Detecting abnormal human behavior by using data from user's portable devices

A method for detecting abnormal human behavior, comprising:

• preprocessing data including:
• defining time intervals applicable to a user based on historical data from sensors;
• personalizing the time intervals based on data actively input by the user or real-time data from the sensors;
• constructing words for each of the time intervals, wherein a word is constructed as a vector that includes one of the time intervals and a plurality of discrete sensor-based feature levels, wherein each of the discrete sensor-based feature levels is mapped to a range of values of a sensor-based feature that are extracted from the data from the sensors;
• and
• constructing a text document from the words based on the time intervals and the discrete sensor-based feature levels;
• extracting a routine for each of the time intervals by extracting the words that most frequently appear in the text document using topic modeling to obtain a routine vector;
• and
• generating an alert if the routine extracted in a current time interval deviates from a past routine extracted for a past time interval prior to the current time interval.

A system for detecting abnormal human behavior, the system comprising:

• a memory that stores computer-readable code;
• and
• a processor operatively coupled to the memory, wherein the processor is configured to implement the computer-readable code to:
• define time intervals applicable to a user based on historical data from sensors;
• personalize the time intervals based on data actively input by the user or real-time data from the sensors;
• construct words for each of the time intervals, wherein a word is constructed as a vector that includes one of the time intervals and a plurality of discrete sensor-based feature levels, wherein each of the discrete sensor-based feature levels is mapped to a range of values of a sensor-based feature that are extracted from the data from the sensors;
• construct a text document from the words based on the time intervals and the discrete sensor-based feature levels;
• extract routines for each of the time intervals by extracting the words that most frequently appear in the text document using topic modeling to obtain a routine vector, wherein the routines are characterized by the discrete sensor-based feature levels associated with the most frequently appearing words;
• and
• generate an alert if the sensor-based feature extracted in a current time interval deviates from a past routine extracted for a past time interval prior to the current time interval.