Data Augmentation for Motion Sensor Machine Learning Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems face bottlenecks in creating machine learning models for motion and microphone sensors due to the scarcity and high cost of properly labeled data, especially for specific gestures or activities, which limits their ability to recognize data across different conditions.
Innovation Solution
The implementation of data augmentation techniques using transformation functions, such as absolute value and squaring, to create condition-agnostic data representations, allowing machine learning models to be trained on data from one condition and deployed on embedded devices to recognize data from multiple conditions, thereby reducing data collection burdens and enhancing model performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If data is collected from multiple conditions to train machine learning models, then model accuracy and generalization improve, but data collection time and cost increase significantly
Solution Approach 1:
The patent creates virtual copies of sensor data by applying transformation functions to data from a single condition. These transformed copies simulate data that would have been collected from multiple conditions, enabling model training without actual multi-condition data collection. The transformation functions generate synthetic data variations that mimic the statistical properties of real multi-condition data.
Solution Approach 2:
The patent applies parameter transformations to the sensor data, such as scaling, rotating, and other mathematical transformations that modify the data parameters. These parameter changes create diverse training samples from a single condition, effectively simulating multiple conditions without physically changing the collection environment.
2Measurement precision
If manually labeled data is used for training machine learning models, then model precision improves, but the labeling process becomes cumbersome and time-consuming
Solution Approach 1:
The patent implements self-service data labeling by using transformation functions that automatically generate labeled training data. The system itself performs the labeling task through mathematical transformations, eliminating the need for manual human annotation. The transformation process inherently creates the labels based on the transformation applied, making the system self-sufficient for data preparation.
3Adaptability or versatility
If extensive data collection is performed to cover all possible conditions, then model robustness across different conditions improves, but resource requirements and system complexity increase
Solution Approach 1:
The patent creates a universal data transformation framework that can handle multiple conditions through a single set of transformation functions. Rather than building separate data collection systems for each condition, the system uses one unified approach that generates all necessary training variations through mathematical transformations, making the system multi-functional without increasing physical complexity.
Data Source
AI summary
Disclosed embodiments provide data augmentation techniques in which collected sensor data (for example, data from a motion sensor or a microphone) related to a gesture or an activity is used to simulate a unified data representation by using one or more transfer functions. The collected sensor data is for a particular condition. The unified representation is agnostic to the condition in which the gesture or activity is made. The unified representation is used to train a machine learning model (MLM). The MLM is then deployed on an integrated circuit chip of an embedded device. Live sensor data received by the embedded device is then transformed and input to the MLM, and the MLM then performs a prediction by, for example, recognizing a gesture made by the user of the embedded device.