Fault Diagnosis Based on Chemical Sensor Data with an Active Deep Neural Network

Area of Science:

• Chemical Engineering
• Data Science
• Artificial Intelligence

Background:

• Big sensor data offers potential for chemical process fault diagnosis.
• Accurate fault diagnosis is crucial for security, stability, and reliability in chemical processes.
• Labeling chemical sensor data is expensive and time-consuming.

Purpose of the Study:

• To present a deep neural network (DNN) with novel active learning for chemical fault diagnosis.
• To develop a method combining deep learning and active learning for efficient fault diagnosis.
• To address the challenge of consecutive fault diagnosis in chemical processes.

Main Methods:

• Utilized deep learning with stacked denoising auto-encoder (SDAE) for unsupervised feature learning from raw sensor data.
• Employed a novel active learning criterion (Best vs. Second Best and Lowest False Positive) for model fine-tuning.
• Implemented a supervised approach by adding learned features to a Softmax regression layer for diagnosis.

Main Results:

• The proposed DNN with active learning achieved superior diagnosis accuracy on industrial datasets.
• Significant improvements were observed in accuracy and false positive rate compared to existing methods.
• The method demonstrated effectiveness in reducing the need for extensively labeled sensor data.

Conclusions:

• The integrated deep learning and active learning approach enhances chemical fault diagnosis.
• The novel active learning criterion is particularly effective for chemical process data.
• This method offers a more efficient and accurate solution for chemical fault diagnosis using big sensor data.