Methodology for optimizing convolutional neural networks for fast production processes

Abstract An individualized production is needed to manufacture products with a batch size of up to 1 at low costs. To ensure that the individual steps in the production process are executed for each individual product, the products have to be classified beforehand. For complex problems, machine vision can be used in conjunction with machine learning (ML) to classify the individual components. Here, models of Convolutional Neural Networks (CNNs) in par-ticular achieve high classification accuracies. However, they require high computational cost, which makes it more challenging to execute CNNs in real-time within a PLC runtime environment and synchronize them with motion control tasks. In this paper, a methodology is presented using a production-oriented application example to minimize the inference time of a CNN models for image classification while maximizing the classification accuracy. The presented methodology demonstrates how computationally cost intensive CNNs can be optimized for real-time image recognition in coordination with machine and motion control tasks. The execution times as well as the achieved accuracies of the CNNs are measured. The results show that the CNNs trained on a synthetic CAD dataset are able to reliably classify individual products with an accuracy of 100 % in less than 185 µs (with an image size of 39 × 26 × 1 pixel).