Inertial Sensor Data Segmentation via Wavelet Discrepancy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in efficiently dividing inertial sensor data into segments that correspond to a single activity, particularly due to difficulties in accurately identifying transitions between different activities.
Innovation Solution
The method involves performing wavelet transforms on a sliding window of inertial sensor data to generate augmented data, calculating a discrepancy value between the first and second halves of the window, and identifying transitions based on these discrepancy values to effectively segment the data into single activity segments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional segmentation methods are used to divide inertial sensor data into activity segments, then the process is simple, but the accuracy of identifying activity transitions is poor
Solution Approach 1:
The patent introduces an intermediary mechanism (discrepancy calculation between first and second halves of sliding window) to detect activity transitions. This intermediary approach enables accurate transition identification by comparing temporal discrepancies without requiring complex classification models, thus resolving the contradiction between accuracy and complexity.
Solution Approach 2:
The patent divides the inertial sensor data stream into segments using a sliding window approach, further dividing each window into first and second halves for comparison. This segmentation strategy enables precise activity transition detection by analyzing local temporal patterns, achieving high accuracy while maintaining computational efficiency.
2Reliability
If manual annotation of training sets is performed to ensure single activity per segment, then the training quality is high, but the time and effort required is excessive
Solution Approach 1:
The patent implements self-service by enabling the system to automatically generate high-quality training sets through its own discrepancy-based segmentation mechanism. The system autonomously identifies activity transitions and creates properly segmented training data without requiring manual annotation, thus achieving high reliability while eliminating time-consuming manual work.
Solution Approach 2:
The patent performs preliminary action by pre-segmenting the inertial sensor data into activity-specific segments using the discrepancy method before training the classification model. This preliminary segmentation ensures that the training set contains pure single-activity segments, improving training quality while reducing the need for subsequent manual correction or verification.
3Measurement precision
If complex classification models are used to identify activity segments, then the segmentation accuracy is high, but the computational cost is high
Solution Approach 1:
The patent extracts the essential feature for activity transition detection (temporal discrepancy in the sliding window) and uses it directly for segmentation, rather than feeding all raw inertial sensor data into complex classification models. This extraction approach achieves high segmentation accuracy while significantly reducing computational energy consumption by processing only the most discriminative features.
Data Source
AI summary
A method of operating an inertial sensor module includes receiving a stream of inertial sensor data representing activity of a user of an electronic device and generating a plurality of wavelet sub-bands by performing a wavelet transform on the inertial sensor data. The method includes identifying a wavelet sub-band of highest energy from the plurality of wavelet sub-bands, generating augmented inertial sensor data by combining the wavelet sub-band of highest energy to the inertial sensor data, and identifying a first transition in the activity of the user based on the augmented inertial sensor data.