Orthonormal Matrix Change Detection for Noisy Multivariate Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Change detection in noisy multivariate time-series data faces challenges in interpretability and robustness due to nuisance noise variables, with existing methods degrading significantly under such conditions.
Innovation Solution
The method employs a regularized maximum likelihood approach based on the von Mises-Fisher distribution for feature extraction, reducing it to a trust-region sub-problem and using a parametrized Kullback-Leibler divergence for scoring, which computes an orthonormal matrix to capture fluctuation patterns and filter out noise, ensuring global optimality and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If direct density-ratio estimation approaches are used to compute change scores, then the calculation can be performed without explicit feature extraction, but the method suffers from lack of interpretability and significant performance degradation under nuisance noise variables
Solution Approach 1:
The patent segments the change detection process into two distinct stages: (1) feature extraction stage where orthonormal matrices are computed to capture fluctuation patterns and filter noise, and (2) scoring stage where parametric models compare features across time windows. This segmentation allows explicit feature representation for interpretability while using robust statistical methods to maintain reliability under noise
Solution Approach 2:
The patent introduces orthonormal matrices as an intermediary representation between the raw multivariate time-series data and the final change score. These matrices serve as a mediator that captures essential fluctuation patterns while filtering out nuisance noise variables, enabling both interpretability through explicit feature representation and robustness through statistical filtering
2Loss of information
If explicit feature extraction is performed for parametric models, then interpretability is improved, but the process becomes more complex and requires additional computational steps
Solution Approach 1:
The patent changes the parameter representation from raw multivariate data to orthonormal matrix parameters that capture fluctuation patterns. By transforming the data into this parameter space, the patent achieves interpretability through explicit feature representation (the orthonormal matrices) while the mathematical properties of orthonormal transformations preserve information and enable efficient computation, balancing complexity and interpretability
Data Source
AI summary
A method includes capturing multivariate time-series data comprising two or more data sets from a system captured over a past time period and a present time period, applying at least two sliding time windows to the multivariate time-series data in determining respective data matrices, computing an orthonormal matrix for each of the data matrices, wherein the orthonormal matrix is a signature of fluctuation patterns of a respective data matrix, computing a difference between at least two of the data sets in the past and the present time periods through the orthonormal matrices, and detecting a fault in at least one of the systems by comparing the difference to a threshold.