Typical Behavior Patterns Extraction and Anomaly Detection Algorithm Based on Accumulated Home Sensor Data

TLDR

The paper proposes a two‑component method for extracting typical home‑sensor behavior patterns and detecting anomalies in daily life. The method accumulates room‑environment sensor data, segments it, assigns behavior labels with HMM and k‑means, builds daily HMMs to compute likelihoods, and then models typical behavior sequences with probabilistic densities estimated by sequential discounting Laplace and EM for anomaly detection. Anomalies are flagged when new data significantly deviate from the modeled typical behavior, and the approach was validated on long‑term sensor recordings from an elderly person's home.

Abstract

In this paper, we propose a method consists of two components, behavior patterns extraction and anomaly detection algorithm in daily life. To begin with, sensor data are accumulated in a room environment and behavior description labels are assigned for each data segment using HMM(hidden Markov model) and k-means method. An HMM is composed every day based on sensor data segments of the day. The behavior description label at each time segment is determined by likelihood of the segment computed using the HMM. In anomaly detection step, typical behavior sequences are acquired using probabilistic density of behavior occurrence and behavior successive time. Each probabilistic density is composed based on accumulating labeled- data using sequential discounting Laplace estimation and sequential discounting expectation and maximization algorithms. When a new datum comes, if typical behavior data change largely, the data is detected as anomaly. The proposed method is verified by a long-time activity detection sensor data taken at a house of elderly person.