Establishing AI Quality Forecast Model
Conventional quality control involves mitigation after the fact with testing and analysis on product quality. However, the time-consuming testing and random sampling make it difficult to facilitate real-time quality monitoring. When irregularities occur, quality downgrade continues. The second plant of OPTC has constructed a real-time quality forecast model, which is able to make quality forecast 2 to 3 hours in advance comparing to the conventional model, and make swift parameter adjustments accordingly. If production proceeds at the rate of 200 metric tons per hour, detecting irregularity 3 hours in advance may prevent approximately 600 metric tons of defective products, which translates to approximately NT$15 million in possible loss. Minimizing loss due to product downgrading may increase the efficiency in raw material use and avoid waste.
The current model delivers approximately 95% accuracy.
- Establishing Quality Forecast Model: Four AI algorithms are conducted using historical data for training on modeling, including deep neural network, support vector machine, random forest and convolutional neural network.
- AI Model Verification: The best quality forecast model is verified and selected using actual data.