A Soft Sensor for Multirate Quality Variables Based on MC-CNN.

Area of Science:

• Industrial Process Control
• Machine Learning
• Data Science

Background:

• Data-driven soft sensor modeling is crucial in industrial, chemical, and biochemical processes.
• Existing multi-input multi-output (MIMO) sensors often neglect the temporal variations in sampling rates between process and quality variables.
• Inconsistent sampling rates among quality variables pose a significant challenge for accurate modeling.

Purpose of the Study:

• To propose a novel deep learning (DL) model, the multitemporal channels convolutional neural network (MC-CNN), to handle varying sampling rates in soft sensor modeling.
• To develop a method that effectively predicts multiple quality variables simultaneously, considering their different temporal resolutions.
• To address the limitations of existing MIMO sensors in industrial applications.

Main Methods:

• A deep learning (DL) model based on a multitemporal channels convolutional neural network (MC-CNN) was designed.
• The MC-CNN architecture features a shared network for temporal feature extraction and parallel prediction networks for individual quality variables.
• A modified backpropagation (BP) algorithm was employed to exclude unsampled data points (blank values) from the training process.

Main Results:

• The proposed MC-CNN model demonstrated effectiveness in predicting multiple quality variables in two industrial case studies.
• The method successfully managed and utilized data with differing sampling frequencies.
• The modified BP algorithm prevented interference from missing data during model training.

Conclusions:

• The developed MC-CNN provides an effective solution for soft sensor modeling in industrial processes with asynchronous sampling rates.
• This approach enhances the accuracy and reliability of quality variable prediction in complex industrial environments.
• The study validates the utility of deep learning for handling temporal data complexities in process monitoring.