Can AI be Useful in FMEA?

In short, artificial intelligence is potentially a very powerful tool to improve Failure Modes and Effects Analysis—FMEA—whether it is applied in DFMEA or PFMEA or any other variation of the technique. I’ve often thought about this myself over many years, starting as a student of technical communication in the early 1970's--far earlier than when the term “AI” entered the jargon of daily discourse.

As a result, I have dwelled long and hard about how computer driven databases could make FMEA more powerful and less of a chore.

However, there are a non-trivial number of concerns that must be addressed to make AI useful, and not just another way to complete FMEA work without much effort.

9 Issues That Should Concern Us All

1. The AI system must be carefully trained using deductive methods of performing FMEA studies, including functional matrix construction and correct syntax for functions and failure modes. This will require a level of linguistic and syntactical considerations that almost no FMEA practitioners, engineers, or AI specialists currently understand at an adequate level to build into an AI system.
2. If the AI system is simply fed thousands of existing FMEA studies, many of which are superficial, logically unsupportable, and have little or no value, then the AI will replicate these bad examples. Knowing which FMEA studies are worth emulating is just as challenging as the linguistic issues cited above, and perhaps even more so.
3. Even if these first two issues can be overcome (and there’s no doubt in my mind that, with sufficient research and effort, these two issues are within the grasp of current FMEA understanding and AI technology), there are much larger concerns that must be front and center regarding further automation of FMEA.
4. Existing FMEA software, and to some degree computer aided design software, have automated a significant amount of actual engineering design work. While in theory this is a great idea, in practice it has become something else, namely the delegation of important design decisions to machine databases. Humans no longer consider many important elements in design, and, insofar as the understanding of safety (in particular) and risk (in general) are concerned, this is NOT a positive feature.
5. This really means that technologists are, to an ever-increasing degree, either not required to and frequently have little interest in developing a deep-seated understanding of the actual decision making that goes into a design or a process. While this is an argument that is as old as writing itself (“once young people can write down our legends, they will no longer understand them the way that the elders who have memorized these tales do”), the level of delegation to “deus ex machina” decision-making for critical technology is truly concerning and more than a bit frightening.
6. Those who receive the output are virtually certain to accept it with limited or even non-existent critical review. The potential for safety disasters is very large indeed.
7. Worse, even if good solutions to these concerns can be developed, will the underlying assumptions and baseline information used to develop an AI system be ensconced as standards used for coming generations—baseline information that may be made obsolete by innovations and/or changing technologies? Will future generations be able to untangle what is all but certain to be engineering disasters that will result from this situation?
8. Ultimately, this all boils down to a simple question that FMEA attempts to answer: how do you know what you don’t know? This can be elucidated using deductive methods (and also with inductive thinking, although that is more unpredictable and less reliable) but will an AI-based FMEA system be able to do just that? I have significant doubts about that.
9. Finally, the threat of hacking and distortion of both specific results and the underlying machine knowledge is a non-trivial threat. Of course, this exists even without AI as long as electronic systems are used to make use of FMEA-derived information, but the potential for this to go unrecognized is probably far greater when AI is involved.

Some Final Thoughts

Of course, I don’t want these drawbacks to be used as an excuse for not pursuing a tie-in between AI and FMEA. There’s no question that finite element analysis is far superior to a complete breakdown of engineering mechanics for load, deflection and yield testing. Computer aided design has, by and large, been a huge step forward in design engineering and analysis, and tying AI and FMEA together could accomplish a similar leap.

I must say, though, that I see a non-trivial amount of sloppy engineering being turned into products through careless use of CAD and CAE, and then compounded through perfunctory and check-the-box DFMEA and poor design verification. A similar situation exists with PFMEA, but in a more complex and more difficult-to-see ways.

So, overall, I see this as I do most potential technological advancements. There’s plenty to be gained, but also more than a bit of critical knowledge that can be lost at the same time. If the issues I’ve raised above can be handled properly, the gain almost certainly will outweigh the risks.

There’s a great deal to be done in that vein to get the benefit, though, and so far, the application of computed engineering with respect to FMEA has fallen rather short—see The Lighthouse: FMEA Software—The Pros and Cons (Mostly Cons) for some of my views in this area. (I think my earlier estimate of 20 years for AI to be useful in FMEA is plain wrong now, as AI has grown exponentially in the past 24 months—the time frame might now be 2 years or even less if someone will take on this task.)

This really comes down to a simple proposition: will the concept of “quality” be applied to the usage of artificial intelligence in FMEA? Or will a quick hit for short-term revenue from sales of the AI system be the target? That’s an issue that applies to AI across the board, but it’s one of overriding importance.