The Internet of Tires, Railcars, and Other Things
The first trip I took in my new Honda CRV was down I-81 through the Virginia Blue Ridge. It was a beautiful drive, until I noticed an orange alert icon on my dashboard. I had no idea what it meant, so pulled off at a rest stop to look at the manual. Turns out it was a low tire pressure alert (would have been nice if the icon looked remotely like a tire). I looked carefully at my brand new tires, which all looked fine, until I saw the nail sticking out of the sidewall of one of the back tires. Quick visit to a gas station at the next exit, and I drove home going no more than 45 mph on a temporary spare tire.
This incident sold me on the safety benefits of sensors in equipment. Until …
Next month, driving to work on a brand new replacement tire, I saw the light go on again. I pulled over, looked at the tires, saw nothing wrong, and continued to work. Took the car to Gene’s Tires, the local tire shop, after work. Gene looked at the tires, pronounced them fine, and sighed. “You have no idea,” he said, “how many times a week this happens. Those sensors are wrong most of the time.”
Fast forward to today, when I have gotten accustomed to the drill. Light goes on. I look at the tires. Even measure the pressure (yes, I now travel with an old fashioned tire gauge). Reset the light, and keep going. I estimate a ratio of 20 false readings to one legitimate one. Talking to colleagues at work, I realize this is no surprise to anyone but me.
What does this have to do with chemicals? This incident reminds me uncomfortably of a conversation I had several years ago with a chemical company that routinely ships hazardous chemicals by rail. Railcars can be equipped with a number of sensors in addition to GPS tracking devices. These sensors will relay information over cellular networks about things like the temperature inside the railcar, whether the hatch on a tank car is open or closed, whether the car’s handbrake is set, and whether the railcar has experienced any impact. All this information could be critical to determining whether there is risk of a chemical spill or other type of release. You might think companies would be anxious to equip all their railcars with such sensors, but in fact, with the exception of railcars that carry extremely toxic materials like chlorine, most railcars carrying chemical shipments are not so equipped. Why? One reason is that the sensor readings are not reliable enough.
If a hatch registers as open somewhere along a rail route where the railcar is not being loaded or unloaded, the shipper who receives the sensor reading is obliged to find out what is going on. This might require sending a helicopter to check – a pretty expensive proposition if the railcar turns out to be just fine, which is apparently what this company has experienced many times. The consequence is that many/most of their railcars are not sensor equipped because the costs – including the cost of responding to false alerts – simply does not justify the benefit.
All this leads me to think more broadly about the Internet of Things. In the vast amount of data we receive from equipment in, for example, manufacturing plants, how much of it is accurate? By some estimates, as much as 40 % of data from “edge” devices connected to IoT networks is inaccurate. So what’s a poor data analyst to do?
The solution I came up with for my car turns out to be similar to one of the methods for identifying bogus IoT data. I used another “sensor” to validate the readings of the onboard sensor. Turns out redundancy is a common approach – to cluster sensors that report either the same or related data and compare readings to determine whether a beyond-threshold reading from one of the sensors is accurate.
Another approach is to use a completely different kind of sensor. In the railcar hatch example, some companies have added webcams to each railcar carrying hazardous substances. The cameras take a picture of the top of a railcar whenever a hatch open or impact reading is received. The picture is used to validate the sensor reading and determine whether emergency action is required.
Then there is the approach of combining data from multiple sources to draw conclusions that can improve over time through machine learning. What happens if the energy consumption of a pump seems to be fluctuating but the flow and temperature readings are within range? Is one of the sensors wrong or does this pump draw power differently under different circumstances? Machine learning can help to improve identification of “normal” behavior for this pump and decrease the number of bogus alerts.
All of this should remind us that getting value from IoT applications depends not just on devices and on intelligent deployment of the devices, but on the analytic capacity of the platform that manages and drives reaction to the device data. Check out SAP’s Leonardo offerings for some examples of the business value of IoT analytics.