A Hybrid LSSVR/HMM-Based Prognostic Approach

Zhijuan Liu, Qing Li, Xianhui Liu, Chundi MuSensors 2013 · DOI

TL;DR

This paper introduces a novel hybrid approach for predicting system failures and estimating remaining useful life (RUL) by combining Least Squares Support Vector Regression (LSSVR) for feature trend prediction and Hidden Markov Models (HMMs) for health state classification. The method demonstrates effectiveness in forecasting faults well in advance and accurately predicting RUL using bearing vibration data.

What Problem Does This Paper Address?

In critical systems like aircraft or industrial machinery, predicting future failures (prognostics) and estimating the remaining useful life (RUL) is crucial for efficient maintenance, safety, and cost reduction. While Condition-Based Maintenance (CBM) is gaining traction, accurate prognostics remain a challenging research area due to subtle fault symptoms and system complexities. Existing methods often struggle with accurately assessing health states over time and predicting future degradation trends. This paper aims to bridge this gap by proposing a more robust and integrated approach to prognostics.

Methods

The core of this work is a hybrid LSSVR/HMM framework. The process involves:

  • Feature Extraction: Vibration signals are segmented into small windows, and linear predictive (LP) coefficients are extracted from each window to form feature vectors. This step aims to capture the essential characteristics of the signal while reducing noise.
  • HMM Training (Off-line): For each identified health state (e.g., normal, different fault severities), a Hidden Markov Model (HMM) is trained using the extracted feature data corresponding to that state. HMMs are well-suited for modeling sequential data and classifying underlying states that are not directly observable.
  • Feature Trend Prediction (On-line): Least Squares Support Vector Regression (LSSVR), an extension of Support Vector Regression (SVR), is employed to predict future trends of the extracted features. LSSVR is chosen for its efficiency in handling non-linear time-series data. The algorithm is enhanced with online data updating (adding new data, deleting old data) to maintain model relevance.
  • Health State Classification and RUL Prediction (On-line): The predicted future features from LSSVR are fed into the trained HMMs. Log-likelihood probabilities are calculated for each HMM, and the HMM with the highest probability indicates the predicted future health state. By tracking the transition to the final "failure" state, the Remaining Useful Life (RUL) is estimated.

Datasets

The proposed approach was evaluated using experimental bearing vibration data from the Case Western Reserve University Bearing Data Center. This dataset includes vibration signals from bearings with different fault diameters (0.007, 0.014, and 0.021 inches) and normal operating conditions. The data was sampled at 12 kHz. The authors constructed a lifetime data sequence by combining data from these different health states for their simulations.

Key Insights and Ideas

  • Synergistic Combination: The paper highlights the power of combining a regression technique (LSSVR) for trend prediction with a classification technique (HMM) for state assessment. LSSVR excels at forecasting feature behavior, while HMMs provide a robust framework for interpreting these features within distinct health states.
  • Online Learning for Adaptability: The inclusion of an online data update mechanism in LSSVR (adding new data, deleting old data) is crucial for adapting to evolving system conditions and maintaining prediction accuracy over time.
  • Early Fault Detection: The results suggest that the LSSVR/HMM approach can forecast faults significantly earlier than they might otherwise be detected, providing valuable lead time for maintenance planning.
  • Distinguishable Feature Spaces: The study demonstrates that features extracted from vibration signals differ across various health states, validating their utility for distinguishing between normal operation and different levels of fault severity.

Key Results

  • The LSSVR/HMM approach successfully predicted future fault features that were consistent with actual trends.
  • Fault diagnosis based on HMMs was shown to be efficient, correctly identifying fault-free data with the normal state HMM and faulty data with the corresponding fault state HMM.
  • The RUL prediction results showed that the predicted RUL was generally less than the actual RUL, indicating that the method can forecast faults well in advance.
  • Prediction errors for RUL tended to decrease over time in the simulation, approaching one time unit, though the authors note real-world data might yield less precise results.

Limitations and Future Work

  • Data Availability: The authors note that collecting lifetime data is challenging, and the bearing data center did not provide complete lifetime data. They constructed a lifetime sequence from available data, which might not perfectly replicate real-world scenarios.
  • Simulation vs. Reality: While simulation results were promising, the authors acknowledge that real-world prediction results might not be as precise due to the perfect nature of simulated data and its shorter duration compared to actual lifetime data.
  • Future Work: The authors suggest future work should focus on improving prediction accuracy and expanding the application of the LSSVR/HMM approach to other types of systems and fault scenarios. They also mention the potential for optimizing LSSVR parameters using methods like Particle Swarm Optimization (PSO) to enhance training speed and performance.

Summary updated

Abstract

n a health management system, prognostics, which is an engineering discipline that predicts a system's future health, is an important aspect yet there is currently limited research in this field. In this paper, a hybrid approach for prognostics is proposed. The approach combines the least squares support vector regression (LSSVR) with the hidden Markov model (HMM). Features extracted from sensor signals are used to train HMMs, which represent different health levels. A LSSVR algorithm is used to predict the feature trends. The LSSVR training and prediction algorithms are modified by adding new data and deleting old data and the probabilities of the predicted features for each HMM are calculated based on forward or backward algorithms. Based on these probabilities, one can determine a system's future health state and estimate the remaining useful life (RUL). To evaluate the proposed approach, a test was carried out using bearing vibration signals. Simulation results show that the LSSVR/HMM approach can forecast faults long before they occur and can predict the RUL. Therefore, the LSSVR/HMM approach is very promising in the field of prognostics.
BibTeX
@inproceedings{liu2013hybrid,
  author = {Zhijuan Liu and Qing Li and Xianhui Liu and Chundi Mu},
  title = {{A Hybrid LSSVR/HMM-Based Prognostic Approach}},
  booktitle = {Sensors},
  year = {2013},
  doi = {10.3390/s130505542},
}
Figure