Not all revolutions unfold at the same speed. And even the same revolution sports multiple timelines. Take the 4th industrial revolution: Depending on your type of industry, your location, your workforce and your company size, your speed and degree of digitization might vary strongly. The situation becomes even more granular when looking at the building blocks that make up the 4th industrial revolution.
One exemplary building block of high interest to us, the global specialty chemicals company Evonik, is machine learning (as part of the field artificial intelligence)…, which – one could reasonably infer from related conferences, TED-talks or (digital) magazines – in some form or another is promising to solve basically any challenge we will face down the road. Adopting these practices in a company that stands to benefit from the promised gains should be straightforward, right? Well, (large) companies in the non-digital, traditional industry space often have well-tuned, complex and strongly ingrained systems and methods in place, which do not typically lend themselves to fast change. Hence, when we at Evonik tried to leverage our already available but dispersed expertise in machine learning, we got caught in a loop of slow progress:
To break the shown cycle, we worked through four basic steps, which meant not answering all (unknown) questions before the fact:
1. Increase transparency: Network and create a company internal overview on internal Machine Learning (ML) expertise, experience, and potential applications.
2. Accept uncertainty: Make a conscious decision that your initial investment in ML might turn out to be an apprenticeship premium.
3. Identify pilot(s): Start on a low complexity use case in a non IP critical environment, with a low complexity application.
4. Keep moving: After the initial use case many questions will have been answered, but new ones will arise. Go for a minimum viable solution to start generating value without limiting future options.
If you accept that your initial ML pilot does not necessarily save you money, you shorten the decision time for implementation tremendously. That doesn’t mean that the first pilot should not be well thought out. Quite to the contrary. But the selection criteria are different: Instead of the use case with the highest value we chose for a low complexity application (health monitoring) for a relatively simple unit (single compressor) and a non IP critical case (chilled water system) with an external predictive maintenance provider.
“If you accept that your initial machine learning pilot does not necessarily save you money, you shorten the decision time for implementation tremendously”
By doing this we were able to quickly learn about the used technical approaches, we didn’t have to worry about data security and finally the data infrastructure was not a limiting factor (low frequency data with no demands on latency). By choosing wisely we were able to implement this first use case within 4 weeks.
So, we didn’t save money (yet), but we learned a lot and progressed to the next set of questions for ML applications that now finally addressed scalability and value for the company:
1. How do we build ML powered asset health monitoring (predictive maintenance) for Evonik?
2. How do we leverage ML towards goals like energy or yield optimization (manufacturing space), logistics as well as R&D and market trends?
For both these questions we had to consider how we perceive ML in the mid- to long-term. What is our strategic need? Is there potential value through ML for all areas within the asset lifecycle and the supply chain (refer to below depiction)?
To guide through our thought-process, lets choose the example of the manufacturing space: Evonik generates significant value through specialized production know how. Consequently, domain know how is critical to understanding data-based results, including results from ML algorithms. Hence Evonik wants to have significant in-house expertise in ML for the process/chemical industry in order to be able to continue to improve its operations without being overly dependent on external providers. And even though we do not see predictive maintenance based on ML as our main objective, it provides a very good entry-point into applying ML at Evonik and allows to grow our internal capabilities.
Based on the initial pilots and the mid- to long-term strategic analysis discussed above, we decided to build our own internal predictive maintenance center based on commercially available ML software. We built the center on-premise, within the company firewall, on scalable IT infrastructure and – most importantly – on an open IT architecture that allows us to test, add or change the calculation and visualization engines according to the best practices of the future.
In parallel, we have started leveraging our ML experience to venture into new fields, as diverse as forecasting accounts receivables and yield optimization in our plants. With each new challenge we grow both our internal capabilities as well as the understanding of the commercial ML market and its capabilities.
In short: Bypass paralysis through analysis, and instead dare an initial (paid) jump into the unknown. What you will encounter might very well be different from your expectation, but it will allow you to move forward nonetheless.