Performance evaluation of Jensen–Shannon divergence-based incipient fault diagnosis: Theoretical proofs and validations

Early fault detection and estimation in nowadays’ complex systems are mandatory to ensure prognosis operations, good conditional maintenance, and safety. Kullback–Leibler Divergence (KLD) and Jensen–Shannon Divergence (JSD) are two measures characterized by high sensitivity for the evaluation of minor differences between probability distributions. KLD has been widely used and shown good detection capability for incipient fault diagnosis, but is limited by environmental noise. Recently, new fault diagnosis schemes for incipient fault detection based on JSD were proposed to cope with the noise influence. Nevertheless, no theoretical study has proved this efficiency. In this paper, we propose to derive the theoretical proofs of performances either for fault detection and estimation. Afterward, this is validated through simulated and experimental data for crack diagnosis. We give the limits and prove that for incipient faults in high noise levels, JSD has a great benefit without major constraints.