DISA

When? October 23, 14.00-16.00
Where? D2272 or online, https://lnu-se.zoom.us/j/61126230818?pwd=2ZJJtYeoTNmKaweBshZDLbN0BradBq.1&from=addon

Abstracts

Self-Adaptive Federated Domain Generalization for Privacy-Preserving Fault Diagnosis – Mehdi Saman Azari
Fault diagnosis is at the core of predictive maintenance, helping prevent costly breakdowns in industrial machines. Deep learning has made diagnosis more accurate, but it works best when a machine operates under the same conditions it was trained on. In practice, machines run at different speeds and loads, and they often meet conditions that never appeared in the training data. Since no data from these unseen conditions is available in advance, the model has to generalize to them without having seen them. One way to achieve this is to train on data from many machines, but gathering all data in one place conflicts with privacy, intellectual-property and cybersecurity requirements. Federated learning lets machines train a model together without sharing raw data, yet current methods lose accuracy compared with centralized training and cannot adapt on their own during operation.

In this talk, I present SAFDG, a framework that combines federated learning with domain generalization techniques and a self-adaptive control loop (MAPE-K). Federated domain generalization helps the model learn fault patterns that stay the same across working conditions, so it can handle conditions it has never seen without have a direct access to raw data of other machine. The control loop lets each machine monitor its own state and decide by itself whether to update its model locally, when it has enough data, or together with other machines, when it faces a new condition with too little data. Evaluated on four bearing datasets under 24 unseen conditions, SAFDG outperforms existing privacy-preserving methods and comes within 3 percentage points of centralized training.

Visual Analytics for Large Language Model (LLM) Evaluation – Darius Coelho
As large language models (LLMs) continue to advance, evaluating their reasoning at scale remains a key challenge. We introduce a visual analytics framework that combines automated evaluation with interactive exploration to help researchers assess model performance. By supporting simultaneous analysis across multiple models and benchmark datasets, our approach enables scalable comparisons, pattern discovery, and targeted case investigations to better understand LLM behavior.