GPS Sensor Web Time Series Analysis Using SensorGrid Technology
Abstract
We present a method for performing signal detection and classification on real-time streams of GPS sensor web data. Our approach has two parts. The first is a hidden Markov model fitting methodology that enables us to robustly describe the statistics of the data. The second is the SensorGrid technology which allows us to manage the data streams through a series of filters tied together with a publish/subscribe messaging system. In this framework, the HMM algorithm is viewed as a filter. The sensor web data we use in this work comes from the Southern California Integrated GPS Network (SCIGN), which produces a number of data products. In this work, we use the real-time (1Hz for most stations) three-dimensional position information. This data is collected from a system which is not only noisy but also poorly understood; driving forces on the system derive not only from the physical processes of the solid earth but also from external factors, including atmospheric effects and human activity. Fitting an HMM to time series allows us to describe the statistics of the data in a simple way that ascribes discrete modes of behavior to the system. By matching incoming data against the statistics of previously learned modes, we can perform classification according to the best match. In addition, we can perform signal detection across the entire sensor web by correlating mode changes in time; a significant number of mode changes across the network or within a certain sub-network is an indication of an event that is occurring over a wide geographical area. For most applications, reliable HMM fitting results are achieved by using a priori information to form constraints that reduce the number of free parameters. For GPS data, however, this information is not available as the underlying system is not well understood. As a result, we use the regularized deterministic annealing expectation-maximization (RDAEM) algorithm to perform the fit. This method provides high-quality, self-consistent model fits without using a priori information (although it does not exclude the use of such information where available), at the cost of some additional computation time. We refer to the software implementation of this algorithm for HMMs as RDAHMM. To integrate this HMM technology with the GPS data streams, we used the the SensorGrid architecture. This provides a service oriented approach to support coupling real-time sensor messages with scientific applications in a Grid environment. Real-time data processing is supported by employing filters around publish/subscribe messaging system. The filters are small applications extended from a generic Filter class to inherit publish and subscribe capabilities. The measurements obtained from the sensors are usually in a proprietary binary format and need to be converted into specific formats to be used by various applications. Our approach allows chaining of several filters to achieve format conversions and data processing. In addition, we developed a proxy service to expose the filters as Web Services. This service provides basic functions to start/stop filters and get metadata descriptions. Using this service filter chains can be created and deployed remotely in a workflow environment. Currently we have basic XML schemas to describe filter metadata and filter chains. We developed and tested several filters to provide real-time access to GPS messages. Currently the system is continuously running for 7 GPS networks, about 70 GPS stations; we examine the performance of the RDAHMM and SensorGrid technologies on this data.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2006
- Bibcode:
- 2006AGUFMIN22A..05G
- Keywords:
-
- 0520 Data analysis: algorithms and implementation;
- 0555 Neural networks;
- fuzzy logic;
- machine learning;
- 0933 Remote sensing;
- 1294 Instruments and techniques;
- 3270 Time series analysis (1872;
- 4277;
- 4475)