Soft sensors for industrial fault detection using multi-scale fusion temporal convolutional autoencoders
Huanqi Sun1, Weili Xiong1, Zhongmei Li2
1School of Internet of Things Engineering, Jiangnan University, Wuxi 214122, PR China.
ISA Transactions
|December 25, 2025
Summary
This study introduces an advanced autoencoder-based temporal convolutional soft sensor for industrial process monitoring. The model effectively extracts multi-scale features and detects faults using reconstruction errors, improving monitoring accuracy.
Area of Science:
- Industrial Process Monitoring
- Machine Learning Applications
- Data Science
Background:
- Intelligent sensors generate vast process data, enabling soft sensor development for complex industrial monitoring.
- Existing methods struggle with multi-scale feature extraction and dynamic data evolution.
Purpose of the Study:
- To propose an enhanced autoencoder-based temporal convolutional soft sensor model.
- To effectively capture multi-scale features and dynamic evolution in industrial process data.
- To improve the robustness and accuracy of industrial process monitoring and fault detection.
Main Methods:
- Developed a filter temporal convolutional network with adaptive filter-response normalization for enhanced multi-scale feature extraction.
- Constructed a multi-layer filter temporal convolutional autoencoder for efficient feature extraction and data reconstruction.
- Implemented a multi-scale feature fusion module with a channel attention mechanism for adaptive temporal feature integration.
- Established a statistical metric based on reconstruction errors and Kullback-Leibler divergence for fault detection control limits.
Main Results:
- The proposed model effectively extracts multi-scale features and captures dynamic process evolution.
- Adaptive filter-response normalization and channel attention enhance model generalization and robustness.
- Accurate process data reconstruction and improved fault detection capabilities were achieved.
- Validated through applications in wastewater treatment and multiphase flow processes.
Conclusions:
- The enhanced autoencoder-based temporal convolutional soft sensor demonstrates superior performance in industrial process monitoring.
- The method offers effective multi-scale feature extraction, robust temporal feature integration, and reliable fault detection.
- The model provides a valuable tool for complex industrial process analysis and control.
