Improving Manufacturing Quality: Deep Learning-Powered Sealing Defect Detection in Industrial Manufacturing

In the context of industrial manufacturing, quality control is a topic of primary importance due to its influence on multiple aspects of the production process. Ideally, an efficient and reliable quality control process in a mass production line ensures a steady throughput of goods that meet or exceed the production requirement parameters at each step of the manufacturing process, hence impacting significant aspects such as safety, durability, reliability and functionality of the product. This thesis, conducted within the framework of the MANAGE 5.0 (MANufacturing Automotive Green Evolution 5.0) project, investigates the application of state-of-the-art deep learning techniques in the context of the detection and localization of sealant application defects in next-generation car production lines. After introducing a thorough review of state-of-the-art deep learning approaches for anomaly detection and localization, the first part of this thesis examines the production line environment, analyzing the underbody of the vehicle and the challenges it poses for the gathering of the data. The implementation of automatic frame extraction based on template matching showed a significant acceleration in the process of gathering frames from the videos filmed during visits at the production site, while keeping strong similarity and alignment to the desired templates. At the same time, in order to increase the availability of defective data, a consistent and numerically relevant dataset of defective samples was manually generated: this led to the creation of a fair and representative dataset, ensuring accurate model selection and validation. In the second part of the thesis, the effectiveness of the considered state-of-the-art deep learning techniques on the novel dataset is evaluated and compared, revealing promising results across a number of metrics. These findings underscore the potential of these algorithms in the industrial settings. In the framework of this thesis and the MANAGE 5.0 project, a fixed system setup for continuous data gathering and, in a second phase, real-time inference was proposed for the future developments of the project, ensuring data consistency and the collection of a vast array of normal and defective samples for training the models, as well as providing a first prototype of a real-time defect localization system. While particularly focusing on the detection of sealing defects in industrial manufacturing, the work and the results of this thesis lay the groundwork for the future scenarios of the MANAGE 5.0 project such as the implementation of extended reality technologies to assist plant operators in defect detection and correction.