Building a Corporate Memory for the Capitalization and Reuse of Maintenance Intervention Experiences :A Hybrid Approach Based on Transformers and Case-Based Reasoning

Abstract

This thesis addresses the problem of the loss of knowledge and know-how in the field of industrial maintenance, and the drawbacks of corrective maintenance on the availability of production lines, taking the rotary kiln of a cement plant as a practical case. We propose an intelligent predictive-maintenance system based on a hybridisation of the forms of knowledge exploitation. On the knowledge-extraction side, a Transformer autoencoder learns the normal behaviour of the rotary kiln from an input formed of sixteen sensors, using six months of historian data, and detects the precursory signs of failure through the reconstruction error. On the knowledge-engineering side, case-based reasoning capitalises the interventions (deemed valid by the operators) into a corporate memory, providing an experience feedback that reuses the solutions of past failures similar to the current breakdown. On the knowledge-management side, our approach shifts the majority of interventions from corrective to preventive main- tenance, guaranteeing the availability and reliability of the production lines. The proposed hybrid approach is implemented as a pipeline (Kafka, InfluxDB, Grafana, Prometheus) with graded and deduplicated alerts, together with an active-learning loop, driven by the operator, that retrains the model. The ultimate objective is the incremental construction of a corporate memory that solves the problem of expert unavailability.