Pseudo Waveform Generation for AI Discriminator Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing technologies face challenges in generating accurate AI models for inspecting abnormal waveforms, particularly when there is a limited amount of abnormality data available, leading to overfitting and insufficient accuracy in determining normality and abnormality.
Innovation Solution
The system generates a large amount of pseudo abnormal waveform data from a small amount of actual abnormal waveform data, using techniques such as noise addition, expansion, and contraction, to create diverse pseudo waveforms that do not impair the physical meaning of the original data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a small amount of abnormal waveform data is used to train the AI model, then the model training can be performed, but the model accuracy and reliability are insufficient due to overfitting
Solution Approach 1:
The patent creates pseudo abnormal waveform data by copying and transforming existing abnormal waveform data through various signal processing techniques including noise addition, magnitude scaling, time-shifting, and splicing operations. This copying approach generates synthetic training data that preserves the essential characteristics of real abnormal waveforms while providing sufficient data volume for reliable model training
Solution Approach 2:
The patent applies parameter changes to the existing abnormal waveform data by modifying signal characteristics such as adding different types and levels of noise, changing magnitude scales, applying time shifts, and adjusting frequency components. These parameter transformations generate diverse pseudo data samples that maintain physical meaning while increasing data variety for robust model training
2Reliability
If more abnormal waveform data is collected to improve model accuracy, then the model performance can be enhanced, but the time required for data collection and the delay in startup are increased
Solution Approach 1:
The patent performs preliminary data preparation by generating pseudo abnormal waveform data in advance using signal processing techniques. This preliminary action creates a sufficient training dataset before actual model training begins, eliminating the need for prolonged data collection periods and enabling faster startup of the inspection system
Solution Approach 2:
Instead of waiting to collect additional real abnormal waveform data over time, the patent copies and transforms existing abnormal waveform samples to generate synthetic training data. This copying approach immediately provides sufficient training data volume without the time delay associated with waiting for natural occurrence of abnormal conditions during data collection
3Quantity of substance
If data augmentation techniques are applied to generate pseudo abnormal waveforms, then the training data volume is increased, but there is a risk of impairing the physical meaning of the original data
Solution Approach 1:
The patent carefully controls parameter changes during pseudo data generation by applying bounded transformations such as adding noise within specific amplitude ranges, scaling magnitudes by controlled factors, and applying time shifts that preserve waveform characteristics. These controlled parameter changes increase data volume while maintaining the physical meaning and diagnostic value of the original abnormal waveform patterns
Solution Approach 2:
The patent copies real abnormal waveform data and applies transformations that preserve the essential physical characteristics and diagnostic features of the original signals. By carefully selecting and controlling the copying and transformation operations, the generated pseudo data maintains fidelity to the physical phenomena being measured while providing sufficient data volume for training
Data Source
AI summary
Provided is a generation system including: one or more processors which acquire waveform data; specify an intention of a user; and generate pseudo waveform data from the waveform data acquired in such a manner that an intention of a user specified is reflected. Provided is a method for generating a waveform evaluation model executed by a computer, including: acquiring waveform data; specifying an intention of a user; generating pseudo waveform by generating pseudo waveform data from the waveform data acquired in the acquiring the waveform data in such a manner that the intention of the user specified in the specifying the intention is reflected; and executing learning by executing machine learning using the pseudo waveform data generated in the generating the pseudo waveform to generate a waveform evaluation model which outputs an evaluation result of input waveform data.