Problem
Within MPC cooler manufacturing, paste coating had the greatest influence on quality and defects. Real-time inspection was not automated, and there was no way to predict how coating condition would affect quality.
Data collected
Unstructured data from a 3D profile vision sensor was collected alongside structured data from the coating robot PLC. Coating volume for segments L1–L12 on both A and B faces, plus equipment setpoints (SV) and measured values (PV), were stored as time series.
Modules applied
C² and D² handled collection and standardization, with an OPC server normalizing protocols before raw data was stored in a TSDB. A² used LightGBM to predict optimal coating volume and quality, with V² performing vision inspection.
Result
An autonomous control system derives the optimal process value and writes it back to the equipment setpoint. The value produced by the AI inference server is protocol-converted, delivered to the PLC, and the result of the change returns to the system.
It did not stop at prediction
Predicting quality is something many sites manage. The question is who applies that prediction to the equipment, and when. If a person has to watch a screen and change the setting by hand, the system only works while that person is there.
In this case we built a path for the AI’s optimal value to return to the equipment.
How the data circulates
- Structured and unstructured data collected from the 3D vision sensor
- Protocols normalized at an OPC server, publish module configured
- Subscribe module receives it; raw data stored in a TSDB
- After preprocessing, the AI inference server derives the optimal process value and a real-time quality prediction
- The result is sent to the PLC, converted to command protocol, and the SV is changed
- The outcome of the change returns and feeds the next decision
Why LightGBM
It handles both regression and classification, so optimal value prediction and quality prediction run on one algorithm. Training is fast and memory use is modest, which suits repeated retraining beside equipment on the floor. Its weakness is overfitting on small datasets, so collection ran long enough to validate properly.