International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 10 | Oct 2025
p-ISSN: 2395-0072
www.irjet.net
AI-Based Predictive Maintenance for Underground Power Cables Using Deep Learning Adhyay Mahesh Deshmukh1 1Student, Department of Electrical Engineering, Dr. Babasaheb Ambedkar Technological University, Lonere,
Maharashtra, India. ---------------------------------------------------------------------***--------------------------------------------------------------------interpretation, manual feature extraction, and extensive Abstract - In this paper, we present the use of Artificial Intelligence (AI) and Deep Learning (DL) for predicting maintenance of underground power cables. Traditional diagnostic methods, like Partial Discharge (PD) analysis, are common for finding insulation defects. However, these methods rely heavily on experts’ interpretations, making them slow and less effective in real-time situations. To tackle these problems, we use advanced DL models, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) models. These models automatically extract features from PD waveforms and Phase Resolved Partial Discharge (PRPD) patterns. They show higher accuracy, can adapt to noisy signals, and generalize better than traditional methods. This allows for better classification of fault types and more accurate estimation of defect locations. Field trials and studies using reflectometry underline their ability to cut downtime, improve maintenance schedules, and prolong the lifespan of power assets. Overall, using DL in PD diagnostics is a significant shift from reactive to predictive maintenance, helping to support reliable and cost-efficient underground power systems.
post-processing. These drawbacks limit scalability and realtime use, especially in large-scale distribution networks that require quick, automated decision-making. Recent studies have applied classical machine learning to PD classification, but these models are still sensitive to noise, data imbalance, and manually designed features.
Key Words: Artificial Intelligence, Deep Learning, Predictive Maintenance, Underground Power Cables, Partial Discharge, Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Condition Monitoring, Fault Diagnosis.
1.1 Problem Statement
This paper explores how to integrate deep learning (DL) models, particularly Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory (LSTM) networks, into a framework for predictive maintenance of underground cables. By learning hierarchical representations from raw PD waveforms and PRPD images, DL methods can lessen reliance on manual preprocessing, increase robustness against noisy environments, and generalize across various defect types. We present a methodology for data acquisition, preprocessing, model design, and evaluation, and we validate our approach using reflectometry-based localization and representative fieldlike datasets.
In this paper, we tackle the important issue of localized insulation defects in underground power cables. These defects lead to partial discharges (PD), which often happen before costly, unplanned failures. Traditional PD diagnostics rely on expert interpretation and manual feature extraction. This limits scalability and reliability in noisy, real-world conditions.
1.INTRODUCTION Underground power cables are essential for modern urban and industrial power distribution systems. They offer aesthetic, safety, and space-saving benefits compared to overhead lines. Despite these advantages, cable networks can develop hidden insulation defects over time due to manufacturing flaws, mechanical stress, moisture ingress, or electrical aging. These defects often show up as partial discharges (PD), which are transient localized ionization events that can lead to severe insulation failures. Therefore, finding PD activity early and accurately is crucial for preventing unplanned outages, cutting repair costs, and prolonging asset life.
Although deep learning (DL) methods have shown promise for automated PD classification, practical use is still held back by a lack of labeled field data, class imbalance, sensor quality issues, and placement sensitivity. There are also challenges with interpretability and inconsistent performance across different cable types and environments. This work presents a complete framework that merges DLbased PD detection and classification with reflectometrybased localization and data enhancement strategies. Our approach aims to improve early fault detection accuracy and provide reliable spatial localization for targeted repairs. It reduces the need for human experts and increases resilience to noise and changing cable conditions, all while keeping computational and explainability concerns in mind.
Traditional PD diagnostic methods, such as time-domain pulse analysis, phase-resolved partial discharge (PRPD) patterns, and reflectometry, have been effective in labs and real-world scenarios. However, they usually need expert
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 32