Experimental Evaluation of Hybrid Fibre–Wireless System for 5G Networks

This article describes a novel experimental study considering a multiband fibre–wireless system for constructing the transport network for fifth-generation (5G) networks. This study describes the development and testing of a 5G new radio (NR) multi-input multi-output (MIMO) hybrid fibre– wireless (FiWi) system for enhanced mobile broadband (eMBB) using digital pre-distortion (DPD). Analog radio over fibre (A-RoF) technology was used to create the optical fronthaul (OFH) that includes a 3 GHz supercell in a long-range scenario as well as a femtocell scenario using the 20 GHz band. As a proof of concept, a Mach Zehnder modulator with two independent radio frequency waveforms modifies a 1310 nm optical carrier using a distributed feedback laser across 10 km of conventional standard single-mode fibre. It may be inferred that a hybrid FiWi-based MIMO-enabled 5G NR system based on OFH could be a strong competitor for future mobile haul applications. Moreover, a convolutional neural network (CNN)-based DPD is used to improve the performance of the link. The error vector magnitude (EVM) performance for 5G NR bands is predicted to fulfil the Third Generation Partnership Project’s (3GPP) Release 17 standards.


Introduction
The fifth-generation (5G) of wireless networks aims to enhance the energy efficiency, end-to-end latency and capacity of networks. The radio access network has become more centralized as the number of base stations (BS) has increased exponentially [1]. Because it enhances scalability and decreases network maintenance, a centralized radio access network (C-RAN) reduces capital costs. A fronthaul (FH) connects baseband units (BBU) and distant radio heads in order to support a C-RAN. As shown in Figure 1, optical fronthaul (OFH) links, such as radio over fiber (RoF) links, constitute the mainstay in OFH. Due to its costeffective and easy paradigm, which improves the network's range, this transport approach is a significant option for high-capacity wireless transmission for transporting analogue or digital signals in OFH.

Figure 1.
Block diagram explaining backhaul connection to the baseband unit (BBU) connected to optical front haul (OFH), and then its transmission from the transceiver station. The base station further connects it to the remote radio heads (RRHs). Extending the network coverage via fibre and wireless infrastructure is shown w.r.t application scenarios such as picocell, stadiums, buildings, etc.
A-RoF suffers from signal degradation and has a network expansion range of only a few tens of kilometres. The reach of the network for D-RoF is efficient for longer lengths [11], but due to the greater resolution bandwidth of the analogue digital converter (ADC) and the requirement of additional signal processing the system's cost increases. In fronthaul networks, inadequate spectral efficiency and high demand for data traffic are also undesirable [12,17]. The Common Public Radio Interface (CPRI) bottleneck is bypassed in S-DRoF by employing sigma-delta modulation (SDM), which works as a 1-bit ADC. The high quantization noise is caused by a 1-bit operation, which necessitates the use of a bandpass filter on the receiver side, increasing the complexity of these systems. As a result, the A-RoF system is a better alternative than D-RoF and S-DRoF [18] since it is simple, inexpensive and already has a widespread infrastructure.
Because of the limitations imposed by the optical and microwave components, analogue fronthaul applications are limited. Many linearization methods have been proposed to address these issues [4][5][6][7][8][9], but these methodological solutions have been questioned due to limited bandwidth, high method complexity and the requirement of returning the output signal to the base station (feedback), which is a lengthy process [18]. Block diagram explaining backhaul connection to the baseband unit (BBU) connected to optical front haul (OFH), and then its transmission from the transceiver station. The base station further connects it to the remote radio heads (RRHs). Extending the network coverage via fibre and wireless infrastructure is shown w.r.t application scenarios such as picocell, stadiums, buildings, etc.
A-RoF suffers from signal degradation and has a network expansion range of only a few tens of kilometres. The reach of the network for D-RoF is efficient for longer lengths [11], but due to the greater resolution bandwidth of the analogue digital converter (ADC) and the requirement of additional signal processing the system's cost increases. In fronthaul networks, inadequate spectral efficiency and high demand for data traffic are also undesirable [12,17]. The Common Public Radio Interface (CPRI) bottleneck is bypassed in S-DRoF by employing sigma-delta modulation (SDM), which works as a 1-bit ADC. The high quantization noise is caused by a 1-bit operation, which necessitates the use of a bandpass filter on the receiver side, increasing the complexity of these systems. As a result, the A-RoF system is a better alternative than D-RoF and S-DRoF [18] since it is simple, inexpensive and already has a widespread infrastructure.
Because of the limitations imposed by the optical and microwave components, analogue fronthaul applications are limited. Many linearization methods have been proposed to address these issues [4][5][6][7][8][9], but these methodological solutions have been questioned due to limited bandwidth, high method complexity and the requirement of returning the output signal to the base station (feedback), which is a lengthy process [18].
The 5G network is expected to provide Gbit/s of connection, acting as a connector for total connectivity and coverage. Enhanced mobile broadband is one of the most widely utilised scenarios (eMBB). By utilizing revolutionary technological solutions, such as new radios (NR), 5G enables multi-services and a wide range of capabilities. Because it affects the coexistence with other systems, 5G NR enables the used multiplexing mechanism [19,20]. Multiple waveforms for multicarrier transmission have been proposed for 5G due to this coexistence capability (including 4G systems), including cyclic prefix-based orthogonal frequency division multiplexing (OFDM), filtered orthogonal frequency division multiplexing (F-OFDM) and generalised frequency division multiplexing (GFDM), which have been framed by the 3GPP for the 5G NR Release 15 that describes the eMBB scenario.

Problem Statement
In order to overcome the issues explained, the fibre-wireless (FiWi) technology based on RoF connections can be exploited to support 5G communications by jointly transporting all radio frequency (RF) signals and using the optical fibre features as an RF transport solution to accomplish the work items of 5G bands stated in 3GPP [21].
Multi-input multioutput (MIMO) is a conceivable method for most new mobile/wireless networks, which are pushed by today's users' high data rates. When compared to the performance of a single-input single-output (SISO) system, the MIMO approach is expected to improve transmission distance, data rate and reliability. As a result, it is a primary objective that the MIMO approach must be addressed when designing any FiWi system to enable wire free connection for next generation (NG) broadband networks.

Objectives
The experimental implementation of a 5G NR FiWi system employing an analogue optical transportation network built on RoF technology with a connection length of 10 km is shown in this study with the motivation to address wired and wireless link reach for the considered scenarios. Apart from this, a linearization technique is evaluated to enhance the performance. In eMBB situations, we implement 2 × 2 MIMO to increase the multiband 5G NR signal transmission range. As a proof of concept, the supercell scenario is accomplished at a 20 m distance using 3 GHz, while the femtocell scenario is realized at a distance of 20 m using the 20 GHz band. The following are the paper's key contributions: 1.
Combination of a 2 × 2 MIMO and an RoF-based hybrid FiWi system, as well as testing with a 5G NR transceiver at 3 and 20 GHz to show the multiband propagation and extend the wired transmission range through a wireless link.

2.
In a supercell situation, convolutional neural network (CNN)-based digital predistortion (DPD) is used to improve the performance.

3.
In line with the C-RAN architecture, experimental assessment is made of multiband 5G NR transmissions using a 10 km analog FH.
The remainder of the paper is divided into the following sections. The implementation of a MIMO 5G NR FiWi system is discussed in Section 2. Section 3 explains the experimental testbed followed by the results in Section 4, and the article is concluded in Section 5.

Literature Review
The utilizaThe RoF systems integrated with the fibre-wireless approach are discussed in this section including linearization methods for the performance improvement in RoF networks. As seen in Table 1, the linearization of OFH has been a major study topic. The desire for greater linearization to obtain higher performance has switched attention to machine learning (ML). Because machine learning is such a novel area, a current literature evaluation has been conducted for all methodologies, with a focus on machine learning methods. Table 1 summarizes a comprehensive literature evaluation for the alleviation of RoF system deficits.

5G New Radio Hybrid MIMO-Based Fibre-Wireless System
The proposed RoF system is integrated into the wireless system to obtain a hybrid fibre-wireless system. The proposed system has been explored utilising an eMBB scenario in this article. A supercell with a frequency of 3 GHz and an indoor femtocell with a frequency of 20 GHz are tested in this study.
The following are the four primary components of this experimental system: A. Generation of signal and its utilization in the system. B. Digital pre-distortion. C. Analog RoF-based OFH for the transmission of radio frequency (RF) signals. D. A fibre-wireless system with a total of 2 × 2 MIMO channels.

Signal Generation and Utilization
A Mach-Zehnder modulator (MZM) modulates a 1310 nm wavelength to drive a 1310 nm distributed feedback laser. The RF1 signal is applied to one of the MZM's arms by the first vector signal generator (VSG1). At 20 GHz, RF1 is a 5G NR signal. The 3 GHz flexible waveforms' (OFDM, GFDM, F-OFDM) signals from the 5G transceiver correspond to RF2. The biasing voltages VBIAS1 and VBIAS2 are chosen in such a way that the observed spectrum is free of the high peak to average power ratio and high-frequency peaks in the measured frequency selection.

CNN Based Digital Pre-Distortion
For VCSEL-based RoF connections [60] and DFB-based RoF links [53], DPD has been widely investigated. DPD is a useful tool for reducing nonlinearities in OFH connections. DPD has been proposed using a Volterra-based memory system and a generalised memory polynomial [1,2,60]. Machine learning approaches have recently been presented as a method for DPD. When compared to Volterra-based conventional designs, neural networkbased DPD appears to yield a considerably greater reduction in signal impairments. As a result, this research employs deep neural network-based DPD. Both carrier frequencies are analysed using DPD In the following subsection, a detailed explanation of the working principle of the proposed CNN-based DPD is discussed.

Convolutional Neural Network (CNN)-Based DPD Method
Convolutional neural networks (CNN) cascades the RoF connection that develops a RoF NN model and train it to imitate the original RoF link because the output of an RoF connection is known. After we have created the appropriate CNN DPD model, we can backpropagate through this RoF CNN and tweak the parameters. The first layer of a CNN is the convolutional layer. It extracts features from input data and generates a feature map with the help of a kernel (convoluted data).
A convolutional operation consists of the kernel or filter (kernel matrix), input data (input matrix) and feature map. Assuming that the investigated RoF connection has the H(n) function and y(n) as an output signal and that a baseband signal x(n) must be sent over it, DPD attempts to compute the inverse transfer function of this RoF link, represented byÎ −1 , whose output is then denoted byx(n).
This becomesx (n) =Î −1 (x(n)) while as G represents the gain. The CNN determines theÎ −1 used for pre-distortion in this case.
Because the idealx(n) is unknown, direct training for creating the CNN for DPD is not possible, as shown in Figure 2. and output as training data to determine the error and then use a loss function. We then reverse promote it through to train .

Training Algorithm
The CNN DPD model training pseudocode is given below in Algorithm 1. Firstly, we train an emulated RoF CNN model that uses the input and output of the original RoF connection, and after this model training is achieved, we connect the CNN to the actual RoF link to this (post distorter) DPD block and commence the pre-distortion procedure when this training is completed (see Pseudocode 1). It is represented in the pseudocode that the CNN is established on ( ) with updates on that is performed on the pre-distortion block used by the terms of ( ( )) as shown in Figure 2. The loss evaluation function is mean square error (MSE), the optimization function is ADAM, the activation function is ReLu, and the values are monitored via backpropagation. Z number of iterations are carried out to enhance the performance.

Radio Frequency Transport Using Analog Optical Fronthaul
A 10 km single-mode fibre (SMF) transmits the optical signal. The signal is subsequently retrieved using a photodetector (PD). The PD sends the received electrical signal to the diplexer (DPX) The diplexed signal must be separated, which is performed through amplification. The signals are then transferred to a 5G transceiver and a vector signal analyzer (VSA) for post-processing and performance testing. The second CNN first replicates the RoF link. For a generic RoF connection,x(n) is the input and y(n) G is the output. As a result of this, the CNN can now determine an estimated transfer function I. Once the RoF CNN model is created, the model weights are set, and it is then coupled to the CNN DPD model. We now utilise the original input, x(n), and output as training data to determine the error and then use a loss function. We then reverse promote it throughÎ to trainÎ −1 .

Training Algorithm
The CNN DPD model training pseudocode is given below in Algorithm 1. Firstly, we train an emulated RoF CNN model that uses the input and output of the original RoF connection, and after this model training is achieved, we connect the CNN to the actual RoF link to this (post distorter) DPD block and commence the pre-distortion procedure when this training is completed (see Pseudocode 1). It is represented in the pseudocode that the CNN is established on I(x(n)) with updates onÎ that is performed on the pre-distortion block used by the terms ofÎ −1 (x(n)) as shown in Figure 2. The loss evaluation function is mean square error (MSE), the optimization function is ADAM, the activation function is ReLu, and the values are monitored via backpropagation. Z number of iterations are carried out to enhance the performance.

Radio Frequency Transport Using Analog Optical Fronthaul
A 10 km single-mode fibre (SMF) transmits the optical signal. The signal is subsequently retrieved using a photodetector (PD). The PD sends the received electrical signal to the diplexer (DPX) The diplexed signal must be separated, which is performed through amplification. The signals are then transferred to a 5G transceiver and a vector signal analyzer (VSA) for post-processing and performance testing.

MIMO-Based FiWi System
After receiving the optical signal at the photodiode, it is then converted into an electrical domain signal followed by the duplex operation and is ready for reception on wireless channels. The duplexed signal is ready to be transmitted at the wireless channels after it is received at the photodiode. Because we have two bands of signals, MIMO with a dimension of 2 × 2 should suffice to capture the signal and assess the wireless link's performance.
As a result, the FiWi system contains two RRHs, each with its transmitter. As a result, the RX1 and RX2 receivers receive signals y 1 and y 2 that are represented as where x 1 is the first broadcast waveform and x 2 denotes the second, N 1 and N 2 denote the channel noise. With Tx i to Rx j , H ij is the channel coefficient. Intersymbol interference (ISI) and optical fibre chromatic dispersion add to the MIMO system's complexity. As a consequence, the frequency domain equalisation (FDE) technique is applied, resulting in MIMO complexity being reduced.

Experimental Setup
To validate the proposed FiWi 2 × 2 MIMO equipped with the CNN DPD technique, we employed a 5G NR scenario at 3 GHz and 20 GHz, both comprising of 50 MHz bandwidth as demonstrated in the previous work [23,24], but there was no integration of fibre-wireless, and also there was no DPD employed. The system represented in Figure 3 uses a dual drive Mach Zehnder modulator (MZM) with two distinct RF-driven signals and a 1310 nm DFB laser. The 5G transceiver transmits RF2, a 3 GHz flexible waveform signal, while the vector signal generator (VSG), labelled VSG, sends RF1, a 5G NR waveform at 20 GHz. The DPD procedure is divided into three steps.
Firstly, the up-converted signal at their respective carrier frequencies of 3 and 20 GHz passes over 10 km of standard single-mode fibre (SSMF) and is collected by a photodetector (0.69 A/W) that receives and converts the signal back to the electrical domain. Because the multiband must be separated individually, an amplification phase is implemented. A diplexer (DPX) separates the 3 GHz and 20 GHz frequencies. The signals are then passed via a series of vector signal analyzers (VSA). Each VSA output is sent to the post-processing block to be evaluated for performance. This phase is carried out without the use of DPD, which implies that the output is assessed without the use of DPD.
In the second phase, referred to as the DPD training phase, the DPD technique depicted in Figure 3 is applied, and training is performed until the error converges.
To put it another way, DPD backs up the idea underlying inverting the amplitude and phase responses recorded at the EA electrical amplifiers. Depending on the user's wants and comparative criteria, CNN approaches can be employed.
To achieve synchronization for the received waveforms, we employ the positioning reference signal (PRS) provided in the 5G NR framework (both input and output). The bandwidth of PRS is expected to be 20 MHz/106 resource blocks. The received and output reference broadcast signals are coupled in the time domain, and the power delay profile (PDP) is evaluated via threshold block to establish the first path of arrival [61].
The pre-distorted baseband signals are supplied into the DPD block in the third phase, where VSGs upconvert them to their carrier frequency before being fed into the optical link. Before being forwarded to the DPD training phase, the signal generated at the photodiode is passed via a diplexer DPX to isolate the different multibands. During the DPD validation stage, we flip the switches in the other manner. Pre-distortion and then transmitting the frames to the VSG complete the evaluation for 5G NR frames. We find that real-time operation in the adaptation is unnecessary since the nonlinearities of the RoF connection slowly change owing to thermal effects and component ageing. Table 2 lists the parameters used, which have previously been discussed in [1,2] and other state of the art [18] studies. 2 lists the parameters used, which have previously been discussed in [1,2] and other state of the art [18] studies. Figure 3. 5G NR fibre-wireless system experimental block schematic. The functions are as follows: A: Selection switch for DPD or direct post-processing for performance evaluation; B: switch selection for synchronisation block or a performance post-processing block. C: B is turned towards synch and connected to train block, while C will be used to synch the DPD trained version. D: Training follows the synchronisation block. DPD: The training is enabled/disabled by DPD. E: The method of time synchronisation (TS) is necessary so that Tx and Rx waveforms are in tight synchronization. F: Required for checking DPD inputs before transmitting to VSGs. CNN has been picked for DPD training.    Table 3 shows the CNN settings that result in optimised NN performance. The parameters that constitute the design of the planned convolutional neural network are listed in Table 3. A trial and error test is used to pick the parameters [18]. The table's last part computes the NN's expressions in terms of its coefficients to determine its complexity.

Experimental Results
The experimental findings of the tests deployed as described in the preceding section are explained. One way of measuring the accuracy of coefficient estimates for various designs is the mean square error (MSE). The MSE is 27 dB when no DPD is employed, whereas CNN reduces it to 20 dB.
The suggested method is compared and the magnitude of the errors is presented in the form of error vector magnitude (EVM). We utilise GMP techniques as a baseline architecture to compare with our recent work [18].

Error Vector Magnitude
The standardized performance metric in 3GPP is the error vector magnitude (EVM), which gives the optimum constellation location for each received symbol. The EVM value of the device is the root mean square (RMS) of all error vector magnitudes between the received symbol positions and their nearest ideal constellation sites.
where M gives constellation symbols, S m represents the constellation's real symbol associated with the symbol "m" and S 0,m denotes the real symbol linked with S m . 3GPP has set the EVM standard limit for the 256 QAM modulation rate to be 3.5% [62,63]. The EVM is displayed in Figure 4a for flexible waveforms without DPD, with the GMP DPD method followed by the CNN DPD method. For all flexible waveform designs, it is notable that the CNN DPD approach achieves a better reduction than GMP. meeting the 3GPP criteria, which means that the dynamic range is affe becomes fundamentally important to improve the performance via D that EVM is within the limits from −8 to −6 dBm. If we consider the Fthe dynamic range in the 3GPP limit is 2 dB, whereas F-OFDM with D the dynamic range to 7 dB from −12 dBm to −6 dBm within 3GPP limit   In Figure 4b, we compare a case called back-to-back (B2B) where a laser and a photodiode are connected via a patch cord of less than 1 m SMF. It can be seen that B2B has 2% EVM, which signifies that this is the least EVM obtained if we remove the presence of fibre or wireless integration into the system. Similarly, we have added the results of a generic RoF case from our previous study [18] and compared it with the proposed hybrid FiWi link. It is noticeable that the hybrid FiWi system has a 5% EVM. The performance of the system is further enhanced using CNN-based DPD, bringing the EVM value to 3.5% Additionally, we show EVM results for the received optical power for the F-OFDM and 5G NR waveforms in Figure 5. The received optical power for the 5G NR waveform is not meeting the 3GPP criteria, which means that the dynamic range is affected. Therefore, it becomes fundamentally important to improve the performance via DPD. DPD ensures that EVM is within the limits from −8 to −6 dBm. If we consider the F-OFDM waveform, the dynamic range in the 3GPP limit is 2 dB, whereas F-OFDM with DPD has extended the dynamic range to 7 dB from −12 dBm to −6 dBm within 3GPP limits.    In addition to EVM, the performance is also measured in terms of power spectral density (PSD). The PSD measures the ratio of power in the useful band to the power in the adjacent bands. It is also known as adjacent channel leakage or power ratio (ACLR/ACPR). The trends in Figure 6 show that the PSD for a case without DPD has a −25 dBc value. The 3GPP has set these limits to be −35 dBc. To improve the performance, DPD with GMP and CNN is utilized and it is signified that CNN improvement is better than GMP, which has been discussed in [18].

PEER REVIEW 12
In addition to EVM, the performance is also measured in terms of power spectral density (PSD). The PSD measures the ratio of power in the useful band to the power in the adjacent bands. It is also known as adjacent channel leakage or power ratio (ACLR/ACPR). The trends in Figure 6 show that the PSD for a case without DPD has a −25 dBc value. The 3GPP has set these limits to be −35 dBc. To improve the performance, DPD with GMP and CNN is utilized and it is signified that CNN improvement is better than GMP, which has been discussed in [18]. In Table 4, the EVM and PSD at 0 dBm are expressed for the DPD utilized.  In Table 4, the EVM and PSD at 0 dBm are expressed for the DPD utilized.

Conclusions
This article validates an unprecedented DPD controlled 5G-NR-based hybrid fibrewireless system equipped with 2 × 2 MIMO. A 10 km SMF is used to send a 5G NR signal at a distant radio head targeting an eMBB scenario with a femtocell (3 GHz) and a supercell (20 GHz) for eMBB applications. The hybrid fibre-wireless system via the proposed CNNbased DPD achieves a significant performance that meets the 3GPP specifications. The proposed CNN DPD approach reduces EVM to 3.2% and PSD to −41 dBc. The integration of fibre-wireless has great potential, and the experimental results suggest that this can be a potential solution for the transport and access network of future communication networks. Future work includes the optimization of fibre-wireless infrastructure and improvement in the linearization methodologies for better performance.

Data Availability Statement:
The data presented in this study are available on request from the corresponding author.

Conflicts of Interest:
The authors declare no conflict of interest.