Nihon Kikai Gakkai ronbunshu (May 2020)
Lifetime modelling for mechanical equipment by utilizing time-series data based-on the damage-based survival analysis (A trial for applying the model to the equipment in a chemical plant)
Abstract
A variety of health assessment technologies for mechanical systems utilizing time-series sensor data has been developed and are recently being applied to the maintenance work as a solution to the predictive maintenance. A majority of these technologies are the condition-based way which premises the existence of condition monitoring sensors such as accelerometers for the vibration monitoring of rotating mechanical elements. In the present study, authors suggested a load history-based methodology for identifying a descriptive and stochastic model of the useful life of mechanical systems to realize the predictive maintenance under constraints of sensing conditions. The methodology (Damage-based survival analysis, DbSA) was based on a parametric survival analysis by the maximum likelihood estimation assuming the Weibull distribution of the useful life. A random variable of the probability distribution was converted from the elapsed time to cumulative value of a function of time-series sensor data and parameters of this function were optimized to minimize the dispersion of the probability distribution by the particle swarm optimization. DbSA was applied to a historical record of a clogging problem in a strainer in a chemical plant and its time-series process data to demonstrate the usefulness. An identified damage-based lifetime model exhibited less than 50% smaller dispersion (coefficient of variance) compared to the timed-based probability distribution. In addition, an identified function composed of the process data implied an effect of the impurity generation to the clogging problem. If the identified model was applied to a dataset which was not used to the model identification, it was indicated that 3 of 8 cloggings were occurred when the damage-based failure probability was more than 50% although the time-based probability did not reach to this level at any time.
Keywords
