Thu, Mar 21, 2019 @ 10:00 AM - 11:30 AM
Time and Location: 3/21 10 am - 11:30 am - PHE 223
PhD Candidate: Yuan Shi
Craig Knoblock (Chair)
T. K. Satish Kumar
Daniel Edmund O'Leary (external member)
Title: Learning to Adapt to Sensor Changes and Failures
Many software systems run on long-lifespan platforms that operate in diverse and dynamic environments. As a result, significant time and effort are spent manually adapting software to operate effectively when hardware, resources and external devices change. If software systems could automatically adapt to these changes, it would significantly reduce the maintenance cost and enable more rapid upgrade. As an important step towards building such long-lived, survivable software systems, we study the problem of how to automatically adapt to changes and failures in sensors.
We address several adaptation scenarios, including adaptation to individual sensor failure, compound sensor failure, individual sensor change, and compound sensor change. We develop two levels of adaptation approaches: sensor-level adaptation that reconstructs original sensor values, and model-level adaptation that directly adapts machine learning models built on sensor data. Sensor-level adaptation is based on preserving sensor relationships after adaptation, while model-level adaptation maps sensor data into a discriminative feature space that is invariant with respect to changes.
Compared to existing work, our adaptation approaches have the following novel capabilities: 1) adaptation to new sensors even when there is no overlapping period between new and old sensors; 2) efficient adaptation by leveraging sensor-specific transformations derived from sensor data; 3) scaling to a large number of sensors; 4) learning robust adaptation functions by leveraging spatial and temporal information of sensors; and 5) estimating the quality of adaptation.
Additionally, we present a constraint-based learning framework that performs joint sensor failure detection and adaptation by leveraging sensor relationships. Our framework learns sensor relationships from historical data and expresses them as a set of constraints. These constraints then provide a joint view for detection and adaptation: detection checks which constraints are violated, and adaptation reconstructs failed sensor values. Our framework is capable of handling multi-sensor failures which are challenging for existing methods.
To validate our approaches, we conduct empirical studies on sensor data from the weather and UUV (Unmanned Underwater Vehicle) domains. The results show that our approaches can automatically detect and adapt to sensor changes and failures with higher accuracy and robustness compared to other alternative approaches.
Location: Charles Lee Powell Hall (PHE) - 223
Audiences: Everyone Is Invited
Contact: Lizsl De Leon