Just before Christmas we read a deconstruction of 10 myths about predictive analytics published by SAP. It was interesting but very focused on marketing. In our sector, such myths are busted in a different way.
Before we myth-bust, let’s deal with the interesting caveat that’s always put in here by horizontal BI vendors, which limits predictive algorithms discussions to marketing data. That’s down to metadata constraints; to make the data meaningful it needs to be put in context. Not so with manufacturing, or oil & gas data. As we’ll discuss in a blog post soon, there’s tons of metadata within the process historian or DCS, showing current plant running modes. The less structured project data is also metadata rich. We can tell what data is related.
Let’s bust some myths.
Myth #1: Predictive analytics is easy
Well sure. It’s very hard. If it was easy everyone would be doing it, right? But hard for who? It’s not meant to be easy for SAP or us – we’re vendors. It’s meant to be easy for end-users.
Our end-users don’t have to manually configure relationships. Process data, less structured project and operations data is all metadata rich. We can tell what data is related. (Now, there are lots of kinds of relationships; from statistical correlations to physical connections. That’s for another day.)
The metadata exists to remove the requirement to manually configure relationships. This is a myth we can bust. With the right UX design and the right algorithms, predictive analytics can be easy for the end-user.
Myth #2: Scientific evidence is proof
When it comes to marketing, sure, scientific evidence is bunk. But we’re not in marketing land any more.
What really matters to our customers can be measured. Precisely. In terms of tonnes/hr, or safety incidents, or defect arrival rates.
Sure, our customers have complex plants. Recommendations can be made by algorithms that don’t always give the results wanted. That’s why we have machine learning, adaptive algorithms or even human beings running a trusted change control process. (Gasp!)
So myth #2 busted. We can measure success.
Myth #3: Only what you can measure matters
You can measure everything, particularly in oil & gas, manufacturing and petrochemicals. I’m sure you’re thinking that’s not the case, but really, it is.
Physical properties are easy: sensors are cheap. Those parts of the production process you can’t get a sensor into you either (i) don’t care about the physical property therein or (ii) you can calculate it.
Employee fatigue? We can monitor brainwaves to look for tell-tale signs (but hey, everyone knows 3am is the time for an accident to happen). You just need to decide what is worth measuring for you.
If you can’t measure it – either because of budget or physical constraints – it’s probably been measured in a study before. There is no more studied world than that of manufacturing.
Myth #4:Correlation = causation
Really, we don’t find this problem. Customers are not stupid. They know that stock markets and average skirt lengths are not related. They know their plant; that the behaviour of two furnaces are or aren’t related, even if the statistics suggest they are.
Sometimes correlation does equate to causation. Sometimes it doesn’t. The people who run the plant, or project, can call out the false positives easily.
We don’t so much bust this myth as say…well, obviously, right?
Myth #5: Predictions are perfect
Of course they’re not. Again, we don’t find this a problem. In the real world, users expect there to be a little bit of ‘wriggle’ room. There is a risk that key workers believe something they shouldn’t and make an incorrect decision, but in practice they are supported by process experts and have their own experience to rely on.
So, of course, predictions aren’t perfect. That doesn’t stop them being very useful.
Myth #6: Predictions are forever
Well, clearly not. For us, a one shot prediction is useless as the plant, or project conditions change constantly. That’s why we’re into real-time algorithms which can constantly recalculate and make new predictions.
Of course, there’s the outside chance of an unpredicted event. That’s life. There are management processes and automation systems to deal with these eventualities.
Myth #7: You need a skilled consultant to implement predictive analytics
Well, yes and no. During the product development lifecycle we use significant skillsets such as process engineers, mathematicians, UX designers and IT developers so the customer doesn’t have to.
This means that when we come to implement it, we need a little time from the data scientists to validate the data and the IT guys to hook up to the data-sources. That’s it. After that, you do need the internal knowledge, particularly process engineers, plant management and operators.
So, you do need external consultants. You do not have all the skills in house.
But you also need to deal with a vendor that minimises this through good product design.
Myth #8: Predictive analytics is mostly a machine problem
We’re IT guys, so for us it is mostly a machine problem. But we’re not the important people. The customer is; the end-user is.
The key challenges are at the front and back of the process, not the machine driven algorithms in the middle
At the front of the process is data acquisition and tagging so as to give the data context. As we describe above, we have lots of data already within systems which we can use as metadata. We also have smart people who can give us additional context if we give them the appropriate UX to do so.
At the back of the process is the change control and actions taken. Every asset and organisation is different, which means the impact on operational processes has to be taken into account. Failing to do this produces a system that looks great for a few months then falls into disuse.
So,even in an industrial data application, people and processes represent a key part of the solution.
Myth #9: Predictive analytics are expensive
Fancy getting started? How about $12k/£15k for us to optimise your alarm system?
But forget the expense; what about the value?
For some basic operational insights for, say, a bulk chemical manufacturing process, at the same price point we regularly prove ROI > 3000%.
Myth #10: Insights = action
As we say above, the algorithms are just part of a bigger machine, which includes the operational processes which actually effect change within the organisation.
But the challenge here is no different to that posed by all BI solutions; there are plenty of reports generated that no one ever reads, or no action is ever taken on.
All we can do is build a solution that makes taking action very easy; indeed, we built a simple Actions workflow into the system so we could capture tasks and make sure they were managed to completion. (We also built it with an open API so other solutions could leverage it.)
It really comes down to implementing a solution, not a piece of software. That’s why, though the software is superb, we focus every day on service. You should demand this of all of your vendors.
Well, that’s quite enough; each myth dealt with.
We hope that’s useful and we’d be interested to hear your thoughts. As always if you have any questions or suggestions, head over to our LinkedIn group and share your thoughts.