Abnormality Prediction Using Probability Density Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing abnormality prediction methods for devices, such as gas engines, struggle when applied to newly introduced or unfailed devices, as they rely on data from similar devices that have experienced abnormalities, and individual differences among devices can lead to inaccurate predictions.
Innovation Solution
A prediction apparatus and method that acquires operation data, estimates probability densities using a variational Bayesian method, and creates a prediction model excluding individual differences, allowing for accurate abnormality prediction in devices of the same model by converting operation data into probability densities and learning from corrected data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a prediction model is created using data collected when an abnormality has occurred in the device, then the prediction accuracy for devices with abnormalities is improved, but the method cannot be used for newly introduced devices or devices where abnormalities do not occur
Solution Approach 1:
The patent creates a virtual copy of the device's probability distribution characteristics by learning the probability density function from normal operation data. This virtual model serves as a reference that can be applied to newly introduced devices without requiring actual abnormality data from each specific device, thus enabling prediction while maintaining versatility.
Solution Approach 2:
The patent transforms the prediction approach by changing from using raw operation data directly to using probability density values derived from the operation data. This parameter transformation allows the model to capture the statistical characteristics of normal operation, making it applicable to newly introduced devices while maintaining prediction accuracy for detecting deviations from normal behavior.
2Adaptability or versatility
If a prediction model is created using data collected from other devices of the same model in which an abnormality has occurred, then the method can be applied to newly introduced devices, but individual differences among devices reduce prediction accuracy
Solution Approach 1:
The patent transforms operation data into probability density values that represent the statistical distribution characteristics. By doing so, it creates a standardized parameter representation that reduces the impact of individual device differences while preserving the essential operational patterns, enabling accurate cross-device prediction.
Solution Approach 2:
The patent develops a universal prediction model based on probability density functions that can be applied across multiple devices of the same model. The model learns the common statistical characteristics of normal operation that are universal across devices, allowing it to function effectively for both existing and newly introduced devices despite individual variations.
3Ease of manufacture
If machine learning algorithms are simply applied to create prediction models, then the model creation process is simplified, but individual differences among devices result in models classified by device-specific characteristics rather than general abnormality patterns
Solution Approach 1:
The patent introduces an intermediate transformation step that converts operation data into probability density values before applying machine learning. This parameter change acts as a feature normalization process that removes device-specific variations, allowing simple machine learning algorithms to achieve high prediction accuracy by focusing on general abnormality patterns rather than device-specific characteristics.
Data Source
AI summary
Provided is a prediction system that predicts whether a prescribed event will occur in a device, without being affected by differences among individual devices. The prediction system comprises: a data acquisition unit which acquires operation data representing the operation status of a device; a probability density estimation unit which estimates the probability density of the operation data; and an abnormality prediction unit which predicts whether an abnormality will occur in the device on the basis of the probability density estimation results of the operation data and a prediction model.