Retrieval-Augmented Prompt Generation for Asset Anomaly Alerting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional monitoring systems for predictive maintenance fail to capture complex, multivariate patterns in telemetry data, leading to missed detections or delayed maintenance, especially with increasing sensor data volumes and complexity, and require significant computational resources.
Innovation Solution
A system that uses a description generator and large language model to generate contextual descriptions of data outliers based on telemetry data, leveraging an embedding-based knowledgebase to suggest resolution suggestions, incorporating statistical analysis and machine learning for improved accuracy and explainability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If static threshold-based rules or manual diagnostics are used to detect anomalies, then the system is simple to implement, but the system fails to capture complex multivariate patterns leading to missed detections
Solution Approach 1:
The patent introduces an embedding model as an intermediary component that transforms raw telemetry data into meaningful representations. This embedding model acts as a mediator between the complex multivariate data and the anomaly detection logic, enabling the system to capture complex patterns without requiring overly complex detection rules. The embedding layer converts high-dimensional sensor data into compressed representations that preserve essential patterns while reducing computational complexity.
Solution Approach 2:
The patent replaces traditional mechanical threshold-based rule systems with a machine learning-based embedding approach. Instead of using fixed threshold comparisons and manual diagnostic rules, the system employs trained embedding models that automatically learn complex patterns from data. This substitution enables the system to detect multivariate patterns that would be impossible to capture with static rules, significantly improving anomaly detection accuracy.
2Reliability
If the number of sensors and measurable parameters increases to improve monitoring coverage, then the system captures more comprehensive data, but the volume and complexity of data increase making manual inspection prohibitive
Solution Approach 1:
The patent extracts essential patterns from large volumes of multivariate sensor data through the embedding model. Instead of requiring manual inspection of all raw sensor readings, the embedding model extracts and consolidates the most important patterns and anomalies into a reduced representation. This extraction process maintains comprehensive monitoring coverage while making the data feasible for automated processing and analysis.
Solution Approach 2:
The patent transforms the parameter representation of sensor data by changing from raw high-dimensional measurements to embedded low-dimensional representations. The embedding model applies parameter transformations that preserve essential information while reducing dimensionality, enabling the system to handle increased sensor coverage without proportionally increasing operational complexity.
3Speed
If traditional rule-based monitoring is used, then computational resources are minimized, but the system cannot satisfy latency constraints for real-time detection
Solution Approach 1:
The patent performs preliminary action by pre-training embedding models offline using historical telemetry data. This preliminary training phase captures complex patterns and relationships in the data, storing them in the embedded model structure. During real-time operation, the pre-trained model can quickly process new data without requiring intensive computational resources, thus satisfying latency constraints while minimizing online computational energy consumption.
Data Source
AI summary
A system, method, and computer-program product includes receiving a machine-generated alert indicating a data outlier for a physical asset; obtaining, from a computer database, a dataset including observed data for the physical asset within a predefined temporal window of the machine-generated alert; generating, via a description generator, a contextual description of the data outlier based at least on the observed data for the physical asset; searching an embedding-based knowledgebase for a subset of embedding representations of the plurality of embedding representations within a similarity threshold of an embedding representation of the contextual description; obtaining, from the computer database, a subset of the plurality of reference artifacts that correspond to the subset of embedding representations; and generating, via a large language model, a resolution suggestion for resolving the data outlier based at least on the subset of reference artifacts obtained from the computer database.