A Health Monitoring Method for Wind Power Generators with Hidden Markov and Probabilistic Principal Components Analysis Models

Riku Sasaki, Naoya Takeishi, Takehisa Yairi, Koichi Hori, Kazunari Ide, Hiroyoshi KuboPHM 2017 · DOI

TL;DR

This paper introduces a novel data-driven health monitoring system for wind power generators, combining Hidden Markov Models (HMMs) and Probabilistic Principal Component Analysis (PPCA) into a unified HM-PPCA model. This approach effectively captures the complex, multi-state, and correlated nature of sensor data to detect faults early, even before they lead to malfunction.

What Problem Does This Paper Address?

Large-scale engineering systems like wind power generators are complex, with numerous interconnected sensors generating highly correlated, multi-dimensional time-series data. Manually inspecting this data for anomalies or impending faults is often impractical. Existing fault detection methods may struggle with the non-stationary nature of this data (due to operational state changes like starting/stopping) and the fact that faulty behaviors are often not well-understood beforehand. This paper aims to address this by developing a robust, data-driven method that can learn "normal" operational patterns and flag deviations indicative of faults.

Methods

The core of the proposed method is the Hidden Markov and Probabilistic Principal Component Analysis (HM-PPCA) model. This is a hybrid approach that leverages the strengths of two established techniques:

  • Probabilistic Principal Component Analysis (PPCA): Used to model the correlations among multiple sensor variables and reduce their dimensionality by identifying underlying latent variables. It can also handle the clustering of data.
  • Hidden Markov Models (HMMs): Used to capture the temporal dependencies and state transitions within the data. This is crucial for systems that operate in distinct states (e.g., running, stopped, transitioning).

In HM-PPCA, the latent variables from PPCA are themselves modeled by a Markovian process, effectively creating a probabilistic framework that accounts for both spatial correlations (via PPCA) and temporal dynamics (via HMM). The model is trained on data without known faults. Fault detection is then performed by comparing the reconstruction of new data using the learned model against the original sensor readings; large reconstruction errors signal potential faults. The authors use an Expectation-Maximization (EM) algorithm for parameter estimation and the Viterbi algorithm for inferring the most likely sequence of hidden states.

Datasets

The paper evaluates the proposed HM-PPCA method on two types of datasets:

  • Synthetic Data: Created to imitate real-world wind turbine data, including simulated running, stopped, and transition states, with added noise and artificial faults (gradual increases in sensor values). This dataset was used to demonstrate the method's ability to detect simulated faults.
  • Real-world Wind Turbine Data: Obtained from strain sensors attached to the turbine blades of a wind power generator. This dataset comprises 12-dimensional time-series data and was used to validate the method on actual operational data, including a scenario where a fault occurred on a turbine blade.

Key Insights and Ideas

  • Synergistic Power of HM-PPCA: Combining PPCA's ability to model correlated variables with HMM's temporal state modeling provides a powerful framework for understanding complex, dynamic systems like wind turbines.
  • Early Fault Detection via Reconstruction Errors: The core idea of detecting faults by monitoring how well the learned "normal" model can reconstruct incoming data is effective. Significant deviations in reconstruction error directly correlate with the occurrence of faults.
  • Handling Non-Stationarity: The HMM component is critical for dealing with the inherent state changes in wind turbine operation, allowing the model to adapt to different operational modes without misinterpreting them as faults.
  • Unsupervised Approach: The method's reliance on an unsupervised approach is a significant advantage, as it doesn't require extensive pre-labeled fault data, which is often scarce and difficult to obtain.

Key Results

The experiments demonstrated the effectiveness of the HM-PPCA method:

  • Synthetic Data: The reconstruction errors on synthetic data showed a clear and significant increase when artificial faults were introduced, indicating successful fault detection.
  • Wind Turbine Data: Similar to the synthetic data, the reconstruction errors on real-world wind turbine data also exhibited a sharp rise in the latter part of the test data, coinciding with the occurrence of a fault on the turbine blade. This confirms the method's applicability to real-world scenarios.
  • State Transition Analysis: The estimated state transition matrix for the real-world data revealed that transitions between the same states were overwhelming, with rare transitions between running and stopped states, providing insights into the system's typical operational dynamics.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Online Parameter Updates: The current method estimates parameters offline. Future work could focus on developing mechanisms for online updates of parameters to adapt to gradual system degradation over time.
  • Scalability to More Sensors: While the method was tested with 12-dimensional data, applying it to systems with an even larger number of sensors presents a challenge that warrants further investigation.
  • Gradual Degradation: The current fault detection primarily focuses on distinct fault events. Investigating how to detect more subtle, gradual degradation of system components would be a valuable extension.

Summary updated

Abstract

In this work, we propose a data-driven health monitoring method for wind power generators, which learns an empirical model from the time-series sensor data and detects irregularities or faults in the turbines and blades. Our main objective is to predict any symptoms of faults as early as possible before the generators fall into malfunction. The data obtained from the wind power generators are strongly correlated multidimensional time-series with multiple states. In this study, we take such features into account and develop the probabilistic model for them, namely, hidden Markov and probabilistic principal component analysis. Once the model is learned with the data that contain no faulty events, it can be used to detect faults in new data by comparing the original sensor values and reconstructed ones. In this research, we apply this method to synthetic data and real-world wind turbine data and show the results of experiments to confirm the availability of the proposed method.
BibTeX
@inproceedings{sasaki2017health,
  author = {Riku Sasaki and Naoya Takeishi and Takehisa Yairi and Koichi Hori and Kazunari Ide and Hiroyoshi Kubo},
  title = {{A Health Monitoring Method for Wind Power Generators with Hidden Markov and Probabilistic Principal Components Analysis Models}},
  booktitle = {PHM},
  year = {2017},
  doi = {10.36001/phmap.2017.v1i1.1766},
}
Figure