Your questions answered: Current issues in industrial analytics
Mike Malone, principal process engineer at Toray Plastics America, answered additional webcast questions about data engineering and machine learning as they relate to analytics.
CFE Media hosted a webcast on “Current issues in industrial analytics.” Presenters included Mike Malone, principal process engineer at Toray Plastics America. After the webcast, Malone provided brief answers to a few attendee questions.
Question: What one thing might you have done differently as to the data engineering aspect of the project?
Malone: Spending more time at the beginning understanding the nomenclature of the underlying SQL script that gets the data from the manufacturing system into the self-service data analytics tools. In this case we caught some formatting problems early on and did not have to back track/re do too much. We prepared and executed a couple of small batches of data to ensure the result we were looking for, and happy with, before scaling up the automatic data allocations.
Question: To adequately identify trends that lead to failure, a large amount of information is necessary. What are the best data storage solutions for monitoring equipment that could be integrated into an existing facility?
Malone: At Toray, we have found that the Process Data Historian is the best solution for a centralized storage of all process data being collected from over 10 unique data collection devices/control systems in our plants. The result is a single time series stream of data where smart data compression algorithms are employed to make efficient use of available storage. We have consistently prioritized the value of historical data over the cost or space of said data storage.
Here is a short article that discusses understanding your data historian choices and options:
Question: What do you see as the most significant hurdle or roadblock to implementing reliable machine AI or the like?
Malone: The biggest challenge is going to be eliminating process blind spots in the machine learning (ML) model. All data that defines and impacts the performance of a machine or manufacturing process will need to be accessible in the machine learning model in real time. Missing even one process variable, even if somewhat random, will make it difficult for the machine model to learn what is normal and what is not and accurately predict future trouble. Even today we still have some degree of qualitative process or equipment evaluations that are not easily captured by a sensor in the field.