Related Experiment Videos
A Layer-Wise Data Augmentation Strategy for Deep Learning Networks and Its Soft Sensor Application in an Industrial
IEEE Transactions on Neural Networks and Learning Systems
|December 17, 2019
Summary
A new layer-wise data augmentation (LWDA) strategy improves deep learning for soft sensor modeling in industrial processes. The LWDA-based stacked autoencoder (LWDA-SAE) enhances prediction accuracy and speeds up convergence.
Area of Science:
- Chemical Engineering
- Data Science
- Artificial Intelligence
Background:
- Inferential sensors are crucial for predicting industrial process quality variables not directly measurable by hard sensors.
- Deep learning offers advanced feature representation for soft sensor modeling but requires extensive data and can suffer from information loss during layer-wise pretraining.
- Existing methods like multilayer perceptron and traditional stacked autoencoders have limitations in handling complex industrial data for soft sensing.
Purpose of the Study:
- To propose a novel layer-wise data augmentation (LWDA) strategy for deep learning networks.
- To develop and detail a LWDA-based stacked autoencoder (LWDA-SAE) model for soft sensor applications.
- To evaluate the performance of the LWDA-SAE model in predicting aviation kerosene boiling points in an industrial hydrocracking process.
Main Methods:
- Development of a layer-wise data augmentation (LWDA) strategy tailored for deep learning pretraining.
- Implementation of the LWDA strategy within a stacked autoencoder architecture, resulting in the LWDA-SAE model.
- Application of the LWDA-SAE model to predict key quality variables (10% and 50% boiling points) in industrial hydrocracking.
Main Results:
- The LWDA-SAE model demonstrated superior performance compared to multilayer perceptron, traditional SAE, and input layer data augmentation SAE (IDA-SAE).
- The proposed LWDA-SAE achieved faster convergence rates during training.
- The LWDA-SAE model exhibited lower learning errors, indicating improved prediction accuracy for aviation kerosene boiling points.
Conclusions:
- The LWDA strategy effectively addresses the limitations of traditional deep learning pretraining for soft sensor modeling.
- LWDA-SAE offers a more efficient and accurate approach for inferential sensing in industrial processes.
- This method holds significant potential for enhancing online quality prediction and process optimization in industries like hydrocracking.