| KEYWORD |
Artificial Intelligence and development of Foundation Models for the Diagnosis of Rotating Machinery via Self-Supervised Learning
keywords ARTIFICIAL INTELLIGENCE, MODELLING AND EXPERIMENTAL TESTS
Reference persons CRISTIANA DELPRETE
External reference persons Ing. Luigi Di Maggio (assegnista di ricerca post-doc)
Research Groups 20-Industrial Systems Engineering and Design
Description The application of Artificial Intelligence (AI) in the field of rotating machinery diagnosis offers significant benefits, but also faces significant challenges. Among these, the difficulty of acquiring labeled data and the limited ability to generalize existing models in different scenarios or on new types of machinery emerge. We also observe a paucity of data regarding damaged machines compared to that available for machines in normal operating conditions.
AI is currently in a phase of rapid evolution, with applications spanning a variety of fields. The most advanced and impactful AI models (e.g. GPT-4) are based on large datasets, but it is mainly their “self-supervised” approach that makes the use of such large datasets practical. Without these self-supervised learning techniques, managing and effectively processing extremely large datasets would be quite complex and likely impractical. This paradigm today seems to offer highly generalizable models, often described as "foundation models" due to their ability to act as a general basis on which to develop more specialized models.
Thesis Objectives:
• Investigate the feasibility of building “foundation models” for the diagnosis of rotating machinery, applying self-supervised learning techniques;
• Explore the ways in which these models can be adapted (fine-tuned) to specific machines, using a limited number of data, mainly relating to machines in normal operating conditions.
Methodology
1. Theoretical study: an in-depth analysis of existing models to understand the self-supervised approach.
2. Data collection and analysis: identification and acquisition of datasets suitable for the diagnosis of rotating machinery. The data acquired concerns an industrial-sized test bench, available at the Mechanics Laboratory.
3. Development of a foundation model: hypothesis for the construction of a self-supervised model capable of learning from unlabeled data or data with limited labels.
4. Creation of a prototype model that applies the principles of self-supervised learning on a small scale in the diagnosis of machinery, in order to demonstrate in a practical way the theoretical concepts and the feasibility of the proposed approach.
5. Fine-tuning on specific machines: test the fit of the model to different machines to optimize the diagnosis.
6. Evaluation and comparison: measure the performance of the model compared to existing methods, analyzing advantages and limitations.
Conclusion and future prospects:
• Summary of the results obtained and discussion on how foundation models can significantly influence the field of machine diagnosis;
• Exploration of potential future developments and applications in other industrial and research sectors.
Deadline 11/09/2024
PROPONI LA TUA CANDIDATURA