A central pillar of AspenTech’s mission is digitalization of the world’s most capital-intensive industries, using the combination of domain knowledge and machine learning analytics. In my mind, for an operating company this is the powerful and necessary blend to find the sources of uptime losses and margin leakage, to identify the causes of such events and to promote the best possible outcomes.
I’d like to tell you, like an arrogant CTO of a startup machine learning company once told me, “Just give me the data and I will sort out the problems.” Unfortunately, it does not work that way. Machine learning is portable across industries, domain knowledge is not – and you need both.
A monitoring solution must separate causation from simple correlation correctly and alert only on true impending issues – see my earlier blogs regarding correlation versus causation. But machine learning is NOT a silver bullet. It needs ‘guiderails’ to find correct answers. The guiderails come from domain knowledge, translated into contextual data limits that establish reasonable expectations of behavior and exclude meaningless correlations that machine learning will find.
Experienced staff can share their understanding of asset behavior under changing circumstances. Insights may also arrive from complex first principle and empirical models that forecast the likelihood of specific results. All in all, that context is crucial to correctly label events, select variables and direct the data cleanup.
Take a Two-phased Approach
First, learn about the process and its failure modes, correctly label the breakdown events and calculate some imperative events (such as the temperature where hydrates form). Use this information as guiderails and cleanse data and subsequent event patterns with an understanding about operating modes. Then when the engineering effort is done switch into data scientist mode. Once there, you’ve supplied the data context: now the algorithms don’t care about your industry or assets.
Experts in new marketing verticals continuously challenge us. They firmly believe their process and equipment are different and we need to know something new. However, unlike alternative condition-based monitoring (CBM) approaches, with Aspen Mtell® nobody needs deep process and engineering experience. Between AspenTech and the customers, the full set of capabilities already exists. We learn the pertinent process and equipment details from each customer… and they learn the transition from engineering to data science from us – the guiderails.
Consequently, in the analytical depths of Mtell, the data, algorithms and patterns do not know from whence they came: what asset type, chemical process, asset class, failure mechanism, causation … nothing! But the data preparation will declare simply and easily the data context that must be known.
Secondly, with guiderails in place the machine learning Agents learn two things:
- Pattern similarities in carefully selected data sets; for example, exposing true normal behavior (unsupervised learning); and,
- the exact patterns and timing leading to an event (date and time) e.g., a failure event that may be manually labeled (supervised learning)
Interestingly, as my arrogant associate asserted, in the data domain it is “just data.” Scales, engineering units and data sources can be diverse and do not matter. There we do not strictly need the rigor of engineering models and the implied complex differential equations. But the data input guiderails do matter. For machine learning to work, you must always do the upfront engineering to get the data right!
New industries and assets make their way to AspenTech’s doors. We talk to your process, mechanical and instrumentation experts, and in short order reveal the knowledge needed to monitor equipment in your industry. Matching appropriate data groupings and sample frequencies with significant events creates the guiderails and leads Mtell’s Agent-based machine learning to discover normal operations, deviations and patterns that lead to degradation and failure for any asset, in any industry, for any failure mode.
Merging domain knowledge and machine learning assures Aspen Mtell delivers extremely early, accurate and trusted alerts.
See how low-touch machine learning is fulfilling the promise of asset performance management in a recent white paper.