Personalized Motion Classifier for Wearables Using Unsupervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generalized classification techniques used in mobile and wearable devices are not user-specific, leading to varying accuracy in motion activity detection due to differences in human size, shape, and weight, resulting in suboptimal performance for individual users.
Innovation Solution
A method that collects and updates motion activity data from sensors to generate a personalized classifier, using a data selection confidence measure to determine the relevance of the data for training, allowing for improved accuracy in motion activity classification without requiring explicit user input.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If generalized classification techniques are used, then the device can provide acceptable performance for any user, but the accuracy varies significantly for individual users due to differences in human size, shape, and weight
Solution Approach 1:
The classifier is made dynamic and adaptive through continuous learning from user-specific motion data. The system transitions from a static generalized classifier to a dynamic personalized classifier that evolves over time by incorporating new motion activity data, thereby adapting to individual user characteristics while maintaining broad applicability
Solution Approach 2:
The system changes the parameters of the classifier by updating it with user-specific training data. The classifier's internal parameters are continuously adjusted based on collected motion data, enabling it to transform from a generic model to a personalized model that accurately reflects individual user patterns
2Measurement precision
If a personalized classifier is created using user-specific data, then the accuracy for individual users improves, but the device complexity and data processing requirements increase
Solution Approach 1:
The system performs self-service by automatically collecting motion data, selecting relevant training samples, and updating the classifier without requiring external intervention. The unsupervised learning mechanism enables the system to autonomously improve its performance, reducing the need for complex manual configuration and data annotation processes
Solution Approach 2:
The system implements feedback loops where classification results are continuously evaluated and used to select high-quality training data. The confidence measure provides feedback on data quality, and the classifier is iteratively updated based on this feedback, creating a self-improving system that manages complexity through structured feedback mechanisms
3Productivity
If all collected motion activity data is stored in the training data set, then the classifier can be continuously improved, but irrelevant or low-quality data may degrade classification performance
Solution Approach 1:
The system applies local quality control by evaluating and selecting training data based on confidence measures. Not all data is treated equally; instead, high-quality data with high confidence scores is selected for training, while low-quality data is excluded. This selective approach ensures that only relevant and reliable data contributes to classifier improvement
4Measurement precision
If the classifier is updated frequently with new data, then the personalized accuracy improves faster, but the processing time and energy consumption increase
Solution Approach 1:
The system applies partial action by updating the classifier selectively rather than continuously. Training is performed only when sufficient high-quality data has been accumulated, as determined by the confidence measure thresholds. This partial updating approach balances the need for personalization with the constraints of processing time and energy consumption
Data Source
AI summary
Motion activity data is collected from at least one sensor. An initial motion activity classifier function is applied to the motion activity data to produce an initial motion activity posteriorgram. Pre-processing and segmenting the motion activity data into windows produces segmented motion activity data from which sensor specific features are extracted. An updated motion activity classifier function is generated from the extracted sensor specific features. Subsequent motion activity data is also collected from the at least one sensor, and the updated motion activity classifier function is applied to the subsequent motion activity data to produce an updated motion activity posteriorgram.