Nerve Network Model for Prediction of System Maintainability

摘要

The design of a kind of new equipment has certain inheritance, i.e. it always inherits part or the whole of one or many types of equipment in service. Therefore, the design characteristics and situation of use and maintenance of the present equipment can be used to predict the maintainability characteristics of a newly developed equipment. According to this clue, this paper sets up a maintainability model of equipment by using back-propagation(BP) multi-layer nerve network. The so-called BP multi-layer nerve network means that each nerve element of the network receives input from its former stage and output to its following stage, there is no feedback in the network. In this back-propagation algorithm, the network not only has input layer node and output layer node, but also has invisible layer node. The invisible layer can be single layer or multiple layers. When a signal is input, it is first propagated to the invisible node, after actuated function. The output signal of the invisible node is propagated to the output layer node, the processed result is then output. According to the characteristics of the nerve network, this paper provides the main procedure of establishing the maintainability nerve network: organization of practical examples; determination of important factors affecting maintainability and determination of the structure of the nerve network. The choosing of the input nodes has relation with the number of important factors considered, the number of output nodes has relation with the number of maintainability characteristic parameters, the number of invisible layer nodes is usually chosen according experience. Then, train the nerve network model with training sample and finally determine the structure of the nerve network model. Then, use the quantitative value of factors affecting the maintainability of the studied object as the input to evaluate the maintainability characteristic parameters of the studied object with the help of the trained and tested nerve network model. Finally, the model is tested and verified in this paper with a practical example, having good result.