Fault Signal Recovery Using Similarity-Based Normal Signal Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current fault signal recovery technologies in early warning systems for plants are inefficient, leading to low accuracy in restoring fault signals to normal signals, which hinders the construction of accurate failure prediction models.
Innovation Solution
A fault signal recovery system that generates a signal subset by removing fault signals, uses a modeling unit to extract feature information and generate recovery models, and a recovery unit to estimate normal signals based on similarity parameters, combining recovery algorithms to optimize the recovery process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If fault signals are removed from the signal set to create clean learning data, then the quality of learning data is improved, but the amount of usable data is reduced
Solution Approach 1:
The patent extracts and removes fault signals from the signal set to create clean learning data subsets. The data processing unit identifies fault signals and separates them from normal signals, extracting only the clean portions for model training. This resolves the contradiction by selectively removing harmful elements while preserving the bulk of usable data.
Solution Approach 2:
The patent discards fault signals from the learning data set but recovers them through the failure prediction model. The model learns from clean data and then predicts fault conditions, effectively recovering the discarded information in a processed, useful form. This allows the system to both remove faulty data and still utilize fault information for prediction purposes.
2Quantity of substance
If all available data including fault signals is used for training, then the quantity of learning data is maximized, but the accuracy of the failure prediction model deteriorates
Solution Approach 1:
The patent segments the signal set into distinct subsets: clean learning data without fault signals, and separate fault signal data. By dividing the data into quality-based segments, the system can train on high-quality clean data while still incorporating fault information through the prediction model, thus maintaining both data quantity and accuracy.
Solution Approach 2:
The patent applies different quality standards to different parts of the data set. Clean signals are used for primary model training while fault signals are handled separately through prediction algorithms. This local quality approach allows the system to maximize data usage while maintaining high accuracy by matching data quality to appropriate processing methods.
3Quantity of substance
If fault signal recovery technology is applied to restore fault signals to normal signals, then all data can be used as learning data, but the accuracy of signal recovery is currently low
Solution Approach 1:
The patent performs preliminary cleaning of the signal set by removing fault signals before model training. Rather than attempting to recover fault signals into normal signals, the system first creates clean training data, then uses the trained model to predict fault conditions. This preliminary separation action avoids the accuracy problems of signal recovery while still enabling comprehensive data utilization.
Solution Approach 2:
The patent creates a copy of the clean signal set for training purposes, separate from the original data containing fault signals. This copying approach allows the model to learn from pristine data while the original data with faults can be used for validation and prediction, avoiding the need for imperfect signal recovery.
Data Source
AI summary
Disclosed is a fault signal recovery system including a data processor configured to generate a signal subset U* by removing, from a signal set U for a plurality of tags, some tags including a fault signal, and a first learning signal subset X* by removing tags disposed at positions corresponding to the some tags from a learning signal set X containing only tags of normal signals, a modeling unit configured to generate feature information F extractable from the first learning signal subset X* and recovery information P on a plurality of recovery models usable for restoring the fault signal, and a recovery unit configured to estimate and recover normal signals for the some tags based on the signal subset U*, the first learning signal subset X*, the feature information F, the recovery information P on the plurality of recovery models, and similarity Z.