Research On Key Technologies For Data-driven Methods Of Prognostics And Health Management Of Equipment

With the development of aerospace,rail traffic,wind power generation,hydroelectric power generation and nuclear power construction in China,the reliability and safety assurance technology of complex equipment is also evolving from condition monitoring of a single component to the prognostics and health management system covering the whole system.As an important part of equipment prognostics and health management system,research on the data-driven technology of equipment health status assessment and remaining useful life prediction is the key to realize intelligent equipment health management and promote the maintenance activities transformation from the preventative to the predictive one.Modern equipment with high integration,complex degradation mechanism and diverse operating environments bring great difficulties to the health management.How to break through the key technology is the key problem to be solved to realize the effective using of equipment prognostics and health management system.Aiming at the problems existing in the key technologies of data-driven prognostics and health management system,and taking the deep learning technology as the main methods,the health status assessment technology and the remaining useful life prediction technology of equipment are deeply studied,and the theoretical research is applied to solve the engineering problems like: the equipment health status assessment questions and remaining useful life prediction questions.The accuracy and practicability of the method is verified by the experiments based on the public data sets,the new deep learning algorithm architecture for data-driven equipment prognostics and health management system provides a new perspective and new method.The main innovations of this paper are as follows:(1)For equipment in engineering application has various types of health status while the existing status assessment algorithm with single structure is hard to observe all the fault type,multi-column convolution neural network was proposed.Through different columns of large,medium and small convolution kernels,different status types can be felt,so as to improve the scope of application and assess the accuracy of model.(2)For there is a long-term relationship between equipment life data and the existing remaining useful life prediction algorithms are difficult to capture this relationship,a new network: CNN-Multi-Head Attention network is proposed for remaining useful life prediction.In this network,the monitoring data collected by the sensor is directly used as the input of the prediction network.In the model,CNN part is used to reduce the feature dimension and the multi-head attentional mechanism is used to capture the correlations between the features,improved the prediction accuracy and training speed.(3)For the present deep learning-based remaining useful life prediction methods are all devoted to the overall trend features,while ignoring the interdependence features during each operation period,a remaining useful life prediction network considering operation period was proposed.In the model,a skip recurrent neural network component based on period is used to capture the interdependent pattern of data during each period which improved the prediction accuracy of the model.(4)The data enhancement method based on time window embedding strategy and multi-sensor data fusion method based on parallel network structure are proposed from the data level and model level respectively,which alleviates the data scarcity problem commonly existing in the field of equipment prognostics and health management to a certain extent.