Industrial Machine Diagnostics Using Synthetic Abnormal Verification Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Diagnostic apparatuses face challenges in verifying the validity of models generated using normal data, as they lack data during abnormal operation, leading to potential over-learning or convergence to local solutions, and the division of data into model generation and verification sets is problematic when only normal operation data is available.
Innovation Solution
The diagnostic apparatus generates abnormal data by applying predetermined changes to normal data, such as adding impulses, fixed values, or frequency components, to verify the validity of the learning model using verification data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If data is collected only during normal operation, then data collection is easier and more frequent, but the model cannot be properly verified for abnormal operation detection
Solution Approach 1:
The patent applies preliminary action by generating abnormal data in advance through data transformation processes before the model verification stage. The verification data generation unit creates synthetic abnormal data by adding predetermined changes (impulses, fixed values, frequency components) to normal operation data, enabling model verification to proceed without requiring actual abnormal operation data collection.
2Measurement precision
If abnormal data is collected to verify model validity, then model verification accuracy improves, but data collection becomes difficult due to rarity of abnormal operations
Solution Approach 1:
The patent applies copying by creating synthetic copies of abnormal data through transformation of normal operation data. The verification data generation unit produces artificial abnormal data samples that replicate the characteristics of genuine abnormal data without requiring actual abnormal operation occurrences, thus enabling accurate model verification while avoiding collection difficulties.
3Measurement precision
If the model is trained to cover entire normal operation range, then normal operation detection accuracy improves, but the model may over-generalize and fail to detect abnormalities
Solution Approach 1:
The patent applies parameter changes by systematically modifying data parameters to generate verification datasets with predetermined changes including impulses, fixed values, and frequency components. These parameter transformations create synthetic abnormal conditions that test the model's ability to distinguish between normal variations and genuine abnormalities, preventing over-generalization while maintaining accurate normal operation detection.
Data Source
AI summary
A diagnostic apparatus of the invention acquires normal data related to an operating state during normal operation of an industrial machine, stores the normal data, generates a learning model by learning based on the stored normal data, and performs an estimation process for normality or abnormality of an operation of the industrial machine using the learning model. The diagnostic apparatus of the invention further generates verification data including at least one piece of abnormal data based on the stored normal data to verify validity of the learning model on receiving a result of the estimation process using the learning model based on the verification data.