Title: Advances in Data-Driven Fault Diagnosis via Distributionally Robust Optimization and Federated Learning

Date: 04/03/2026

Time: 9:30 am – 11:30 am ET

Location: Online at https://gatech.zoom.us/j/97646188995

Committee

1. Dr. Nagi Gebraeel (Advisor), H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
2. Dr. Weijun Xie, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
3. Dr. Yao Xie, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
4. Dr. Kamran Paynabar, H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology
5. Dr. Sameer Vittal, Department of Data Science & Reliability, GE Vernova

Abstract

Original Equipment Manufacturers (OEMs) of capital-intensive assets often seek to offer Machine Learning (ML)-driven fault diagnosis services to their clients as part of lucrative long-term service contracts (LTSCs). While it can be modeled as a classification problem, this task involves unique challenges, which can be divided into Structural Constrains and Data Challenges. The Structural Constrains are caused by the distributed nature of the training data, which cannot be shared from the clients to the OEM (server) due to privacy and bandwidth constraints. Additionally, the local distributions governing the features and labels at each client are heterogeneous, and the OEM does not have prior knowledge of all the unique fault classes observed across clients. The Data Challenges are caused by data collection in industrial settings, which can lead to highly uncertain data with a large number of classes and high imbalance. In this thesis, we aim to leverage techniques from Distributionally Robust Optimization (DRO) and Federated Learning (FL) to address all the challenges of this problem.

Chapter 2 focuses on the utilization of DRO to address the Data Challenges of our problem locally at each client. To that end, we develop a distributionally robust classifier that is characterized by a natively multiclass loss function. The use of DRO in our model allows it to hedge against uncertainty both in the data features and labels. Additionally, the implementation of a natively multiclass loss function allows for accurate modeling of all the fault classes while providing robustness against data imbalance.

Chapter 3 considers the problem of distributionally robust classification from an FL perspective, thereby providing the first step in addressing the Structural Constraints of our problem. However, we make  a key simplifying assumption at this stage. Namely, we assume that clients share a common class set. To that end, we develop a novel Mixture of Wasserstein Balls (MoWB) ambiguity set, which can be seen as a natural extension of the classical Wasserstein ball to the FL setting. We demonstrate that our MoWB ambiguity set allows for natural problem separability, enabling us to propose two distinct federated training algorithms for our model.

Chapter 4 addresses one of the key Structural Constraints of our problem--clients having different but overlapping class sets, while the server lacks prior knowledge of all the fault classes observed across clients. To that end, we propose a federated clustering approach that is unsupervised and does not require any data sharing. Our approach relies on clients performing local Expectation Maximization (EM) steps, and sharing local centroids and uncertainty set radii with the server. The server relies on overlaps between uncertainty sets to infer shared clusters between clients. Therefore, this approach allows the server to learn the class structure of the problem without reliance on client-provided labels.

Finally, Chapter 5 addresses both the Data Challenges and Structural Constraints simultaneously. More specifically, we propose a two stage approach for personalized, federated, distributionally robust classification. The first stage of our approach is a class structure inference stage, where the server can rely on a one-shot version of the method from Chapter 4 to learn the total number of unique classes and class membership. The second stage is a classifier training stage, where we propose a personalized version of the classifier from Chapter 3. In this approach, each client only contributes to the training of model parameters associated with locally-observed classes