A couple of weeks ago I saw some information about AI’s entry points into manufacturing and some of the beneficial effects seen in those technology/process junctions. Several of these areas are germane to the readers of this newsletter, to wit:
- Predictive maintenance
- Quality control
- Robotics
It was very interesting and it prompted some thought about that first bullet point. Although the article did not discuss it, predictive maintenance (and for that matter, quality control) are, as manufacturing activities go, on the passive side. Make no mistake, the subject has proven quite valuable in efficiency, accuracy, and gathering information and making an estimate of a need for planned downtime, for instance.
A handful of years ago Cosen Saws brought out a predictive maintenance function on their saws (which included the blades that ran on them). It was interesting and helpful, the idea being that sensors could help us manage parts replacement/refurb, and that collected data could point the way toward accurate maintenance schedules. All true.
We’ll stay with predictive maintenance, but let’s introduce the real time (or as close to real time as is necessary) AI machine function evaluations that use similar data collection but do something quite different with it. For a number of years, Mitsubishi has offered laser cutters with AI functions. It has the aura of “cool,” because it uses sensory inputs like sight (the color of the cut) and hearing (the sound of it too) to make determinations on the “health” of the cut.
Now here’s the really cool part—it shares that information with the control system of the laser cutter and if things are going wrong, the machine will perform any number of tasks like increasing or decreasing the speed of the cut, or even temporarily abandoning the cut if necessary.
As great as predictive maintenance is, suppose it could be extended. Instead of saying, here are the 12 areas of data collection that are driving a prediction of maintenance on this date at this time for these replaceables and consumables, it could also either a) tell you to take the following actions while running, or b) let you know that the machine itself is taking the following actions while running. In other words, combine predictive maintenance with the instantly reactive intelligence from the laser cutter.
This could have positive effects on duration between maintenance, consumable time, replaceable time (parts and subsystems, I mean). Mostly it takes away the finality of predicted maintenance, because as the AI systems become more fully-featured and autonomous, the predictive maintenance date should move a little more into the future with each bit of efficiency and reaction from the system.
However, this new setup begs some unwanted complexity, namely: 1) where do we put the data? 2) how do we prioritize the actions that either we take based on recommendation, or the machine takes based on its own AI-based programming?
Let’s take these in order. Where to put the data? My vote would be to split the computing between edge and core. You would put the AI and some processing at the edge (close to the work being done and the sensors collecting data about it). You would want enough processing power here to make sense of the sensor inputs and report the recommended actions into the machine’s operating system. I would think that the core operating system could be remotely flash-updated (to use a slightly old term) or done via a portable drive. But it might make sense to modularize the edge computing, because new capabilities are coming out all the time and we would not want the AI features to be hogtied because of an outmoded communication infrastructure or some valuable but unique methodology.
As for prioritizing, there would have to be some parameters inside the AI computing system. Maybe some of the parameters would be about speed, work/time (for example, tool depth setting on a machining center, head speed on a laser cutter, approach and bend speed on a press brake, angle and speed on a saw), and even proximity (how close is that nozzle of the laser to the sheet?). There are a multitude of examples.
These parameters and their values would have to be weighted—it can be done but it is a somewhat herculean task because the units are immiscible. Eventually it would become an exercise in “what don’t you want the most,” perhaps it’s an overheat, a collision, tool failure, etc. You can see where even this simple idea gets complex very quickly. We did not even mention variables such as material quality and its effect on each one of the parameters.
Of the few things I am good at in this world, creating AI systems somehow is not on that short list. But I do think we can approach predictive maintenance differently if we allow AI to drive the data (and build machine tool systems to accommodate core functions fed and led by AI functions) and have a mindset of completely safely delaying maintenance based on actions performed by the system to extend the life of those things needing maintenance.
Coming issues will explore other aspects of AI. I would love to hear your reactions, counterpoints, and opinions on these important topics. Email me at dave@fifthwavemfg.com.