Imagine being at a party and talking to someone about a great restaurant experience you had. You might say the service was amazing and the staff so attentive they imperceptibly attended to your every need.
This level of customized attention, is what our shared mobility customers like Tesloop, try to create for their customers, day in and day out.
Just like a restaurant needs positive customer reviews to stay in business, fleet operators need a good reputation for service. For a vehicle fleet in a user-facing business, cleanliness and reliability are major brand differentiators. Just like restaurants are at risk of going out of business if their competitors offer better food and service, a car sharing operator’s days are numbered if it gets known their fleet is less well maintained than the competition. Because Operation and Maintenance costs are a big fraction of fleet operator’s expenses, any misspending here could make the difference between capturing market share and turning into that one restaurant that used to be here. What was it called? You get the idea.
From a data science perspective this is a really cool problem to work on for several reasons:
First, it is fundamentally a psychometrics problem. What a user judges as satisfactory is fundamentally proprietary and subjective as it is based on opinion and perception. “Good service” at a fine restaurant is subjective, and very different from “good service” at a fast food joint. Fleets of cars are just as unique. A user experience in a luxury electric car is going to be measured differently than a user experience in an affordable urban compact. Being able to develop models based on the subjective experience of users is essential to a fleet maintenance provider’s success.
Second, the economics of this problem are all about minimizing decision risk. Just like with the smoke detector in your home, you have a classic two-by-two decision classification problem:
|There is a fire in your house||There is no fire|
|Smoke detector sounds an alarm||Whew, thanks for warning us!
This was exactly how we wanted a smoke detector to work.
|We might have wasted a few minutes running outside, but it was good prevention. It’s okay if this happens occasionally to prevent catastrophic events.|
|Smoke detector stays off||This is catastrophic. Life and property are at risk and we need to minimize how often this happens.||This is exactly how we want a smoke detector to work.|
This problem of balancing decision-risk is a classic problem in data science and statistics and it boils down to building the perfect smoke detector.
Our fleet operators and maintenance providers manage decision risks like these on a daily basis. If they send maintenance a little too soon, they waste some amount of per-unit cost to prevent having a dirty car go out. If they send it too late, they risk having an unhappy customer and churn a profitable user. Now what if that user had a particularly high customer lifetime value, or felt the need to share that bad experience in an online review that said the car didn’t have enough charge, had trash in it, or was smelly.
With that framing, we can take a closer look at the math involved. First, we build and train a model to predict what good maintenance and customer experience looks like for a fleet operator based on the vehicles’ makeup. This can include the charge level and age, the number of uses it had since its last service and the zip code where it was dropped off. This model is purposely subjective, and tuned to what real users experience with that make and model of car as well as numeric inputs from the car’s use. Now that we have a way to rate and compare outcomes from service inputs, the problem shifts to economic optimization.
The economic goal of this problem is to maximize revenue from good decisions, such as whether to send maintenance a little too soon rather than risk sending it too late. These factors are all represented in a model with the expected value of each outcome, a model for predicting churn based on customer segment, that can optimize when and how maintenance should be triggered. What this means is that we are in effect, building the perfect “smoke detector for fleet maintenance.”
This is all pretty straightforward decision theory. The really exciting stuff is how Ridecell Fleet Ops is tying it to vehicle telematics, which are the stream of on-board sensor data a car collects during operation. By tying these together, we can offer a fleet maintenance provider, like RideKleen, a concrete decision framework to compare various economic outcomes: sending limited service resources to inspect a car because sensor data indicates it might be out of alignment, combining repositioning with cleaning a few hours ahead of schedule to hit peak demand somewhere else, triggering an inspection after some high-acceleration events to make sure there isn’t any damage to a vehicle before it goes out again.
By developing these tools in an integrated data science framework, Ridecell is working to provide efficiency and a path to profitability to fleet operators and servicers alike – powering the future of shared mobility.
Author: Shawn Higbee, Senior Product Manager, Data Science, Ridecell