Probabilistic state–space modeling for robust condition monitoring of industrial equipment

This paper proposes a generic low-cost method for probabilistic condition monitoring of industrial equipment. A computationally efficient recursive data-driven model is constructed that correlates real-time machine measurements with a quantity or several quantities of interest (QoIs), including artificial metrics such as the Remaining Useful Lifetime (RUL). To that end, a probabilistic state–space model (PSSM) is identified, based on a fully instrumented measurement set obtained from a limited set of experiments that can only be obtained in a specialized testing environment. The dataset contains both cheaply available sensory information as well as prohibitively expensive, invasive or artificially constructed sensory signals. To identify a Maximum Likelihood Estimate of the PSSM, we rely on the Expectation–Maximization (EM) algorithm and Sequential Monte Carlo (SMC) estimation techniques. During operation, only the vital, non-intrusive and cheap sensors are used. The PSSM is then repurposed to reconstruct the costly sensory signals, realizing an effective and general purpose virtual sensor. Our methodology demonstrates the capacity to robustly estimate unmeasured physical variables in real-time and artificially constructed prognostic QoIs, such as the RUL, even when working with an incomplete measurement array. We validate the presented methodology for condition monitoring on the C-MAPSS dataset and a solenoid valve (SV) use case. The presented tool has similar predictive capabilities as compared with other state-of-the-art RUL prognostic methods and furthermore provides uncertainty quantification and contextual information with respect to equipment health.