This paper presents an approach for image classification based on an ensemble of convolutional neural networks and the application to a real case study of an industrial welding process. The ensemble consists of five convolutional neural networks, whose outputs are combined through a voting policy. In order to select appropriate network parameters (i.e., the number of convolutional layers and layers hyperparameters) and voting policy, an efficient search process was carried out by using an evolutionary algorithm. The proposed method is applied and validated in a case study focused on detecting misalignment of metal sheets to be joined through submerged arc welding process. After selecting the most convenient setup, the ensemble outperforms other seven strategies considered in a comparison in several metrics, while maintaining an adequate computational cost.

本文提出了一种基于卷积神经网络集成的图像分类方法，并将其应用于工业焊接过程的实际案例研究。该集成由五个卷积神经网络组成，其输出通过投票策略进行组合。为了选择合适的网络参数（即卷积层数和超参数层数）和投票策略，使用进化算法进行了有效的搜索过程。所提出的方法在一个案例研究中得到应用和验证，该案例研究的重点是检测通过埋弧焊工艺连接的金属板的错位。选择最方便的设置后，集成在几个指标的比较中优于其他七种策略，