A Fault Prediction and Cause Identification Approach in Complex Industrial Processes Based on Deep Learning

Area of Science:

• Industrial Engineering
• Machine Learning
• Process Control

Background:

• Production line faults lead to significant losses.
• Traditional methods struggle with complex industrial process data.
• Modern production lines generate ample data for predictive solutions.

Purpose of the Study:

• To propose a novel deep learning approach for fault prediction and cause identification in industrial production lines.
• To overcome limitations of traditional time-series analysis for complex industrial data.
• To improve the accuracy and interpretability of fault diagnosis systems.

Main Methods:

• Data treated as spatial sequences, not traditional time series.
• Improved Long Short-Term Memory (LSTM) encoder-decoder model adapted for branched spatial sequences.
• Attention Mechanism (AM) integrated for fault detection, cause identification, and output as a sequence of fault types.

Main Results:

• The spatial sequence approach overcomes multidimensional challenges and improves prediction accuracy.
• Attention mechanism weights effectively correlate faults with input data, aiding cause identification.
• The proposed method demonstrated higher prediction accuracy compared to established methods in the Tennessee Eastman process.

Conclusions:

• The deep learning approach offers a more effective way to predict and diagnose faults in complex industrial settings.
• Viewing data as spatial sequences enhances the model's ability to handle intricate production processes.
• The attention mechanism provides valuable insights into fault causality, assisting engineers in mitigation efforts.