There is a big problem in Industrial Internet.
The truth is…
Sometimes finding actual applications for data-based insights is extremely difficult.
Industrial production and processes are complicated, which makes implementing analytics in industrial companies that much harder. However, there IS one fairly overlooked low-hanging fruit you could look into.
There’s no value in data in itself. Those are simply facts from the past. The value comes being better able to predict the behaviour of devices, eliminating waste and creatinhg predictive models. The real value is not necessarily in replacing people but supporting them and enabling them to do their work more efficiently.
Machine learning – deriving meaning from data – is an excellent application for advanced analytics.
So, how do you start experimenting with machine learning while minimizing your financial risk and time investment? While there are very few genuine shortcuts, following these steps will remove the confusion from the equation.
1 Figure out what you’ve got
Last week we wrote about turning latent IoT assets active. However, you should also take the maturity your device technology into consideration, as that can affect some connectivity issues and capabilities sensor-wise.
2 Pick your battle
When you’re starting out, it is a VERY good idea to start small. Don’t try to fully automate your entire production in one week, that will fail miserably. Instead, pick a sensible goal that can be met within reasonable timeframe.
A very simple example of a learning algorithm is when office lights dim and brighten based on past actions of employees. This in itself is a very simplistic task (and does not constitute a case for machine learning here), but you can probably draw a connection to similar instances in your plants and fleets.
3 Create a model for the desired outcome
This is where you need to take some time for deeper thinking. Being careful not to equate correlation with causation, look at your production as a whole and ask yourself, “which factors directly or indirectly influence the desired outcome?” and “What are the crucial systems within systems this device or machine interacts with?”
Then create an algorithm and start testing.
4 Build a feedback loop into the algorithm
Wouldn’t it be great if you were done after the first 3 steps? Yes, that would be fantastic. Unfortunately, that is not enough. Your algorithm is probably not going to be flawless from the very beginning, so, you need to teach it. You need to tell it that “when Mike started this last time, it was an accident, so ignore that piece from the data.” The algorithm must learn from the past behaviour of its environment, but also from the feedback from you.
5 Let the data accumulate
Pattern recognition takes time, so be patient in collecting data. Let it run while you teach the algorithm how it should work.
6 Speed up learning with advanced analytics
You can speed up the process by discovering patterns from data you already have. If you have collected data from your devices and machines, you can simply transfer all of that somewhere and perform analytics to it. This will let you iterate on a larger scale and come up with a better model for your desired outcome faster.
In industrial environments, predictive maintenance is one of the most valuable add-ons to your service portfolio. If you are not extending your services downstream to your customers’ customers or upstream to your vendors, you are leaving a whole lot of business potential unfulfilled.
Connecting well-functioning and self-adjusting models to existing processes is good business. If you would like to learn more about the opportunities available to you in this regard, take the first step by contacting us. Let’s get together and see what kind of models might work the best in your situation.
COO, Service Delivery
+358 40 550 2524
Fredrikinkatu 61 a 6. krs