Optimizing Yield by Predicting Product Quality in Real Time
A large plastics manufacturer with more than 36 industrial plants across three continents produces more than 32 billion pounds of thermoplastic resins and petrochemicals annually. Because each customer order is unique, the company must transition manufacturing settings for each order and then test each batch to ensure quality. Lack of real-time visibility into current quality causes significant lag time between orders. Miscalibration in product quality can cause entire orders to fail quality testing, leading to millions of dollars of lost revenue.
By implementing the BHC3 Yield Optimization™ application, the manufacturer reduced average product transition times by more than 30%, allowing the firm to understand exactly when a product moves in and out of spec, saving millions of dollars. BHC3 Yield Optimization machine learning predictions vary considerably less than lab tests during steady-state production periods, providing further confidence in the accuracy of the machine learning models.
About the Plastics Manufacturer
- €15 billion annual revenue
- 32 billion pounds of thermoplastic resins
and petrochemicals produced per year
- 36 industrial plants across three continents
- 7,000+ employees
"In one specific product line we were able to reduce the transition time by 30%.”
Senior Director, Products
Building and deploying an AI-based application
To predict product sample quality using AI, a joint BakerHughesC3.ai and customer team integrated relevant data and applied scalable supervised machine learning algorithms. The team aggregated over 2 billion rows of data from eight sources into a unified federated image. The data was represented as a collection of 25 BHC3 Types to model the relationship and continual process flow between key pieces of equipment such as vent recovery compressor, reactor, and product discharge system. The team then performed analytics to understand the delay between manufacturing settings and the conditions inside the reactor and created over 32,000 time-based expressions to extrapolate information that could not be captured from raw data. After creating and tuning machine learning models, the team optimized train-test splits to ensure an appropriate balance of product types in each segment of data.
- 10 weeks from kickoff to production-ready application
- 2 billion rows of 10 years of data from 8 data sources, measured at 1-minute intervals
- 32,000+ time-based expressions constructed for machine learning
- 8.1 million pounds of product moved from out-of-spec lower-margin product to in-spec higher-margin product on a single production line