Cross-Domain Fault Detection through Optimal Transport

This website contains information about the thesis "Cross-Domain Fault Diagnosis through Optimal Transport", done by Eduardo Fernandes Montesuma as a requirement for obtaining his Bachelor degree in Computer Engineering at Universidade Federal do Ceará, Fortaleza, Brazil.

Abstract

Automatic fault diagnosis systems are an important component for fault tolerance in modern control loops. Nonetheless, the training of such diagnosis systems can be costly or even dangerous since faulty data need to be collected by driving the process to dangerous conditions. A possible solution to the said problem is training an automatic diagnosis system solely on simulation data. However, due to modeling errors, the data acquired may not reflect real process data. This is characterized by a change in the probability distributions upon which data is sampled. This problem is known in the literature as domain adaptation or cross-domain fault diagnosis in our context. Thus this work analyzes the cross-domain diagnosis problem through the point of view of optimal transport. We apply our methodology in a case study concerning the continuous stirred tank reactor (CSTR) system. Our contributions are three-fold:

• 1. we perform a comparative study concerning feature extraction and domain adaptation algorithms
• 2. we analyze the relation between wrongful model specification and the distance between source and target distributions, and its impact on classification performance
• 3. we analyze the impact of modeling errors in the quality of optimal transport plans, and the influence of this latter factor into classification performance.

In summary, we found that optimal transport-based domain adaptation is the best choice for solving the distributional shift problem. In addition, we further verified that an increasing degree of modeling error is correlated with an increase in the distance between source and target distributions. Furthermore, we found experimentally that the latter distance is correlated with a decrease in classification performance, confirming previous theoretical findings. Finally, the degree of modeling error can cause the transportation plan between source and target domain to transfer mass between different classes, harming classification performance.

Overview

We use the pipeline above for performing cross-domain fault diagnosis, which corresponds to unsupervised domain adaptation when data corresponds to the detection of faults in dynamical systems. The pipeline is composed by the following blocks,

• Process: the real process (in our case, a chemical plant). It generates real data, in opposition to simulation data.
• Modeling/Simulation: corresponds to a mathematical model used to simulate the real process. It generates simulation data. Note that the acquisition of simulation data incurs none to very small cost.
• Supervisor: it may be a human, or any a priori knowledge (e.g. knowing that the system is simulated for a specific fault) that gives us the ground-truth label of the sample.
• Feature Extraction: it corresponds to the act of extracting meaningful features from the generated signals. In our experiments we use raw signals themselves (without any feature extraction), autocorrelation features and convolutional neural network latent space features.
• Fautl Classification: this block classifies the features into one of the possible sets of faults. Indeed, it performs the fault diagnosis.

Reference

If you find our work useful in your research, please consider citing it.

            
@thesis{montesuma2021,
    author       = {Eduardo Fernandes Montesuma}, 
    title        = {Cross-Domain Fault Diagnosis through Optimal Transport},
    school       = {Universidade Federal do Ceará},
    year         = 2021,
    type         = {Bachelor's Thesis}
}