Improved similarity-based modeling for the classification of rotating-machine failures

https://doi.org/10.1016/j.jfranklin.2017.07.038Get rights and content

Abstract

Similarity-based modeling (SBM) is a technique whereby the normal operation of a system is modeled in order to detect faults by analyzing their similarity to the normal system states. First proposed around two decades ago, SBM has been successfully used for fault detection in varied systems. In spite of this success, there is not much study performed in the literature regarding its design, that encompasses both similarity metrics and model training. This work aims at contributing with an in-depth study of SBM for fault detection considering these two design aspects. This is done in the context of proposing a novel system to identify rotating-machinery faults based on SBM, that is employed either as a standalone classifier or to generate features for a random forest classifier. New approaches for training the model and new similarity metrics are investigated. Experimental results are shown for the recently developed Machinery Fault Database (MaFaulDa) that has an extensive set of sequences and fault types, and for the Case Western Reserve University (CWRU) bearing database. Results for both databases indicate that the proposed techniques increase the generalization power of the similarity model and of the associated classifier, achieving accuracies of 98.5% on MaFaulDa and 98.9% on CWRU database.

Introduction

Maintenance of critical equipment to ensure high levels of reliability, availability, and performance is one of the major concerns on today’s industrial sector [1]. Unexpected failures can lead to substantial losses, either from the maintenance procedure itself or from the resulting production halts [2]. To achieve an effective and cost-efficient procedure, new maintenance strategies are being devised based on real-time and continuous monitoring, allowing one to detect and classify operational anomalies at an early stage, thus limiting additional system degradation [2]. Applications of such techniques include, for instance, flight paths [3], natural gas and nuclear power plants [4], [5], [6], [7], wind turbines [8], and bearing or rotating-machine faults [9], [10], [11], [12], [13], [14]. Among these equipments, rotating machines are some of the most important, being a key element used in a variety of applications, including airplanes, automobiles, power turbines, oil and gas refineries, and so on [12], [15].

There are many approaches for detecting faults in rotating machines. Most of them consist of extracting features from the vibration signal to assess the machine current condition, in a supervised or automatic manner. Different features are needed to extract useful information relevant to detect faults from the original sources over multiple conditions. These features can be classified considering their domain (time, spatial, time-frequency, frequency) or its computation method (e.g. transform coefficients or aggregated statistics) [16], [17], [18], [19].

An illustrative example is the approach in Yang et al. [20]. There, a system is presented which uses an adaptive resonance theory Kohonen neural network (ART-KNN) for fault diagnosis, having as inputs features derived from the discrete Wavelet transform coefficients. Unfortunately, the fault database used is not publicly available, making its comparison with other approaches impractical.

A methodology for detecting broken and half broken bars using spectral information over a FPGA is presented in [21]. This methodology is latter extended in [22] and in [23], adding the discrete wavelet transform (DWT) coefficients as features, and combining it with discrete frequency transform coefficients. These works also treat the detection of other faults and failures. The broken bar detection problem is also approached by the authors of [24] using motor current signature analysis and mathematical morphology.

The authors of [25] focus on the feature extraction procedure proposing a novel feature extraction scheme which utilizes the generalized S transform and 2D non-negative matrix factorization to detect possible faults. Three classifiers were used to assess the system: k-nearest neighbors (kNN), naive Bayes, and support vector machines (SVM), all achieving good results. A similar approach is presented in [26] using multiscale permutation entropy for feature extraction and an SVM classifier for fault diagnosis. The work of Rauber et. al. [18] also studies the effect of the features in the system performance. It tests multiple features of different types, such as complex envelope spectrum, statistical time- and frequency-domain parameters, as well as wavelet packet analysis, together with a feature selection algorithm. A fault classification database was used as testbed, and three different classifiers (kNN, feedforward artificial neural networks (ANN), and SVMs) were used during the assessment, achieving good performance.

This work proposes an automatic fault detector and classifier that uses similarity-based modeling (SBM) to identify rotating-machine failures such as imbalanced load, (horizontal or vertical) shaft misalignment, and bearing defects (in rolling elements or inner/outer tracks). The similarity model can be used either as an auxiliary model to generate features for the classifier (a random forest classifier in this case) or as a standalone classifier. In this context, new approaches for training the similarity model and new similarity metrics are investigated. Two databases were employed to evaluate the performance of the proposed techniques. The first one is the machinery fault database (MaFaulDa) [27], a relatively new, large database of problematic scenarios of rotating-machine operations [13], [14]. Performance evaluation on this database included continuous monitoring of six vibration sensors, one microphone, and one tachometer [14]. The second database is the Case Western Reserve University (CWRU) bearing database [28]. This database has become a standard reference in the bearing diagnostics field [19], [29] and is used as testbed for comparison between the proposed methodology against other algorithms [18], [25], [26], [30]. Results indicate that the proposed methodology is capable of correctly diagnosing the machine operating states, achieving an accuracy of 98.5% on the MaFaulDa dataset and 98.9% on the CWRU database.

This paper is organized as follows: Section 2 presents the original SBM technique [3], [5], [7], [9], [10], [11], devised for detecting unusual patterns in some system or machine operation. Section 3 describes the proposed modifications to the standard SBM technique that allow the detection and classification of different types of anomalous machine operations in an efficient and robust manner. Section 4 details the MaFaulDa database, used to design and evaluate the system’s performance and the CWRU database, used for comparison. The methodology of performance assessment is described in Section 5. This section also describes the designed system, including the preprocessing and validation procedures. Section 6 discusses the experimental results obtained during the processes of training and selection of the best model, as well the assessment results. Comparisons to other works are also included in this section. Finally, conclusions and discussions emphasizing the main contributions of this paper are provided in Section 7.

Section snippets

Similarity-based modeling (SBM)

The SBM is a simple and yet powerful nonparametric modeling technique that puts together an ensemble of previous state vectors in a single matrix D to represent the normal behavior of a given system, process, or machine. The SBM then evaluates the similarity of the current state vector with all vectors within D to assess the normality or not of the current system operation. This technique was first proposed in [4], and since then has been used in a variety of industrial applications, such as

Proposed SBM enhancements

This section presents the proposed enhancements to the SBM formulation, which include: a generalization of the SBM framework that allows it to operate in a multiclass (more than two classes) scenario; introduction of alternative similarity operations; and the development of a new strategy to compose the matrix D.

Databases used

Two databases were used to evaluate the contributions of this paper. The first one, named machinery fault database (MaFaulDa) [14], [27] is a comprehensive database including multiple types of faults covering different severities and rotation frequencies. This database was extensively used to validate the proposed approach and to search for the models with the best set of parameters based on their performance. The second database is the Case Western Reserve University bearing database [28], the

Experimental methodology

This section describes the experimental methodology employed to evaluate the modified SBM performance in detecting and classifying the ABVT’s faulty scenarios within the databases described in Section 4.

The proposed system follows a modular architecture similar to the ones described in [2], [36] for a condition-based maintenance system. It comprises three blocks (see Fig. 2): the preprocessing module converts the original data to a feature space which is more descriptive for the given

Experiment description

This subsection describes the experiments made during the validation procedure to select the best model for the proposed task considering all the following system variations:

  • Feature types: only spectral features, only statistical features, or both families of features, as discussed in Section 5.

  • Use of full SBM formulation (as given in Eq. (8)) or the AAKR scheme, which considers G=I in this same equation.

  • Choice of similarity function, as presented in Section 3.2, with distinct values of γ

Conclusion

This paper addressed the automatic fault diagnosis in rotating machines. The use of similarity based modeling (SBM) was investigated, either as a stand-alone classification method or in combination with an off-the-shelf classifier, in this case a random forest classifier. The system is evaluated in two databases. One of them is a comprehensive database with multiple faults referred to as MaFaulDa [27]. The other is the CWRU bearing database [28], that is the current standard database for

Acknowledgments

This research was supported by CNPq and Petrobras. The authors would like to thank Dr. Kenneth A. Loparo and the Case Western Reserve University Bearing Data Center for providing the CWRU dataset for this study, and Dr. Wade A. Smith for his help.

References (38)

  • J. Garvey et al.

    Validation of on-line monitoring techniques to nuclear plant data

    Nucl. Eng. Technol.

    (2006)
  • F.A. Tobar et al.

    Anomaly detection in power generation plants using similarity-based modeling and multivariate analysis

    Proceedings of the American Control Conference

    (2011)
  • GuoP. et al.

    Wind turbine gearbox condition monitoring with AAKR and moving window statistic methods

    Energies

    (2011)
  • S.W. Wegerich et al.

    Nonparametric modeling of vibration signal features for equipment health monitoring

    Proceedings of the IEEE Aerospace Conference

    (2003)
  • S.W. Wegerich

    Similarity-based modeling of time synchronous averaged vibration signals for machinery health monitoring

    Proceedings of the IEEE Aerospace Conference

    (2004)
  • S.W. Wegerich

    Similarity based modeling of vibration features for fault detection and identification

    Sens. Rev.

    (2005)
  • LiuJ. et al.

    An enhanced diagnostic scheme for bearing condition monitoring

    IEEE Trans. Instrum. Meas.

    (2010)
  • A.A. de Lima et al.

    On fault classification in rotating machines using Fourier domain features and neural networks

    Proceedings of the Latin American Symposium on Circuits and Systems

    (2013)
  • D. Pestana-Viana et al.

    The influence of feature vector on the classification of mechanical faults using neural networks

    Proceedings of the Latin American Symposium on Circuits and Systems

    (2016)
  • Cited by (54)

    View all citing articles on Scopus
    View full text