HMMM-based Fault Diagnosis Technology Under Complex Conditions

The hidden Markov model (HMM), as a kind of intelligent diagnostic technology, has astrong pattern classification and modeling capabilities on time series of dynamic process. Ithas a good prospect of application in the field of fault diagnosis because of these features,especially in the complex systems with many devices and high reliability requirements.Hidden Markov model can be used not only for state recognition of the steady-state but alsobe used for modeling and recognition in speed-up and speed-down process of rotatingmachinery. The fault diagnosis results of simple machine based on HMM are good. But whilethere are many kinds of faults or equipment is very complicated, the recognition rate of HMMwill always be lower.Funded by the advanced project―Research on Condition Monitoring and Fault DiagnosisTechnologies based on hidden markov model and support vector machine for Equipment ofNuclear Power System‖(National High Technology Research and Development Program―863‖, No.2008AA04Z407) and "Research on Condition Monitoring and Fault DiagnosisTechnologies for Key Equipment of Nuclear Power System‖(National Defense Basic ScienceFund Project, No. B0120060585), the purpose of this study is to address the significantdecline in the recognition rate of HMM model under the large fault types and similar, notobvious features by experimental verification on the nuclear power failures simulator, andprovide technical assurance and theoretical support for the practical application of the HMMintelligent diagnostic techniques in the diagnosis of nuclear power critical equipment, andother complex equipment.The main content of this dissertation are as follows:1. Applying artifical immune theory in the HMM diagnostic system to solve the HMMtraining problem on initial value setting.The algorithm improvements include the re-division of the P collection, the diversityenhancement using best set clonal selection algorithm, dimensionality reduction of B matrix,spatial coverage of the hypersphere, and search algorithm of optimal combination.2. Improve the recognition rate of HMM model under complex conditions.The first method is utilizing HMM/SVM structure to improve the classification ability;another method is to build a Markov chain HMMs network.3. Experiment verification. The experiment was carried out on fault simulator of main pump to validate thealgorithm. The experimental results show that the proposed HMM-based fault diagnosismethod is effective under complex conditions.Through the research of HMM-based fault diagnosis technology under complexconditions,4categories and13states of the pump failures simulator were simulated anddiagnosed. Using BCSA aided HMM training algorithm to improve training efficiency andother improved technologies, the stable process average recognition rate of HMM/SVM canbe increased from86.8%to94.8%, recognition rate of HMM network can also be reached93.4%. Both methods are effective in improving the system performance of recognition underlarge fault types and similar, not obvious features complex conditions in nuclear powerequipment and other complex equipment diagnosis.