Thesis summary :
Berger-Levrault is a major actor in the development of software solutions for the management of public and private services. In the field of CMMS (Computerized Maintenance Management Systems), many processes still rely heavily on human experience and oral transmission: incident analysis, diagnostic capitalization, intervention planning, and decision documentation. Although crucial, this expertise is difficult to formalize and exploit on a large scale. The lack of structured memory leads to redundant errors, loss of efficiency, and heavy dependence on experts. Knowledge graphs offer a unique opportunity to represent, store, and query these past experiences and business rules. They can provide an explicit and verifiable basis for facilitating the execution of industrial actions and capitalizing on knowledge within Berger-Levrault.
The main objective of the thesis is to design and evaluate a knowledge graph focused on maintenance experiences. This structured thesis should:
- a) Model relevant concepts (interventions, validated diagnostics, human decisions, business rules, action scenarios).
- b) Structure and encode past experiences to make them searchable and usable in real-world contexts.
- c) Evaluate the relevance of the graph in terms of its usefulness for assisting interventions and optimizing maintenance processes.
The thesis will therefore focus on three interrelated axes :
- Ontology and experience graph model: Defining an ontology suited to the field of industrial maintenance; Integrating dimensions of knowledge evolution.
- Graph extraction and feeding methods: Explore the extraction of information from technical documents and intervention reports using large language models (LLMs); Study the feasibility of incremental graph updates.
- Evaluation and industrial applications: Several use cases can be studied. For example, searching for similar interventions and suggesting relevant procedures (or sequences of actions) to solve maintenance problems.
