1 A Similarity Based Modeling Approach for Turbomachine Fault Prediction Weijian Tang 1 , Xiaomo Jiang 2 , Haixin Zhao 3 ,Qing Chen 4 , and Yunqing Gong 5 1,2,3,4 Dalian University of Technology, Dalian, Liaoning, 116024, China tangweijian@mail.dlut.edu.cn xiaomojiang2019@dlut.edu.cn zhxstc@mail.dlut.edu.cn chenqingdut@mail.dlut.edu.cn 5 Shenyang Blower Group Co., Ltd, Shenyang, Liaoning, 110022, China gongyunqing@shenguyun.com ABSTRACT Faults in the critical components of a turbomachine usually result in unplanned outage, leading to huge loss of properties and life. Condition monitoring becomes a promising tool to provide automatic early alerting of potential damage in critical components thus ensuring the system safety and reliability while lowering its maintenance cost. This is still a challenging hot topic due to the data imperfection and multivariate correlation, as well as the variation of faults and components in different turbomachines. In this paper, a condition monitoring method based on similarity-based model is proposed to solve these problems in fault prediction of large turbine machinery. Bayesian wavelet multi-scale reconstruction is proposed to address the potential noise in the sensed multivariate time historical data. The advanced signal processing balances the over-denoising and under-denoising of raw multivariate signals. An optimized auto-associative kernel regression (OAKR) approach is developed to represent the healthy status of the turbomachine system and further predict its responses under unknown status. The residual error between the estimated and measured values of the OAKR model will become the larger when the turbine machinery has an early fault. The statistical method of moving window is used to detect the change of mean square error of residuals over the time. When the mean square error exceeds a preset threshold, a fault mark will be given. A comparison study is conducted to demonstrate the effectiveness and feasibility of the proposed methodology by using the real-world data and events collected from a centrifugal compressor. Keywords: Bayesian wavelets, OAKR, turbomachine, fault prediction 1. INTRODUCTION Faults in a large-scale turbomachine such as gas turbine, steam turbine, or compressor usually result in unplanned outage and even huge loss of properties and life in the fields of power generation, oil & gas, and petrochemistry industries. As the quick development of high-performance computing capacity and artificial intelligence (AI) algorithms in the past decade, real-time condition monitoring system (CMS) has become an increasingly important tool in improving the safety, reliability and performance of a turbomachine, while reducing its unplanned breakdown, and lowering its maintenance costs. In the past decades a wide spectrum of data-driven predictive analytics methods like time series forecasting, machine learning and artificial neural network models have been developed to predict faults in condition monitoring of a turbomachine (Caselitz 2015, Bennouna 2005, Zaher 2009). These methods are generally composed of two main steps, 1) establishment of a high-fidelity predictive model to produce the system response and then 2) determination of a decision threshold to produce alarms when the system response is deviated too much from the actual measurement. Besides existing uncertainties in sensor data, both model establishment and threshold determination contain uncertainties, which would impact the fault prediction accuracy of a turbomachine to some degree. Therefore, it has become of key importance to accurately predict the fault for condition monitoring of a turbomachine considering various uncertainties. Recently the auto-associative kernel regression (AAKR) method has been developed as a similarity-based model (SBM) for condition monitoring and fault alerting in large- scale turbomachines (Garvey 2007, Di Maio 2013, Fei 2015, Sairam 2016, Yu 2017, Guo 2011, Brandsæter 2017, Qian 2018, Baraldi 2015). This approach utilizes multivariate historical data collected at normal conditions to establish a system identification model representing the Weijian Tang et al. This is an open-access article distributed under the terms of the Creative Commons Attribution 3.0 United States License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Proceedings of the 6th European Conference of the Prognostics and Health Management Society 2021 - ISBN – 978-1-936263-34-9 Page 398