Research On Key Issues Of Evidence Theory And Its Application In Working Condition Recognition Of Roadheader

The roadheader is a kind of coal mine machinery with the functions of cutting,loading,dusting,handling and operation.The roadheader,which is mainly used to cut down the rock,coal or semi-coal,is one of the important mechanical equipments in the underground production of coal mine.With the development of new sensor and information technology,the real-time state data of roadheader from electric,hydraulic,vibration are obtained in order to monitor.After signal processing,feature extraction and pattern recognition,the roadheader can be recognized the working conditions,diagnosed faults,forecasted and maintenanced,which promote the development of coal mining towards high efficiency,automation and intelligence.Because of the bad working environment of the roadheader the vibration signals are non-stationary,random,and uncertaint.Dempster-Shafer(D-S)evidence theory,which has significant advantages in uncertainty representation,measurement and information combination,has attracted more and more attention of researchers,and has been widely used in information fusion,pattern recognition,decision support systems and other fields.Therefore,it has great theoretical significance and practical value to apply D-S evidence theory to the research of roadheader condition recgnization.D-S evidence theory has many advantages,however there are still some key problems to be solved that restrict its application and popularization.One of the problems is that Dempster’s combnation rule often draw counterintuitive conclusions when dealing with highly conflicting evidences,which fails to achieve effective decision.The other problem is how to automatically and reasonably obtain the basic probability assignment(BPA)function.This thesis studies some key issues and shortcomings of D-S evidence theory,and then proposes some corresponding solutions as well.Finally,the evidence theory is applied to recognize the working conditions of the roadheader,the feasibility and effectiveness are verified by the simulation experiments.The main contents and innovations are as follows:(1)A new eviednce distance based on Baroni-Urbani similarity coefficient is proposed to solve the problem that the combination result in D-S evidence theory is counterintuitive when dealing with high conflict evidence.So evidence can be preprocessed and combinzed using Dempster rule.It is proved that the new distance satisfied nonnegativity,nondegeneracy,symmetry and triangle inequality and is a full metric distance in a metric space spanned by the evidence vectors.Therefore,the new evidence distance is used to measure the degree of conflict of evidence,and the support and reliability of each evidence is calculated based on the weighted coefficient,and at last the weighted average evidence is obtained.The Dempster rule is used to combinize n-1 times for the average evidence,and the final result is obtained.Compared with other methods,the method proposed in this thesis can combine both low conflict and high conflict evidence and has more exact results and faster convergence,furthermore it solved one vote problem and carfocal elements fuzzy paradox.(2)A conditional formula of generating BPA is proposed and proved based on neighborhood rough set to solve the problem that BPA is subjectively acquired in D-S evidence theory.Rough set can extract the relevant rules effectively from historical data of information system without any prior knowledge.Therefore,by dividing the domain according to the conditional attributes of the input sample and the decision attribute,the BPA can be obtained automatically from the historical data,and thus the subjectivity and uncertainty of the BPA provided by expert experience are avoided.Because there are a large number of numerical variables in practical application,information loss may happen if discretization is used.this thesis adopts the neighborhood model of rough set.The weights of BPAs are different because of the different importance of conditional attributes.In the view of algebraic and information,a attribute significancy formula was proposed combinng attribute dependency with mutual information,which is used in attribute reduction of neighborhood rough set.The feature selection and classification of UCI data sets are carried out.It was proved that the new significancy can obtain higher recognition rate with fewer features.Finally,the normalized significancy is used to correct the generated BPAs.The feasibility and rationality of the proposed method are proved by examples.(3)The vibration acceleration signal of roadheader is nonlinear and non-stationary,so the classical signal processing technique is no longer applicable.Variational mode decomposition(VMD)can adaptively decompose the signal into a series of intrinsic mode functions(IMF)that contain a lot of characteristic information.In this thesis,the vibration signal of roadheader is decomposed by VMD,and the instantaneous frequency is obtained by using Teager energy operator.In time-frequency domain,280 Hz and 350 Hz signals appeared repeatedly,which is related to the working frequency of the roadheader.Then the energy of each IMF component,approximate entropy,singular value of time-frequency matrix,peak value,standard deviation kurtosis index of signal are extracted,whicht constitute the characteristic quantity of roadheader working condition recognition.Through the recognition algorithm,it is found that there are some differences between different features for different working conditions,which need to be considered comprehensively.(4)A recognition method based on neighborhood rough set and D-S evidence theory is proposed,which is used to identify four working conditions of roadheader such as no-load,drilling,left-cutting and down-cutting.For the sake of the different vibration characteristics of the roadheader under four working conditions,the proposed method is verified from the following aspects: single direction single feature,single direction multi-feature and multi-direction multi-feature.After many simulation experiments,the propoesd method has a stronger recognition ability,compared with traditional D-S,support vector machine(SVM),k-nearest neighbor(KNN),neighborhood classifier(NEC)and convolutional neural network(CNN).The average recognition rate is 89.31% and 78.61% respectively.