Early Abstraction of Inertial Sensor Data for Long-Term Deployments.
Technische Universität, Darmstadt
[Ph.D. Thesis], (2014)
Available under Creative Commons Attribution Non-commercial No Derivatives, 2.5.
Download (17MB) | Preview
|Item Type:||Ph.D. Thesis|
|Title:||Early Abstraction of Inertial Sensor Data for Long-Term Deployments|
Advances in microelectronics over the last decades have led to miniaturization of computing devices and sensors. A driving force to use these in various application scenarios is the desire to grasp physical phenomena from the environment, objects and living entities. We investigate sensing in two particularly challenging applications: one where small sensor modules are worn by people to detect their activities, and one where wirelessly networked sensors observe events over an area.
This thesis takes a data-driven approach, focusing on human motion and vibrations caused by trains that are captured by accelerometer sensors as time series and shall be analyzed for characteristic patterns. For both, the acceleration sensor must be sampled at relatively high rates in order to capture the essence of the phenomena, and remain active for long stretches of time. The large amounts of gathered sensor data demand novel approaches that are able to swiftly process the data while guaranteeing accurate classification results. The following contributions are made in particular:
* A data logger that would suit the requirements of long-term deployments is designed and evaluated. In a power profiling study both hardware components and firmware parameters are thoroughly tested, revealing that the sensor is able to log acceleration data at a sampling rate of 100 Hertz for up to 14 full days on a single battery charge.
* A technique is proposed that swiftly and accurately abstracts an original signal with a set of linear segments, thus preserving its shape, while being twice as fast as a similar method. This allows for more efficient pattern matching, since for each pattern only a fraction of data points must be considered. A second study shows that this algorithm can perform data abstraction directly on a data logger with limited resources.
* The railway monitoring scenario requires streaming vibration data to be analyzed for particular sparse and complex events directly on the sensor node, extracting relevant information such as train type or length from the shape of the vibration footprint. In a study conducted on real-world data, a set of efficient shape features is identified that facilitates train type prediction and length estimation with very high accuracies.
* To achieve fast and accurate activity recognition for long-term bipolar patients monitoring scenarios, we present an approach that relies on the salience of motion patterns (motifs) that are characteristic for the target activity. These motifs are accumulated by using a symbolic abstraction that encodes the shape of the original signal. A large-scale study shows that a simple bag-of-words classifier trained with extracted motifs is on par with traditional approaches regarding the accuracy, while being much faster.
* Some activities are hard to predict from acceleration data alone with the aforementioned approach. We argue that human-object interactions, captured as human motion and grasped objects through RFID, are an ideal supplement. A custom bracelet-like antenna to detect objects from up to 14 cm is proposed, along with a novel benchmark to evaluate such wearable setups.
By aiming for wearable and wirelessly networked sensor systems, these contributions apply for particularly challenging applications that require long-term deployments of miniature sensors in general. They form the basis of a framework towards efficient event detection that relies heavily on early data abstraction and shape-based features for time series, while focusing less on the classification techniques.
|Place of Publication:||Darmstadt|
|Classification DDC:||000 Allgemeines, Informatik, Informationswissenschaft > 004 Informatik|
|Divisions:||20 Department of Computer Science|
|Date Deposited:||13 Nov 2014 13:07|
|Last Modified:||13 Nov 2014 13:07|
|Referees:||Van Laerhoven, Dr. Kristof and Amft, Prof. Dr. Oliver|
|Refereed:||16 October 2014|