Fault detection based on signal reconstruction with Auto-Associative Extreme Learning Machines

Early fault detection of engineering systems allows early warnings of anomalies and provides time to initiate proactive mitigation actions before the anomaly has developed to a problem that either requires extensive maintenance or affects the productivity of the system. In this paper, a new fault detection method using signal reconstruction based on Auto-Associative Extreme Learning Machines (AAELM) is proposed. AAELM are applied for fault detection on an artificially generated dataset to test the performance of the algorithm under controlled conditions and a real case study based on condition monitoring data from a combined-cycle power plant compressor. The performance of AAELM is compared to that of two other commonly used signal reconstruction methods: Auto-Associative Kernel Regression (AAKR) and Principal Component Analysis (PCA). The results from the two case studies demonstrate that AAELM achieve a smaller reconstruction error, shorter detection delay, lower spillover and a higher distinguishability compared to AAKR and PCA on the evaluated datasets. The obtained results are generalized to elaborate guidelines for industrial users for selecting suitable signal reconstruction algorithms based on their specific requirements and boundary conditions.