Neural Network Wireless Positioning via Augmented Channel State Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning algorithms for wireless positioning rely heavily on supervised learning, requiring large amounts of labeled data and needing frequent retraining due to the changing relationship between channel state information (CSI) and user device location, which is impractical given limited data availability.
Innovation Solution
A method using a neural network model that generates loss function components by comparing location data with estimates based on augmented channel state data, allowing for training with both labeled and unlabeled data sets, and employing data augmentation techniques such as phase rotation, amplitude modification, and Gaussian noise to enhance model robustness and reduce reliance on extensive labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning algorithms are used for wireless positioning, then positioning accuracy can be achieved, but large amounts of labeled data are required and frequent retraining is needed
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network model using unlabeled channel state information through self-supervised learning tasks (such as predicting channel properties or reconstructing channel data). This preliminary training phase prepares the model with general channel characteristics before fine-tuning with limited labeled positioning data, thereby reducing the amount of labeled data needed while maintaining positioning accuracy.
Solution Approach 2:
The patent changes the training parameters and objectives by introducing self-supervised learning tasks that do not require labeled positioning data. The model learns from intrinsic channel state information using tasks like channel reconstruction or property prediction, changing the learning paradigm from supervised to self-supervised, which eliminates dependence on large amounts of labeled data.
2Measurement precision
If supervised learning algorithms are used for wireless positioning, then positioning accuracy can be achieved, but frequent retraining is required due to changing CSI-location relationships
Solution Approach 1:
The patent applies preliminary action by pre-training the neural network model using unlabeled channel state information through self-supervised learning tasks (such as predicting channel properties or reconstructing channel data). This preliminary training phase prepares the model with general channel characteristics before fine-tuning with limited labeled positioning data, thereby reducing the amount of labeled data needed while maintaining positioning accuracy.
Solution Approach 2:
The patent applies self-service by enabling the model to continuously learn from unlabeled channel state information through self-supervised learning. The model can adapt to changing channel conditions by automatically learning from incoming unlabeled data without requiring external retraining with labeled data, thereby reducing retraining frequency and time loss.
3Reliability
If data augmentation techniques are applied to channel state data, then model robustness and generalization are improved, but training complexity increases
Solution Approach 1:
The patent applies parameter changes by introducing data augmentation techniques that modify channel state information through transformations such as phase rotation, amplitude scaling, and noise addition. These parameter changes create varied training samples from limited labeled data, improving model robustness and generalization capability while managing training complexity through efficient augmentation strategies.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
An apparatus, method and computer program is described comprising: generating a first loss function component comprising comparing first location data with first location estimates, wherein the first location estimates are based on channel state data, wherein the first location estimates are generated using a model, and wherein the model comprises a plurality of trainable parameters; generating a second loss function component comprising comparing the first location data with second location estimates, wherein the second location estimates are based on channel state data that have been subjected to a first augmentation and wherein the second location estimates are generated using the model; generating a third loss function component comprising comparing third location estimates based on channel state data and fourth location estimates based on channel state data that have been subjected to a second augmentation, wherein the third and fourth location estimates are generated using the model; and training the trainable parameters of the model by minimising a loss function based on a combination of the first, second and third loss function components.