Problem
When a defect or a customer complaint came in, the process and work status at that moment could be looked up — but not seen as a whole. The bigger problem was the judgment criteria. The tolerance band between equipment set value (SV) and process value (PV) was drawn wide, so when several readings fell outside it at once, there was no way to tell which one related to the defect. Decisively, defects were found in quality inspection after production had finished, by which point the material and the time were already spent.
Data collected
Thirteen equipment signals were collected from HMI/PLC: main tank homo RPM, EC, scraper RPM, paddle RPM, chamber temperatures, jacket temperature, vacuum and viscosity, plus water tank and oil tank temperatures and agitator RPM. These were linked with lot-level quality inspection results from ERP and production date, operator, item and production time from MES, along with item and equipment master data.
Modules applied
Edge gateway filtering rulesets cleaned equipment data at the point of collection, and a time-series labeling tool merged manufacturing operations information and quality results into the time-series equipment data. Decision trees and LSTM were used together — decision trees to interpret which variables affect quality, LSTM to handle time-series prediction. The trained inference model was deployed to a real-time inference engine as a streaming analysis pipeline.
Result
Sub-processes within the manufacturing process are now classified automatically using signals such as RPM, temperature and elapsed work time. Linked with quality data, this yields an optimal work model, and current work is compared against it during production to produce operator guidance. A flow that started by noticing a defect and working backwards became one that detects deviation while production is still running.
Defects you only learn about at the end
In cosmetics manufacturing, defects surfaced in quality inspection after production finished. By then the material and the equipment time were already spent. Nothing could be taken back.
Finding the cause was not easy either. The tolerance band between equipment set values and actual values was drawn wide, so several readings falling outside it at the same time was common. Which of them related to this particular defect was unknowable. To a human reader, they were all simply “out of range”.
Turning what happens inside the tank into data
First, the values coming off the equipment. Main tank homo RPM, scraper and paddle speeds, chamber and jacket temperatures, vacuum, viscosity. Water tank and oil tank temperatures and agitation speeds. Thirteen signals from HMI and PLC.
That alone is not enough, because you have to know what result each of those led to. Lot-level quality inspection results came from ERP, and production date, operator, item and production time from MES, attached to the time-series equipment data.
Letting the process divide itself
A cosmetics manufacturing process looks like one continuous operation, but there are distinct stages inside it — and where one ends and the next begins is not recorded anywhere.
Patterns in equipment data such as RPM reveal those boundaries. Once sub-processes were classified automatically, it became possible to see stage by stage how conditions at each point act on final quality.
Why two algorithms
Decision trees and LSTM were used together, because they fall short in different places.
Decision trees pick out which variables matter and show why in a tree structure, including how two variables combine. That matters when the floor needs an answer to why the model judged as it did.
LSTM suits time-series prediction. It looks ahead with the whole history of how temperature and rotation speed inside the tank have moved over time.
Interpretation from the decision tree, prediction from the LSTM.
Visible during production
The trained model was deployed to a real-time inference engine. While production runs, equipment data streams in and current work is compared on the spot against the optimal work model.
Confirming a defect after the fact and tracing backwards became detecting deviation during production and receiving operator guidance.
Carried out from December 2020 to April 2021 under an AI solution demonstration support program.