Rotor Fault Detection of a Hydrogenerator Using AI

Key Information on Health Monitoring and Fault Diagnosis of Large Hydrogenerators

Large hydrogenerators are synchronous salient-pole generators used in electricity production at hydropower plants. As these machines are expensive and vital equipment for the industry, condition-based maintenance techniques must be applied to extend their healthy lifespan, as multiphysical degradation phenomena can occur because of various types of faults, resulting in higher operating costs ‎[1] – [2]. Early and reliable anomaly detection with low false alarm rates leads to higher efficiency.

For decades, several physical entities have been used to this end, for several types of rotating machines, drawing on different forms of data acquisition ‎[3]. Recent developments in this field have mainly been based on advanced tools such as artificial intelligence techniques.

Vibration Signals in Health Monitoring and Diagnosis: Challenges and Solutions

The importance of vibration signals in monitoring and fault diagnosis is clearly indicated in literature [1] – [3]. Also, there is an absolute need for a real-time health monitoring system that detects faults at early stages without triggering false alarms. However, obtaining labelled data on actual faulty conditions through in situ measurements is challenging due to the inability of shutting down power plants. As a result, research on monitoring and diagnosis of faults in large hydrogenerators using vibration signals has been limited, highlighting the challenges in this field.

Therefore, as shown in Figure 1 real healthy vibration signals collected in situ on a hydrogenerator (74 MVA with 76 poles) were used and divided into three sets (sets 1, 2 and 3). To obtain the fault signature, which represents the rotor inter-turn short-circuit (RITSC) at various levels of severity, a numerical model of the hydrogenerator, using ANSYS® Workbench, was applied: the 2D electromagnetic model results were used as input in the 3D mechanical model. Then, it was injected into set 3 to obtain a dataset containing healthy and faulty signals (set 4).

The resulting database was used in a deep learning technique, the variational autoencoder (VAE), as shown on the up-hand side of Figure 1. It was trained and validated based on the healthy signals of sets 1 and 2, respectively. Set 4 was used to assess the model’s capacity for statistical monitoring and classifying health status.

Classification of Different Health Degradation States in a User-friendly Space

Creating a user-friendly clustering of various degradation states within a system is a crucial technique in facilitating fault diagnosis. Figure 2 displays the latent space obtained from the signals of set 4 after fault injection. The model has a good structure and well-defined clusters, where each color represents a degree of degradation, showing its superior clustering in the R³ User-friendly space. Moreover, the model is sensitive to the varying degrees of faults—healthy and faulty clusters becoming increasingly disjointed as fault amplitude increases—and an upward degradation trend is observed. One can therefore assume that the present model can be considered an effective visual tool for clustering and can potentially be used for fault diagnosis.

Figure 1 Global methodology diagram for health monitoring and diagnosis

Figure 2 Latent space clustering visualization

Early Fault Detection Based on the VAE Model and Vibration Signals

The reconstruction mean square error (MSE) of each signal in set 4 with an increase in severity degree of the fault is compared to the results obtained with the classical root mean square monitoring technique, as presented in Figure 3, where only the healthy and 1RITSC cases are shown. The green and pink horizontal bars represent the fault threshold indicator (FTI) of the VAE model and the RMS technique respectively, based on the signals from set 2.

The error exceeds the FTI in the MSE , in the early stages of the fault (between 0.25% and 0.75%). However, using the RMS technique, it is hard to note a significant variation between the healthy case and the faulty case below 3%, and an alarm is triggered between 3% and 6%, depending on the signal used. However, in practice, in both methods, the fault is detected before 1RTISC (6%), but using the VAE, it is detected earlier than in the RMS method, as can be seen in Figure 3. Therefore, the suggested model represents a powerful tool in detecting faults at earlier stages, as it is more sensitive, proving its suitability for health monitoring.

Figure 3 RMS and MSE variation with the degree of severity in the fault of the six signals in set 4

Potential of the VAE for Health Monitoring and Fault Diagnosis in Large Hydrogenerators

This paper demonstrated the capability of the VAE for early detection of RITSC. The model used was trained and validated on healthy signals collected in situ from a real hydrogenerator. Because insufficient faulty labelled data was available, a set of faulty signals were created by injecting the fault, based on its frequency pattern derived from numerical simulations on ANSYS Workbench®, into a set of healthy signals with varying degrees of RITSC fault severity.

Our findings prove that the proposed approach provides better results in early fault monitoring than the traditional RMS monitoring technique. Also, the model can represent the different states of degradation (healthy, 1RITSC, 2RITSC, and 3IRTSC) in a user-friendly 3D latent space, where each state is represented by a cluster that can be easily identified. The model proved suitable for use in fault diagnosis.

Additional Information

For more information, please read the following research article:

Ibrahim et al., “Non-invasive Detection of Rotor Inter-turn Short Circuit of a Hydrogenerator Using AI-Based Variational Autoencoder,” in IEEE Transactions on Industry Applications, doi: 10.1109/TIA.2023.3281311.