Ph. D. Project
Title:
Machine Learning-based Failure prognosis and decision support optimization for predictive maintenance combining expert knowledge and process data
Dates:
2019/10/01 - 2022/09/30
Supervisor(s): 
Description:
I. BACKGROUND
The manufacture of steel in all its forms (coils, plates, wires, etc.) is a complex process involving multiple successive operations (continuous
casting, hot rolling, cold rolling, annealing, galvanizing, etc.) supported by numerous industrial tools (furnaces, cylinders, engines, etc.). Each of
them is subject to high stresses, for example mechanical and thermal stresses applied continuously or according to production cycles depending on
the plant's speed. Thus, these production constraints lead all installations to deteriorate, leading to a reduction in their performance or even to their
unavailability. In order to reduce the "uncontrolled" losses, which become significant when considering continuous production, especially at the
level of a group like ArcelorMittal, preventive maintenance strategies have been put in place. However, although this strategy makes it possible to
satisfy customer demand by reducing unplanned downtime and ensuring product quality, it has the major disadvantage of replacing equipment
while it still produces sufficient quality and the failure may still be far away ("too early" intervention). In addition, this type of maintenance does
not avoid a certain number of residual failures. An evolution towards "just-in-time" maintenance of a predictive type (predictive in the sense of the
NF standard) is therefore desired at ArcelorMittal in order to deploy actions only according to the real state, and no longer assumed, of the
components.
II. OBJECTIVES OF THE THESE
II.1 Industrial objective
The transition to predictive maintenance is a priori only possible by providing all the data relevant to the construction of algorithms for the
prognostic and decision support processes. In this industrial context, the use of measured data on lines and their history as well as the group's
"business" knowledge of process operation should make it possible, through the modelling of their impact on machine performance, to determine
the best maintenance decision based on the coupling of prognosis and decision support.
This thesis is therefore based on this original combination of process and business data exploitation for the joint development of predictive and
decision support processes in predictive maintenance.
The deployment of this new maintenance orientation on ArcelorMittal's production lines is made difficult because it requires tackling five main
industrial locks: low quality, complex and heterogeneous data, Small Dataset available on failures (rare events), expert "business" knowledge and
its integration into predictive models, massive data in the context of plant modernisation and "industry 4.0" and the application of Machine
Learning methods, optimality of the predictive maintenance decision considering the strong dynamics of the production lines.
II.2 Scientific objective
To provide answers to these industrial locks, the main objective of the thesis is to propose advanced prognostic algorithms coupled with decision
support in predictive maintenance based on Machine Learning algorithms efficiently combining expert knowledge, which the ArcelorMittal group
has on its installations, processes, plants, and process data.
The contributions inherent to this objective are built around several scientific locks resulting from industrial locks, and which materialize in the
form of the following questions:
- How can heterogeneous knowledge of production processes (e. g. machine data, production data and expert knowledge), and maintenance and
their impact on the degradation process of each component or even the production line be formalized within a common framework?
- What are the prognostic algorithms best adapted to the issues formalized in the first point and allowing to provide the necessary elements for the
decision optimization phase in maintenance, such as the level of degradation and the RUL?
- What are the elements to be integrated into the optimization model (in dynamic) for decision support (based on the results of the prognosis) with
regard to maintenance business issues in terms of possible actions and planning and their impact?
- How can Machine Learning algorithms for prognosis and decision support be used efficiently and effectively to support the requirements raised
in the previous questions?
The major scientific originalities of the thesis in relation to the previous locks are therefore:
- The construction of advanced prognostic algorithms to increase its robustness by combining process data and expert knowledge (towards a
prognosis with a relevant degree of confidence).
- The development, by integrating the results of the advanced prognosis, of a decision support model for predictive maintenance based on data and
expert knowledge.
- An approach to optimize this decision in response to the dynamics of components and installations.
- The development of a model to evaluate the effectiveness of predictive maintenance actions implemented by the use of process data "after
intervention" for the purpose of loopback on the optimization approach.
Thus, the major contribution of the thesis is the development of an original approach based on Machine Learning tools that combines data-driven
mathematical models with formalized expert knowledge, for the purpose of implementing prognostic processes for the degradation of industrial
systems and decision support in predictive maintenance. It is planned that all these contributions will be validated on a case of application of a real
ArcelorMittal installation in order to give credibility not only to the scientific but also to the industrial scope of the proposals made (towards
TRL5-6).
Department(s): 
Modeling and Control of Industrial Systems