Research On Fault Diagnosis Based On Information Fusion

Various fields are showing a new trend towards large-scale,complex,high-speed and intelligent direction and the links between various parts of the system are getting closer since the fast development of science technology and the quick progress of modern production.Once a fault happens and it is not dealt immediately,it may trigger more faults,thus affecting the normal progress of industrial production and causing huge economic losses and even more serious casualties.Furthermore,there are many factors causing the fault,the diagnosis results from a single sensor are incomplete,thus it's normal that the system provides a low accuracy of fault diagnosis using only one sensor.Moreover,due to the different mechanisms of different diagnostic algorithms and the limited priori knowledge,it is difficult to make a comprehensive diagnosis of complex faults using only one diagnostic algorithm.Therefore,it is necessary to make a comprehensive and accurate diagnosis of complex faults combining with multi-source information.In this thesis,multi-level information technique is used to deal with fault diagnosis.Firstly,different diagnostic algorithms are used at the feature level to obtain different preliminary diagnostic results;Secondly,the final diagnostic results are obtained by evidence theory at the decision-making level.The main research work is as follows:Firstly,due to the complexity of fault modes,it is hard to obtain the diagnostic model.So this thesis studied the machine learning algorithm based on historical data for fault identification,which can realize the nonlinear mapping from the symptom space to the fault space and avoid the complicated mathematical model.This thesis applied BP neural network,support vector machine and extreme learning machine to establish the model of rolling process fault diagnosis and obtain the preliminary diagnostic results in feature level.The traditional two-class support vector machine is not suitable for the situation since the fault type of rolling process is complex.Therefore,the one versus one based multi-class support vector machine strategy for fault diagnosis has been studied and since standard support vector machine cannot provide posterior probability,a posterior probability output model has been established in the purpose of fusing it in the decision level.Moreover,different parameters of the support vector machine affects the diagnosis results.On this basis,the grid search and cross validation based support vector machine parameter optimization method has been studied.At last,the preliminary results obtained from BP neural network,support vector machine and extreme learning machine are used as evidence in order to provide the basis for the decision level fusion.Secondly,this thesis researched on single fault diagnosis based on the PCR6 theory.With the deepening of research,some prior knowledge acquired in advance can help people diagnose more accurately.Compared with other fusion rules,PCR6 can distribute the conflict of prior knowledge better.But it has a disadvantage that the main focus doesn't converge through example simulation.Aiming at this phenomenon,this thesis proposed a method to improve it and verify its validity by example simulations.Thirdly,aiming at the phenomenon that the improved PCR6 rules cannot deal with single fault evidence mixed with abnormal evidence,this thesis proposed to cluster the evidence.When clustering,fuzzy clustering based on transitive closure and silhouette value is used and when measuring similarity in clustering,this thesis proposes a measure of two-group conflict,and verify the validity of the clustering method and clustering the evidence to deal with the single fault evidence mixed with abnormal evidence.Finally,apply the above approach to deal with multiple faults,since PCR6's main focal element cannot converge in low conflict and it's effective in high conflict,which means that different fusion rules are needed when dealing with single and multiple faults.Aiming at this phenomenon,this thesis proposes to decide fusion rules based on the maximum value of evidence distance and proposes a frame which is suited for both single and multiple faults,and verify its validity in dealing with single faults and multiple faults.