Failure Prediction Using Simulated Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing failure prediction systems face challenges in accurately predicting failures in industrial equipment like motors and gears due to the difficulty in collecting training data for supervised learning, especially for equipment with different operating conditions and configurations, leading to inaccuracies in predicting failure states.
Innovation Solution
A learning apparatus that extracts time fluctuation patterns of feature frequencies from state observation signal data, generates simulated data to create training sets, and uses these to build a classification model for determining failure states, enabling accurate prediction of equipment failures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is used to improve prediction accuracy, then prediction accuracy is improved, but training data collection becomes difficult
Solution Approach 1:
The patent generates simulated failure state data by copying and transforming normal state data through signal processing operations. Specifically, it creates training data by adding noise components and modifying frequency characteristics of normal operation data to simulate various failure conditions, eliminating the need to collect actual failure data from equipment.
Solution Approach 2:
The patent performs preliminary data preparation by pre-generating simulated failure data and pre-processing normal operation data before the actual prediction task. This includes extracting feature frequencies, creating amplitude modulation patterns, and preparing training datasets in advance, so that when prediction is needed, the model can be quickly trained or updated without requiring real-time failure data collection.
2Ease of manufacture
If unsupervised learning is used to simplify data collection, then data collection becomes easier, but prediction accuracy deteriorates
Solution Approach 1:
The patent introduces simulated failure data as an intermediary between normal operation data and the supervised learning model. This intermediary training data bridges the gap by providing labeled failure examples that don't require actual failure occurrences, enabling supervised learning while maintaining ease of data acquisition through simulation rather than direct observation of rare failure events.
Data Source
AI summary
The learning apparatus according to one exemplary embodiment includes: a pattern extractor that extracts a time fluctuation pattern of an amplitude of a feature frequency from state observation signal data up to a first time point, the state observation signal data indicating an operation state of equipment, the feature frequency being associated with a part of the equipment; a training data generator that generates, based on the time fluctuation pattern of the amplitude of the feature frequency, simulated state observation signal data representing the time fluctuation pattern of the amplitude of the feature frequency at and after the first time point, and generates training data including the simulated state observation signal data; and a learner that generates a classification model for determination of a failure state of the part of the equipment using the training data.