Développement d'un système intelligent de détection précoce des défaillances dans les réseaux électriques urbains à l'aide de l'apprentissage fédéré et de l'internet des objets.

Project Abstract

This study presents a novel framework for early fault detection in urban electrical distribution networks by integrating federated learning (FL) and the Internet of Things (IoT) to enhance reliability, scalability, and data privacy. The proposed system deploys a hierarchical sensor network across substations, feeders, and consumer nodes to continuously monitor electrical parameters such as voltage, current, temperature, partial discharge, harmonic distortion, and switching transients. By leveraging edge devices for local feature extraction and model updates, the framework minimizes raw data transmission, thereby addressing latency and privacy concerns while preserving data sovereignty across utilities and third-party partners. The federated learning architecture enables collaborative model training across distributed nodes without sharing sensitive raw data, using secure aggregation and differential privacy to mitigate inference risks and model inversion attacks. The core contribution lies in developing a multi-task FL model that detects incipient faults, such as insulation degradation, transformer oil aging, switchgear warming, and cable moisture ingress, with calibrated uncertainty estimates to support risk-based maintenance decisions. To cope with heterogeneous data distributions and non-IID inputs common in urban grids, the study designs personalized sub-models with a shared global backbone, adaptive aggregation strategies, and transfer learning techniques to quickly adapt to new feeders and environmental conditions. The data fusion strategy combines time-series analytics, spectral features, and spatial correlations across neighboring nodes to improve early warning performance beyond single-point diagnostics. A robust evaluation methodology is developed, including synthetic data generation, real-world pilot deployment in a metropolitan area, and cross-validation against historical fault records. Performance metrics encompass detection latency, true and false positive rates, remaining useful life (RUL) estimation accuracy, and energy efficiency of edge computing operations. The system is validated under various scenario conditions, including extreme weather events, grid reconfiguration, and equipment aging, to demonstrate resilience against cyber-physical threats and communication outages. The research also investigates governance and privacy considerations, proposing a framework for secure multi-party computation, policy-compliant data sharing, and explainable AI to provide interpretable fault indicators and maintenance recommendations to grid operators. From an operational perspective, the solution aims to reduce unplanned outages, extend asset life, and optimize maintenance scheduling, while enabling scalable deployment across large urban networks through modular software microservices, containerization, and cloud-edge orchestration. The study contributes to the body of knowledge on federated analytics in critical infrastructure, offering practical guidelines for integrating FL with IoT sensing in power systems, and establishes a pathway toward autonomous condition-based maintenance that aligns with smart grid modernization objectives. The anticipated impact includes improved grid reliability, enhanced situational awareness for operators, potential cost savings from targeted interventions, and strengthened data privacy protections for stakeholders.

Project Overview

What This Project Is About

---

A straightforward overview of building a smart system to spot early signs of problems in city electrical grids using two modern ideas: learning from data across many devices without sharing raw data (federated learning) and connecting everyday devices (the Internet of Things or IoT).

---

---

The Problem It Addresses

---

The current grids can fail suddenly, causing outages and costly repairs. Traditional methods rely on data from a single location, which may miss early warning signs. This project explores how distributed data and continuous monitoring can help detect issues sooner and reduce outages.

---

---

Objectives of the Project

---

1. Understand the basics of power grids and common failure modes.
2. Learn what federated learning is and why it helps protect data privacy.
3. Explore how IoT sensors collect grid health data.
4. Develop a simple model to predict faults using distributed data.
5. Test the approach on simulated or open data to show potential improvements.

---

---

What You Will Do Step by Step

---

1. Review background material on power systems, ML basics, federated learning, and IoT.
2. Design a data collection plan with IoT sensors or public datasets.
3. Implement a basic federated learning setup across multiple data sources.
4. Build a lightweight fault-prediction model and evaluate its accuracy.
5. Analyze privacy, communication, and scalability considerations.
6. Run experiments comparing centralized vs. federated approaches.
7. Document results and discuss practical deployment steps.

---

---

Expected Outcome

A validated, privacy-friendly fault-prediction framework that uses federated learning with IoT data, showing potential reductions in outage duration and improved proactive maintenance planning.