Description
By monitoring sensor data of physical processes real time, we may
predict and mitigate failures. In the presentation I will illustrate
the new generation of time series prediction problems by two tasks:
- We predict scrap rate in transfer molding to package automotive
electronic components in epoxy for protection.
- We predict connection loss in radio networks based on mobile base
station logs, and independently of mobile device logs.
The difficulty of predicting the above processes lie in the complex
structure of the available data. For example, manufacturing data is
organized hierarchically into individual products, groups, charges,
work shifts, cleaning cycles, etc. And mobile device data is highly
dependent on device, operating system and permission types and vary
widely from device level constants to rapidly changing accelerometer
measurements.
In the presentation I will describe feature and shape based time
series predictive models, which work for multiple time series of mixed
measurement types.