Deep learning based methodology for the development of industrial quality inspection systems

Laburpena

In recent years, the manufacturing industry has gone through what has been called the fourth industrial revolution or Industry 4.0. Apart from still automating industrial processes, the revolution has as well brought new trends like zero-defect manufacturing, non-destructive unitary tests, or complete traceability of every part along the production chain. One of the sectors that have been influenced by this revolution is the solar sector. This sector, as part of the strategic sector of renewable energies, has received large funding from government entities and individual investors that have led to an improvement in technology. This has lowered the prices of panels, which in turn has increased the demand for them making it more necessary to automate the production process. Among all the stages during production, quality control plays a crucial role. In the specific case of the photovoltaic sector, quality control in industrial manufacturing is performed using the Electroluminescence technique which allows practitioners to obtain high resolution images of the photovoltaic cells where defects are highlighted. In contrast to the trend towards automation, in practice, panel inspection is still mostly performed by operators. In recent years, many proposals have been made to automate this quality inspection. However, the proposals made so far show certain limitations for their application in the increasingly dynamic and demanding industrial context. Some of the identified limitations are: the lack of flexibility to changes in production since the proposed procedures have been designed to take advantage of case specific data features. For example, an inspection system might have been designed to take advantage of the high contrast between the light background and the thin and dark longitudinal cracks in the cells. However, a variation in the data like a darker due to a different material composition of the cells or different shapes of the cracks may suppose to redesign the entire inspection system to adapt to the changes. Other proposals contemplate algorithms that require a large number of representative defective samples for training, which are usually difficult to obtain in industrial environments. And finally, some solutions consist of algorithms that can act as black boxes with respect to their interpretability, which together with giving as a result only whether a part is defective or not, can raise doubts about the performance of the inspection system. For these reasons, the objective of the thesis has consisted in designing a methodology based on Deep Learning techniques for the development of inspection systems. The methodology has contemplated techniques that are robust and flexible to changes, but also able to work in industrial environments where there are few available defective samples, and output more interpretable results than a mere classification, for example, the location of defects in the samples. At the same time, the methodology offers ways to obtain inspection models from the very beginning in the production line, and take advantage of their characteristics to obtain more accurate models with almost no need for human intervention.