Gisela Lanza

Micro face gears exhibit complex non-involute flank geometry and lack standardized inspection concepts. In contrast to cylindrical gears, where VDI 2612 provides interpretable profile, flank and pitch characteristics, face gear quality assessment remains fragmented and qualitative. This work presents a feature-based evaluation strategy for micro face gears that transfers cylindrical inspection logic and extends it with dedicated diagnostic features for axial infeed and tangential misalignment. Optical scan data are converted into profile-, flank- and pitch-inspired descriptors within the functional engagement region, establishing a geometric interface between metrology and process parameters in micro-scale hobbing. The strategy is validated on 225 experimentally manufactured face gears with intentionally varied axial infeed, tangential shift and spindle wobble. The results show that axial infeed and tangential shift can be diagnostically separated based on their characteristic geometric responses. For spindle wobble, the proposed features encode the deviation information but require multivariate evaluation for reliable extraction. The findings indicate diagnostic capability suitable for virtual metrology and process correction.

This paper presents the results of the interdisciplinary research project “Tolerance‐free series production of high‐performance concrete modules through transient‐interactive coupling of design and production” within the priority program “SPP 2187”, funded by the German Research Foundation (DFG). The project aims to enable tolerance‐compensating, resource‐efficient serial production of modular precast structures by integrating material behavior, structural design, manufacturing, and assembly. Methods for coupling design and production address final and intermediate construction states, assemblability, and sustainability. Placement strategies and specification heuristics link structural performance and assembly requirements with adaptive heat‐treatment control. The production system is modeled using discrete‐event simulation. Uncertainties are modeled probabilistically and the decision problem is formalized as a Markov Decision Process. Single‐ and multi‐agent deep reinforcement learning systems select module‐specific heat‐treatment durations to reduce dimensional variability. A virtual demonstrator, complemented by experimental and pilot‐scale structures, demonstrates the benefits of the integrated approach for precise, sustainable, industrialized precast concrete construction.

In micro gear manufacturing, quality assessment relies on full-geometry measurements to evaluate functional performance. Due to tolerances down to one micrometer, high-end metrology is essential. Optical systems enable fast, non-contact measurements and can, in principle, be used for full-gear scans. These scans serve as input for single-flank rolling simulations, which assess how geometric deviations affect functional behavior such as transmission accuracy. However, full scans remain time-consuming and often unsuitable for inline inspection due to the trade-off between speed and measurement uncertainty. We address this by proposing a partial-scan workflow, based on the observation that tool-induced deviations propagate periodically across gear teeth. This allows reconstruction of micrometer-accurate point clouds from a subset of teeth. We compare a deep learning-based completion network with an analytical reconstruction, evaluating both geometrically and functionally. While the deep learning approach shows higher geometric fidelity, it falls short in functional accuracy. This reveals a common gap in learning-based methods, where achieving geometric similarity may fail to preserve the underlying functional behavior. Our approach enables faster inspection while maintaining confidence in gear performance.

Remanufacturing is a key process for enabling a circular economy by restoring used products, often referred to as cores, to a like-new condition. It still heavily relies on manual work. One exemplary manual task is the visual inspection of cores before further processing. This manual effort arises from uncertainties, such as varying product conditions, a broad variety of product variants, and the lack of product information to support automation. To address these challenges and to enable the automation of visual inspection tasks, flexible and adaptive inspection systems are required. These systems must be capable of automatically detecting defects and performing product-specific inspections across a wide range of product variants. Therefore, this work introduces a semantic 3D product modelling method that integrates two-dimensional (images) and three-dimensional (point cloud) data. With the resulting semantic 3D model, the semantic information, such as detected defects and components requiring closer inspection, can be encoded on a geometric product model. This model provides the information basis for an adaptive inspection approach. Using semantic (U-Net) and instance segmentation approaches (Mask-RCNN), the proposed method assigns each surface point of the 3D model to a specific component, thereby creating the semantic 3D product model during the inspection process. The results show that the method presented can achieve semantic 3D modelling both in accuracy and model completeness, encoding component information on the geometric product model. Furthermore, we show that the U-Net architecture used to detect components is also able to detect corrosion as one exemplary defect type, enabling the encoding of various semantic information into the semantic 3D product model. This semantic 3D product model then enables targeted individual inspection of these semantically mapped components and defects on the product model in a later stage of an automated inspection procedure.