AI with poor data and extreme events

AI with poor data and extreme events

Oct 16th, 2019

We have been speaking about AI with poor data and extreme events in recent blogs and there is a chapter touching on this in our book.

AI with poor data and extreme events

Big data, machine learning and AI can only work if, well, big data exists. A recent MIT study has shown this.

AI with poor data and extreme events, a risky mix

In mining, dams failures are relatively rare events despite what sensationalistic media and activists state without delving with detailed numeric analyses or by biasing results.

Old reporting on accidents were generally censored, possibly biased and based on expert assumptions. It is indeed difficult to perform flawless root cause analyses in the aftermath of a catastrophic failure when extant documentation may be insufficient.

Those are conditions where machine learning struggles. Thick data (oral transmission, experience, until baby boomers are still available) would help, but it is not fashionable (anymore). Everyone loves new buzzwords!  And new technologies.

New technology means new type of data that have a hard time reconciling with older technology/equipment.

AI should learn from failures, but these are happily rare events

Even if we start monitoring every site to death, using IoT etc. machine learning will struggle with dams. As AI can only learn from failures, not from “business as usual”.

Let’s use a Go play example: AlphaGo beat the professional go player Lee Sedol.

The machine uses a neural network algorithm and machine-learning techniques to develop a sense for the game, because it cannot compute all possible choices available a priori.

To learn how to play better and develop a sense for the game it needs to learn how to play, so it needs to play a large number of games to the finish and then see which ones encompassed good moves.

Now if one would stop every game half way, then the machine wouldn’t know if what it did was good or not. The machine could not learn.

That is exactly the case with Tailings Dams and AI. Indeed, we know that most dams will be observed for the longest time with no catastrophic failure, hence the AI will not be able to learn. Mind-you this occurs for us humans as well and we use to say: “no failure, no progress”.

Mis-measurements cloud the issue

My old professor used to say that many sophisticated monitoring systems are only very expensive thermometers. My experience over three decades has confirmed this. And not only as it relates to temperatures.  I am please to note I am not alone!

In Tailings and Mine Waste, 1999, a paper Characterization of pore pressure conditions in upstream tailings dams, T.E. Martin stated for example:

It has been the author’s experience that, more often than not, piezometric monitoring systems associated with upstream dams are inadequate, and in some cases nonexistent. Such inadequate monitoring systems can easily lead to misinterpretation of saturation and pore pressure conditions, with potentially disastrous results.

A 1997 Doctoral Thesis in Sweden discusses the temperature conundrum and monitoring systems inadequacy as follows:

Sweden has 143 large dams and 117 of them are embankment dams. Most of the embankment dams were constructed between 1950 and 1980. One large Swedish dam, Noppikoski, has failed and this happened in 1985. The reason was overtopping due to non-functioning spillway gates. Deterioration has occurred in 26 embankment dams (Nilsson 1995a). Damage has mainly been detected by direct observation, and only in a few cases have indications of inadequate performance been given by the surveillance system. Consequently, both international and Swedish experience show a need to further develop and improve surveillance systems for embankment dams.

Closing remarks

Now, let’s be clear. At Riskope we are all for monitoring. Because we believe that classic monitoring and satellite monitoring of dams are key to the future.

What we are saying here? Simply that poor monitoring, misleading monitoring of the past will not give AI good information to build alerts in the future.

Long time will have to go by before AI can be usefully deployed or predict anything.

In the interim, there is no universal panacea to the ailments of the tailings dams world-wide portfolio coming from IoT, big data and AI alone.

Contact us to discuss rational and efficient integrations. And to learn how convergent quantitative risk assessment will help.

Tagged with: , , ,

Category: Risk analysis, Risk management

Leave a Reply

Your email address will not be published. Required fields are marked *

Riskope Blog latests posts

  • Strategic operational and tactical planning for tailings hydro dams
  • 16-09-2020
  • We are learning that structural integrity of hydro dams is also becoming an issue. Thus, similarly to tailings dams, we…
  • Read More
  • Dam risks informational gap
  • 9-09-2020
  • Dam risks informational gap impacts to a certain extent all dams: hydro, tailings and finally, toxic dumps. There are thousands…
  • Read More
  • Comparison of ORE2_Tailings estimated dam failure probabilities
  • 2-09-2020
  • The comparison of ORE2_Tailings estimated dam failure probabilities is an important element of Riskope continued commitment to test the validity…
  • Read More
  • Get in Touch
  • Learn more about our services by contacting us today
  • t +1 604-341-4485
  • +39 347-700-7420

Hosted and powered by WR London.