This project was developed as my bachelor's degree final thesis. Detailed information can be found in the pdf document.
- Technologies(#technologies)
- Motivation
- Data Collection
- Data Validation
- Data Generation
- Results
- Further Steps
- SRSLTE
- MGEN
- Python 3.8
- PyTorch
To satisfy the demands of 5G systems, there is a tendency to Network Functions Virtualization (NFV) and Software Defined Networks (SDN). The integration of Artificial Intelligence (AI) in this paradigms, such as C-RAN, will make use of Artificial Intelligence techniques to make short time scale predictions of UE SNR levels, among other channel state parameters. However, it can be challenging to properly train the AI models without sufficiently large and representative datasets. This work analyzes how to enlarge said datasets through advanced learning techniques, allowing an optimal analysis of the SNR properties even when the available samples are much lower than the real ones, eventually leading to more scalable algorithms.
It is also possible to extrapolate this work to other fields as the Data Generation part is general for any time series (with an appropriate hyperparameters tuning).
Retrieving data from a LTE base station is a rather trivial process in a real scenario, but having access and permits to them is not that much. And even if they are accessible, the data will correspond to multiple users and terminals, so keeping track of a particular one is a costly procedure.
Moreover, datasets of controlled metrics of individual users are scarce . Thus, for the purpose of this project, original datasets are generated. The metrics of interest in this datasets are the signal to noise ratio (SNR), the modulation coding scheme (MCS), the decoding time, and the buffer state report (BSR).
The available testbed for this study is based on the open source LTE library srsLTE, which provides a fully working Software-Defined Radio User Equipment (srsUE), an eNodeB (srsENB), and an Evolved Packet Core (srsEPC).
The testbed consists in two different Ubuntu 16.04.6 LTS devices, which, both have srsLTE installed, and their network cards are connected by a RF cable. One of the devices runs srsUE to work as a user’s terminal, and the other device runs both srsENB and srsEPC to work as the rest of the LTE network. In this work the data is collected in the second device as the scope of the study only covers the uplink parameters.
The request of the transmission gain modification is send by executing the script set_txgain.py. The open source software MGEN is used to generate traffic. It creates real time traffic patterns that can be configured as desired, as later is broadly discussed. Another tool used is iPerf, which utility is to measure the maximum bandwidth and packet loss, among other parameters. It is therefore very useful to analyse the functioning of the testbed.
The crucial characteristics to understand the testbed behaviour for this work are the maximum bandwidth and the performance of the signal to noise ratio at the receiver with respect to the transmission gain set at the emitter. Hence, this study only considers an uplink channel (and another downlink channel), and a given distance between the UE and eNB antennas, it is possible that the measured characteristics differ when this conditions change.
In order to know which is the maximum available bandwidth, the tool iPerf is used. The procedure to achieve that goal consists in executing the srsue command in the computer 1, and the srsepc and srsenb commands (in two different consoles) in the computer 2. After the UE is successfully attached and the connection is established, an iPerf UDP server is ran in another console in the computer 2 and an iPerf UDP client is executed in the computer 1. In the mentioned iPerf client, the bitrate is sequentially set to increasing values every try. As the client bitrate increments, the received bitrate at the server reaches a boundary of approximately 15 Mbps. This implies that when the client bitrate is set to values above that boundary, the packet loss increases proportionally.
From that experience it is safe to say that the maximum achievable bandwidth is around 15 Mbps, and that the packet loss due to bandwidth overflow is negligible if the bitrate is set to values below 12 Mbps.
Both computers of the testbed (with their respective peripherals) are static and always in similar conditions. Thus, for the signal to noise ratio (SNR) at the receiver to change, the transmission gain at the srsUE device has to be modified. In this case the transmission gain can be set in the srsUE configuration file as it remains constant during the capturing periods, but by convenience it is set using the Python code fixed_txgain.py before capturing.
The ground truth data has to be diverse enough so that it is possible to check that the synthetic data can be generally valid. For that purpose, the real dataset is composed by a combination of eight traffic patterns and nine transmission gain patterns.
This pattern generates messages of a fixed sized and send them in a given interval that varies following a Poisson distribution. In this work the packet size is set to the maximum possible for UDP traffic: 8192 Bytes, and the sent packets per second is set to 31, 77, or 153, corresponding to 2 Mbps, 5 Mbps, or 10 Mbps, of uplink bitrate.
As the name of this pattern announces, it generates bursts of messages of another kind of MGEN pattern at a given average time interval. The distribution of this time interval can also be chosen as exponential or uniform. The bursts last a time interval that can also be chosen fixed or random. In this study the configuration used is: bursts appear following a exponential distribution with mean 5 seconds, the pattern of the bursts is a Poisson pattern with the same parameters explained in the previous point, and the duration of the bursts follows a exponential distribution with average on 3 seconds.
This pattern uses a tcpdump binary file and "clones" it, meaning that takes the packet sizes and its correspondent timestamps from the tcpdump file and generates a traffic with the same characteristics. In this work the tcpdump files are created using WinDump, which is the equivalent of tcpdump for windows, capturing the traffic of video-calls using two different applications: Skype and Hangouts.
In a real scenario signal to noise levels may vary over time as the user terminal approximates to the base station or as they get farther away, interference may also not be stationary, or even the climatic conditions can effect. Hence, in this case the physical conditions of the testbed cannot be changed to simulate these real scenarios, the transmission gain is modified for that purpose.
The Python code used to modify the transmission gain following these patterns is in the file set_txgain.py
This pattern tries to simulate the approaching and distancing of the UE and eNB. It does so by following the form of a stepped triangular wave. The transmission gain is initialized within the range of gain levels, and every 10 seconds the gain increases by 1dB, until it reaches the maximum level of the range, in which case, starts to decrease by 1dB every 10 seconds, until it reaches the minimum value of the range. Then the process repeats until it is manually stopped.
This pattern emulates the sudden appearance (or disappearance) of interferences. It consists in, every 10 seconds, increasing the transmission gain by a integer value from −5dB to 5dB that is generated randomly following a uniform distribution. If the new transmission gain exceeds the upper or lower boundary of the range it is set to that boundary.
This simulates a more stationary scenario where the previously discussed conditions do not have a mayor impact on the SNR, meaning that the transmission gain is set before starting capturing, and remains at the same value until the end of it.
The capture procedure consists of 4 different steps:
- Connect srsLTE
- Set transmission gain
- Send MGEN traffic
- Monitoring
More details in the report.
Before working with the dataset to generate synthetic data, it is crucial to determine the possible correlation between the different parameters of the collected data as it can facilitate the task by generating parameters together, or in the absence of this correlation it can difficult the process.
In this case, the parameters of interest are mainly the signal to noise ratio (SNR), the buffer status report (BSR), the modulation and coding scheme (MCS), and the decoding time. These parameters are referred to the uplink, as are worthier of been synthetically generated than been captured, in contrast with the downlink case.
Plots and descrpition of scenarios in the report.
In this work, a Wasserstein generative adversarial network is used. It is formed by a CNN generator and a CNN critic, due to the fact that these neural networks take into consideration the time dependency between samples and the relationship between the different input parameters. RNNs were also tested, as they also fulfill these requisites, but their performance was poor due to the impossibility of using GPUs to run them in the last PyTorch release.
The stability of traditional GANs is overcome by Wassertein generative adversarial networks (WGAN), thus this algorithm is chosen over the traditional one for the study. The WGAN algorithm is well explained in this article and its main difference with the traditional GAN algorithm is that there is no longer a discriminator model but a critic model. This model has to be trained till optimality so that it avoids modes collapse. For this work the implementation of WGAN is made following the model in here. The architecture of the model is the same as the one depicted in 4.1, the only change is that the "discriminator" module is in this case named "critic". The full configuration of the generative model is in
Every combination of recurrent neural networks and convolutional neural networks, as generator or critic, is tried. The best results are achieved by the utilization of CNN as generator and critic, and are the ones presented in this section. This chapter considers a visual comparison between the synthetic data (generated after at least 4 hours of training) and real data. Additionally, it is compared the mean root mean square (RMS) and mean peak to average ratio (PAR) of 100 generated time series with 100 slices of the real time series.
Tables and analysis in the report.
This work purposes a WGAN model to generate different mobile network parameters as SNR, BSR, or decoding time. Although promising results are obtained for said parameters, some limitations are addressed as the model struggles with emulating all metrics simultaneously and low variance time series.
It is possible to say that the combination of CNNs as generator and critic performs better than any combination with RNNs, as they lead to a less efficient (and therefore slower) trainings.
Generally, parameters generated using datasets of Clone patterns outperform the other cases in being similar to real data. This is a positive aspect given that Clone patterns are the nearest to real use cases: they replicate actual traffic scenarios. The SNR patterns have not shown to be very useful as the sudden changes of the SNR values cannot be simulated with the purposed generative model. However, by using them, more diverse scenarios have been created for the other parameters. Smoother evolving SNR patterns might be more appropriate, but the repercussion of high frequency requests of transmission gain modification should then be studied in detail.
It would be helpful to have a higher time resolution original dataset, in order to analyse profoundly the time evolution of the metrics. Other further steps may include the utilization of peak to average ratio metric in the training process, the implementation of a WGAN with gradient penalty, or the utilization of automated hyperparameters tuning algorithm.
Machine Learning models based on: https://github.com/proceduralia/pytorch-GAN-timeseries/tree/master/models
Contact author: Mario Rodríguez Ibáñez [email protected]
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent