Supervised Failure Model Creation Using Unsupervised Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Asset data platforms face challenges in predicting failures when there is limited information available regarding prior failure occurrences, leading to unpredictable failures, costly downtime, and safety issues.
Innovation Solution
An approach using unsupervised modeling techniques to derive label data for supervised failure models, involving phases such as determining deviation bounds, classifying assets, defining anomaly thresholds, identifying failure occurrences, and categorizing prior failures, enables the creation of predictive models even with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning techniques are applied to predict failures, then prediction accuracy is improved, but the requirement for labeled historical failure data increases
Solution Approach 1:
The system performs preliminary unsupervised learning analysis on historical operating data before supervised learning can be applied. This preliminary action identifies potential failure patterns and creates initial labels, enabling subsequent supervised learning to achieve high prediction accuracy even when complete labeled failure data is unavailable
Solution Approach 2:
Unsupervised learning models serve as an intermediary between raw operating data and supervised learning algorithms. These intermediary models process unlabeled data to generate failure labels and insights, bridging the gap between available data and the requirements for accurate supervised prediction
2Reliability
If more historical operating data is collected for analysis, then failure prediction capability is improved, but data processing complexity increases
Solution Approach 1:
The data processing system is segmented into multiple specialized components: unsupervised learning modules for initial pattern detection, supervised learning modules for refined prediction, and anomaly detection systems. Each segment handles specific aspects of the data, reducing overall system complexity while maintaining high prediction capability
Solution Approach 2:
The system applies partial action by using unsupervised learning on all historical data to identify key patterns, then focuses supervised learning only on the identified failure cases. This approach processes the necessary amount of data for reliable prediction without the excessive complexity of analyzing every data point with full supervised learning
Data Source
AI summary
The example systems, methods, and devices disclosed herein generally relate to generating create a supervised failure model for assets in the given fleet that is configured to receive operating data as inputs and output a prediction as to the occurrence of a given failure type at the asset. In some instances, a data analytics platform may create and use an unsupervised failure model for a subset of the assets, use the respective unsupervised failure models to detect a set of anomalies that are each suggestive of a prior failure occurrence, from the set of anomalies, identify a subset of anomalies that are each suggest of a prior failure occurrence of the given failure type, and create the supervised failure model using failure data for the identified subset of anomalies.