Prescriptive Asset Alerts Using LLM Feedback and Historical Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing condition monitoring systems require explicit configuration by system designers or users, which is complex and challenging, especially for novice users, and often perform poorly with small numbers of user-identified failure examples, leading to a gap between data patterns and prescriptive messaging.
Innovation Solution
A machine-agnostic method using a trained large language model that processes user feedback and historical data to provide prescriptive messaging without the need for explicit configuration, by filtering and analyzing operational data patterns across multiple assets using similarity and relatedness scores, and incorporating feedback to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If predefined rules or thresholds are used for prescriptive messaging, then the system can provide automated failure analysis, but the system requires complex configuration by system designers or users
Solution Approach 1:
The system performs self-configuration by automatically learning failure patterns and their causes from historical data without requiring manual rule definition. The machine learning model autonomously identifies correlations between operational parameters and failure modes, eliminating the need for expert configuration while maintaining automated prescriptive messaging capabilities
Solution Approach 2:
Manual rule configuration is replaced with automated machine learning algorithms that learn patterns from data. Instead of requiring users to manually define thresholds and rules, the system uses computational models to automatically discover and update failure detection logic based on historical operational data
2Adaptability or versatility
If learning algorithms are used with user-identified failure examples, then the system can adapt to specific failure patterns, but the algorithms perform poorly with small numbers of examples
Solution Approach 1:
The system performs preliminary learning by training on extensive historical operational data and failure examples during system setup and initialization. This pre-training phase allows the model to learn general failure patterns before deployment, so that when actual failures occur, the system can provide accurate prescriptive messaging even with limited new examples
Solution Approach 2:
The system merges multiple data sources including operational parameters, maintenance records, and failure examples into a unified training dataset. By combining these diverse data sources, the system creates a more robust learning foundation that improves detection accuracy even when individual failure examples are scarce
3Productivity
If maintenance management systems are integrated to collect failure information, then the system can automatically gather historical data, but the integration complexity increases
Solution Approach 1:
The system employs universal data collection mechanisms that can interface with multiple types of maintenance management systems through standardized protocols. The learning model is designed to process data in various formats and structures, allowing it to adapt to different data sources without requiring complex custom integration for each system type
4Adaptability or versatility
If the system provides prescriptive messaging without explicit configuration, then the system becomes more scalable and machine-agnostic, but the system requires advanced algorithms to learn patterns autonomously
Solution Approach 1:
Manual pattern definition is replaced with automated machine learning algorithms that autonomously learn failure patterns from data. The system uses computational models to identify correlations and patterns without human intervention, enabling machine-agnostic operation across different asset types while maintaining scalability
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
A data processing system and methods for evaluating an operational condition of an asset are described. The asset is one of a plurality of assets that are monitored by a condition monitoring system. The condition monitoring system automatically issues alert messages based on the operational condition of the monitored assets. The methods use a trained large language model to generate output messages that provide a summary of contextually relevant content from feedback from previous alert messages in order to assist a user in evaluating an operational condition of an asset.