Research On Association Rule Mining Algorithm In Intelligent Prognostics

Intelligent prognostic is an essential part of intelligent maintenance system. The understandability of intelligent prognostic results has an effect on precision and efficiency of decisions made from equipment maintenance personnel. However, the prognostic results usually turn up in form of large data. For example, the knowledge about artificial neural network distributes in all connection weights and thresholds. Especially, when neural network structure is complicated, the size of weights and thresholds matrix will be very large, which makes it difficult to understand and use these data directly. In this paper, we extract association rules from these data to find underlying information and knowledge about prognostics from the point of data mining, which convert prognostic results data into knowledge in form of association rules and therefore improve understandability of prognostic results.Support vector regression (SVR) has been widely used in remaining useful life prediction in process of intelligent prognostics. However, prediction accuracy for SVR has a very close relationship with the selection of regularization parameter C, insensitive coefficientεand core function parameterθ. Given the disadvantages of conventional methods, we propose a methodology for parameters selection based on association rules in this paper. At first parameters sample data are constructed, then a data mining tool CBA is applied to extract association rules from the constructed data sample. Finally those extracted rules are checked. Experimental results show that rules with highest precision can be chosen for parameters selection which is a new way for SVR parameters selection.Given that CBA tool pays more attention to data normalization, a new algorithm MAMP based on multiple pruning strategies for class association rules mining is proposed in this paper. Firstly MA algorithm is applied to extract association rules that meet both minimum support and minimum confidence. Then relevant analysis, redundancy analysis and data-covering analysis are performed on the association rules, and the most interesting rules will be reserved. Experimental results show that MAMP algorithm has a perfect pruning performance that it can keep least number of association rules as well as keep high precision. Additionally, MAMP algorithm can deal with general data sample. In other words, it has no special request on data normalization. Therefore, it can be applied to deal with the data yielded in process of intelligent prognostics.Performance evaluation is a key part in intelligent prognostics. Given that the modeling of material fatigue performance evaluation based on fracture mechanics is complicated, we experimentally construct a new model for fatigue performance degradation assessment based on Bayesian network in this paper. In our model we directly employ Bayesian network to study fatigue data obtained from fatigue tests, therefore it becomes unnecessary to study the process of fatigue crack initiation and propagation during fatigue failure. Performance evaluation results show that the proposed model make great progress comparing with conventional models. The model reduces modeling complexity, and considers the effect of more factors on fatigue performance degradation. Additionally it provides a possible solution to undertake online monitoring on equipment's performance.In the last section of this thesis, MAMP algorithm is applied to extract association rules from material fatigue performance degradation assessment results, which turns data-formed performance evaluation results into knowledge in form of association rules, and its understandability is improved. Therefore maintenance personnel can directly understand and apply those rules to judge the stage that equipment's performance is on, and make certain maintenance decisions.