Context
Predicting product quality with a soft sensor has been a long-standing research topic in the process industries. Quality variables such as viscosity, molecular weight, concentration, and impurity levels are notoriously difficult to measure in real time during production. Lab results often come back hours later, and any off-spec product made in the meantime is effectively a loss. For plants that produce many specialty grades on the same line, this problem is even more acute: data on each grade is limited, and off-spec material tends to concentrate around grade transitions, before the process settles. Building a separate model for each grade is rarely practical.
Approach
We developed an edge AI soft sensor that estimates the target quality variable in real time, from process measurements already being collected in the plant. The model is built on a class of lightweight machine-learning techniques (reservoir computing) well suited to time-series signals and edge deployment. It is designed to handle the realities of multi-product operations: limited data per grade, frequent grade transitions, and high-dimensional process inputs. Final tuning to the customer's specific plant, product portfolio, and operating conditions is done in collaboration with the customer's engineering team.
Outcome
The soft sensor provides real-time estimates of the quality variable, allowing the plant team to detect deviations without waiting for lab results. This helps reduce off-spec production and shorten response times during grade transitions.
This may apply to your operation too
If you run a multi-product chemical or materials plant where lab delays drive off-spec losses, or where new grade introductions go through long stabilization periods, this approach may be worth considering. Whether it fits your specific situation can usually be assessed in a single short conversation. Technical implementation details are discussed individually with each customer.