Anomaly detection for compressor systems under variable operating conditions

The operating conditions of compressor systems used in shale gas fields are variable. To enhance the performance of anomaly detection methods, it is crucial to capture the running state inside compressor and set an adaptive threshold. This paper proposes an anomaly detection framework for compressor systems under variable operating conditions, using multi-source variables, based on batch-normalized variational autoencoders (VAE) and optimized extreme value theory (EVT). Firstly, the multi-source input variables are obtained by combining secondary variables constructed based on thermodynamic principles and primary variables from the programmable logic controller (PLC) system. Then, the anomaly scores are obtained based on the batch-normalized VAE. Finally, an adaptive threshold is established based on the optimized EVT for anomaly detection. The method is validated using two real datasets, since all of the performance metrics on both datasets exceeded 96 %, which indicates that the proposed method can accurately identify anomalies in compressor systems under variable operating conditions. In addition, the effectiveness of multi-source data and adaptive EVT-based threshold are also discussed. The results show that multi-source data can more directly reflect the working state inside compressors. And the EVT-based threshold can accurately follow the fluctuation of anomaly scores, to provide dynamic criteria for the model.