by Sergey Sviridov
HELSINKI - AI is still considered by some market players as either a Messiah, capable of solving all the problems of mankind, or a kind of monster that will take away jobs from people and conquer the world. In reality, both perceptions fall wide of the truth. A lot will depend on tech growth rates in the near future, but hardly anybody doubts the promising outlook for AI.
According to a
forecast by Gartner, AI augmentation
will generate $2.9 trillion in business value and recover 6.2 billion hours of
worker productivity by 2021. And whereas, in 2017, the AI market was estimated at
$ 4.8 billion, by 2025 it is projected to increase by almost 20 times
to $89.8 billion. The AI that is being used now in industry and production is
capable of addressing local tasks.
It must be recognized that there are a number of intrinsic factors limiting AI opportunities and the scale of AI implementation. Data seeking, gathering and processing take up 60–70% of all AI project implementation time.
Roadblocks: data storage, collection, and management
Enterprises currently lack a unified system for data storage and processing: data is collected in separate and often nonintegrated systems. Approaches to storage organization are also different.
Furthermore, the process of data management (its quality and completeness) is also of great importance. Many enterprises simply do not collect the data required for implementing AI algorithms.
Thus, for the moment, the pace of AI implementation is restrained by the fact that enterprises are not ready to invest in data storage, collection and management, and to incorporate the data into key business processes.
In addition, there is the complication of sector specificity and differing technological processes – it is impossible to introduce standardized AI solutions across the petrochemical and metallurgical industries, for instance. Sometimes an individual approach is required to each industrial installation.
To give an example,
each blast furnace is custom-made, with characteristics that are individually
selected to address specific production tasks. Life cycle and obsolescence levels
also vary. The same challenges arise with oil refineries, where phasing
limitations can be unique for each enterprise.
As we know, not everybody likes AI, it is impossible to standardize it, and its current level of development does not enable adjustment to the various tech solutions in complex production systems.
That is why everything still needs to be explained in layman’s terms, teaching the program how to handle each process separately. In addition to information on the work of the enterprise, AI needs data on the business and physico-chemical processes relevant to each type of production.
The value of an efficient data compilation process
Experienced process engineers often rely on intuition in the metal production process. For example, they use ferroalloys to ensure the required chemical formula. Ferroalloy steel has improved physical and mechanical properties.
As evidenced in practice at different production sites across a single enterprise, there can be fluctuations in the amount of added ferroalloy. This greatly influences the cost of raw materials and could be optimized if the decisions on ferroalloy proportions were not based on a standardized AI approach rather than an engineer’s intuition. But the problem here is the absence of an efficient data compilation process: mechanisms are often generated in somebody’s imagination and never formalized.
Can AI learn by watching a process engineer at work and collecting data on how he makes decision depending on external factors? Yes, it can and should. But AI decisions will be limited by these human-learned schemes.
If AI stumbles across an unfamiliar situation that it hasn’t yet been seen, it could give an incorrect recommendation and will be forced to transfer control back to the process engineer.
It is important to note that several major suppliers of industrial equipment, including Siemens and Mitsubishi, who initially instructed subcontractors to apply a data collection system in their equipment, have a certain advantage when it comes to AI implementation.
On the whole, the
lack of unified standards and operating procedures are thwarting the process of
digitalization. Thus, at the end of the day, the rise of the machines will be
stopped by human thirst for intuition-based decision making and a total breakdown
of the production process.
AI’s ability to critically evaluate is strictly limited by its rigid framework. That is why AI solutions are implemented only in carefully chosen production stages and plants. We are not talking about intellect at this point, only about its roots. For the moment, AI is not capable of anything – neither learning nor understanding transient factors – without human beings, because the industry has been created by humans for humans.
AI may replace humans in the very distant future, but for now, when finding itself in an uncertain place the program defers to the human operator. And this is right, because safety should be put at the heart of the whole process of AI implementation. Incorrect AI decisions should be immediately cut off by human beings. It is our hope that research will lead to AI capable of aggregating information and taking correct and safe decisions in critical situations.
Sergey Sviridov is Head of R&D at the ZYFRA Group