Wearable Posture Identification Using Sensor Fusion and Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional action estimation techniques using acceleration sensors struggle to accurately detect posture and activity states, particularly when distinguishing between torso and arm movements, which limits their effectiveness in health management and disease prevention.
Innovation Solution
A sequential posture identification system that combines acceleration and biological signal information using machine learning to classify postures and activities into dynamic and static patterns, enabling accurate identification and continuous learning to adapt to changes in health status.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single triaxial acceleration sensor is attached to the arm for action estimation, then the device complexity is reduced, but the measurement precision of posture and activity states deteriorates
Solution Approach 1:
The system segments the measurement task by using multiple sensors placed at different body locations (arm, torso, head) to capture distinct movement patterns. Each sensor monitors specific body parts, and their data are integrated to achieve comprehensive and accurate posture detection, resolving the contradiction between simple device configuration and precise measurement.
Solution Approach 2:
The system merges data from multiple acceleration sensors and gyro sensors into a unified posture estimation framework. By combining measurements from different body parts and integrating them with gyroscopic data, the system achieves high-precision posture and activity detection while maintaining reasonable device complexity through coordinated sensor fusion.
2Device complexity
If action estimation is performed using only acceleration sensor data, then the device complexity is minimized, but the reliability of health management information deteriorates
Solution Approach 1:
The system merges acceleration sensor data with gyro sensor data to create a more reliable estimation of posture and activity. This multi-source data integration compensates for limitations of individual sensors, providing more accurate and trustworthy health management information while keeping the overall device complexity manageable through systematic data fusion.
Solution Approach 2:
The system incorporates feedback mechanisms where estimated posture and activity information is continuously refined based on incoming sensor data. The machine learning model learns from accumulated data to improve the reliability of health management assessments over time, creating a self-improving system that enhances accuracy without proportionally increasing device complexity.
3Ease of operation
If conventional action estimation techniques are used, then the ease of operation is maintained, but the productivity of health management and disease prevention deteriorates
Solution Approach 1:
The system employs machine learning models that automatically learn and adapt to individual user patterns without requiring manual configuration or intervention. The posture estimation and health management functions operate autonomously, continuously improving their accuracy through self-learning while maintaining ease of operation and significantly enhancing the productivity of health management and disease prevention efforts.
Solution Approach 2:
The system dynamically adjusts estimation parameters and model configurations based on learned user-specific characteristics and changing health conditions. By adapting parameters such as sensor weighting, threshold values, and classification criteria, the system maintains ease of operation while maximizing health management effectiveness through personalized and context-aware analysis.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
In a wearable device attached to a subject, an acceleration information measurement unit that measures acceleration information, and a biological signal information measurement unit that measures biological signal information of the subject, are provided. From the measured acceleration information and biological signal information, first feature data corresponding to a first predetermined period and second feature data corresponding to a second predetermined period are extracted. By machine learning based on the first feature data, a dynamic/static activity identification model, a dynamic-activity identification model, and a static-activity identification model, for the subject, are generated. By combination of results of determination based on each of the identification models, a posture and an activity of the subject are identified. Correspondence information, which associates the identified posture and activity with the biological signal information of the subject, is generated.