Improving interpretability for fault diagnosis in complex chemical processes using combined regularization and linearization in neural network modeling

Chemical industrial systems are continuously subject to faults that can cause accidents and affect production and the environment in many ways. Their diagnosis after detection is crucial for operations. Artificial Neural Networks (ANNs) have been used for both tasks in complex chemical processes, but their usual nonlinear operations make it difficult to find interpretations for fault diagnosis. Several techniques have been used to extract interpretability from ANNs. In this work, we propose a method that combines L1 and L2 regularizations with linearization as a way to reduce model complexity, favoring interpretability. The results were compared with PCA and logistic regression. The commonly used benchmark problem Tennessee Eastman Process (TEP) was used as a case study. We found that the combined approach greatly increases the sparsity of input weights in ANNs, isolating the features most related to faults. The results are superior mainly for faults that are hard to detect. This contributes to an earlier and more efficient fault diagnosis, which is crucial to recover or at least mitigate potential losses in chemical industrial processes.