Machine State Diagnosis Using Normalized Autoencoder Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing diagnostic methods for industrial machines, particularly those using autoencoders, struggle to fully learn and reflect small-scale feature data, such as displacement or frequency components with low amplitude, which are crucial for accurate state diagnosis.
Innovation Solution
A diagnostic apparatus that normalizes data to create a uniform distribution for each feature, allowing for effective learning and inference by expanding the range of possible values for each frequency component, thereby addressing the issue of small-scale data handling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If auto encoder is used as a model for diagnosing the state of an industrial machine, then the model can learn from time-series data of physical quantities, but feature parts having a small scale such as displacement or frequency components with small amplitude values may not be fully learned
Solution Approach 1:
The patent applies parameter changes by normalizing the scale of feature data through statistical transformation. Specifically, it uses z-score normalization (subtracting mean and dividing by standard deviation) to transform features with different scales into a unified scale, ensuring that small-scale features like displacement and frequency components are treated equally with large-scale features during model learning.
2Measurement precision
If frequency component analysis is performed and amplitude values at several frequencies are used as learning data, then the model can process spectral features, but frequencies with small amplitude values may not be reflected in the model
Solution Approach 1:
The patent transforms the scale of frequency component data by applying statistical normalization. This ensures that frequency components with small amplitude values are scaled appropriately so they can be effectively learned by the model alongside frequency components with larger amplitude values, preventing loss of important diagnostic information.
3Adaptability or versatility
If different scales of feature data are used without normalization, then the model can process diverse features, but the learning process may be biased towards features with larger scales
Solution Approach 1:
The patent applies parameter changes by normalizing the scale of feature data through statistical transformation. Specifically, it uses z-score normalization (subtracting mean and dividing by standard deviation) to transform features with different scales into a unified scale, ensuring that small-scale features like displacement and frequency components are treated equally with large-scale features during model learning.
Solution Approach 2:
The patent creates equipotentiality in the feature space by transforming all features to have the same statistical scale. This ensures that no single feature dominates the learning process due to its original scale, allowing the model to learn from all features equally regardless of their original magnitude.
Data Source
AI summary
A diagnostic apparatus acquires data relating to a state of a machine, performs preprocessing on the acquired data, and modifies the preprocessed data by using a statistic relating to the state of the machine. Then, the diagnostic apparatus performs a process of machine learning by an auto encoder on the basis of the modified data and learns or diagnoses the state of the machine. Thus, this diagnostic apparatus provides a learning and inference method capable of uniformly handling a scale of data indicating a feature.