Uncertainty in Steady-State Diagnostics of a Current-Pressure Transducer: How Confident are We in Diagnosing Faults

Current-Pressure IP transducers are effective pressure regulators that can vary the output pressure depending on the supplied electrical current signal, and are commonly used in pneumatic actuators and valves. Faults in current-pressure transducers have a significant impact on the regulation mechanism, and therefore, it is important to perform diagnosis to identify such faults. However, there are different sources of uncertainty that significantly affect the diagnostics procedure, and therefore, it may not be possible to perform fault diagnosis and prognosis accurately, with complete confidence. These sources of uncertainty include natural variability, sensor errors gain, bias, noise, model uncertainty, etc. This paper presents a computational methodology to quantify the uncertainty and thereby estimate the confidence in the fault diagnosis of a current-pressure transducer. First, experiments are conducted to study the nominal and off-nominal behavior of the IP transducer however, sensor measurements are not fast enough to capture brief transient states that are indicative of wear, and hence, steady-state measurements are directly used for fault diagnosis. Second, the results of these experiments are used to train a Gaussian process model using machine learning principles. Finally, a Bayesian inference methodology is developed to quantify the uncertainty and assess the confidence in fault diagnosis by systematically accounting for the aforementioned sources of uncertainty.