Industrial AI — 3 Barriers to Adoption & How to Overcome Them
How to foster adoption of AI among industrial executives & overcome barriers in their minds.
All of us are steadily exposed to a lot of media diet which emphasizes both, the advantages and pitfalls of AI. Either way, it is the new reality. It has been silently and steadily penetrating multiple facets of every industry.
This steady state AI penetration has largely been in the consumer and enterprise space, and has had tangible and intangible impact on outcomes — changing behavior, shifting business models and disrupting marketplaces.
However, we’re also seeing significant penetration in industrial, manufacturing, energy, engineering and such spaces.
Take oil and gas production for example. We now have sensors tracking every event in the production stage, especially in wells.
This sensor data allows companies to score and predict well health, which ensures minimal to no stoppage in oil production, owing to increased efficacy of preventative maintenance. This translates into increased productivity as a function of the increase in the number of barrels produced per month.
Another less glamorous example would be adhesive manufacturing, which has a lot of operational parameters. Even with advancements in manufacturing technology, traditionally, batches can only be evaluated after production, leading to greater wastage and losses from bad batches.
However, by tracking variables in the early stages of production coupled with sufficient data from previous batch productions, batch quality can be predicted. This in turn, allows for corrections early on in the manufacturing process, leading to tremendous savings and reduction of wastage.
These solutions aren’t just limited to manufacturing and production. They can be applied to real time operation of heavy machinery as well, as in the case of ship engines. Normally, in the event of a shutdown, sensors trigger a warning one minute before imminent shutdown.
However, if sensors are deployed across various components of the engine, coupled with predictive AI, shutdown can be predicted as early as an hour in advance. This is especially useful in naval security, like in the case of the coast guard or patrol boats, where any downtime can have ramifications on the security of a nation.
In this article, we wanted to share practical insights on the adoption of industrial AI and the barriers to it, as it can significantly impact the future of a nation’s economy at large, which is powered to a great extent by energy and engineering companies.
At a macro level, AI applications can broadly be grouped into 2 buckets:
Consumer AI Applications: Cross-sell recommenders, sentiment analyzers, market mix modelers, diabetic retinopathy predictors, etc. fall in this category.
Industrial AI Applications: Process chemical yield predictors, down-hole drilling inefficiency predictors, motor downtime failure predictors while fracking, etc.
We at Flutura have been quite fortunate to have had a ringside view of many practical industrial AI applications in the last 5 years, which have scaled massively beyond just ‘innovation PoCs’.
Interestingly, we noticed a difference in rhythm across both these sectors and asked ourselves 2 simple questions:
1. What discriminates industrial AI adoption rates from consumer AI and why do those differences matter?
2. What can be done about it?
Our from-the-trenches experience helped us understand the head-start that consumer AI has had over industrial AI, and quantify the 3 primary barriers to adoption of the latter:
1: Labelled Data Availability for Industrial vs Consumer AI Models
Lets face it, machine learning algorithms need to hog a lot of labelled data before the model tunes into real world behavior — be it modeling consumer behavior or machine behavior.
If one takes facial recognition as a problem to solve using deep learning neural nets, there is a ton of data to learn from sources like MNIST. A point to note here is that models had to process at least 3 years worth of data before facial recognition accuracies met desirable thresholds.
If we take a similar problem such as detecting product quality image anomalies in a diaper manufacturing plant, or crack detection and progression on sub sea structures, the foundational job of creating labelled data sets need to be initiated. Additionally, these labelled data sets would have to be sufficiently large (at least 1–3 years worth) to train models to a high level of accuracy.
This presents a problem — if the decision makers at a company have an open mind to delayed gratification, these projects take off, whereas if they want a here-and-now solution, these projects get stalled.
How do we address this?
Industrial executives must be made aware that access to labelled data will be a source of competitive advantage. Labelled industrial process and equipment data will be a tool for survival in the soon-to-be hyper competitive marketplace where access to algorithms becomes democratized and access to labelled data becomes the ‘moat’.
Using anecdotal evidence can be quite useful here — take the case of cancer detection startups dependent on radiology data. There are two layers to the solution, the data layer and the algorithm later, the latter of which is democratized. Symphony AI had a monopoly over cancer patients’ journey data from the American Cancer Society, which gave them an edge over everybody else.
Ring-fencing data is the name of the game, algorithms no longer are. In short, DATA IS KING.
Additionally, to ease executives in, a simpler barrier to cross (if labelled data doesn’t exist) is to demonstrate immediate business value in terms of anomaly detection (if applicable) and get a buy-in to simultaneously begin collecting labelled data and establishing process for the long term.
2: Difference in Perception of Dollar-Value Unlocked in Industrial AI
“Forget the fancy jargon, show me the money” is a constant feedback we heard from executives across Houston, Tokyo & Dusseldorf.
Industrial mindsets are used to perceiving value along electro-mechanical dimensions — ‘I can see what horizontal drilling does, I can see how adding vibration and shock sensors reduces warranty liability’. This is more tangible than the digital dimension where you hear — ‘I can’t see what I can’t perceive’ — as a result of which, executives aren’t sufficiently satisfied with the usual answers to the question ‘show me the money’.
What can be done about it?
We find ‘engineering curves’ as a good tool to manifest tangible value additions to skeptical industrial mindsets. Engineers look at graphs which show the before and after of a process, and in this instance before and after industrial AI solutions were deployed.
For drilling contractors, the rate of drilling can be visualized with a ‘Depth vs Time’ graph at each rig state, which can be a construct for making them perceive the value of AI in moving the dwell time at each drilling state, unlocking millions of dollars of efficiency realized across hundreds of rigs.
Every industry segment must have an engineering efficiency curve on which dollar-value perception of industrial AI can be mapped. Find it and you will nail it.
3: Difference in Industrial & Consumer Mindsets
Industrial executives carry a lot on their shoulders. A small mistake could in some cases mean life or death of the humans in close contact with risky industrial operations.
Any change in industrial process is always slow and cautious, mainly because of the need to ensure there is no risk to personnel. A refinery for example, requires a battery of field tests before any implementation.
This is a major difference between consumer and industrial AI — in consumer AI false positives may be acceptable at times, but not in industrial.
The solution:
Having experienced many successes and failures, they need to be empathized with and gently guided into looking at operations through new lens of industrial AI.
Reliability becomes the operative word whether it is a guarantee of prediction rates for complex fracking equipment or top drives or downhole stick shift failure. Establishing the accuracy of the predictive model and ensuring it is well within the margin of error that executives are comfortable with, goes a long way here.
What all of this means is that the adoption of industrial AI to unlock massive economic value is just getting started. There is always a long horizontal line in the adoption curve before a massive spike in adoption.
We need to be empathetic to the needs of industrial executives as they manage the risk-vs-reward equation in a world which is accelerating and changing at a rate never before seen in human history.
This article was written by Derick Jose, co-founder of Flutura, a decision sciences company developing cutting edge industrial AI solutions for energy and engineering industries.
