The Analytics of Things is the Payload
In our last posting, we complained about all the noise out there about the IoT infrastructure, the edge devices, the tech, and tried to give you some real nuts-and-bolts advice on getting started. Once you have grabbed a batch of, visually explored, and developed some norms around what’s valuable, what’s noise, what’s garbage, we laid out some simple steps for automating the process for wrangling the data.
Once you’ve freed up 50-75% of the time you were otherwise spending on that wrangling process, think of all the things you could be doing to create real insights from that data. Here are three tangible steps to start getting to valuable insights:
- Fuse sensor data to other enterprise data
- Simple Analytics of Things you can start with
- Running experiments with IoT data
What can I learn about the device?
If you have multi-channel sensor data coming off a device that is deployed in the field, or running in a factory, most likely you will want to understand the device’s behavior in context of its peer devices and the overall infrastructure it’s operating in. You will eventually want to model its behavior as a machine, but first, you may want to just get the basics.
- What is the make, model, manufacture of the device?
- How long has it been deployed and where?
- What is the maintenance history of the device?
- What is its service usage history? Where has it been used? To do what? Has it been used continuously?
- What is the bill of material components that make up this device?
- How has this device changed over time (as repaired/maintained vs as built) because of successive repairs?
- What is the mean time between failures of it, and its peer devices? eg. If this is a truck, and we are tracking the fuel pump, what is the “typical” life of the fuel pump?
- What events have happened that would potentially shorten its life?
To do that you need to collect some relevant data. Notice, the questions above are simple, factual questions that need to be answered, and the data available before you can even answer the more complex questions.
Fusing Sensor Data to Business Context Data
The simple, factual questions that need to be answered will dictate what data elements you need to go find. You will find those either in business systems like your Enterprise Resource Planning (ERP) system, your Supply Chain or Transportation Management System (TMS), the Manufacturing Execution System (MES), to name a few types. Or if you’re lucky, your data and analytics and IT groups have already built a data lake or data warehouse where this data can be sourced from. Wherever you get it, you need to put all those data elements (or features of the device) in one place where they can be organized for analytics and data science exploration on demand. And you have to link it, fuse it, and basically make it so that you can ask any question (like the examples above) of the data.
Analytics of Things Starter Fuel
With a list of the basic questions that you want to ask, the basic analytics and reporting should be easy, right? Well, at least repeatable, you ask? Yes! As a matter of fact, analytics professionals have been hard at work developing repeatable analytics that list all the elements needed, how to organize them, what starter / sample analytic reports might look like so that you don’t have to reinvent the wheel. Two solution accelerators from Teradata, for example, can help with a ready-made list of needed data elements that answer questions like the ones above. These accelerators can help show you how to map the data needed, fuse it to the sensor data, as well as design sample reports and metrics with example definitions and prototypes:
- Condition-Based Monitoring and Maintenance (CBM): for analytics of things that are deployed out in the field, and part of a fleet of equipment
- Manufacturing Performance Optimization (MPO): for analytics of things that are operating within a manufacturing process
With that jump start, you may want to do some simple context reporting and dashboarding on the device, the machine(s), the fleet, or the operation your sensor data is coming from.
Move from simple BI to more advanced analytics of things experimentation
Once you can ask any simple question of the data, and you also have visually explored what data, what features are important, you can begin to design experiments around more complex analytical questions, like:
- What is this particular machine’s estimated time to failure? Which, of a fleet of machines, is the most likely one to fail today?
- What is the best timing of, and repairs to include in proactive maintenance on my production line given the cost of planned downtime, the likelihood of unplanned downtime of process equipment, and the cost of early part replacement?
- What is the most likely set of parts that fail together, so that when I send a service engineer out to repair it, he will have the right replacement parts?
- What is the best distribution of service parts inventory across the globe given the service history, MTBF, and usage profile of all the machines in use by our customers?
For that, now we can really get down to some hard science.
The Experimental Method in Data Science
You’ve all seen it, the CRISP-DM approach to data mining. I’d like to propose that this typically hypothesis driven (tops down) approach must be balanced with a discovery driven (bottoms up) approach to experimentation.
This fusion of deductive and inductive, tops down and bottoms up, allows for innovation to come through and allow data scientists to go where the data take them. More on this topic in the next posting, “RACE-ing to Analytics Value”
Cheryl Wiebe is the west Americas lead for Teradata’s Think Big business focused on the Advanced Analytics Center of Expertise, and she leads the Analytics of Things practice across the various business units of Teradata.
She collaborates with business and advanced analytics professionals and clients on how they can accelerate their capabilities and professionalize their advanced analytics practices. She also advises many clients who are initiating their Internet of Things journey on how to align business strategy with analytics investments to get started on the road to being Competitors in Analytics of Things.
Cheryl openly shares her views and opinions publicly, and was asked to join the faculty of the International Institute of Analytics (IIA) in 2015 to begin talking about advanced analytics specifically for manufacturing companies and the Analytics of Things.
Jeff Cohen is a data scientist in Teradata’s Advanced Analytics COE with specializations in time series data analysis and sensor data qualification. Not a stranger to getting his figuratively and literally dirty, his background in physics, experimental design, and civil infrastructure inspection has afforded him a broad variety experiences ranging from the application of machine learning algorithms to multiple sensor systems to actually being the person in the field installing sensitive sensor arrays.