A Scalable Algorithm for Identifying Multiple-Sensor Faults Using Disentangled RNNs

Area of Science:

• Industrial Process Control
• Machine Learning Applications
• Sensor Systems Engineering

Background:

• Sensor fault detection and isolation (SFD-SFI) are crucial for industrial operations.
• Increasing system complexity poses challenges for traditional SFD-SFI methods.
• Analytical redundancy, based on models, is key for SFD-SFI.

Purpose of the Study:

• To develop an advanced SFD-SFI method using a disentangled recurrent neural network (RNN).
• To address the 'smearing-out' effect where faults propagate to non-faulty sensors.
• To enable efficient and accurate identification of faulty sensors.

Main Methods:

• Utilized a disentangled recurrent neural network (RNN) architecture.
• Incorporated a probabilistic model for residual generation.
• Developed a novel procedure for sensor fault identification with linear computational complexity.

Main Results:

• The proposed RNN effectively mitigates the smearing-out effect in sensor fault detection.
• The probabilistic residual model enables accurate faulty sensor identification.
• The algorithm demonstrates linear computational complexity relative to the number of sensors.

Conclusions:

• The novel disentangled RNN offers a robust solution for SFD-SFI in complex industrial systems.
• The method provides efficient and accurate fault identification, outperforming traditional approaches.
• Empirical validation on petrochemical plant data confirms the architecture's effectiveness.