Failure prediction for multivariable and multi-failure-mode complex systems based on performance degradation

This paper investigates the failure prediction problem for multivariable and multi-failure-mode complex systems based on performance degradation. In our treatment, neural network is employed to simulate system performance degradation and to predict the health states of functional modules. BP network and Learning Vector Quantization (LVQ) network are used simultaneously to simulate and to predict future possible failures, the outputs of the BP network are the inputs of the LVQ network. Historical test data, environment condition data, etc., are used to train the BP network to simulate the performance degradation on system level. Historical failure records, performance degradation data, etc., are utilized to train the LVQ network to predict future possible failure on functional module level. Simulation results show that future system performance degradation and possible failure can provide important information for decision-making for preventive maintenance planning, maintenance budgeting and spare accommodation.