PROJECT DESCRIPTION
The customers' products rely on lab tests to measure performance during manufacture. These lab tests were leaving the customer vulnerable to production blind spots that caused notable quantities to be produced of unknown quality. The vulnerability arose from intervals between tests and the delay from when product performance is determined vs. when it was reported. This project gave real-time estimates of product performance as it happened for increased awareness and faster response to problems, reducing risk and increasing yields. It also gave insights into key product performance drivers, use of alternative raw materials to improve profitability and process improvement and directly control quality on-line.
In any manufacturing process where product samples are taken and sent to the QC lab for testing, or having on-line instrument(s) that has long cycle times or are far downstream, the operators and engineers are running blind between tests. There can be a large investment of questionable value if something "goes wrong" in between lab tests. Perhaps a new vendor or lot of raw materials was started, or a shift change occurred and an operator made an ill fated seemingly small adjustment to the process. Large quantities are produced before the first lab test comes back reporting poor results and by the time corrective action is taken and progresses through the manufacturing process and a new test sample is taken, tested and reported even larger quantities of materials can be rejected. If that lab test also reports bad results, a small fortune can be lost in rework, recycling, recovery or out-right scrap.
Of course, there must be a better way, and there is... we put a virtual sensor on-line that uses process conditions and materials characteristics to estimate the product characteristics. This is particularly useful when product characteristics are not directly measurable, nor can be calculated directly, such as "wet burst strength" or "softness" or smells or tastes. Such things are intractable from a theoretical basis. However, we knew that raw material characteristics and process conditions do define product performance and we had the lab test results that are associated to those materials and conditions. Data-driven models were built that relate causes (materials and process) to the effects (quality test results). We put the models on-line using the materials and process conditions as inputs to obtain performance estimates. As the operator "turned the knob" on the process, they IMMEDIATELY saw the expected performance. We analyzed the models using "sensitivity analysis" to see and better understand key drivers of product performance. Response curves and surfaces graphically show the effect that a raw material characteristics and process conditions have on quality. The models were also used in an Excel work book to do off-line "what-if" analysis. In some cases, the models were of sufficient quality to use in our "iImprove" product to obtain recommended setpoints for controllable process conditions, given other non-controllable factors, that would deliver the desired product quality. In some cases we directly controlled quality by "closing the loop" with the process control system to automatically write setpoints to attain the desired quality.
PROJECT SUMMARY
- Title: Performance Prediction
- Client: Many
- Challenge: Customer wants to know product performance before instrument or test results are available.
- Skills: Data Access, Cleansing, Validation, Modeling, Real-Time Prediction, Performance Monitoring, Reporting, Project Management
- Brief:
Predict product characteristics before lab or instrument results are available so that operators and engineers can respond with process adjustments in real-time, manually or automatically with closed-loop control.