In order to further increase the performance of neural networks in the field of optical quality assurance of soldered joints, a hierarchical classifier can be used instead of a single network. The global expansion of the classifier enables the inspection task to be distributed over several subnetworks, which results in higher accuracy. Since the individual sub-models only concentrate on the identification of certain characteristics, categorical problems can be solved more effectively.

This article presents a classification concept based on deep learning as an additional optical test method for real-time visualization and analysis of electrical assemblies in the production environment. For this purpose, a neural convolutional network is used to identify the quality of the solder joint of surface-mounted chip components in the inspection images. The concept can be used to increase the detection performance of the solder joint inspection systems.

Multi-component sensors for force and moment are commonly used in different areas as robotics, crash tests or material testing. The dynamic behaviour of such sensors may differ significantly from the static behaviour. To analyse the dynamic characteristics, we developed an improved test set-up for periodic excitation. It consists of an electrodynamic shaker, adapting elements and different acceleration references. The first experiments show good results for periodic excitation up to 1000 Hz.

For optical measurement systems, camera noise is the dominant uncertainty factor when optically cooperative surfaces are measured in a stable environment. In this work it will be shown that the resulting spatial noise can be estimated by means of a noise model. The capability of this approach will be demonstrated for a fringe projection system. This provides a valuable tool for a statistical comparison of different evaluation strategies, which is shown for two different triangulation procedures.

Deflectometric measurements of specular surfaces when performed with two reference planes lead to both data for surface points and surface normals. The deviations in the surface points are usually several orders of magnitude larger than those for the surface normals. In this paper we propose a method to fuse these data to increase the accuracy of the surface points. The results show that the accuracy can be increased by more than an order of magnitude.