Concentration and Moment Inequalities for Vehicle Sensor Data

Vehicle sensor data is hard to deal with¶

I have been considering how we can apply concentration and moment inequalities to ‘models’ built on vehicle sensor data to determine whether the ‘part’ modeled by the suite of sensors is healthy or not. To do this, we of course use a sequential testing framework, as regularly checking to see whether the vehicle is still healthy introduces a great deal of false positives due to the multiple testing problem.

A particular issue we must contend with is that the sensors themselves don’t have known uncertainty bounds, or bounds on higher moments for that matter, to which I currently have access. Further complicating the problem is the fact that any models built on top of sensor data will introduce further uncertainty, but perhaps in different ways. So finding ways of characterizing or bounding tail behavior of ‘healthy’ sensor suites becomes quite difficult.

Another complicating factor is that not all sensors are independent - either of one another or of vehicle line or other parameters.