Tuesday, 09 May 2017
You could accurately call a single variable trend line an “analytic”. In the same way, you could use the word to describe a predictive algorithm extracted from a neural network-driven correlation analysis. And that causes problems for those of us in the realms of providing or implementing technology.
When confronted with high-value problems or opportunities, our customers (in industries as diverse as food manufacture, metals and utility-scale power generation) found two lines of thinking that helped them work out what capabilities they needed and where to get them.
Setting the nature of the investigation, and understanding where you are in your journey:
Are we trying to more finely characterise a performance or reliability question that we already understand? Or are we trying to discover effects and relationships we haven’t considered?
This isn’t always an either/or question – generally it should be one, then the other. Two years ago, a specialty metals manufacturer expressed an interest in multivariate analytics for process analysis – aimed at the second of the two questions above. Two years on, they’re still making gains in everything from energy consumption and process reliability thanks to super-special tools like (wait for it…) Excel. In their case, more advanced tools would have been premature – additional cost and capability on the tool side wouldn’t have gotten them any farther in terms of practical results to date.
There will come a point when the questions become more complex, and Excel will run out of steam (hard to do multidimensional sensitivity analysis on 20 or more factors in a spreadsheet!). At that point, tools that make it easy to rapidly perform “discovery” experiments on complex data sets will offer value. To illustrate the flip side of our metals processor’s experience: at a power plant, a team of engineers spent months analysing data with spreadsheets and a stats package trying to develop a predictive failure signature for a turbine. This meant a labour-intensive process of selecting small combinations of variables based entirely on informed opinion (or “guessing what might have mattered”, in layman’s terms), and applying techniques that could only yield conclusions if risk indicators were “self-contained”, rather than interlinked.
Borrowing the wisdom developed by those customers and others, we now guide our customers to consider the following:
- Have the obvious lines of investigation dictated by the process or machine science already been evaluated?
- Are there more than a few variables to consider?
- Will those variables come from a mix of systems?
- Will there be a need to solve for optimisation of competing outcomes (quality vs. energy consumption, for example)?
Turning insights into action - what do we do with "new truths" the tools have delivered? Is it enough to have a human able to describe new insights and incorporate them into improved recipes, procedures and algorithms? Should analytic outputs be incorporated into the broader systems stack to automate the use of those insights? Whether your problem sets are simple or difficult, plan for how their conclusions will be consumed, assessing any or all of:
- How a model developed “offline” can be incorporated into a real-time comparison engine
- Especially if that model is based on analysis across datasets from different types of system (ie: material profile data, process/machine event data, maintenance records, quality records, ambient condition data)
- Ease of configuring write-back to automation layer systems
- Methods for triggering system-to-system processes and events outside the automation layer
- Delivering visualisation to supplement reporting (both regular and exception-based)
A fully-fleshed reporting and analysis plan will incorporate a mix of traditional reporting, Business Intelligence and statistics tools – advanced analytics shouldn’t be considered a replacement for the capabilities they provide. What our customers’ experiences show is how to understand where the limits of those tools may lie, and ways to assess if and how new solutions can provide a useful supplement. We hope what they learned will help you understand your potential path forward, and provide some useful ways to assess the different options you find.