Microlocation Classification Using Semi-Supervised Sensor Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern mobile devices face challenges in efficiently determining position indoors and recommending applications based on location due to limitations in GPS resolution and reliability, especially in areas without satellite connectivity.
Innovation Solution
A semi-supervised machine learning model is employed to classify sensor data points into clusters, allowing mobile devices to predict locations and recommend or perform actions based on historical usage patterns, using sensor measurements from stationary signal sources like Wi-Fi routers and Bluetooth devices.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If GPS coordinates are used to determine device location, then outdoor positioning is achieved, but indoor positioning accuracy and resolution are insufficient
Solution Approach 1:
The patent introduces sensor measurements (Wi-Fi, Bluetooth, accelerometer, gyroscope, magnetometer) as intermediary data sources to determine device location when GPS is unavailable indoors. These sensors act as mediators between the device and the environment, capturing electromagnetic signals and physical motion patterns that can be processed to infer location without requiring satellite connectivity.
Solution Approach 2:
The patent replaces the GPS satellite-based electromagnetic system with a local sensor-based detection system for indoor positioning. Instead of relying on distant satellite signals, the device uses local electromagnetic fields from Wi-Fi routers, Bluetooth devices, and physical sensors to determine position, substituting the GPS mechanical system with a distributed sensor network approach.
2Measurement precision
If sensor data is collected continuously to improve location accuracy, then positioning precision increases, but energy consumption increases
Solution Approach 1:
The patent implements periodic action by collecting sensor data at specific intervals and triggering data collection based on contextual events rather than continuous monitoring. The system periodically updates location models and processes sensor data only when necessary, such as when the device detects it has entered a new microlocation or when user interaction is anticipated, thereby reducing overall energy consumption while maintaining positioning accuracy.
Solution Approach 2:
The patent dynamically adjusts sampling parameters based on environmental conditions and device state. The system modifies data collection frequency, sensor activation, and processing intensity according to the current context, such as reducing sampling rate when the device is stationary versus moving, or adjusting based on available power levels, thus optimizing the balance between measurement precision and energy consumption.
3Measurement precision
If machine learning models are trained with extensive tagged data to improve classification accuracy, then model performance increases, but data processing complexity increases
Solution Approach 1:
The patent segments the machine learning system into modular components: separate models for different applications, distinct processing pipelines for different sensor types, and divided training data organized by microlocation and application category. This segmentation allows the system to manage complexity by handling smaller, specialized models rather than one monolithic system, making the overall system more maintainable and scalable.
Solution Approach 2:
The patent creates a universal machine learning framework that serves multiple functions: location classification, application recommendation, and user behavior prediction. The same core architecture and training mechanisms are applied across different applications and contexts, allowing the system to achieve high accuracy across diverse tasks without proportionally increasing complexity, as the underlying infrastructure is shared and reusable.
Data Source
AI summary
A semi-supervised machine learning model can provide for classifying an input data point as associated with a particular target location or a particular action. Each data point comprises one or more sensor values from one or more signals emitted by one or more signal sources located within a physical area. A tagged sample set and an untagged sample set are combined to train the machine learning model. Each tagged sample includes a respective data point and a label representing a respective location/action. Each untagged sample includes a data point but is unlabeled. Once trained, given a current data point, the machine learning model can classify the current data point as associated with a particular location/action, after which a target object (e.g., other device or application to be used) can be predicted.