Detecting Anomalies in Smart Homes: More (models) are better than One!

The Smart Home System

Smart home systems are based on a number of predictive algorithms for making “smart” decisions in order to maximise comfort and safety, improve health and wellness, as well as minimise energy usage and reduce the number of repetitive tasks performed by users [1]. These predictive systems rely on techniques such as occupancy and anomaly detection for making accurate decisions in the smart home system.

 

Fig. 1 Smart Home Environment

The Anomaly Detection Problem

Anomaly detection methods are used in a variety of different smart home subsystems such as security and alert systems, elderly care, temperature control and energy management systems. These techniques depend on sensory data streams obtained from a number of different IoT devices (e.g temperature, contact, motion, cameras etc.) installed in the smart home to monitor the changes in the environment. Moreover, the main task of the anomaly detection algorithm is the identification of patterns in the sensory data streams that do not conform to expected behaviour [2, 3].

The dynamically changing environment (an environment where observations have different statistical distributions at different times of the day, week, month, year i.e. at different environment states) inherent to many smart home systems, imposes the biggest challenge for anomaly detection in these type of systems. Many traditional anomaly detection methods assume static monitoring environments and are based on static models for modelling normal behaviours which don’t consider the changes in the monitoring environment. However, in a dynamically changing environment such as the typical smart home environment, these approaches could introduce a high false positives rate and significantly lower the anomaly detection accuracy.

To solve this problem, different statistical models need to be created for each of the different environment states. Models corresponding to the different environment states can easily be generated by extracting and using appropriate subsets from the sensory data streams known as switching data streams. The models in this context must also be able to continuously learn “new” normal behaviours from the switching data streams (i.e. online training). In other words, the created models need to be constantly updated by adding to their training the most recent sensory data in order to produce accurate results and predictions.

The final step of solving the anomaly detection problem involves designing a switching mechanism that will designate the appropriate anomaly detection model to be used by the inference engine. The selection of the appropriate model will depend on the current environment state obtained from the live sensory streaming data.


Smart Home Applications

The importance of the approach presented above can be highlighted through a number of anomaly detection applications within the smart home environment. In smart home anomaly detection problems such as temperature control, different dynamic models are necessary to accommodate the temperature variations at different times of the day and during different seasons throughout the year. Furthermore, by using multiple energy consumption models, the energy consumption and costs could be significantly reduced, and the effectiveness of smart grid systems improved.

All in all, by considering all of the different environment states and using the appropriate models, we could significantly improve the accuracy of the anomaly detection methods that will lead to even “smarter” decisions within the smart home system.

Resources
  1. A.Dixit and A. Naik, “Use of Prediction Algorithms in Smart Homes.”, International Journal of Machine Learning and Computing, 2014.
  2. A.Amitai, W. Gilad, and F. Lev, “Change and Anomaly Detection Framework for Internet of Things Data Streams”, 2016.
  3. M.Salehi, C.A. Leckie, M. Moshtaghi, and T.Vaithianathan, “A Relevance Weighted Ensemble Model for Anomaly Detection in Switching Data Streams”,Pacific-Asia Conference on Knowledge Discovery and Data Mining, pp 461-473, 2014.

Thanks to Simon Vajda, Rodney Pilgrim and Elodie Thilliez for proofreading and providing suggestions.