Sistema de inspeção visual de conformidade do produto: redes neurais convolucionais

Abstract

On mass production, where a large amount of raw material and finished goods are delivered, the quality assurance is of great importance, being the difference to deliver a high-quality product in large quantities. Yet, in injected parts, problems during the process can happen and even small faults might impact negatively in the function of the product. Still, the lack of standardization and a difficult in inspection turns this assurance hard to obtain, and tools regarding inspection ease and allow to achieve the necessary quality, being artificial intelligence one of the strongest tools. Therefore, the present dissertation, using artificial intelligence – more specifically convolutional neural networks – and programming using TensorFlow and a deploy with Arduino, C# and python communication, seeks to integrate a physical prototype with image analysis, giving opportunities to make the assembly line more robust with real-time inspections, comparing three different neural networks and discussed the advantages and differences between them, also adopting as a solution a pouca yoke system (error-free). As a result, the neural networks show up to 88% of precision and a low occurrence of false positives, as a simplified interface and a low implementation cost equipment, with ease of use. Additionally, including database interfaces and greater analysis on the results, it is possible to compare between different time stamps, being applied to predict maintenance on the injection molds depending on the approval yield, or even traceability to problems encountered on field (after the finished good is delivered).