Sitech Services B.V. pay-off

Inside the internship of Alex

Finding the right Model in Condition Based Monitoring:
a needle in the haystack problem

In chemical industry we focus on asset health to ensure that the Safety, Health, Environment and Quality (SHEQ) of the plant is maintained. New options arise with the abundance of sensory capabilities through IIoT. The data that these solutions produce are not insightful in itself. A meaningful way of understanding continuity of the chemical process, is to build a “fingerprint” for it, based on Machine-Learning techniques. Failure and fouling of assets can then be identified as a deviation from this fingerprint.

Alex Eijssen started his internship at Sitech Asset Health Center on this premise, and investigated the options for modeling such a fingerprint. As a master’s student in Econometrics and Operation Research at the University of Maastricht, he realized that the vast landscape of algorithms is too big to consider in its entirety, for production purposes. A model is the result of not just a particular algorithm, but what also matters is how this algorithm is configured (hyper-parameters) and how the data is pre-processed before training. So, something smarter was required. Instead of training everything, it should be possible to ‘search’ the best model iteratively.

When we consider AutoML, massively complex meta-algorithms are used to produce the best models. Although AutoML solutions are a great method to get ‘good’ models quickly, it tends to be difficult to customize them for specific applications, e.g. CBM. For this reason, Alex investigated the basic probabilistic search method that is based on a Gaussian process. For the details on his work, please check out his Medium article.

With his internship, Alex was able to prove the value and feasibility of using a search algorithm on top of pre-existing ML methods, that were developed in-house. AutoML, therefore, is achievable in specialized, niche settings like Condition Based Monitoring (CBM). We found out that we needed to start from what we already had, and further customize the search with our knowledge, instead of relying on the generic approaches that most off-the-shelve AutoML solutions provide.

Share this page

Subscribe to our digital newsletter