The Research Of Diagnosis Method Based On HHT And D-S Theory For Centrifugal Pump

Fault diagnosis is a new multi-subject-crossed technique, it has been developed rapidly in last twenty years, and it has brought huge benefit. Due to the fact that fault signals of Centrifugal Pump present nonstationary properties, it is essential to choose appropriate signal processing methods that are suitable for nonstationary signals to extract fault feature. The time-frequency analysis methods such as windowed Fourier transform (WFT), Wigner-Ville distribution (WVD) and wavelet transform have their own limitations. Recently, a novelty time-frequency analysis method, Hilbert-Huang transform (HHT), which is suitable for nonstationary signals, has been put forward and confirmed to be superior to the other signal processing methods in many applications.Information fusion is a subject formed recently, it has been researched and applied in many fields. But, it is in starting stage of fault diagnosis. There are lots of available information in fault diagnosis. Only when the available information is used, can the precision and credibility be improved. So fault diagnosis is a process of disposing information. In fault diagnosis of Centrifugal Pump, problems will appear, such as: lots of data to be processed various faults, difficulty of obtaining knowledge, and low ratio of identifying faults. Information fusion is presented to solve the above problems in this paper. HHT,neural networks and evidence theory are combined to diagnose faults, which improves the agility, efficiency and accuracy of the fault diagnosis.Firstly, this text has described Hilbert-Huang Transform technology characteristic, form structure and concrete processing method. Time-frequency analyzes for Centrifugal Pump vibration and cavitations signals has been achieved using HHT. It can simultaneously analyze the change of signal in time-frequency domain using Hilbert-Huang spectrum of signal. It is found its energy distribution has great difference for signal in different frequency scope. For the sake of describingthese difference quantitatively, the energy ratio of signal in different frequency scope as characteristic variable is defined.Secondly, the topological structure of neural networks and learning method are expounded; especially the characteristic of RBF neural network is introduced, and an online algorithm is proposed which can structure the neural network dynamically.Finally, the basic principle, formatting rule, reasoning process of evidence are introduced in detail. We design the diagnostic method based on neural net and D-S inference. Started with the characteristics of the vibration signals from the waiting diagnostic system, the two symptom field of fault characteristics are divided, the individual fault diagnosis is preliminarily done in the two subnets. the diagnostic result from three subnet are converted to basic probability assignment. Information fusion is token in time and spatial domain by D-S inference. By experiment data, the feasibility and availability of this diagnostic method are verified, it can increase the accuracy of fault diagnosis.