Detecção e classificação de múltiplos componentes em linha de montagem automotiva usando deep learning

Abstract

The industrial production process capable of simultaneously producing multiple types of products on the same manufacturing line defines an industrial model called flexible manufacturing. An example of flexible manufacturing is observed in the automotive industry, which manufactures different car models on the same plant, each composed of its own set of components. In this type of environment, it is common for part sets to be initially separated by human operators and subsequently manipulated by robots. As this process depends on human perception, it is susceptible to errors of conformity, a situation in which elements of manufacture may not belong to the same model of the manufactured product. This task has recently been complemented by artificial perception from computer vision systems (CSVs), camera-based systems and an embedded computer for image acquisition, processing and classification. These systems can be used to check for defects and to attest to the conformity of the components. Despite the practical relevance, the efficiency of a conventional CSV depends on the control of lighting, closure and stop of the line for the fixation of the parts, which makes the solution expensive and reverses most of its benefits. In addition, traditional methods have manual parameter assignment steps dependent on a specialist. Recently, the literature has addressed methods based on deep learning (DL) to overcome these problems. This approach seeks to imitate the human capacity for learning and pattern recognition, through the abstraction of characteristics of a set of images. This feature adds robustness to SVC, reducing the need for environmental control and human perception for parameter definition. This paper proposes a supervised learning deep text alternative for detecting multiple components in supervised flexible automotive assembly lines for object detection and classification. When supervised, these algorithms require a set of images as a training base. In addition to the images, the region and class of each object must be identified. This identification consists only of defining a region that delimits the target object and its respective class. Still, the solution does not require interventions in the operating environment, neither the stop of the production line for image collection and analysis. The approach is illustrated by two case studies conducted on an actual vehicle assembly plant.