(614c) Remaining Useful Life Prediction for a Reciprocating Compressor By a Data-Driven Model Based on an Operation Condition-Corrected Health Indicator

Reciprocating compressors are vital components in the refinery and chemical plants used for high discharge pressure. Reciprocating compressors usually cost several times to maintain than other critical rotating assets in the plants because they operate at high speed and harsh conditions (e.g., high pressure and corrosive gases), and they have many moving parts, leading to frequent failures. Bearings are one of the most critical parts of reciprocating compressors, and their failure can result in unscheduled downtime in process operation, leading to enormous economic loss. The prognostics and health management (PHM) of bearings are crucial to reducing production losses and avoiding machine damage by estimating the remaining useful life (RUL) of a bearing. An accurate estimation of a time-to-failure of a bearing can help find the optimum time for performance inspection and maintenance plans, improving the reliability of process operation. Most of the bearing's RUL estimation techniques can be roughly classified into model-based and data-driven methods [1]. The model-based method aims to build mathematical models to describe the physics of the system and failure mechanisms. This approach can be highly accurate if the failure mechanisms are well explained, and the physics of models remain consistent across systems. However, it is often difficult to employ the approach because failure mechanisms are generally too complex and stochastic to understand the underlying physical principle of damage. In contrast, the data-driven method aims to find the degradation mechanism from historical data by utilizing statistical/machine learning or deep learning. Therefore, this approach is well suited for complex systems. However, since the data-driven method requires historical data, its usage can be limited to the systems that have sufficient historical data for modeling training. It sounds easy, but in reality it is not often to obtain the whole failure data because the degraded bearings usually get replaced before reaching complete failure. This study aims to develop a new framework for the RUL estimation of bearings with minimal historical data. In the data-driven method, it is essential to extract the good health indicators (HIs) and determine the time to start prediction (TSP) for accurate prediction of bearing's RUL. In this study, we propose a new health indicator based on k-nearest neighbors (k-NN) distance of five time and frequency domain features extracted from vibration data collected from bearings. The features extracted under healthy conditions are used as the training set from which the k-nearest neighbors are selected. It is assumed that the k-NN distance of features extracted during the degradation period is greater than that of features extracted during the healthy period. Instead of using a feature such as root mean square (RMS) or kurtosis as a HI, the k-NN distance can take advantage of mutual information from multiple features extracted from time-domain and frequency-domain. Also, operation conditions (e.g., temperature, pressure, speed) can affect vibration signals collected off bearings. To develop a reliable HI, we propose to model their effects under healthy operations and apply the model to remove their effects during monitoring and RUL estimation. The 3-sigma approach is a well-known method to determine the TSP [2]. Therefore, in this study, the three-sigma approach is employed to choose the TSP based on the new HI (i.e., k-NN distance). Since bearings' degradation evolution is a cumulative process without self-recovery, once started, the bearing's degradation trend looks like exponential growth [2,3]. Therefore, the exponential model is used to predict the RUL of bearings based on the proposed operation-corrected kNN-based composite HI. The proposed method is validated on the real industrial reciprocating compressor dataset, and its prognostic performance is compared with existing prognostic methods. The proposed approach is promising for industrial applications.

[1] Heng, A., Zhang, S., Tan, A. C., & Mathew, J. (2009). Rotating machinery prognostics: State of the art, challenges and opportunities. Mechanical systems and signal processing, 23(3), 724-739.

[2] Li, N., Lei, Y., Lin, J., & Ding, S. X. (2015). An improved exponential model for predicting remaining useful life of rolling element bearings. IEEE Transactions on Industrial Electronics, 62(12), 7762-7773.

[3] Gebraeel, N., Lawley, M., Liu, R., & Parmeshwaran, V. (2004). Residual life predictions from vibration-based degradation signals: a neural network approach. IEEE Transactions on industrial electronics, 51(3), 694-700.