Inverse Covariance Matrix for Source Estimation in Sensor Arrays
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for estimating the number of sources in a mixture are inadequate in handling correlated and colored noise, non-Gaussian noise, and complex network geometries, often leading to underestimation or incorrect identification of source numbers due to assumptions about noise structure and eigenvalue distribution.
Innovation Solution
A method that calculates the eigenvalues of a covariance matrix and uses a threshold-based approach to differentiate between signal and noise eigenvalues, allowing for the estimation of the number of sources in the presence of correlated or colored noise, without relying on specific noise assumptions, and can detect a number of sources greater than the number of sensors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional statistical tests (chi-square, AIC, MDL) are used to estimate the number of sources, then the method works under white Gaussian noise assumptions, but the estimation becomes inaccurate when noise is colored or non-Gaussian
Solution Approach 1:
The invention changes the fundamental parameter being analyzed from eigenvalues of the covariance matrix to eigenvalues of the inverse covariance matrix. This parameter transformation allows the method to work effectively with colored and non-Gaussian noise by capturing the inverse relationship between signal and noise subspaces, making the estimation reliable across diverse noise conditions without requiring white Gaussian assumptions
Solution Approach 2:
The core innovation inverts the conventional approach by using the inverse covariance matrix instead of the covariance matrix itself. This inversion reverses the roles of signal and noise subspaces, allowing the method to identify signal eigenvalues as the largest ones in the inverse matrix, thereby achieving accurate source estimation under colored and non-Gaussian noise conditions where traditional methods fail
2Measurement precision
If eigenvalue-based methods are used to separate signal and noise, then the separation is clear in asymptotic cases, but statistical fluctuations make separation difficult for finite sample sizes
Solution Approach 1:
By transforming to the inverse covariance matrix, the invention amplifies the separation between signal and noise eigenvalues. The inverse operation enhances the contrast between the largest eigenvalues (signal) and smaller eigenvalues (noise), making the separation more distinct and easier to detect even with limited sample sizes, thereby improving measurement precision without requiring large quantities of samples
3Adaptability or versatility
If empirical criteria are used to classify eigenvalues, then the method can handle non-ideal noise conditions, but strong assumptions are made about covariance matrix structure
Solution Approach 1:
The parameter transformation to the inverse covariance matrix provides a theoretically grounded alternative to empirical criteria. Instead of making strong assumptions about covariance matrix structure or noise characteristics, the inversion method naturally adapts to colored and non-Gaussian noise by its mathematical properties, reducing the need for structural assumptions while maintaining versatility across different noise conditions
4Measurement precision
If the number of sources is estimated using conventional methods, then the estimation is straightforward for white noise, but underestimation occurs when signals are correlated or have weak angular deviations
Solution Approach 1:
The inversion of the covariance matrix reverses the impact of signal correlation on eigenvalue distribution. In the inverse matrix, correlated signals produce distinct large eigenvalues that are easier to separate from noise, preventing underestimation. This approach counteracts the harmful effect of signal correlation and weak angular deviations that plague conventional eigenvalue-based methods
Data Source
Figure 1~2
AI summary
The invention relates to a method and system for determining the number of incident sources in an array that comprises C sensors receiving N observations, said method comprising at least the following steps: calculating the matrix B and the eigenvalues {?1,..., ?N} thereof obtained from a signal received on C sensors; classifying the eigenvalues {?1,..., ?N} so as to obtain ?1=... = ?N; initializing from i to i= M + 1 and i = i - 1; calculating the mean and the standard deviation of the noise eigenvalues; calculating the mean of the N - i lowest eigenvalues of the matrix B; and calculating the standard deviation of the N - i lowest eigenvalues of the matrix B. If ?1 > ?moy + ? s, then said eigenvalue belongs to the signal space, and the number of sources present in the mixture is equal to i, where ? is a threshold that makes it possible to monitor the probability of a false alarm. If ?1 > ?moy + ? s, said eigenvalue belongs to the noise space, and the steps are repeated.