WiFi Sensing via Subspace Eigenvectors for Activity Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems for inferring context information, such as human activity or presence, from wireless communication signals face challenges in sustaining sensing performance beyond experimental conditions and require improved methods for reliable detection.
Innovation Solution
The approach involves generating statistical channel correlation matrices, eigen-decomposing them to distinguish between signal and noise subspaces, and using machine learning models trained with signal and noise subspace eigenvectors to indicate presence or activity within a defined physical space, leveraging differential unitarity measures to track changes in eigenbases.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing sensing systems use simple WiFi signal monitoring, then device complexity is low, but sensing reliability deteriorates beyond experimental conditions
Solution Approach 1:
The patent segments the complex signal processing task into distinct functional components: channel state information extraction, subspace decomposition (signal vs noise), eigenvector analysis, and machine learning classification. This segmentation allows each component to be optimized independently while maintaining overall system reliability.
Solution Approach 2:
The patent introduces intermediate representations (correlation matrices, eigenvectors, subspace projections) that bridge the raw WiFi signals and the final activity recognition. These intermediaries transform the raw data into meaningful features that machine learning models can process, improving reliability without requiring direct complex analysis of raw signals.
2Productivity
If the system processes all channel data including noise subspace, then information completeness improves, but computational efficiency deteriorates
Solution Approach 1:
The patent extracts and separates the noise subspace from the signal subspace through eigenvalue decomposition. By identifying and removing the noise subspace (associated with small eigenvalues), the system retains only the most informative signal components for activity recognition, improving computational efficiency while maintaining information completeness through the preserved signal subspace.
Solution Approach 2:
The patent applies partial action by processing only the signal subspace components (associated with large eigenvalues) rather than all channel data. This selective processing achieves sufficient information for reliable activity recognition without the computational burden of analyzing every data point, including noise.
3Measurement precision
If the system uses only receive-side data, then device complexity is reduced, but sensing accuracy may deteriorate
Solution Approach 1:
The patent applies self-service by using only the receive-side channel state information to infer human activity. The system leverages the inherent richness of receive-side data (which contains spatial and temporal characteristics of reflected signals) to achieve accurate activity recognition without requiring complex transmit-side processing or additional sensors.
Data Source
Figure 1~2
Figure 3~4
Figure 5A~6
AI summary
An apparatus, method and computer is described comprising: generating statistical channel correlation matrices at each of a plurality of time instances from a plurality of multi-dimensional arrays or tensors, wherein each of the plurality of arrays or tensors describes a channel impulse or frequency response of a multiple-input, multi-output wireless transmission system at a respective time instance; eigen-decomposing said correlation matrices; determining a signal subspace extent at each time instance by distinguishing between a signal subspace and a noise subspace; extracting eigenvectors from the eigen-decomposed correlation matrices at each time instance for at least the determined signal subspace; and training a machine learning model for generating an output indicative of presence or activity within a defined physical space using at least the determined signal subspace extents and the signal subspace eigenvectors and/or generating an output indicative of presence or activity within a defined physical space based on at least the determined signal subspace extents and said signal subspace eigenvectors.