Incipient fault detection and estimation based on Jensen-Shannon divergence in a data-driven approach

摘要

Most data-driven diagnosis methods that are designed to detect faults, rely on measuring the mean and variation shifts. However, for incipient fault detection, these statistical criteria are slightly varying and are difficult to be accurately evaluated to reach good performances. Indeed, such faults are more likely to induce slight changes on the probability distribution rather than particular parametric changes. Therefore, the Jensen-Shannon Divergence(JSD), characterized by high sensitivity in measuring minor changes between probability distributions, is proposed in this paper. Its efficiency for detection and estimation is theoretically studied and validated by simulated data considering an auto-regressive (AR) system designing a multivariate data-driven process. The superior detection performances are demonstrated and compared with other more traditional statistical tests such as the Hotelling's T~2 and the Squared Prediction Error (SPE) in the Principal Component Analysis (PCA) framework. Minor crack detection based on eddy-currents testing (ECT) experimental data are evaluated to highlight the performances of our proposal. The results show that JSD can detect minor cracks (0.01 mm~2 to 0.04mm~2) which were not possible when using the baseline impedance signal measurement. For the fault severity estimation, the accuracy of the theoretical model derived for Gaussian distributed signals is shown with an AR system. The maximum relative estimation error obtained in the worst faults severity conditions is smaller than 2.75% when the Signal to Noise Ratio (SNR) is larger than 25dB and smaller than 2.15% when the Fault to Noise Ratio (FNR) is larger than -21 dB. Application for the fault severity estimation on the ECT data validates the effectiveness of this fault estimation model.