DC6 Project: GPU-accelerated Edge computing for Federated Learning reasoning in industrial environments    

Doctoral Candidate

Alexandre Niyomugaba, MSc

Alexandre Niyomugaba holds a Master's degree in Embedded and Mobile Systems from The Nelson Mandela African Institution of Science and Technology (NM-AIST). He has a background in intelligent systems design and academic contribution. His expertise spans AI-based reasoning systems such as IoT systems and federated learning, applied to monitoring, anomaly detection, and fault prediction, contributing to edge computing advancements, closely aligned with the goals of the TUAI project

Alexandre_Niyomugaba_Self-presentaion.pdf

Main Supervisor: Dariusz Mrozek (SUT)

Co-supervisors: Francesco Piccialli (UNINA), David Camacho (UPM), Jia-Chun Lin (NTNU)

R&D cooperation: AIUT

Objectives: develop Federated Learning (FL) framework with direct GPU acceleration at the edge devices, enabling efficiency training, updating and testing the prediction, and reasoning processes that are performed at the Edge of a network.

The GPU-accelerated Edge-based Federated Learning (FL) approach will be applied to enable collaborative model training across multiple decentralized industrial AGV edge-based without sharing the raw, sensitive local data. This will make the models more robust, privacy-preserving, and relevant to each local environment, while keeping the data traffic and latency minimal. FL will serve as an orchestrator to ensure that model updates from different Edge nodes are efficiently aggregated into a high-quality global model while maintaining data privacy, and that the resulting global model provides fast, accurate, and reliable inferences locally. Measures of model convergence, data heterogeneity, and communication overhead are key to evaluating these aspects and to ensuring that the solutions are both scalable and secure. Research challenges include addressing the limitations in the performance and efficiency of distributed anomaly detection and fault prediction industrial models when deployed in federated industrial edge networks where resource awareness is critical. 

Expected Results: include a novel programmatic library for AI-based inferencing within the distributed industrial environment of AGV vehicles with Edge modules that operate on automated production lines. Decreasing the amount of data that is sent in industrial environments that requires early anomaly detection and fault prediction via AI-based reasoning. Accelerating the detection and prediction processes to support real-time predictive maintenance in smart manufacturing.

Applied research: The project will focus on optimizing Edge IoT services for effectively processing the sensor data by GPU-based processing including the data accumulators used in the event of a network disconnection will be verified by industrial projects proposed by AIUT company with focus on various approaches, multi-threading arrangements and GPU memory types for the optimal execution of the analytical processes for industrial data.

Planned secondments: UNINA(4 months); UPM(4 months); NTNU (4 months)                   

Enrolment in Doctoral degree: SUT