ML Model Susceptibility to Signal Degradation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine-learning models used for prognostic-surveillance in critical systems often 'follow' signal degradation, leading to missed anomalies due to poor signal-to-noise ratios and small model sizes, which can result in undetected failures and costly consequences.
Innovation Solution
A system that characterizes the susceptibility of inferential models to follow signal degradation by training them on time-series signals, introducing degradation, and using a Following metric (FM) to assess susceptibility, suggesting parameter adjustments such as changing training vectors, filtering noise, or modifying monitored signals to mitigate this issue.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If an ML model is trained on time-series sensor data with poor signal-to-noise ratio, then the model can operate with smaller size and lower computational requirements, but the model becomes susceptible to the Following phenomenon where predicted values follow degradation in real signals, causing missed anomaly detections
Solution Approach 1:
The patent applies preliminary action by introducing artificial degradation into training data before model training. This allows the model to learn the expected degradation patterns in advance, so that during actual surveillance operations, the model can distinguish between normal degradation and anomalous deviations, preventing the Following phenomenon while maintaining small model size
Solution Approach 2:
The patent changes training parameters by adding controlled degradation amplitudes and varying training data characteristics. This enables the model to adapt to degradation patterns without increasing model complexity, resolving the contradiction between small model size and reliable anomaly detection
2Loss of time
If an ML model is trained with insufficient training data or poor signal-to-noise ratio, then training time and data requirements are reduced, but the model generates predicted values that follow real signal degradation, making residual analysis ineffective
Solution Approach 1:
The patent introduces artificial degradation patterns during the training phase, allowing the model to learn expected degradation behaviors beforehand. This preliminary exposure enables the model to maintain good generalization performance with limited training data, preventing the Following phenomenon while reducing training time requirements
Solution Approach 2:
The patent uses artificial degradation as an intermediary element during training. This intermediary allows the model to learn degradation patterns without requiring extensive real-world degraded data, bridging the gap between limited training data and robust anomaly detection performance
Data Source
AI summary
The disclosed embodiments relate to a system that characterizes susceptibility of an inferential model to follow signal degradation. During operation, the system receives a set of time-series signals associated with sensors in a monitored system during normal fault-free operation. Next, the system trains the inferential model using the set of time-series signals. The system then characterizes susceptibility of the inferential model to follow signal degradation. During this process, the system adds degradation to a signal in the set of time-series signals to produce a degraded signal. Next, the system uses the inferential model to perform prognostic-surveillance operations on the set of time-series signals with the degraded signal. Finally, the system characterizes susceptibility of the inferential model to follow degradation in the signal based on results of the prognostic-surveillance operations.