Multi-method fault detection considering uncertainty through MC dropout for enhanced voting

Creating a robust process monitoring system is critical for energy systems to ensure satisfactory performance in terms of quality of service and cost. Most frameworks for fault detection have focused on (sub)systems with data that leverage accompanying fault labels for supervised training. However, this approach cannot be leveraged in systems with fewer fault occurrences and/or unknown fault patterns. Creating a robust framework that can detect faults in energy systems with residual methods trained on data with verified normal operation is necessary. This paper addresses this challenge by proposing a framework consisting of an ensemble of neural networks that predict each process variable and vote on a final value based on uncertainty calculations enabled by Monte Carlo dropout. Thereafter, it detects anomalies with residuals informed by the variables’ uncertainty and ranks them based on fault probability by employing the Chernoff bound. Finally, the anomalies are categorized based on their influence on identified key performance indicators. This enables maintenance crews to prioritize their time to focus on faults with a higher probability of being a fault and impacting operational key performance indicators. The proposed methodology was applied to the Tennessee Eastman process benchmark dataset to facilitate direct comparison with other fault detection approaches. It demonstrated a satisfactory trade-off between minimizing false alarm rates and ensuring prompt detection of known faults, outperforming several existing methods. Additionally, the methodology was tested on a case study within the building domain, where it achieved effective fault categorization and demonstrated robust performance.