Study On Ordinal Feature Analysis For Fault Severity Recognition

Fault diagnosis technology has attracted further research and wide application inthe past decade. With the development of artificial intelligence, pattern recognition,machine learning and computer technology, fault diagnosis technology is developingtowards to intelligent direction. Detecting fault severity levels is a challenge in the fieldof fault diagnosis, which can help users to grasp running condition, plan reasonablemaintenance strategy and scheme.Fault severity levels can be represented as “slight fault”,”moderate fault”,“severefault” and so on, there are ordinal structures between different severity levels. In thefield of pattern recognition, the fault level identification can be considered as ordinalclassification. Some fault features have monotonic relationships with fault severitylevels, such features are monotonic features, that is, the feature values changemonotonically with severity levels. Monotonic features can reflect the monotonic trendwith fault severity levels, which provide straightforward and simple diagnosticinformation for fault severity level identification. However, there are some features arenon-monotonic with the severity levels. The former is monotonic feature, and the latteris no-montonic feature.Only a small part of sensor data is relevant with the fault severity levels. A largenumber of irrelevant features may increase the computational complexity and reduce theidentification ability. Consequently, it is needed to select relevant features from all faultfeatures using feature selection algorithm, remove irrelevant and redundant features. Inthis paper, from the aspect of ordinal classification, the feature selection algorithm forfinding the optimal feature subset from the orginal feature space is discussed, which canimpove generalization capability of fault severity level identification. Firstly assumethat all features are monotonic, fault severity level identification can be viewed asmonotonic classification, then considering only some features have mononotnicrelationships with fault severity levels. The main contributions are listed as follows:Firstly, we design a feature selection algorithm for monotonic classification basedon distance metric. The basic idea is to estimate the quality of feature subsets accordingto how well they distinguish between instances that are near to each other. Define thedistance metric as feature evaluation function, a feature selection algorithm are designed.The experimental results show its effectiveness.Secondly, we design a robust feature selection algorithm for monotonicclassification. Most of real classification tasks are non-monotonic consistent, we should design a robust feature selection algorithm. Rank entropy computed based onprobability distribution function is a robust metric. We use rank mutual information toevaluate features and combine it with minimum Redundancy Maximum Relevance(mRMR) search strategy for searching a feature subset, and validate the effectivenessand robustness of the feature selection algorithm using experiments.Thirdly, we design a hybrid feature selection algorithm for ordinal classification. Infault severity level identification, just some features are monotonic with the severitylevels, instead of all. In this work, a hybrid feature selection algorithm is proposed, inwhich monotonic features and non-monotonic features are handled individually. Wedefine hybrid classification consistency assumption, design a feature evaluationfunction and present a hybrid feature selection algorithm for ordinal classification. Thisalgorithm has been tested its performance with experiments.Finally, the hybrid feature selection algorithm is applied to identify the gear cracklevels. To extract fault features from the vibration data, the fault dataset for each gearcrack fault levels are collected. This paper firstly select revelant features from theorginal fault features, then apply stochastic hybrid ordinal decision tree to identify faultlevels.This paper discusses the feature evaluation function and selection algorithm forfault severity level identification. Based on dividing the fault features intonon-monotonic features and monotonic features, different feature evaluation functionsand feature selection algorithms are designed for finding feasible solutions for faultseverity level identification.