[home] [Personal Program] [Help]

---

Session: Poster session I
Session starts: Thursday 24 January, 15:00

---

Marc Mertens (Thomas More University of Applied Sciences - Mobilab, Kleinhoefstraat 4, 2440 Geel Belgium)
Glen Debard (Thomas More University of Applied Sciences - Mobilab, Kleinhoefstraat 4, 2440 Geel Belgium)
Bart Vanrumste (eMedia Research Lab at Campus GroepT and STADIUS at the Department of Electrical Engineering (ESAT) from the KU Leuven)
Jesse Davis (Department of Computer Science, KU Leuven, Belgium)

Abstract:
Since the population is ageing, more elderly people will depend on the support of others. Automatic monitoring systems that can provide information regarding the status of elderly living alone, such as a gradual or sudden change of behavior can aid the caregiver in assessing the health status and self-reliance. Several sensor types can be used in such a monitoring system. In our study, we monitor the usage patterns of electrical appliances. By linking usage of appliances to Activities of Daily Living (ADL), behavioral patterns can be constructed. The first step in this approach is the recognition of used appliances. This abstract describes a comparison of classifiers which can be used to discern between different appliances. It builds on an earlier exploration using the J48 tree classifier as described in [1]. The dataset consists of electrical current profiles of several appliance types (e.g., TV sets, coffee machines, vacuum cleaners, electrical heaters, etc.), sampled at 5kS/s. From this data, several features were extracted such as Irms, Imax, correlation of the electrical current with a perfect 50Hz sinewave, etc. A total of 109 instances of one period were collected. The data set is split up randomly into a training set of 85 and a test set of 24 instances. We compared four classification algorithms on accuracy; the above mentioned J48, Naive Bayes, Nearest Neighbors (kNN, k=1..3) and Logistic Regression. Comparing accuracy results, it is shown that kNN (n=1) gives the best result (95.8%) and in order of decreasing accuracy: J48 (90%), kNN (n=2 or 3) (87.5%), Naïve Bayes (79%) and finally Logistic Regression (75%). kNN (n=1) giving the best result, indicates that a near perfect prediction can be done if we first measure the relevant appliances in the house and use those as a training set for detection. With this experiment, we showed it is possible to recognize electrical appliances from their electrical current signature in an accurate way. For future work, we will analyze appliances running at the same time. We acquired four datasets in real living environments and will use this classification method to construct ADL patterns based on appliance usage. [1] MONITORING ACTIVITIES OF DAILY LIVING THROUGH DETECTION OF USED ELECTRICAL APPLIANCES Marc Mertens, Glen Debard, Bert Bonroy, Els Devriendt, Koen Milisen, Jos Tournoy, Jesse Davis, Tom Croonenborghs, Bart Vanrumste, BME2013