A chemical manufacturer is challenged by variability in quality of the final goods they produce. They believe that the large volumes of process and machine data that they collect for the different phases in their processes should be able to identify problematic batches in early phases of a process. Efforts to use this data to improve yields, however, had been largely ineffective.
The manufacturer collects process batch data for different phases (up to 8 phases in one process) of selected processes that include:
In addition, the manufacturer collects machine data for the given Machine ID over the period of batch execution that includes:
The quality of the final product is measured at the end of each process.
For the most part, this data is used in forensic investigations to help determine what may have caused a particular sequence of bad batches long after negative financial impacts have already been incurred.
There is a strong desire to analyze the data in real-time so that interventions could be made as early as the end of the first process phase.
The process phase execution data, final product quality data, and the machine telemetry data (that come from the Manufacturing Execution Systems – MES) are captured in a custom-built process historian managed by the customer. A simple integration agent is built from the Falkonry C# Agent DevKit and is deployed at the customer site. This agent connects to a Falkonry server deployed in the customer’s Azure cloud.
A member of the customer’s process efficiency team creates a Falkonry Datastream for one phase of the process. The history window that is provided to the Datastream represents a few thousand batch executions of the particular phase for a single machine. For each batch, there is the following data:
The user creates an assessment that represents to predicted final product quality of a batch, and then creates models based on a portion of the history window using the Falkonry UI. These models used a “batched window” approach where each batch ID represents one window or one value of the assessment. These windows in the training data were labeled with the appropriate quality metric values that were supplied as “Facts”.
Once the user creates the models, she was able to test how well the models predicted quality outcomes for other periods in the provided history window. Satisfied that the predicted values were sufficiently predictive for the two quality metrics, the model is made ‘active’ and is used to produce assessments for streaming data.
The assessment results generated by Falkonry provide a prediction of final product quality at the end of each phase completion. This result is used by the process monitoring application in the following ways:
The use of Falkonry ready-to-use machine learning then improves machine utilization, reduces downtime and maintenance costs, while also supporting early diagnosis of raw material issues.
Industrial Predictive Analytics
1307 S. Mary Ave. #215
Sunnyvale, CA 94087 USA
+1 (408) 461-9286
© 2013-17 All rights reserved. Falkonry Inc.
All rights reserved. Falkonry Inc