Active features extracted by deep belief network for process monitoring

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Fault Detection

Background:

• Deep learning excels at feature extraction for process monitoring.
• Previous methods often use all extracted features, potentially including detrimental ones.
• The impact of individual feature validity on monitoring performance remains under-researched.

Purpose of the Study:

• To analyze the feasibility of individual deep belief network features for process monitoring.
• To propose and validate the concept of active features (AFs) that actively indicate faults.
• To improve fault detection performance by utilizing only advantageous features.

Main Methods:

• Feasibility analysis of each feature extracted by deep belief networks.
• Proposal of 'active features' (AFs) with clear fault expression.
• Utilized Euclidean distance for sample dissimilarity and moving average for noise reduction.

Main Results:

• Demonstrated the concept of active features (AFs) for enhanced process monitoring.
• The proposed AF-DBN method showed superior fault detection rates compared to existing advanced techniques.
• Validation performed on both a numerical process and the benchmark Tennessee Eastman (TE) process.

Conclusions:

• Not all deep learning-extracted features are beneficial for process monitoring.
• Active features (AFs) provide a more effective basis for fault detection models.
• The AF-DBN method offers a feasible and superior approach to industrial process fault detection.