Skip to Main content Skip to Navigation

Contribution to the development of a methodology coupling natural language processing and machine learning to react to production disturbances.

Abstract : In the age of Industry 4.0 (I4.0), exploiting data stored in information systems offers an opportunity to improve production systems. Datasets stored in these systems may contain patterns that machine learning (ML) models can recognise to react more effectively to future production disturbances. In the case of industrial maintenance, data are frequently collected through reports provided by operators. However, such reports are often provided using free-form text fields, resulting in complex unstructured data; therefore, they may contain irregularities such as acronyms, jargon, and typos. Furthermore, maintenance data often present asymmetrical distributions, where certain events occur more frequently than others. This phenomenon is known as class imbalance, and it can hinder the training of ML models as they tend to recognise the more frequent events better, ignoring rarer incidents. Finally, when implementing I4.0 technologies, the inclusion of humans in the decision-making process must be ensured. Otherwise, companies may be reluctant to adopt new technologies. The work presented in this thesis aims to tackle the general objective of harnessing maintenance data to react more effectively to production disturbances. To achieve this, we employed two strategies. First, we performed a systematic literature review to identify the research trends and perspectives regarding the use of ML in production planning and control. This literature analysis allowed us to understand that predictive maintenance may benefit from the unstructured data provided by operators. Additionally, their usage can contribute to the inclusion of humans in the implementation of new technologies. Second, we addressed some of the identified research gaps through case studies that employed data from real production systems. These studies harnessed the free-form text data provided by operators and presented class imbalance. Hence, the proposed case studies explored techniques to mitigate the effect of imbalanced data; moreover, we also suggested the use of a recent architecture for natural language processing called transformer.
Complete list of metadata
Contributor : Samir LAMOURI Connect in order to contact the contributor
Submitted on : Wednesday, January 5, 2022 - 4:36:15 PM
Last modification on : Sunday, June 26, 2022 - 12:36:27 PM
Long-term archiving on: : Wednesday, April 6, 2022 - 9:21:27 PM


Files produced by the author(s)


  • HAL Id : tel-03513071, version 1


Juan Pablo Usuga Cadavid. Contribution to the development of a methodology coupling natural language processing and machine learning to react to production disturbances.. Automatic. HESAM Université préparée à : École Nationale Supérieure d’Arts et Métiers, 2021. English. ⟨tel-03513071⟩



Record views


Files downloads