High latency incident detection with machine learning : A case study of an LTE private network in a port terminal
Ibrahim, Amr (2025)
2025
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-04-11
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202504103587
https://urn.fi/URN:NBN:fi:tuni-202504103587
Tiivistelmä
Reliable and low-latency wireless communication is a cornerstone for industrial environments nowadays, especially in an automated port terminal, where disruptions and high latency can impact operational efficiency and safety. Thus, this thesis work investigates the capability of machine learning to detect high-latency peaks on the device level so that the Nokia operations team can be informed early about them and hence troubleshoot the network as early as possible before dramatic impacts occur.
This machine-learning problem is a supervised classification-learning problem. LTE Key Performance Indicators (KPIs) were used as training features, where these KPIs are measured on a cell level every 15 minutes in a private LTE network deployed in an automated port terminal. Conversely, the latency on a device level every 10 seconds was collected from the same network and utilized as the model label. The main challenge of the model training is the class imbalance of the label, where only 0.14% of the data samples contain high-latency labels. Class balancing techniques were applied to the data such as down-sampling the majority class and over-sampling the minority class. After that, linear, tree-based, and Support Vector Machines (SVM) training models were selected to train the data.
The final outcome was achieving 0.83 recall and 0.5 precision in detecting high-latency incidents. Furthermore, the features used by the tree-based algorithms during the training were analyzed. It was found that the cell-level average latency, the Physical Resource Block (PRB) usage, residual Block Error Rate (BLER), and re-transmission rate in the Radio Link Control (RLC) layer were the strongest features being used to detect high-latency short-time peaks on the device level.
This thesis is split into three parts. Firstly, the theoretical part in Chapter 2 and Chapter 3. Chapter 2 demonstrates the foundation of private cellular networks, types of deployments, applications, key enablers, and challenges. Moreover, this chapter illustrates the sources of latency in an LTE network and how it can be minimized in a private LTE network. In Chapter 3, data analysis, collection, and preprocessing, along with, types of machine learning, algorithms, model training, and evaluation are showcased. Secondly, in Chapter 4, the methodology followed in this work is explained in terms of the case study description and the tools utilized. Thirdly, the key outputs are reported in Chapter 5.
This machine-learning problem is a supervised classification-learning problem. LTE Key Performance Indicators (KPIs) were used as training features, where these KPIs are measured on a cell level every 15 minutes in a private LTE network deployed in an automated port terminal. Conversely, the latency on a device level every 10 seconds was collected from the same network and utilized as the model label. The main challenge of the model training is the class imbalance of the label, where only 0.14% of the data samples contain high-latency labels. Class balancing techniques were applied to the data such as down-sampling the majority class and over-sampling the minority class. After that, linear, tree-based, and Support Vector Machines (SVM) training models were selected to train the data.
The final outcome was achieving 0.83 recall and 0.5 precision in detecting high-latency incidents. Furthermore, the features used by the tree-based algorithms during the training were analyzed. It was found that the cell-level average latency, the Physical Resource Block (PRB) usage, residual Block Error Rate (BLER), and re-transmission rate in the Radio Link Control (RLC) layer were the strongest features being used to detect high-latency short-time peaks on the device level.
This thesis is split into three parts. Firstly, the theoretical part in Chapter 2 and Chapter 3. Chapter 2 demonstrates the foundation of private cellular networks, types of deployments, applications, key enablers, and challenges. Moreover, this chapter illustrates the sources of latency in an LTE network and how it can be minimized in a private LTE network. In Chapter 3, data analysis, collection, and preprocessing, along with, types of machine learning, algorithms, model training, and evaluation are showcased. Secondly, in Chapter 4, the methodology followed in this work is explained in terms of the case study description and the tools utilized. Thirdly, the key outputs are reported in Chapter 5.