Intelligent Contextual Data Stream Monitoring K. Kolomvatsos1, C. Anagnostopoulos2, and S. Hadjiefthymiades1 1University of Athens, Department of Informatics & Telecommunications 2University of Glasgow, School of Computing Science 8th ACM International Conference on PErvasive Technologies Related to Assistive Environments (PETRA 2015) July 1-3, 2015, Corfu, Greece

Outline Monitoring Systems The Proposed Mechanism The Prediction Module The Fusion Module The Consensus Module The Fuzzy Logic Module Experimental Results

Monitoring Systems (1/2) Monitoring System (MS): monitors specific contextual parameters. The MS generates alerts when any abnormalities are present. Alerts are derived by triggering mechanisms that continually check the fulfillment of certain conditions. Inputs are received by a set of sensors. Sensors send their observations at specific intervals to MS. MS based on observations derives the Degree of Alert (DoA) that indicates if measures should be applied.

Monitoring Systems (2/2) MS minimizes false alerts as those can affect the performance of the entire system. If measures are applied in a wrong interval, they will not benefit the performance of the response. We propose a combination of: multivariate contextual data fusion techniques time-series forecast and consensus theory

The Proposed Mechanism (1/2) Let a team of N sensors. Sensors observe the same phenomenon and report data to pre- defined intervals. When the MS receives the reports, it decides if an event is identified, thus, it derives alerts without any human intervention. The MS is not based only on a single sensor observations, as false alarms could be derived. Sensors could report invalid observations due to various reasons like resources status and environmental characteristics. The MS relies on the opinion of the majority before it decides the initiation of an alert.

The Proposed Mechanism (2/2) The proposed MS includes the following components: Prediction module (PM): is based on sensors historical values, predicts future observations. Fusion module (FM): it eliminates the outliers data and provides the final fused measurement. Consensus module (CM): it produces the Degree of Consensus (DoC) which denotes the unanimity in the opinion of sensors (experts). Fuzzy Logic module (FLM): combines the outputs of the FM and the CM and derives the Degree of Alert (DoA). The DoA provides a support for deriving an alert to users or applications. The decision is made based on (1) the current aggregated measurements and (2) the consensus of the team.

The Prediction Module For each sensor Si a time ordered set of past values HSi is maintained. We consider the history of the latest observations. We predict the missing measurement through a linear combination of the historical measurements with real-valued prediction coefficients. The set of coefficients are estimated to minimize the prediction error between the predicted and the actual measurement. We get a number of equations, (a.k.a Yule-Walker equations) for estimating the coefficients. In our effort, we adopt the Levinson-Durbin algorithm.

The Fusion Module (1/4) Data fusion combines contextual data from all sensors to derive reliable fused measurements. We adopt the cumulative sum (CumSum) algorithm for outliers detection and the linear opinion pool algorithm for deriving the final aggregated value. The CumSum algorithm detects if there is any change in the distribution of a contextual time series corresponding to a sensor. The algorithm is a change-point detection technique based on the cumulative sum of the differences between the current value at instance t and the overall average up to t. Slopes depict jumps in the times series, thus, corresponding to outliers. We adopt a two-side detection scheme where an observation is inferred as an outlier when it deviates above a target threshold h+ or below a target threshold h-.

The Fusion Module (2/4) Two signals are the outputs of the CumSum algorithm: the first is related to the above detection signal while the second corresponds to the below detection signal. When the time series deviates from the thresholds, detection signals are set to 1. The detected outliers are eliminated and, thus, the mechanism at t is based on the opinion (measurements) of those sensors that do not produce outliers. The proposed MS adopts a linear opinion pool scheme for the remaining values.

The Fusion Module (3/4)  

The Fusion Module (4/4)  

The Consensus Module  

The Fuzzy Logic Module The Fuzzy Logic Module has two inputs: the fused measurement f the DoC The output is the DoA value. We consider that DoA  1 indicates that the danger is at high levels; the opposite stands when DoA  0. We consider three linguistic values: Low, Medium, High. Low value represents that the fuzzy variable takes values close to the lower limit while a High value depicts the case where the variable takes values close to the upper level. A Medium value depicts the case where the variable takes values close to the average. We consider triangular membership functions.

Experimental Results (1/2) We examine the MS performance related to the Rate of False Alerts (RFA). As RFA  1, the MS results a lot of false alerts. As RFA  0, the MS minimizes the rate of false alerts. We experiment with two datasets: The MHEALTH dataset (Banos, O., et al., 'mHealthDroid: a novel framework for agile development of mobile health applications', In 6th IWAAL, 2014) We consider the provided electrocardiogram signal and assume 1,000 measurements for each volunteer The second dataset is retrieved by the Intel Berkeley Research Lab (http://db.csail.mit.edu/labdata/labdata.html) We get 15,000 measurements assuming that N = 15 sensors produced 1,000 reports

Experimental Results (2/2) We inject p faulty measurements into datasets. We compare the proposed MS with the Single Sensor Alerting (SSA) mechanism. The SSA mechanism delivers an alert when at least one sensor reports a value over a predefined threshold

Thank you!