Predictive Diagnostics of Power Generating Equipment Failures Based on Neural Network Technology

This paper is covers basic problems of creating a predictive diagnostic system for assets and units of a heating power plant based on a fully connected neural network. The research confirms frequent situation when the network cannot be properly trained against typical failures due to lack of statistical data associated with each type of failure. In order to solve the problem we implement the idea of utilizing additional data that belongs to pre-fault time period and/or corresponds to warning alarms thrown before the failure itself. The primary analysis shows that existing databases from industrial enterprises contain significant amount of pre-fault data. We introduce a procedure of creating and preprocessing the training set based on pre-fault data measurements. The procedure includes basic stages of detecting cross-correlation between measured parameters, defining each parameter variation/range and choosing the data most applicable for neural network training. The result of training the network and testing the diagnostic system on real data taken from the turbine's feeding pump located at the heat power plant. Based on the real result demonstrated we certify positive outcome of the designed predictive diagnostic system and share some future improvements to be done.