In anticipation of his upcoming Predictive Analytics World for Manufacturing conference presentation, Staying Ahead of Failure: Parametric Data and Analytics in High-Tech Manufacturing, we interviewed Field Cady, Senior Data Scientist at Think Big Analytics. View the Q-and-A below to see how Field Cady has incorporated predictive analytics into manufacturing at Think Big Analytics. Also, glimpse what’s in store for the PAW Manufacturing conference.
Q: In your work with predictive analytics, what behavior do your models predict?
A: I have addressed many different problems with different clients – we see quite a range. Within the manufacturing space we mostly focus on predicting failures of components early in the pipeline, so that they can be removed and not cause problems downstream. Identifying duds early is the main thrust, but the models can also be dissected to try and find what is causing problems in the first place. Another interesting problem we have worked with is component matching. If you’re not careful it’s possible to combine two good components into a whole that doesn’t work well because of random variation. Like maybe a screw is on the small end of the acceptable range and a nut is on the large end; put them together and you won’t have a tight fit.
Q: How does predictive analytics deliver value at your organization? What is one specific way in which it actively drives decisions?
A: We are a Big Data consulting firm, so predictive analytics and the software behind it is our bread-and-butter. Among our manufacturing clients (at least in my experience) analytics is used less to drive business decisions and more to monitor and optimize assembly lines. The earlier you can identify a faulty component, for example, the less money you will waste on it in later stages. It’s often a machine learning model that is making those calls in real time. Analytics is also used offline to try and diagnose problems and inefficiencies in the pipeline.
Q: Can you describe a successful result, such as the predictive lift (or accuracy) of your model or the ROI of an analytics initiative?
A: Most client results we’ve had I can’t talk in detail about. But I can say that we’ve made some shockingly accurate predictions of particular types of component failure. Pass/fail criteria for components are usually based on only a few parameters, so it’s amazing what you can do with a machine learning model that takes dozens of features into account.
Q: What surprising discovery have you unearthed in your data?
A: Within manufacturing we’ve been constantly impressed by how many ways physical hardware can go wrong that you don’t see in other horizontals. For example, some incomplete data joins ended up being so because the component IDs had been physically stamped into the components, and the OCR read them wrong.
Q: Sneak preview: Please tell us a take-away that you will provide during your talk at Predictive Analytics World for Manufacturing.
A: My talk is about using Big Data to “keep our finger on the pulse” of a large, complex manufacturing pipeline. Oftentimes these pipelines go into expensive “red alert” modes when the failure rate at some stage becomes too high, and everybody scrambles to figure out what the problem is. Those crises can cost tens of thousands of dollars an hour. We were able to create a system that passively monitors all stages of the pipeline, giving early warning signals for possible problems and making it easier to diagnose issues if there is another yield crisis.
Don’t miss Field Cady’s conference presentation, Staying Ahead of Failure: Parametric Data and Analytics in High-Tech Manufacturing at PAW Manufacturing, on Tuesday, June 9, 2015, from 3:55-4:40 pm. Click here to register for attendance.