Targeted Cough Detection Using Edge-Cloud User Profiling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing cough detection technologies face limitations due to reliance on speech datasets, model dependency issues, and challenges with triplet loss functions, leading to inefficiencies and high computational complexity, particularly when differentiating between coughs of a targeted user and others, and adapting to changes in self-supervised models.
Innovation Solution
A method involving a targeted cough detector with multiple stage detectors, including an edge detector and a cloud detector, utilizing machine learning architectures to preprocess audio data, differentiate between coughs of a targeted user and others, and maintain privacy by filtering sensitive sounds locally.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If unobtrusive remote sensing is used to monitor coughs, then patient monitoring convenience is improved, but the system cannot differentiate between coughs of the targeted user and others
Solution Approach 1:
The patent applies local quality by creating user-specific cough profiles that capture individual characteristics of each user's cough sounds. The system processes audio data to generate embeddings that reflect local (user-specific) patterns in cough acoustics, enabling differentiation between users while maintaining unobtrusive monitoring. This resolves the contradiction by making the monitoring both convenient and user-specific.
Solution Approach 2:
The patent introduces an intermediary mechanism - a similarity computation module that compares detected cough embeddings against stored user profiles. This intermediary layer enables the system to identify which user produced a cough without changing the unobtrusive nature of the audio recording, thus maintaining monitoring convenience while adding source identification capability.
2Measurement precision
If self-supervised triplet loss embedding models are used for cough detection, then feature learning capability is improved, but computational complexity and training convergence difficulty increase
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing cough embeddings for each user during an enrollment phase. These pre-computed profiles serve as reference data that simplifies subsequent detection operations. The complex triplet loss training is performed once during profile creation, rather than continuously, reducing ongoing computational complexity while maintaining high feature extraction accuracy.
Solution Approach 2:
The patent creates copies of user cough characteristics in the form of stored embedding profiles. Instead of re-processing raw audio data during detection, the system works with compressed representation copies (embeddings) that capture essential cough features. This copying approach maintains measurement precision while significantly reducing computational complexity during operation.
3Adaptability or versatility
If audio data from multiple users is processed together, then system versatility is improved, but the ability to identify targeted user coughs deteriorates
Solution Approach 1:
The patent applies segmentation by separating user-specific cough characteristics into distinct embedded profiles. Each user's cough patterns are segmented and stored independently, allowing the system to handle multiple users while maintaining the ability to identify which profile matches a detected cough. This segmentation enables both multi-user versatility and accurate targeted user identification.
4Measurement precision
If centralized cloud processing is used for cough analysis, then detection accuracy is improved, but computational resources and privacy concerns increase
Solution Approach 1:
The patent segments the processing architecture into edge computing components (local embedding extraction and similarity comparison) and cloud components (profile storage and management). The computationally intensive real-time analysis is performed locally at the edge device, reducing cloud computational requirements and energy consumption while maintaining detection accuracy through local processing of audio data against stored profiles.
Data Source
Figure 1~2
Figure 3~4
Figure 5a
AI summary
The present invention relates to a method to detect cough of a targeted user. The method comprises processing a cough audio clip with a targeted cough detector and generating based thereon a result indicative of whether the cough audio clip contains a cough sound of the targeted user. There present invention also relates to a corresponding system.