Project Experience

Project Examples

1. 01

   Virtual Sensor for Product Quality and Chemical Composition

   SituationQuality characteristics and chemical composition of the final product were determined through manual sampling and laboratory analysis. Results were therefore available only at individual points in time and with a delay. Between two samples the trajectory remained unknown, and deviations only became apparent once the affected material had already been produced.

   ApproachModel-based estimation of the quality variables from live process and plant data. Selection of input variables based on process relationships, model building on historical operating data, and calibration against the existing laboratory values. The estimate closes the gaps between samples and runs continuously. On the same basis, a predictive model was developed that calculates expected quality from the control parameters currently set.

   OutcomeThe quality trajectory is visible continuously rather than at isolated points. Deviations are detected as they develop, not after laboratory analysis. The prediction makes it possible to assess the effect of parameter changes before intervening, instead of reading it off a sample afterwards.

   My contributionFeature selection based on process relationships, model building and validation against laboratory values, and design of the predictive model for use in process control.

2. 02

   Camera-Based Process Monitoring at Tuyeres

   SituationMonitoring of the tuyeres in pig iron production was carried out manually at fixed intervals. Personnel had to enter the hazard zone to do so, and between inspection rounds process disturbances went undetected until they showed up in plant behaviour.

   ApproachContinuous optical monitoring: selection of camera and optics for the thermal and optical conditions at the tuyere, setup of data acquisition, preparation and annotation of the image data, and training and validation of the model. The model was quantised for deployment on embedded hardware. Downstream classical image processing and rule logic translate the model output into status messages and alerts in the control room.

   OutcomeCritical process states are detected while they are still manageable. This reduces unplanned downtime and protects the plant from damage that occurs when disturbances are only noticed through their consequences. Operators also gain a continuous statistical view of plant condition, and the hazard zone no longer has to be entered for routine monitoring.

   My contributionSelection of optics, data preparation and annotation supervision, model selection, training and validation, quantisation for embedded inference, and design of the downstream image processing and process logic.

3. 03

   Automated Surface Inspection of Forged Parts

   SituationOn a high-volume line, surface defects were identified by manual visual inspection. Results depended heavily on the individual inspector, their condition on the day, and the lighting, and subtle defects were assessed inconsistently. A conventional supervised approach was not viable: defects were rare, the range of defect types was not fully known, and every new defect type would have required fresh data and retraining. Good parts, by contrast, were available in large numbers.

   ApproachInverting the problem. Instead of learning defects, the model learns the normal state from images of good parts and flags deviations from it. Implemented as a few-shot method based on pre-trained feature extractors, requiring no annotated defect images. Images are captured with high-resolution cameras under defined lighting; evaluation runs patch by patch across the part surface to capture fine features that would be lost in a whole-image assessment.

   OutcomeSurface defects are detected with consistent sensitivity, independent of inspector and shift. The system supports manual inspection by flagging suspect parts for review rather than taking over the decision entirely. New defect types are detected without having been learned beforehand, and commissioning of further variants is faster because only good parts need to be captured.

   My contributionMethod selection, design of camera and lighting setup, implementation of the patch-based evaluation, validation, and definition of decision thresholds in coordination with quality assurance.

4. 04

   Digital Fingerprinting for Material Flow Tracking

   SituationCast intermediate products with rough as-cast surfaces were handled unsorted in internal logistics. Individual pieces could not be unambiguously tracked across the process chain, so quality findings at the end could not be traced back to production parameters at the start.

   ApproachMarking-free identification using the surface texture itself: a digital fingerprint is generated from the cast structure using SIFT and related feature descriptors and linked to existing process numbers. Individual pieces can thus be re-identified without physical marking, even after they have left their original sequence.

   OutcomeMaterial flow is traceable even in unsorted logistics. Quality assessments can be linked to the corresponding production parameters, which is what makes statistical root cause analysis of scrap possible in the first place.

   My contributionConcept, selection and tuning of the feature descriptors, and design of the matching procedure for the conditions of the cast surface.

5. 05

   Automated Reading of Barcodes and Product Data

   SituationIn intralogistics, barcodes and product information had to be captured in varying formats and positions. Fixed reading equipment reaches its limits here because the position and form of the marking vary.

   ApproachA two-stage method: object detection locates the marking in the image and crops the relevant region; a downstream open-source barcode model was fine-tuned on the code types and imaging conditions present.

   OutcomeMarkings are read reliably regardless of position and form. Product data is thus available automatically within the material flow.

   My contributionSelection of camera and lens, model selection and fine-tuning, and design of the processing chain.

Industries and Application Areas