91ÊÓƵÍøÕ¾ÎÛ

Machine monitoring happened fast. Once the enabling technology was in place to allow real-time gathering and aggregating of performance data from CNCs (MT Connect, for example, is an important enabler), machine shops quickly embraced the capability. Now, the digital dashboard showing pie charts or green/yellow/red displays of machine tool status information has become a commonplace sight in shops, though it was practically nonexistent not many years ago. The change happened so fast that we can reasonably expect the next step after machine monitoring to happen quickly as well. I believe that step is liable to be machine learning.

After all, there is far more data for the machine tools and processes to disclose than just what pie charts show. As machine monitoring systems get better at capturing that data, and as the information available from ERP systems is drawn in as well, the amount of data will soon grow beyond what machine shop personnel can reasonably make sense of day-to-day. However, large data sets are both the resource and requirement for machine learning. Once a few enabling technologies are in place, it seems reasonable to expect that this form of artificial intelligence will help manufacturers comb their performance data automatically to find opportunities to optimize processes or costing through improvements that otherwise might be too subtle or counterintuitive to see.

And yet, data is not wisdom. To get ready for a future in which machine learning is a commonplace tool in industry, a book I’ve found worth reading is The AI Delusion by economist Gary Smith. Contrary to the title, Dr. Smith does not discredit AI, but instead questions the level of faith many place in it. AI simply looks for correlations in big data sets, that’s all. Humans finding correlations in data may or may not find something useful, he notes, and the same is true of AI. When AI comes to manufacturing facilities, it will be another tool in the shop — an effective tool but nothing more.

Some interesting points about AI and data analysis I drew from Dr. Smith’s book include the following:

• Looking for correlations in data and expecting to act on them can be dangerous. Any large data set is going to have correlations that look predictive, but in fact do not mean anything. Patterns are part of randomness. If we find a correlation between two variables (say, discovering the rise in grooving insert tool wear correlates to a rise in job cost overruns), then we should consider the correlation meaningful only if we had good reason to believe the variables were related before looking at the data.

• Big data can be bad data. A big data set can be full of meaningless input that gets in the way of the few measures that are valuable.

• We can always rationalize why a seeming correlation found in data might make sense. This doesn’t make it true. As a test, assume the data suggest the opposite of the prediction you see. Watch how easily you can rationalize that correlation, too.

• The problem described in the points above has already played out in other fields. Seemingly convincing correlations that are actually meaningless are the reason why we’ve seen so many popular medical findings reversed (various substances that have been linked to cancer, for example). In manufacturing, the use of machine learning to mine data sets is going to reveal the same kind of noise.

• The new computational tools are only tools, and our words for them are only analogies. Machines using “artificial intelligence” are not really intelligent. “Machine learning” does not lead to a machine truly learning anything at all. People are the ones who learn (hopefully).