MonitorX

The aim of this case was to develop algorithms and models for fault detection and prediction of remaining useful life (RUL) of rotating equipment based on high frequency sensor data (kHz resolution and larger) from vibration sensors (accelerometers), acoustic emission sensors and microphones. Such high frequency data is normally not available in power plants. Thus, the model development was carried out with publicly available test data from roller bearings used in laboratory tests.

The case resulted in a set of models for analysis of the vibration data for the purpose of RUL estimation and anomaly detection and classification. The modelling of RUL and anomalies does not use the monitoring data (i.e. the vibration data) directly, but the data is pre-processed in the feature extraction modelling step in order to calculate a parameter (feature, health indicator) that can be tracked and trended over time. Health indicators can be calculated in time and frequency domain, or combination of both (such as continuous wavelet transform - CWT). Mean of the signal, kurtosis, maximum amplitude, root mean square (RMS) and square mean root (SMR) are examples of features.

RUL estimation is based on trending the health indicator, as illustrated in the figure. In the beginning, illustrated in the left diagram for the (current) time 426 ¹ , the uncertainty regarding further development of the indicator is quite large. This is indicated with the blue lines representing different possible trajectories for the further health indicator development. The red line is the mean (expected) development, representing the mean lifetime of the bearing, given the observations until current time. The uncertainty decreases when more data become available, as illustrated in the middle and right diagram. The predicted lifetime will be updated step by step, as more data becomes available. Consequently, the RUL estimate, which is the difference between current time and predicted lifetime, can also be updated step by step.

Lifetime prediction and RUL estimation, and updating step by step over time, illustrated for three examples representing different times in early life (left, current time = 426), middle of the life (middle, current time = 1126) and close to end of life (right, current time = 1226). Courtesy of J. Yuan [1].

The approach of anomaly classification (i.e. the classification of anomalies in different states from slight to large/significant) allows for monitoring the development of degradation from a condition as good as new (no degradation, no anomalies, normal behaviour) to a condition with major degradation (large anomalies). An example is illustrated in the figure below, where the evolution of the condition over the bearing lifetime is illustrated from condition state (working condition) 1 to condition state 7.

Anomaly classification (classification in condition states 1 to 7) and evolution of condition over bearing life [2].

¹ The time may be measured e.g. in hours, or operational hours, or as data/indicator count, meaning that 426 represents the 426th time that the health indicator that is calculated. If, e.g. the health indicator is calculated every 6th hours of operation, 426 corresponds to 2556 hours of operation time.

[1] J. Yuan and K. Wang, "Twin Exponential Degradation Model for Online Remaining Useful Life Prediction," Paper draft, May-2019.
[2] J. Yuan, K. Wang, and T. M. Welte, "Deep Learning Approach to Multiple Features Sequence Analysis in Predictive Maintenance," in Advanced Manufacturing and Automation VII, Changshu, China, 2017.