Quality Prediction Model Transfer Learning for Manufacturing Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for predicting the quality of products in manufacturing settings face challenges with reduced man-hours and costs, leading to decreased estimating accuracy of quality prediction models, which can result in production failures due to low accuracy models.
Innovation Solution
A quality prediction system that utilizes transfer learning by converting learning data from one manufacturing equipment to another to enhance the number of samples for model generation, while also identifying non-contributory adjustment items and acquiring additional data to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If the number of experiments is reduced to lower costs and man-hours, then the costs and time required for model generation are reduced, but the estimating accuracy of the generated quality prediction model decreases
Solution Approach 1:
The patent applies transfer learning by copying knowledge from a source domain (first manufacturing equipment) to a target domain (second manufacturing equipment). Learning data collected from the first production line is transferred and adapted to generate the quality prediction model for the second production line, thereby reducing the need for extensive experiments on the target line while maintaining model accuracy through knowledge transfer.
Solution Approach 2:
The patent transforms learning data by adjusting parameters to match between source and target domains. The data transfer part converts the first learning data to match the second learning data format and distribution, enabling the model to generalize across different manufacturing equipment while requiring fewer target-domain experiments.
2Quantity of substance
If learning data from different manufacturing equipment is transferred and combined, then the number of samples for model generation is increased, but the complexity of data processing and domain adaptation increases
Solution Approach 1:
The patent introduces a data transfer part as an intermediary component that bridges the source domain and target domain. This intermediary performs domain adaptation by transforming the first learning data to match the second learning data, facilitating knowledge transfer while managing the complexity of integrating data from different manufacturing equipment through a dedicated transformation layer.
Data Source
AI summary
The present invention reduces costs relating to generation of a model, and suppresses a decrease in the estimating accuracy of the generated model. A quality prediction system of one aspect of the present invention acquires first learning data collected in first manufacturing equipment, acquires second learning data collected in second manufacturing equipment, and in order to implement transfer learning, converts the acquired first learning data to match the second learning data, uses the converted first learning data and the second learning data to implement machine learning of a quality prediction model, uses a trained quality prediction model to specify an adjustment amount for each adjustment item of the second manufacturing equipment so that a quality index predicted for a second product satisfies a quality standard, and outputs the specified adjustment amount of each adjustment item of the second manufacturing equipment.