# US 10,503,967 B2 · Fast behavior and abnormality detection

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 10,503,967 B2
- Document ID: US10503967B2
- Status: Granted
- Kind: B2
- Application number: 14/949,634
- Filed: 2015-11-23
- Published: 2016-05-26
- Granted: 2019-12-10
- Earliest claimed: 2014-11-21
- Assignees: The Regents of the University of California
- Applicants: The Regents of the University of California
- Inventors: Majid Sarrafzadeh; Foad Dabiri; Hyduke Noshadi
- Web page: https://patentgrep.com/patent/US10503967B2

## CPC

- G06N 20/00: Machine learning
- G06V 40/25: {Recognition of walking or running movements, e.g. gait recognition}
- G06F 2218/00: Aspects of pattern recognition specially adapted for signal processing
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

A system includes an interface configured to receive time series data representing information from a plurality of sensors, and a processor configured to construct a behavior model based on the time series data. The processor identifies features in the time series data, divides the time series data of each of the identified features into segments, and extracts feature components from the segments. The processor further constructs a plurality of state graphs, each state graph including components connected by weighted edges, constructs a behavior graph, wherein the state graphs form vertices of the behavior graph, clusters the state graphs in the behavior graph; and selects a representative state graph from each cluster, wherein the behavior model includes the selected state graphs.

## Claims

### Claim 1

A system, comprising:
an interface configured to receive first time series data representing information from a plurality of sensors;
a processor configured to construct a behavior model by executing unsupervised machine learning based on the first time series data as training data, wherein the processor constructs the behavior model by executing instructions to:
identify features in the first time series data;
divide the first time series data of each of the identified features into segments;
extract feature components from the segments;
construct a plurality of state graphs, each state graph including components connected by weighted edges;
construct a behavior graph, wherein the state graphs form vertices of the behavior graph;
cluster the state graphs in the behavior graph; and
select a representative state graph from each cluster, wherein the behavior model includes the selected state graphs,
wherein the processor is further configured to receive second time series data and identify a performed motion by comparing a state graph constructed from the second time series data to the behavior model.

### Claim 2 (depends on 1)

The system of claim 1, wherein the time series data from the plurality of sensors represents data from a plurality of wearable sensors.

### Claim 3 (depends on 2)

The system of claim 2, wherein the time series data represents body movement.

### Claim 4 (depends on 3)

The system of claim 3, wherein the segments of the time series data include time intervals representing foot motion states of airborne, grounded, takeoff and landing.

### Claim 5 (depends on 4)

The system of claim 4, wherein the feature components include at least one of maximum applied pressure or progression of applied pressure.

### Claim 6 (depends on 1)

The system of claim 1, wherein the sensors are a first set of sensors; and the second time series data are received from a plurality of second sensors.

### Claim 7 (depends on 1)

The system of claim 1, wherein to construct the state graphs includes to determine similarities between components, and assign a weight to an edge between two components based on a similarity value between the two components.

### Claim 8 (depends on 1)

The system of claim 1, wherein the selected state graphs are vertices of the behavior model.

### Claim 9 (depends on 1)

The system of claim 1, further comprising identifying a group of classes representing the behavior model.

### Claim 10 (depends on 9)

The system of claim 9, wherein each class represents a known motion.

### Claim 11

A method, comprising:
training a prediction model by executing unsupervised machine learning, comprising
receiving time series data representing information from multiple sensors;
identifying features in the time series data;
extracting components of the features;
identifying similarities between the components;
constructing state graphs, each state graph including similar components;
constructing a behavior graph of similar state graphs;
organizing the state graphs in the behavior graph into clusters; and
selecting a representative state graph from each cluster, the representative state graphs constructing the prediction model; and
predicting a behavioral or physiological condition based on the prediction model.

### Claim 12 (depends on 11)

The method of claim 11, wherein executing unsupervised machine learning further comprises assigning a class to each representative state graph.

### Claim 13 (depends on 12)

The method of claim 12, wherein each class represents one of a known motion or an identified abnormal motion.

### Claim 14 (depends on 11)

The method of claim 11, further comprising receiving real-time sensor data; constructing a real-time state graph from the real-time sensor data; and identifying a known motion by comparing the real-time state graph to state graphs of the behavior graph.

### Claim 15 (depends on 11)

The method of claim 11, wherein the time series data represents measurements from multiple subjects.

### Claim 16 (depends on 11)

The method of claim 11, wherein the time series data represents measurements from one subject.

### Claim 17 (depends on 11)

The method of claim 11, wherein executing unsupervised machine learning further comprises identifying principal components of the extracted components, and constructing the state graph based on the principal components.

### Claim 18 (depends on 11)

The method of claim 11, wherein extracting components includes determining salience vectors to describe the time series data.

### Claim 19 (depends on 11)

The method of claim 11, wherein the time series data represents pressure sensors in a shoe, and the model represents gait.