Research On Fault Diagnosis Based On Improved RBF Neural Network

With the progress of technology and the development of modern mass production,mechanical equipment now is developing towards complexity, precision, integration,informatization and automation. Mechanical equipment is widely used in such a seriesof engineering fields as Aero-Space, nuclear reactor, heat and power plant, chemicalindustry, etc... But due to the complicated structure, the close connection amongcomponents, the bad working condition, the difficulty of data collection, the slow faultresponse and the low accuracy of failure diagnosis, the fault diagnosis of mechanicalequipment has become a hot topic in current research. In this thesis, the authorintroduce the theory of Rough Sets (RS) and neural network into the fault diagnosis ofmechanical equipment, and puts forward a new method of diagnosis, which is based onthe RS and Improved Quantum-behaved Particle Swarm Optimization (IQPSO) tooptimize the RBF network.First, it clarifies the research contents and significance of the fault diagnosis. Itexpounds the main methods, the developing trend of fault diagnosis, and also exploresthe feasibility of the combination of Rough Sets and neural network.Second, as the higher correlation and redundancy of the fault samples, it introduces RoughSets into pretreatment of faulty data, using improved HORAFA-A reduction algorithmto remove redundant information, firstly, it should add core of the decision table toinitial reduction set, choose the maximum of attribute weight frequency to the candidatereduction constantly and at the same time delete all the elements in the current reductionuntil the matrix is empty. In order to find optimum of reduction, this paper adds theattribute which has the best identifiable capacity to the candidate reduction when thereare more than one attributes having the maximum weight frequency and adds theconverse eliminate action only on the non-core attributes. And the experiment verifiesthis improved method.Third, based on the precocious trend of QPSO algorithm, it proposes an improvedalgorithm of QPSO. In this algorithm, the particles not only take their own positionsinto consideration, but also learn from the surrounding particles. So it solves suchproblem that the particles are in the trouble of premature because of their failure todiversify and shows that this algorithm does better than PSO and QPSO in convergencespeed and global searching ability by the experiments of three benchmark functions. Then, it brings IQPSO algorithm in optimization of RBF neural network parameter andgives the coding rules of particles and the detailed steps of optimization.At last, through the global searching ability of IQPSO, it optimizes the widthvector, center vector and weight of neural network, thus it sets up a fault diagnosismodel with stable structure and high convergence speed. The diesel engine valveinstitution experimental result shows that comparison with other models, this faultdiagnosis model has some qualities, such as high convergence speed, low forecastingerror rate, predictive accuracy and etc...