Low-Label Wi-Fi Indoor Localization with Variational Autoencoders

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Indoor positioning methods face challenges in accurately determining device location due to the complexity of indoor environments and the high cost and difficulty of generating up-to-date radio maps using Wi-Fi fingerprinting, particularly when only a small fraction of training data is labeled.

Innovation Solution

A semi-supervised deep learning technique using variational autoencoders (VAEs) that requires only a small amount of labeled data to train a classifier for indoor positioning, utilizing encoder and decoder neural networks to embed signal strength values in a latent space and predict locations from unlabeled data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Wi-Fi fingerprinting is used for indoor positioning, then positioning accuracy is improved, but the cost and effort of data collection and annotation increases significantly

Engineering Contradiction:
Improvepositioning accuracyVSAvoiddata collection efficiency
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent applies semi-supervised learning to use only a small fraction of labeled data (partial action) combined with abundant unlabeled data to achieve positioning accuracy comparable to fully supervised methods, thereby reducing the expensive data annotation effort while maintaining performance

Inventive Principle:
Principle #16Partial or excessive action

2Measurement precision

If a dense radio map is generated for accurate positioning, then positioning accuracy is improved, but the time and resources required for data collection increase

Engineering Contradiction:
Improvepositioning accuracyVSAvoiddata collection time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary unsupervised pre-training on abundant unlabeled data to learn the underlying data distribution and features before fine-tuning with a small amount of labeled data. This preliminary action on unlabeled data reduces the time required for complete data collection while maintaining accuracy

Inventive Principle:
Principle #10Preliminary action

3Measurement precision

If more labeled training data is collected, then prediction accuracy is improved, but the annotation cost and effort increases

Engineering Contradiction:
Improveprediction accuracyVSAvoidannotation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces variational autoencoders as an intermediary unsupervised pre-training stage that processes unlabeled data to learn useful representations. This intermediary step acts as a bridge, allowing the system to leverage abundant unlabeled data without requiring expensive annotation, thereby reducing annotation complexity while improving prediction accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP3798918B1Using semi-supervised variational autoencoder for wi-fi-based indoor localization
Publication Date: 2025.10.29 NAVER CORP
  • EP3798918B1 patent drawingFigure 1
  • EP3798918B1 patent drawingFigure 2
  • EP3798918B1 patent drawingFigure 3~4

AI summary

Methods of training predictors for the location of a computing device in an indoor environment are provided. The methods comprise receiving training data comprising labelled data and unlabelled data. A method of training a predictor comprises training a variational autoencoder, wherein the variational autoencoder comprises encoder neural networks, which encode signal strength values in a latent variable, and decoder neural networks, which decode the latent variable to reconstructed signal strength values, and training a classification neural network that employs the latent variable to generate a predicted location. Another method of training a predictor comprises training a classification neural network together with a variational autoencoder, wherein the classification neural network receives signal strength values of the training data as input and outputs a predicted location to decoder neural networks of the variational autoencoder.