Investigations on Real-Time Image Recognition with Convolutional Neural Networks on Industrial Controllers

Abstract

Modern manufacturing systems are collecting more data as a result of progressive digitization. In the application area of individualized production, the collected data can be used, e.g., with the help of image processing systems, to control required workflows by classifying individual objects. Machine learning (ML) methods can be used to evaluate and classify the captured images. As soon as the acquired objects shift rotationally or translationally within the acquisition area, a reliable classification of the objects with a simple multilayer perceptron (MLP) becomes difficult. Convolutional Neural Networks (CNNs) use methods that can offer a solution to these problems. This paper investigates whether it is possible to use computation intensive CNNs for image recognition, in real-time and in coordination with machine and motion control tasks. These CNNs are compared to less computation intensive MLPs. In this paper, an innovative approach is presented to show how trained CNNs can be integrated into the real-time environment of a programmable logic controller (PLC). Based on two application examples, several CNNs with different network architectures are integrated into a PLC runtime environment on a standard industrial PC (soft-PLC). The execution times of the different networks are measured and compared.