Automatic Fault Classification Using PCA and Temporal Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing early event detection systems rely heavily on manual localization of faults, as they struggle to automatically translate anomaly detection into meaningful information for operators, and current clustering methods fail to account for the temporal closeness of data points and noise, leading to incorrect cluster representations.
Innovation Solution
A computer-implemented method and system that uses a three-phase clustering algorithm to automatically classify faults by grouping data points based on sensor contributions to prediction errors, incorporating a PCA model to identify abnormal behavior and an agglomerative hierarchical algorithm to determine clusters, thereby reducing dependence on human operators and improving fault localization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If clustering techniques are used to group excursions representative of events based on sensor residuals, then fault classification capability is improved, but the accuracy of cluster representation deteriorates when temporal closeness and noise are not accounted for
Solution Approach 1:
The patent applies preliminary action by performing temporal preprocessing of sensor residuals before clustering. Data points are grouped into temporal windows based on their time proximity, and noise filtering is applied within these windows. This preliminary temporal organization ensures that the subsequent clustering algorithm receives pre-processed data that already accounts for temporal relationships, thereby improving cluster representation accuracy while maintaining fault classification capability.
2Reliability
If all sensor residuals are used to represent excursions in N-dimensional space, then comprehensive fault detection is improved, but clustering results deteriorate due to distortion from unrelated sensors
Solution Approach 1:
The patent applies the taking out principle by extracting and removing unrelated sensors from the clustering process. After initial clustering, the system identifies sensors that do not contribute meaningfully to fault detection (unrelated sensors) and excludes them from the final clustering analysis. This extraction of irrelevant data elements improves clustering accuracy by eliminating distortion from unrelated sensors while maintaining comprehensive fault detection through the initial use of all sensors.
Solution Approach 2:
The patent applies local quality by applying different processing treatments to different sensors based on their relevance. Sensors are evaluated individually, and those identified as unrelated to specific fault types receive different weighting or are excluded from particular clustering operations. This localized differentiation allows the system to maintain comprehensive monitoring while improving clustering precision for relevant fault indicators.
3Measurement precision
If manual event localization is used, then operator expertise can identify faults, but dependence on human operators increases and automation decreases
Solution Approach 1:
The patent applies self-service by enabling the system to automatically perform fault localization without human intervention. The clustering algorithm autonomously identifies fault patterns, localizes events, and classifies anomalies based on the processed sensor data. This self-service capability eliminates dependence on manual operator analysis while maintaining high accuracy in fault localization, directly addressing the contradiction between automation extent and localization precision.
Data Source
AI summary
A computer implemented method, system and program product for automatic fault classification. A set of abnormal data can be automatically grouped based on sensor contribution to a prediction error. A principal component analysis (PCA) model of normal behavior can then be applied to a set of newly generated data, in response to automatically grouping the set of abnormal data based on the sensor contribution to the prediction error. Data points can then be identified, which are indicative of abnormal behavior. Such an identification step can occur in response to applying the principal component analysis mode of normal behavior to the set of newly generated data in order to cluster and classify the data points in order to automatically classify one or more faults thereof. The data points are automatically clustered, in order to identify a set of similar events, in response to identifying the data points indicative of abnormal behavior.