A novel anomaly detection method for turbomachinery based on back-constrained Gaussian process latent variable model and adaptive threshold

Accurate anomaly detection (AD) from sensor data of conventional turbomachinery is of great significance to the prognostics and health management of equipment. Effectively detecting abnormal events during the equipment operation brings benefits for health management tasks such as fault warning and fault diagnosis. Existing time series anomaly detection methods however have shortcomings such as fixed threshold settings. Meanwhile, they rarely consider the highly nonlinear characteristics and abnormal event nature of turbomachinery data. To this end, this paper proposes a new AD method for turbomachinery based on back-constrained Gaussian process latent variable model and adaptive threshold strategy. The comparison results against two representative AD methods on the real-world turbomachinery data show that: (1) the proposed method showcases superiority in terms of detection effect and stability, and the model upper limit and stability can be improved by up to 2% in terms of the best-Fc1 and mean-Fc1 criteria; and (2) the proposed model is more efficient for AD with small and moderate amount of data, with the average training time cost reduced by 80%. The above experiments demonstrate the necessity of considering the nonlinear characteristics and abnormal events of data and help improve the quality of operation and maintenance of turbomachinery.