Improved Process Diagnosis Using Fault Contribution Plots from Sparse Autoencoders

Development of model-based fault diagnosis methods is a challenge when industrial systems are large and exhibit complex process behavior. Latent projection (LP), a statistical method that extract features of data via dimensionality reduction, is an alternative approach to diagnosis as it can be formulated to not rely on process knowledge. However, LP methods may perform poorly at identifying abnormal process variables due a “fault smearing” effect - variables unaffected by a fault are unintentionally characterized as being abnormal. The effect occurs because data compression permits faulty and non-faulty variables to interact. This paper presents an autoencoder (AE), a nonlinear LP method based on neural networks, as a monitoring
method of a simulated nonlinear triple tank process (TTP). Simulated process data was used to train the AE to generate a monitoring statistic representing the condition of the TTP. Sparsity was introduced in the AE to reduce variable interactivity. The AE’s ability to detect a fault was demonstrated. The individual contributions of process variables to the AE’s monitoring statistic were analyzed to reveal the process variables that were no longer consistent with normal operating conditions. The key result in this study was that sparsity reduced fault smearing onto unaffected variables and increased the contributions of actual faulty variables.