Titulo Estágio
Development of active machine learning methods for prescriptive maintenance processes in Industry 4.0
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
CERN; DEI-FCTUC
Enquadramento
Prescriptive Maintenance (PM) processes in Industry 4.0 (I4.0) aim at detecting the onset of anomalous conditions, diagnosing their causes, predicting the remaining useful life of the machinery and proactively managing its failures. Fault diagnostics and prognostics results are used to guide effective condition-based and maintenance strategies for increasing productivity, optimizing operating performance, reducing lifecycle costs, extending operating periods between maintenance and reducing downtimes, frequency and severity of unanticipated failures.
Internet of Things enabled PM is a key opportunity in I4.0. It requires advanced Machine Learning (ML) capabilities for fault diagnostic and prognostics and must face two major challenges:
i) decision making at the asset level, which requires to anticipate the potential scenarios generated from the possible operation and maintenance actions, and to minimize their effects on the system operation;
ii) disclosing and transferring (massive) sensitive data to the cloud for analytics exposes data to third parties and requires administrative/legal burden to protect IP and ownership, thus leaving data in a limbo.
The emerging Federated Learning (FL) technology leverages federations of devices that are cooperating for training a machine learning model. FL addresses the issues of decision making at the asset level and data ownership, by ensuring that the data used for model training never leaves the industrial equipment responsible for its production.
Active Learning may be a better solution than Passive Learning (traditional ML) to the aforementioned problem. Instead of passively learning from a fixed dataset, the key idea behind active learning is to focus on labeling the instances that are most uncertain or informative to the model's current state. By selecting these instances strategically, the algorithm can learn more effectively with fewer labeled examples compared to traditional supervised learning approaches.
The objective of the thesis work project is the development of Active Machine Learning methods for the prediction of the remaining useful life of industrial machineries based on vibration analysis and equipment process data. FL technology based on CERN’s (European Organization for Nuclear Research) existing FL platform allows the networked machinery to identify the optimal prescriptive operation and maintenance policy at the system level, based on the failure time predictions obtained at the components level.
The goal is to showcase the benefits of the technology based on an evolved implementation of CERN existing FL platform. Due to confidentiality reasons established by CERN at this thesis proposal stage, the dataset to be used for training, validation, and testing of the application cannot be disclosed. However, it can be confirmed that the data originates from CERN's database, specifically from the Large Hadron Collider (LHC). The thesis will be developed in collaboration with CERN where the master thesis student will stay for some periods (pending signature of a collaboration agreement and available funding).
Objetivo
1. Machine learning, active learning and Federated Learning methodology investigation and development of algorithms.
2. Analysis of real and synthetic data, development, and pilot case examination, with software implementation (Python, Matlab or C) of the methods explored.
3. Study of active and federated machine learning applications for modeling complex predictive maintenance (i.e., cryogenic system of the CERN LHC).
Plano de Trabalhos - Semestre 1
1. Initial analysis of the information available for the development of the methods (literature review, theory, algorithms) [Sept – Nov];
2. Analysis of the possible solution methods for the machinery fault diagnostics and prognostics (physical models, datasets) [Nov – Dec];
3. Selection of the most promising solution methods [Dec – Jan];
4. Thesis proposal writing [Dec – Jan].
Plano de Trabalhos - Semestre 2
5. Development and implementation of the selected solutions [Feb – Apr];
6. Implementation over CERN’s Federated Learning platform [Apr – May];
7. Analysis of the obtained diagnostics and prognostic results for a selected case study [May].
8. Scientific article writing [May – Jun]
9. Thesis writing [Feb – Jul]
Condições
The thesis will be developed in collaboration between the Department of Informatics Engineering of the University of Coimbra (DEI-UC)/Center for Informatics and Systems (CISUC) and CERN (European Organization for Nuclear Research). The master thesis student will stay at CERN for some periods (pending signature of a collaboration agreement and available funding -- the scholarship value will follow the CERN standards).
Observações
## Advisors
Luís Macedo (DEI/CISUC), Luigi Serio (CERN), Diogo Santos (CERN).
## Bibliography
Russell, S. J., & Norvig, P. (2020). Artificial intelligence: A modern approach (4th ed.). Pearson.
Settles, B. (2009). Active Learning Literature Survey (Computer Sciences Technical Report1648). University of Wisconsin--Madison .
Macedo, L. (2024 - forthcoming). AI Paradigms and Agent-based Technologies. In Panagiotis (Eds.), Human-centered AI: An Illustrated Scientific Quest. Springer.
Orientador
Luís Miguel Machado Lopes Macedo; Luigi Serio; Diogo Santos
macedo@dei.uc.pt 📩