Health condition of rolling element bearings has a significant impact on rotational machinery systems. The failure of bearings may cause a total breakdown of machinery and may even lead to a fatal accident. Therefore, the early bearing fault detection and identification arises as a critical mission in the frames of predictive maintenance, receiving the last years extensive attention. Nowadays advances signal processing techniques are combined with high level machine learning approaches focusing towards automatic fault diagnosis. On the other hand in real industrial conditions, the available data belong mainly in the healthy operating condition and the faulty datasets are rather limited. Therefore the standard machine learning approaches, which are based on the training of a number of classes cannot be realistically used. In order to overcome this limitation semi supervised/anomaly detection techniques have been proposed which are based on the training of the methodology exclusively on healthy data. Before the application of the anomaly detection methodology, a diagnostic indicator, which accurately tracked the degradation of the system, should be extracted. A plethora of diagnostic indicators have been proposed including time domain indicators (RMS, kurtosis, skewness etc.) as well as frequency domain indicator (amplitudes at specific fault characteristic frequencies: BPFI, BPFO, BSF etc.). Signal processing and spectral analysis methods such as the Squared Envelope Spectrum (SES) are also used to extract features which can detect bearing faults accurately[1]. The anomaly detection focuses on setting a threshold at the indicators in order to separate the anomalous samples from the normal ones. Support Vector Machine (SVM) is one of the most popular semi-supervised methods since it can effectively isolate anomalies by fitting a hyperplane and projecting features into a higher dimension [2]. On the other hand, the minimization of the false alarms and misdetections for bearing anomaly detection is still a challenging task.

In this paper the SVDD anomaly detection techniques is combined with advanced diagnostic indicators which are based on cyclostationarity. Cyclic Spectral Correlation (CSC) and Cyclic Spectral Coherence (CSCoh) have been proved as powerful tools in signal cyclostationary analysis [3]. They represent the potential fault modulation information into frequency-frequency domain bivariable maps. The integration over the cyclic frequency, leads to the estimation of the Enhanced Envelope Spectrum (EES), which demonstrates the modulation frequencies and their harmonics. Due to the periodic mechanism of the bearing faults' impacts, the EES can provide a clearer detection of bearing faults comparing to SES. Therefore, the sum of the amplitudes of the harmonics of the bearing characteristic fault frequencies of EES are extracted as diagnostic indicators. Meanwhile, the semi-supervised Support Vector Data Description (SVDD) is used as a detector. Instead of a hyperplane, SVDD fits a hypersphere for fault isolation and can be extended in the nonlinear case with a kernel trick [4]. In this paper, the EES- based diagnostic indicators and the SVDD detector are combined and the methodology is tested and evaluated on experimental data for bearing fault detection. The results demonstrate the efficacy of the method presenting high detection rate with low false alarm and misdetection rate.

References

[1]. Mauricio, A., Qi, J., & Gryllias, K. (2019). Vibration-Based Condition Monitoring of Wind Turbine Gearboxes Based on Cyclostationary Analysis. Journal of Engineering for Gas Turbines and Power, 141(3), 031026.

[2]. Widodo, A., & Yang, B. S. (2007). Support vector machine in machine condition monitoring and fault diagnosis. Mechanical systems and signal processing, 21(6), 2560-2574.

[3]. Antoni, J. (2007). Cyclic spectral analysis of rolling-element bearing signals: Facts and fictions. Journal of Sound and vibration, 304(3-5), 497-529.

[4]. Tax, D. M., & Duin, R. P. (2004). Support vector data description. Machine learning, 54(1), 45-66.