Variable Influence Indicator for Anomaly Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting patterns related to diseases or failure modes in machines, such as Principal Component Analysis (PCA) and regression techniques, are limited by being uni-modal and struggling to handle complex data densities and missing data, making it difficult to accurately determine variable contributions to anomalies.
Innovation Solution
A method using a mixture model in graphical form to calculate Variable Influence Indicators by setting evidence on selected variable nodes, generating a new graph, and evaluating the magnitude of these indicators to determine the directional change and contribution of variables, allowing for more intuitive handling of complex data and missing values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If Principal Component Analysis (PCA) is used for anomaly detection, then variable contributions to abnormality can be indicated, but the method is restricted to uni-modal data and does not provide an intuitive method for handling missing data
Solution Approach 1:
The patent changes the fundamental parameters of the analysis method by transitioning from PCA's linear correlation-based approach to a Bayesian Network approach that models conditional dependencies. This allows the system to handle complex, non-linear data densities and missing data by using probability distributions and conditional probability calculations instead of fixed linear relationships.
Solution Approach 2:
The patent introduces an intermediary computational framework (Bayesian Network with variable influence indicators) that mediates between the raw data and the anomaly detection results. This intermediary layer handles the complexity of missing data and non-linear relationships by using probabilistic inference, allowing the system to work with incomplete information while maintaining accurate variable contribution assessment.
2Measurement precision
If regression techniques are used to calculate residuals for variable contribution, then the magnitude of residual provides a measure of contribution, but it is difficult to directly compare different variables and outputs can be misleading when multiple variables contribute to anomaly
Solution Approach 1:
The patent creates a universal variable influence indicator that serves multiple functions simultaneously: it measures individual variable contributions, enables direct comparison across different variables through standardized scoring, and handles multiple contributing variables through probabilistic aggregation. The Bayesian Network framework provides a unified approach that works regardless of the number or type of variables involved.
Solution Approach 2:
The patent transforms the variable contribution analysis from a single-dimensional residual magnitude approach to a multi-dimensional probabilistic framework. By introducing conditional probabilities and variable influence indicators that consider relationships between variables, the system adds dimensional depth to the analysis, enabling more accurate comparisons and interpretations of variable contributions.
3Reliability
If traditional anomaly detection methods are used, then abnormal situations can be detected, but diagnostic information containing patterns of association between input variables and anomaly is difficult to extract
Solution Approach 1:
The patent implements feedback mechanisms where the Bayesian Network model continuously learns from data patterns and updates variable influence indicators. The system provides feedback about which variables are most influential in causing anomalies, and this information feeds back into the diagnostic process to improve pattern recognition and association detection over time, preventing information loss about variable-anomaly relationships.
Data Source
AI summary
A method of determining the influence of a variable in a phenomenon includes extracting a selected variable for analysis and conducting a sequence of graphical operations that includes other variables in the phenomenon. Calculating a variable influence indicator for the selected variable and repeating the steps for other selected variables enables an evaluation among the selected variables to determine their influence in the phenomenon.