifications. Over the last decade, coherent
detection （相干探测） and digital signal processing (DSP) techniques have been the cornerstone （基石） of optical transceivers in fiber-optic communication systems. Advanced modulation formats such as 16 quadrature amplitude modulation (16-QAM) and above together with DSP-based estimation and compensation of various transmission impairments such as laser phase noise have become the key drivers of innovation. In this context （在这种情况下）, parameter estimation and symbol detection are naturally regression and classification problems, respectively, as demonstrated by examples in Fig. 1(c) and (d). 

Currently, most of these parameter estimation and decision rules are derived from probability theory and adequate understanding （充分理解） of the problem’s underlying physics. As high-capacity optical transmission links are increasingly being limited by transmission impairments such as fiber nonlinearity, explicit statistical characterizations of inputs/outputs become difficult. An example of 16-QAM multi-span dispersion-free transmissions （多跨无色散传输） in the presence of fiber nonlinearity and inline amplifier noise （内联放大器噪声） is shown in Fig. 2(a).The maximum likelihood decision boundaries in this case are curved and virtually impossible to derive analytically （几乎不可能用解析的方法推导出来）.Consequently, there has been an increasing amount of research on the application of ML techniques for fiber nonlinearity compensation (NLC).

Another related area where ML flourishes （繁荣昌盛） is short-reach direct detection systems that are affected by chromatic dispersion (CD，色散), laser chirp（激光啁啾） and other transceiver components imperfections（收发机部件缺陷）, which render （导致） the overall communication system hard to analyze.

Optical performance monitoring (OPM) is another area with an increasing amount of ML-related research. OPM is the acquisition （得到） of real-time information about different channel impairments（损害） ubiquitously （无处不在） across the network to ensure reliable network operation and/or improve network capacity. Often, OPM is cost-limited so that one can only employ simple hardware components and obtain partial signal features to monitor different channel parameters such as OSNR, optical power, CD。In this case, the mapping between input and output parameters is intractable from （是棘手的） underlying physics/mathematics, which in turn warrants ML. An example of OSNR monitoring using received signal amplitudes distribution is shown in Fig. 2(b).

Besides physical layer-related developments, optical network architectures and operations are also undergoing major
paradigm shifts under the software-defined networking (软件定义的网络SDN) framework and are increasingly becoming complex, transparent and dynamic in nature. One of the key features of SDNs is that they can assemble large amounts of data and perform so-called big data analysis to estimate the network states as shown in Fig. 3.

This in turn can enable (i) adaptive provisioning of resources such as wavelength （波长等资源的自适应供应）, modulation format, routing path, etc., according to dynamic traffic patterns and (ii) advance discovery of potential components faults（故障） so that preventative maintenance （预防性维修） can be performed to avoid major network disruptions （中断）. The data accumulated in SDNs can span from physical layer (e.g., OSNR of a certain channel) to network layer (e.g., client-side speed demand) and obviously have no underlying physics to explain their interrelationships （相互关系）. Extracting patterns from such cross-layer parameters naturally demands the use of data-driven algorithms such as ML.

We will first introduce artificial neural networks (ANNs) and support vector machines (SVMs) from communication theory and signal processing perspectives.

Fig. 17 shows the use of a DNN as well as a conventional ANN on signal’s eye-diagrams to monitor OSNR. In the first approach, the eye-diagrams are directly applied as images at the input of the DNN, as shown in Fig. 17(a), and it is made to automatically learn and discover OSNR-sensitive features without any human intervention. The extracted features are subsequently exploited by DNN for OSNR monitoring. In contrast, with conventional ANNs, prior knowledge in optical communications is utilized in choosing suitable features for the task, e.g., the variances of “1” and “0” levels and eye-opening can be indicative of OSNR. Therefore, these useful features are manually extracted from the eye-diagrams and are then used as inputs to an ANN for the estimation of OSNR as shown in Fig. 17(b). For completeness（完整性）, Fig. 17(c) shows an analytical and non-ML approach to determine OSNR by finding the powers and noise variances that best fit the noise distributions of “1” and “0” levels knowing that they follow Rician distribution. In this case, a specific mathematical formula or computational instruction is pre-coded into the program and there is nothing to learn from the input data.

Fig. 18 compares the underpinning philosophies （基本哲学）of the three different approaches discussed above. Note that, in principle, there is no hard rule on how many layers are needed for an ML model in a given problem. In practice, it is generally accepted that when more underlying physics/mathematics of the problem is used to identify and extract the suitable data features as inputs, the ML model tends to be simpler.

APPLICATIONS OF ML TECHNIQUES IN OPTICAL COMMUNICATIONS AND NETWORKING

Fig. 24 shows some significant research works related to the use of ML techniques in fiber-optic communications. A brief discussion on these works is given below.

Optical Performance Monitoring (OPM)

Optical communication networks are becoming increasingly complex, transparent(透明的) and dynamic. Reliable operation and efficient management of these complex fiber-optic networks require incessant and real-time information of various channel impairments ubiquitously across the network, also known as OPM. OPM is widely regarded as a key enabling technology for SDNs. Through OPM, SDNs can become aware of the real-time network conditions and subsequently adjust different transceiver/network elements parameters such as launched powers, data rates, modulation formats, spectrum assignments, etc., for optimized transmission performance. Unfortunately, conventional OPM techniques have shown limited success in
simultaneous and independent monitoring (同时独立监测) of multiple transmission impairments since the effects of different impairments are often difficult to separate analytically. Another crucial OPM requirement is low complexity since the OPM devices need to be deployed ubiquitously (无所不在) across optical networks. ML techniques are proposed as an enabler for realizing low complexity (and hence low cost) multi-impairment monitoring in optical networks and have already shown tremendous potential.

Fiber Nonlinearity Compensation (NLC)

光纤非线性补偿(对于VLC而言，就是LED的非线性补充与信道的非线性补偿)

 

Proactive Fault Detection

主动式故障检测

Reliable network operations are essential for the carriers to provide service guarantees （服务保证）, called service-level agreements (SLAs，服务水平协议), to their customers regarding system’s availability and promised quality levels. Violation （违反） of these guarantees may result in severe penalties （刑罚）. It is thus highly desirable to have an early warning and proactive protectionmechanism incorporated （主动保护机制）into the network. This can empower network operators to know
when the network components are beginning to deteriorate （恶化） and preventive （预防的） measures can then be taken to avoid serious disruptions （中断）.

Conventional fault detection and management tools in optical networks adopt a rigid approach （刚性方法）where some fixed threshold limits are set by the system engineers and alarms are triggered to （被触发到） alert malfunctions （警报故障） if those limits are surpassed. Such traditional network protection approaches have the following main drawbacks: (i) These methods protect a network in a passive （被动的） manner, i.e., they are unable to forecast the risks and tend to reduce the damages only after a failure occurs. This approach may result in the loss of immense （巨大的） amounts of data during network recovery process once a failure happens. (ii) The inability （无能为力） to accurately forecast the faults leads to ultraconservative （超保守） network designs involving large operating margins （边缘） and protection switching paths which in turn result in an underutilization of the system resources. (iii) They are unable to determine the root cause of faults. (iv) Apart from hard failures (i.e., the ones causing major signal disruptions), several kinds of soft failures (i.e., the ones degrading system performance slowly and slightly) may also occur in optical networks which cannot be easily detected using conventional methods.

Software-Defined Networking (SDN)

Software-defined networking approach centralizes （集中） network management by decoupling （解耦） the data and control planes. SDN technologies enable the network infrastructure （基础设施） to be centrally controlled/configured in an intelligent way by using various software applications.

总的来说，这篇综述写得还是不错的。内容太多了。第一次看只是大概瞄了一下，没有细看。后面会有博客专门精度这篇paper

 

Deep Learning Framework for Wireless Systems: Applications to Optical Wireless Communications

this article presents deep learning (DL) techniques that identify an efficient optical transceiver pair. In particular, an unsupervised learning framework based on an autoencoder (AE) is investigated to design an optical wireless system.AE techniques have recently been adopted to RF communication designs. However, it is not straightforward to bring the machine learning structure into the OWC design since the effect of lighting constraints （约束条件） has not been properly studied. Therefore, additional processing is required to control behaviors of neural networks (NNs) to construct OWC systems. Since most state-of-the-art DL techniques have focused on unconstrained problems in classification and generative model applications, it is highly challenging to introduce complicated constraints into NNs in general.

This article provides an overview of recent DL approaches for various optical wireless setups such as multi-colored systems and on-off keying (OOK)-based OWC. Subsequently, a convolutional AE (C-AE) structure is proposed for image sensor communication (ISC) where the information is conveyed by spatially separated LED arrays, and a receiver is implemented with an optical image sensor. Finally, concluding remarks and implementation challenges for DL-based communication systems are addressed.

Convolutional Autoencoder

In the C-AE, several hidden layers are implemented with convolutional layers, which have proven powerful in handling a 2D image input.

this can be viewed as a 2D convolution operation that slides a 2D window filter of the same weights over the input matrix. This computation helps extract spatially correlated features of the image input such as edges and lines.More complicated features can be learned with the aid （帮助） of multiple convolution filters having different weights, which provide several 2D output matrices.A pooling layer can be added to a convolutional layer to reduce the output dimension by sampling one element over the predefined 2D region. With the pooling layers, NNs become robust to minor spatial changes （微小的空间变化） in the input image.

At the encoding network of the C-AE, the message is first mapped to a one-hot vector, which is a zero vector except for the bth element equal to 1, and then is processed by several fully connected layers.To yield a 2D OOK intensity matrix, convolutional layers are adopted to an output vector of the fully connected layers by reshaping into a matrix. Each element of the output matrix of the encoding network is mapped to the OOK transmit intensity of each LED.

Through the optical channel, which contains the signal-dependent noise as well as random image rotation and blur effects, the receiver obtains 2D images capturing the transmit LED array at each data transmission.

The decoding network, which includes multiple convolutional layers followed by fully connected layers, retrieves the transmitted message from the received image.

For implementing the OOK modulation, the parameterized sigmoid activation is adopted at the end of the encoding network with the aid of the annealing-based multi-stage training strategy （基于退火的多阶段训练策略）

 

Visible Light Communication Using Deep Learning Techniques

Deep learning (DL) techniques have the potential of making communication systems more efficient and solving many problems in the physical layer.

An autoencoder (AE) is an unsupervised DL technique (AE是unsupervised learning？) and is considered in this work to realize the data recovery process that includes the VLC transmitter, a layer characterizing the optical channel and the VLC receiver. The model is trained to minimize the error in recovery of the transmitted data symbols.

An AE is a type of a deep neural network (NN) which comprises of an encoder and a decoder and serves the purpose of dimension reduction by making use of unsupervised learning. Encoder is a component of AE which compresses （压缩） the given input into few number of bits. Space represented by these compressed bits is called latent space（隐空间）. The decoder is a component of an AE which reconstructs the input using the compressed bits in the latent space. AE avoids features which are redundant （冗余） and include only the features which the model needs. It enables capturing statistical dependencies between different elements of the signal and optimizes the loss while reconstructing the signal using stochastic gradient descent (SGD，随机梯度下降) algorithm during backpropagation

 

A Model-Driven Deep Learning Method for LED Nonlinearity Mitigation （缓解） in OFDM-Based Optical Communications

The nonlinearity of light emitting diodes (LED) has restricted the bit error rate (BER) performance of visible light communications (VLC). In this paper, we propose model-driven deep learning (DL) approach using an autoencoder (AE) network to mitigate the LED nonlinearity for orthogonal frequency division multiplexing (OFDM)-based VLC systems.

First, a deep neural network (DNN) combined with discrete Fourier transform spreading (DFT-S) is adopted at the transmitter to map the binary data into complex I-Q symbols for each OFDM subcarrier. Then, at the receiver, we divide the symbol demapping module into two subnets in terms of nonlinearity compensation and signal detection, where each subnet is comprised of a DNN. Finally, both the autocorrelation （自相关） of the learned mapping symbols and the mean square error of demapping symbols are taken into account simultaneously by the cost function for network training.

Recently, optical orthogonal frequency division multiplexing (O-OFDM) has been widely employed in VLC system due to its high spectral effciency and resistance to inter symbol interference (ISI) resulting from the optical diffuse （弥漫的） channel

in this paper, we formulated the LED nonlinearity mitigation problems as a DL task, and proposed an AE-based nonlinearity mitigation scheme, abbreviated as NC-net, for the OFDM-based VLC systems.

 

MIMO detection using a deep learning neural network in a mode division multiplexing optical transmission system

In this paper, a novel MIMO detector has been designed using a supervised Deep Learning Neural Network (DLNN) and has been implemented successfully in a Mode Division Multiplexing (MDM) optical transmission system.

Existing communication systems are designed based on splitting the entire signal processing into a chain of multiple independent blocks. In such a system, each block performs a pre-defined function. These functions, such as source coding, channel coding, modulation, channel estimation, and equalization, are completed isolated from each other. This technique of individual block optimization does not provide the best end–end performance. However, DLNN can do so even in complex communications scenarios that is difficult to represent using tractable mathematical models （易处理的数学模型）.

In this paper, a novel MIMO detection technique based on Deep Learning Neural Network (DLNN) has been designed and implemented successfully in an MDM optical transmission system.

这篇讲基于DL的MIMO detcetion的，感觉用处不大

 

Self-Taught Anomaly Detection With Hybrid Unsupervised/Supervised Machine Learning in Optical Networks

这是一篇Highly-Scored Paper，要好好看看了。

This paper proposes a self-taught anomaly （异常） detection framework for optical networks.上面有一篇综述对DL在光通信1的应用作了挺多介绍的，其中这篇就是应该就是算Proactive Fault Detection（主动式故障检测）了

 

DeepRMSA: A Deep Reinforcement Learning Framework for Routing, Modulation and Spectrum Assignment in Elastic Optical Networks

深度强化学习与光通信结合。跟上面一篇论文一样，出自相同的作者，IEEE fellow。这两篇论文都会好好拜读