Signal Separation via Frequency Domain Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional blind signal separation techniques face challenges in efficiently using information from multiple sensors for signal separation, particularly in solving the permutation problem and requiring precise sensor position information, which is difficult to obtain especially when sensors are randomly disposed.
Innovation Solution
The method involves transforming mixed signals into the frequency domain, normalizing complex vectors to eliminate frequency dependence, clustering these normalized vectors to generate clusters dependent on signal source positions, and using these clusters for signal separation without requiring precise sensor position information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If conventional blind signal separation techniques are used, then signal separation can be performed without direction information, but the permutation problem cannot be accurately solved and sensor position information is required
Solution Approach 1:
The patent transforms the signal processing from time domain to frequency domain, adding a dimensional perspective. By performing STFT (Short-Time Fourier Transform) on the mixed signals, the system analyzes signals at multiple frequency points simultaneously. This dimensional change enables the clustering of frequency responses across different frequencies, which provides sufficient information to solve the permutation problem without requiring sensor position data.
Solution Approach 2:
The patent introduces frequency response clustering as an intermediary mechanism. Instead of directly solving the permutation problem from raw sensor data, the system creates clusters of frequency responses that serve as intermediate structures. These clusters capture the characteristic relationships between sensors and sources across frequencies, enabling accurate permutation solving without direct reliance on sensor position information.
2Measurement precision
If precise sensor position information is obtained, then signal separation accuracy is improved, but the complexity of the system increases and calibration is required
Solution Approach 1:
The patent enables the system to self-determine sensor-position-independent features through frequency response analysis. Instead of requiring external calibration or position input, the system automatically extracts clustering features from the frequency responses of the mixed signals. This self-service approach eliminates the need for manual sensor position calibration while maintaining signal separation accuracy.
Solution Approach 2:
The patent changes the parameter space from physical sensor positions to frequency domain characteristics. By transforming the problem from spatial domain (sensor positions) to frequency domain (frequency responses), the system achieves the same information extraction goal without requiring precise position parameters. This parameter transformation eliminates calibration requirements while preserving separation accuracy.
3Measurement precision
If information from all sensors is used for signal separation, then separation performance is improved, but the complexity of processing increases
Solution Approach 1:
The patent merges the frequency response information from all sensors into unified clusters. Instead of processing each sensor independently and then combining results, the system clusters frequency responses across all sensors simultaneously at each frequency point. This merging approach efficiently utilizes information from all sensors while reducing processing complexity through the clustered representation.
Data Source
AI summary
A frequency domain transforming section transforms mixed signals observed by multiple sensors into mixed signals in the frequency domain, a complex vector generating section generates a complex vector by using the frequency-domain mixed signals, a normalizing section generates a normalized vector excluding frequency dependence of the complex vector, and a clustering section clusters the normalized vectors to generate clusters. Then, a separated signal generating section generates separated signals in the frequency domain by using information about the clusters and a time domain transforming section transforms the separated signals in the frequency domain into separated signals in the time domain.