Horizontal Data Augmentation Strategy for Industrial Quality Prediction.

Area of Science:

• Industrial Process Control
• Machine Learning Applications
• Data Science

Background:

• Neural network-based soft sensor technology is crucial for industrial optimization and prediction.
• Existing methods may face challenges with data availability and quality for accurate predictions.
• Soft sensors are vital for real-time monitoring and quality control in complex industrial settings.

Purpose of the Study:

• To propose a novel horizontal data augmentation strategy, Combined Autoencoder Data Augmentation (CADA).
• To develop a CADA-based Convolutional Neural Network (CADA-CNN) soft sensor model.
• To evaluate the CADA-CNN model's effectiveness in industrial debutanizer and steam volume processes.

Main Methods:

• Implemented a Combined Autoencoder Data Augmentation (CADA) strategy for data enhancement.
• Developed a CADA-CNN soft sensor model integrating the augmentation strategy.
• Validated CADA feasibility using Spearman correlation coefficient.
• Compared CADA-CNN performance against Artificial Neural Network (NN), Support Vector Regression (SVR), and Convolutional Neural Network (CNN) models.

Main Results:

• The CADA strategy provides highly available data for prediction models.
• The CADA-CNN model achieved lower prediction error compared to NN, SVR, and CNN.
• The proposed model exhibited a better distribution of prediction errors.
• Spearman correlation confirmed the feasibility of the CADA strategy.

Conclusions:

• The CADA strategy effectively enhances data for soft sensor applications.
• The CADA-CNN model offers a significant improvement in prediction accuracy and reliability for industrial processes.
• This approach provides a robust solution for soft sensor modeling challenges in industrial settings.