View Proposal

---

Proposer

John See

Title

(PV)³: Federated Learning for Privacy-Preserving Solar PV Defect Detection

Goal

The aim of this project is to implement a robust federated learning framework for solar PV defect detection. More extensively, the project will also investigate the feasibility of various aggregation strategies, heteogeneous models, and their impact to performance and deployment

Description

This project will establish a federated learning environment for solar PV panel defect detection, where different publicly available datasets are treated as independent clients with their own local data distributions. By way of simulation, each client will train a defect detection model locally and send model updates to a central server, where they will be aggregated using techniques (e.g. classical FedAvg) to improve the global model. The project will compare federated learning with conventional centralised training and investigate the impact of heterogeneous client data on detection performance. Project will potentially collaborate with an industry partner Airlytic.

Resources

There are open-source federated learning frameworks that can be utilised (though customisation level varies). Also require reasonable HPC compute, and a machine that can be used for demonstration of deployment. Publicly available datasets to be used as starting point (data from industry partner can be explored)

Background

Solar PV panels are increasingly deployed for renewable energy generation, but defects such as cracks, hotspots, delamination and damaged cells can reduce their efficiency and operational lifespan. AI-based computer vision methods can automate defect detection, but training effective models often requires large and diverse datasets collected from different sites and organisations. Sharing such data centrally may raise concerns regarding data privacy, confidentiality and ownership. Federated Learning (FL) provides an alternative by allowing different organisations or data owners to train models locally without directly sharing their datasets. Their locally trained models can then be aggregated by a central server to produce a more generalised global model. The availability of several publicly accessible but heterogeneous solar PV defect datasets provides an opportunity to simulate such a distributed environment and investigate whether federated learning can improve defect detection while preserving data locality.

Url

Difficulty Level

Moderate

Ethical Approval

None

Number Of Students

1

Supervisor

John See

Keywords

federated learning, defect detection, ml models

Degrees

Bachelor of Science in Computing Science