Semantic segmentation applied for error recognition in the production line of V8 engines cylinder heads

Abstract

Foundry parts face currently the issue of cavities, due to their high cooling speed. When these parts are used on production of car motors, this peculiarity becomes even more important, since motors are exposed to high fluid pressures. Given the motor production is a vital aspect for the company, specially when produced in large scales, it is essential to inspect them after machining and assembly, which may occasionally originate scratches as well. Usually, this is done by a worker manually, who inspects several parts and is subject to human limitations. Artificial Intelligence, on the other hand, arose recently on the quality field due to the possibility of automating this process, só as reducing the error rates. This study aims to assist the worker on the recognition of scratches and cavities through the use of Artificial Intelligence, specifically Semantic Segmentation, while keeping it compact for further implementation on other production lines, só as intuitive enough to allow the worker without previous system knowledge to utilize it. In order to achieve that, the project components was divided into Hardware and Software. For the first one, a metal inducer sensor was placed on the production line, which is activated every time a cylinder head came. The activation was sent to the analog pins of an Arduino, which by his turn, sent a signal to a computer serial port. This serial port was read by a python file, which then entered a loop to shoot a photo, when the signal was sent. Regarding the Software, it was divided into four independent modules or programs. First, photos were manually taken and labeled in the format of masks. After that, in the second module, the model was trained and its history was saved. The third module was responsible for shooting a photo and predicting based on the trained model. Finally, the fourth module concerned the intuitive goal by designing a Graphical User Interface. The results fulfill the project’s initial conditions and needs, after a thorough assessment of the output images. Initially the images were manually labeled, presenting a human failure in itself and not extracting all the meaningful features. Given that, not only the system learned to extract new features missing on labelling, but also several classes were mispredicted. The final results are visible and sharp detections, but with different colors than expected, which in short, reduce the overall score, but do not affect the effectiveness of the work. Taking into account the diversity of photo environments, the model surpassed expectations, presenting only a few limitations, which does not represent a threat to the worker.