The manufacturing of hot water tanks requires multiple welding processes. The quality of the
welds is crucial for the durability of the product. It is often assessed by visual inspection, which is
time-consuming and prone to error. One solution to this problem is the use of machine learning,
which is a growing technology in the manufacturing industry. This research aims to develop systems
for detecting and diagnosing welding faults using machine learning by exploiting the signals captured
during automatic welding and images of the weld. Fault detection is difficult in the context of hot
water storage tanks for many reasons: the complexity of welding dynamics, the variety of welding
processes, and the wide range of faults. The approaches proposed here treat these challenges with
the aim of developing systems that can be used in real-time. After studying the feasibility of
detection, we propose an approach based on One-Class SVM and distance substitution kernels.
This approach only requires raw data of conforming welds and detects anomalies based on their
distance from normality, which facilitates the generalization. Moreover, we propose a diagnostic
approach based on classification. Another contribution is proposed, which extends the random
kernel transform to the problem of anomaly detection and explainability. In addition, we propose
a system detecting defects from welding images composed of an acquisition system and a neural
network capable of locating and classifying defects.

Jury :  
Jairo  Cugliari                 Maître de conférence HDR     Université de Lyon 2                           Rapporteur
Emmanuel Ramasso      Maître de conférence HDR     ENSMM                                              Rapporteur
Gilles ROUSSEL                     Professeur                     Université du Littoral-Côte d’Opale     Examinateur
Marianne Clausel                    Professeure                   Université de Lorraine                         Examinatrice
Anis Hoayek                            Maître de conférence    Ecole des mines de Saint-Etienne      Examinateur
Mireille Batton-Hubert             Professeure                   Ecole des mines de Saint-Etienne       Directrice de thèse
Erwan Stephens                      Pilote projets IA               elm.leblanc. Bosch                 Invité.