Usually I do not blog until the evening times here in Minnesota; but today I stumbled across a very good blog by a very good blogger I follow David White which I had to push out there. He blogs at www.industrial-iot.com. The blog I read is here. He talks about machine learning in industrial applications and I especially like this quote:
"In the long run, a weaker algorithm with lots of training data will outperform a stronger algorithm with less training data. That’s because machine learning algorithms naturally adapt to produce better results based on the data they are fed, and the feedback they receive."
This has really been true during our trials and tribulations of developing a product. In the earlier days often we were trying to put a single feed sensor on a line and drive value from that. We spent a ton of time on our algorithm of that single sensor and got it to be pretty good. Even with that really good algorithm, the data would quickly get skewed, generate false alerts and leave customers unhappy. This was painful to go through. In reality, we needed to have several more sensors feeding in to a robust data platform (like GroveStreams) which then we use all the data streams to interact with each other to provide the valuable information. As a result of these streams interacting we can then teach the sensors how to be more accurate. If we use several different sources of data I have now taught devices to be much more accurate over time as David White suggests. This is a huge breakthrough for what we are working on accomplishing.
Often times users are on a budget and only wish to try out certain sensors to save money. We need to really resist this urge for this very purpose. All of the streams working together provide a more complete picture of performance and are needed to effectively reach the goals. If we do go halfways at a site, users need to realize the picture painted may only be halfways completed.
No comments:
Post a Comment