Source Separation Using Sparse Time-Frequency Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional source separation techniques are limited by the number of sensors available, making it difficult to separate multiple radioelectric sources effectively, especially in complex situations where signals are mixed in time and frequency, and are costly, large, heavy, and consume significant power, which is a concern in embedded applications.
Innovation Solution
A method that calculates Discrete Fourier Transforms of received signals, divides the time-frequency grid into windows, determines covariance matrices, identifies single-source or dual-source windows, generates directional vectors, and combines estimates from associated windows to improve source separation beyond the number of sensors available.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the number of sensors is increased to separate more sources, then the number of separable sources is improved, but the cost, size, weight, and power consumption increase
Solution Approach 1:
The patent changes the parameter of signal representation from raw time-domain signals to a sparse time-frequency representation (using short-time Fourier transform). This transformation allows the system to exploit the sparse structure of radioelectric signals in the time-frequency domain, enabling separation of more sources than the number of physical sensors by utilizing the temporal and spectral sparsity characteristics of the signals
Solution Approach 2:
The patent transitions from analyzing signals only in the time domain to analyzing them in the time-frequency domain by applying short-time Fourier transform. This dimensional change introduces a frequency dimension that provides additional information for source separation, allowing the system to distinguish more sources by exploiting their different spectral characteristics and time-frequency localization properties
2Measurement precision
If classical source separation techniques are used, then the number of separable sources is limited by the number of sensors, but the computational complexity and power consumption remain manageable
Solution Approach 1:
The patent extracts and exploits the sparse structure of signals in the time-frequency representation by identifying and processing only the significant time-frequency cells where signals are present. This selective processing of sparse regions rather than the entire time-frequency grid reduces computational complexity and power consumption while maintaining the ability to separate more sources
Solution Approach 2:
The patent changes the computational approach by working in the sparse time-frequency domain rather than the dense time domain. By applying thresholding and sparsity constraints in the time-frequency representation, the computational burden is significantly reduced compared to traditional methods, while enabling separation of more sources through exploitation of signal sparsity
Data Source
Figure 1
Figure 2
Figure 3
AI summary
This method of determining the direction of arrival, respectively the direction of arrival and the polarisation, of signals emanating from several radioelectric sources by means of an array of space diversity sensors, composed of P reception pathways, comprises the following steps: a) calculating (106) the successive discrete Fourier transforms of the signal received sampled so as to obtain a P-vectorial time-frequency grid of measurements of the signal sliced into windows; b) over each window (120), b1 - calculating (122) the covariance matrix (C), b2 - enumerating sources so as to identify a mono source of bi source window, b3 - determine (124) the largest eigenvalues, and associated eigenvectors of the matrix; c) discretising the domain of directions of arrival (θ); d) generating directional vectors (U) from the discretised directions of arrival (θ); e) associating (130) the directional vectors (U) with the windows; f) for each set of windows associated with one and the same directional vector (U), obtaining (140) the filtered directional vector of the corresponding source by calculating the associated eigenvector with the largest eigenvalue of a matrix constructed on the basis of the eigenvalues and eigenvectors of the covariance matrices (C) of the windows.