Unsupervised Pre-training for RF Sensing Model Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing mobile communication networks face challenges in providing efficient sensing-related services and generating training/sensing model data, as they rely heavily on annotated/labeled data and lack effective methods for processing sensing data within the network.
Innovation Solution
The method involves unsupervised pre-training of sensing-capable entities within the mobile communication network, selecting a sensing model based on predefined use cases, and fine-tuning the model to reduce data complexity and computational effort, allowing for the generation of sensing model data without requiring extensive labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning with labeled data is used for sensing model training, then measurement precision is improved, but loss of time and productivity deteriorate due to extensive data annotation requirements
Solution Approach 1:
The patent applies unsupervised pre-training to initialize sensing models before fine-tuning with limited labeled data. This preliminary action of pre-training on unlabeled data prepares the model in advance, reducing the time and effort required for subsequent supervised learning and data annotation
Solution Approach 2:
The patent changes the learning paradigm from supervised to unsupervised pre-training, then transitions to supervised fine-tuning. This parameter change in training approach allows the model to learn from unlabeled data first, significantly reducing the amount of time-consuming labeled data annotation required
2Measurement precision
If complex sensing models are used to process sensing data, then measurement precision is improved, but device complexity and computational effort increase
Solution Approach 1:
The patent segments the training process into two distinct phases: unsupervised pre-training phase and supervised fine-tuning phase. This segmentation allows the model to first learn general features from unlabeled data, then specialize with labeled data, reducing overall complexity requirements
Solution Approach 2:
The unsupervised pre-training serves as a preliminary action that prepares the model with general sensing capabilities before fine-tuning. This preliminary training reduces the complexity of the subsequent fine-tuning process and lowers computational requirements for achieving high accuracy
3Measurement precision
If extensive labeled sensing data is collected for training, then measurement precision is improved, but loss of substance and productivity deteriorate due to data scarcity and collection overhead
Solution Approach 1:
The patent changes the data requirement parameter by switching from supervised learning (requiring大量 labeled data) to unsupervised pre-training (requiring only unlabeled data). This parameter change dramatically reduces the amount of labeled sensing data needed while maintaining model accuracy through subsequent fine-tuning
Solution Approach 2:
The unsupervised pre-training allows the model to self-train on unlabeled sensing data without requiring external annotation resources. This self-service capability eliminates the need for extensive manual data labeling, reducing both time and resource overhead
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach reduces the need for labeled data, simplifies data processing, and enables faster convergence of object detection models, thereby improving the efficiency and cost-effectiveness of sensing-related services in mobile communication networks.
Implementation Method 1
the at least one sensing-capable or sensing-enabled entity or functionality transmits a radiofrequency sensing signal, wherein, as a result of the radiofrequency sensing signal being transmitted, a radiofrequency sensing reception signal is received
Data Source
Figure 1~2
Figure 3~4
Figure 5~8
AI summary
The invention relates to a method for providing at least one sensing-related service and/or for generating training and/or sensing model data by means of processing sensing-related data that is generated within or as part of a mobile communication network, wherein the mobile communication network comprises at least a first sensing-capable or sensing-enabled entity or functionality and a second sensing-capable or sensing-enabled entity or functionality, wherein the at least one sensing-capable or sensing-enabled entity or functionality transmits a radiofrequency sensing signal, wherein, as a result of the radiofrequency sensing signal being transmitted, a radiofrequency sensing reception signal is received by either the first sensing-capable or sensing-enabled entity or functionality or the second sensing-capable or sensing-enabled entity or functionality or both, wherein by means of the radiofrequency sensing reception signal information about the environment of the first and/or second sensing-capable or sensing-enabled entities or functionalities is able to be gathered by means of inferring sensing inference data based on training data, wherein, in order to provide at least one sensing-related service and/or in order to generate training and/or sensing model data, the method comprises the following steps: -- in a first step, unsupervised pre-training is performed on either the first sensing-capable or sensing-enabled entity or functionality or on the second sensing-capable or sensing-enabled entity or functionality or on both, -- in a second step, a sensing model is selected in dependence on a predefined sensing use case, -- in a third step, a fine-tuning of the sensing model and/or a fine-tuning of the generation of the training and/or sensing model data is performed.