A novel anomaly detection method for gas turbines using weight agnostic neural network search

Anomaly detection of aircraft gas turbines is a small sample size classification problem. Under the condition of small observation samples, a significant challenge is to determine the scale of the neural network. The inappropriate scale might cause test accuracy to be low due to either overfitting or underfitting problems. In this paper, a novel anomaly detection method based on weight agnostic neural network search is investigated for gas turbines. On one hand, to prevent underfitting, the scale of the network evolves from small to large using inserting node, adding connection, or changing activation until it can reflect the complex relationship between the gas turbine health state and the gas path performance parameters. On the other hand, the number of connections in the network is considered as a term in the objective function, which can avoid overfitting in network evolution. Moreover, this paper proposes a new threshold selection method based on time series tendency turning point, which is effective in distinguishing normal patterns from abnormal patterns under the circumstance of unbalanced data. The effectiveness of the proposed model is verified using real-life monitoring data from CFM56-7B gas turbines.