Year: 2025
Type: article
Abstract: The research seeks to establish a Project Management Framework for the Implementation of AI-Driven Fault Mitigation Systems by utilizing quantitative methodologies such as machine learning, in conjunction with insights from telecommunications engineers. This study integrates empirical data with practical expertise. The research was conducted using data obtained from Savanna Fibre Limited, resulting in a dataset comprising fault logs. Hybrid AI models, specifically a combination of CNN-LSTM with certain physics-based modifications, were trained and tested to detect faults at an early stage, identify the nature of the problems, locate the source of the issues, and assess the potential severity of the faults. Data preprocessing pipelines were developed to tackle challenges including imbalanced classes and sparse records of submarine cable faults, while domain knowledge also played a crucial role in guiding feature engineering and model interpretation. The framework demonstrates impressive performance: it achieves a 94.3% F1-score in fault classification, forecasts issues up to 72 hours ahead with a 92% confidence level, and accurately identifies fault locations within ±25 meters. To enhance its practicality, a versatile deployment configuration integrates model outputs into real-world workflows through CI/CD pipelines and even utilizes AR tools to assist in field repairs, resulting in a 42% reduction in repair times during actual tests. This research indicates that AI-driven, proactive maintenance is not merely theoretical; it is achievable with the appropriate data, interdisciplinary collaboration, and practical testing. Looking forward, there is significant potential to expand this approach for 5G and IoT networks or to refine our management of uncertainty in critical systems.
Source: Machine Learning Research
Authors: Isaac Buyondo, Konstantinos Kiousis
Institution: Berlin School of Business and Innovation GmbH
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7108203960