Injection Molding Control Using Dual Learning Models for Defect Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing injection molding systems require a large amount of training data and manual operation to determine optimal set values, making reinforcement learning impractical due to high defect rates and resource consumption.
Innovation Solution
A learning model generation method that collects and utilizes first and second training data to generate two learning models: one for predicting the quality of molded products and another for determining set values that reduce defect rates, thereby optimizing molding conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If reinforcement learning is used to determine set values, then automation is improved, but a huge amount of training data and resin material is required
Solution Approach 1:
The patent segments the learning process into two distinct models: a first learning model that predicts quality degrees using relatively small training data, and a second learning model that determines set values based on defect degrees and measured values. This segmentation allows each model to be trained with appropriate data quantities, reducing overall resin material consumption while maintaining automation.
Solution Approach 2:
The first learning model acts as an intermediary between the training data collection and the final set value determination. It processes quality predictions that inform the second learning model, enabling the system to achieve automation with reduced direct training data requirements by using this intermediate processing layer.
2Extent of automation
If reinforcement learning is used to determine set values, then automation is improved, but high defect rates occur during learning process
Solution Approach 1:
By dividing the learning system into two models with different functions, the patent reduces the defect rate during learning. The first model learns quality prediction from limited data, while the second model focuses on set value optimization, allowing each to specialize and reduce overall defects compared to a single monolithic reinforcement learning approach.
Solution Approach 2:
The first learning model performs preliminary quality assessment before the second model determines final set values. This preliminary action allows the system to identify potential quality issues early in the learning process, reducing the generation of defective products during training.
3Manufacturing precision
If accurate simulation of molding process is attempted, then manufacturing precision is improved, but training data requirements increase
Solution Approach 1:
The patent achieves accurate molding process simulation with reduced training data by segmenting the learning task. The first learning model handles quality prediction with smaller datasets, while the second model optimizes set values based on defect patterns, allowing accurate simulation without requiring enormous training data volumes.
Solution Approach 2:
Instead of requiring extensive real-world training data for accurate simulation, the system uses measured values from actual molding processes as inputs to the second learning model. This copying approach allows the system to learn from real process data without needing to generate massive amounts of training data through repeated trial and error.
Data Source
AI summary
First training data including a set value related to a molding machine, a measured value obtained by measuring a physical quantity related to molding, and a degree of quality of a molded product generated by the molding machine is collected, a first learning model for outputting a degree of quality of a molded product when a set value and a measured value are input is generated by machine learning based on collected first training data, second training data including a defect degree for each defect type of a molded product, a measured value, and a set value capable of reducing the defect degree is collected, and a second learning model for outputting a set value capable of reducing a defect degree when a defect degree and a measured value are input is generated by machine learning based on collected second training data and a degree of quality output from the first learning model.