Anomaly detection of gas turbines based on normal pattern extraction

The integration of large-scale renewable energy forces the operating conditions of gas turbines to change more frequently to maintain the stability of electricity grid. This makes gas turbines malfunction more easily. In this case, anomaly detection is increasingly important to the safety and reliability of gas turbines. Current research mainly focuses on the case of abundant fault instances. However, fault data are rare or even unavailable in practice, especially for new gas turbines that have just operated for a short time. Thus, this paper proposes a novel anomaly detection method using only normal data based on nonlinear autoregressive with exogenous inputs network and prior knowledge fusion. This paper proposes the concept of gas turbine normal pattern extraction for the first time, extracts the unchanged feature of normal pattern from historical normal data and detects its changes for anomaly detection. Experiment on actual data shows the proposed method obtains 99.96% detection accuracy for fault data and 98.67% detection accuracy for normal data. Meanwhile, comparison between nonlinear autoregressive with exogenous inputs network and other methods including single-hidden layer feedforward network, Elman network and extreme learning machine verifies its superiorities in characterizing dynamic behaviors and normal pattern of gas turbines. Furthermore, comparison between the proposed method and 17 measurable parameter combinations shows it is the optimal parameter combination for anomaly detection. Comparison with three common purely data-driven anomaly detection methods including one-class support vector machine, isolation forest and principle component analysis further verifies its superiorities. This also indicates the superiorities of data-driven methods and prior knowledge fusion to some extent.