Learning Model Input Using Autoencoded Physical-Property Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models for predicting product quality using physical-property data struggle with low accuracy due to incomplete representation of multi-dimensional physical-property characteristics, as numerical values derived from spectrum or image data do not comprehensively cover overall product characteristics.
Innovation Solution
A learning apparatus that utilizes an autoencoder to derive multi-dimensional physical-property relevance data from multi-dimensional physical-property data, incorporating production condition data, and inputs this data to a machine learning model for improved prediction accuracy, with the autoencoder learned using high-quality product data to output feature data or difference data for enhanced representation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If numerical values (slope, periodicity, amplitude, peak height, peak width) are derived from spectrum data as learning input data, then the data processing is simplified, but the accuracy of product quality prediction is leveled off at a relatively-low level because these numerical values do not comprehensively cover overall characteristics of the multi-dimensional physical-property data
Solution Approach 1:
The patent transforms the representation of spectrum data from traditional numerical extraction (1D parameters like peak height, width) to a 2D image-based representation where the spectrum waveform is visualized as an image. This allows the machine learning model to process the complete multi-dimensional characteristics of the spectrum data without loss of information, thereby improving prediction accuracy while maintaining processing efficiency through established image processing techniques
2Loss of information
If multi-dimensional physical-property data (spectrum data with multiple parameters) is used as learning input data, then the comprehensiveness of product characteristics representation is improved, but the complexity of data processing and model input increases
Solution Approach 1:
The patent creates a visual copy or representation of the multi-dimensional spectrum data in the form of an image. Instead of directly feeding complex multi-parameter spectrum data into the machine learning model, the system generates an image copy that preserves all the original information while being compatible with standard image processing and machine learning pipelines, thus reducing system complexity
Data Source
AI summary
There are provided a learning apparatus, an operation method of the learning apparatus, an operation program of the learning apparatus, and an operating apparatus capable of further improving accuracy of prediction of a quality of a product by a machine learning model in a case where learning is performed by inputting, as learning input data, multi-dimensional physical-property relevance data, which is derived from multi-dimensional physical-property data of the product, to the machine learning model. In the learning apparatus, a first processor derives, as learning input data, multi-dimensional physical-property relevance data which is related to multi-dimensional physical-property data. A first processor inputs the learning input data to the machine learning model, performs learning, and outputs the machine learning model as a learned model to be provided for actual operation.