Study On Fault Diagnosis Methods For Rotating Machinery Based On Multi-Sensor Information Fusion

Rotating machinery is one of the most widely used mechanical equipments. If high speed parts of rotating machinery are malfunction, the whole production fails to work properly, and even the disaster accident may take place. Therefore, the research on fault diagnosis for rotating machinery has, not only theoretical significance and academic value, but also society and economy value. Fault diagnosis for rotating machinery consists of signal collection, feature extraction, diagnosis and analysis of fault state. First, multi-sensor signals are de-noised. Second, individual source signal is separated from the mixed signals to extract characteristic information. Third, the rotating machinery faults are diagnosed based on information fusion technology.The innovated achievements of this dissertation about fault diagnosis for rotating machinery are summed up as follows:1) Uncertainty of fault diagnosis for rotating machinery is identified. There are some uncertainties in testing method, work environment, process of data acquisition and characteristic of rotating machinery, some uncertainty of knowledge and conclusion, and lots of unpredictable factors in rotating machinery fault diagnosis. Therefore, uncertainty of fault diagnosis for rotating machinery is determined according to uncertainty principles.2) Targeting at non-stationary vibration signals de-noising of rotating machinery, a new kind of wavelet threshold de-noising method is proposed. The characteristic of the suggested method is analyzed in depth. Through comparing the results of signals de-noising with soft and hard threshold value method, it shows that the proposed method has high signal-noise ratio and smaller RMS error.3) A new method based on EMD, FastICA and threshold algorithm is proposed through researching separated individual source signal from the mixed signals to extract characteristic information. The proposed method solves the problem that extracting features from the mixed vibration signals of rotating machinery obtained by multi-sensor is difficult. Furthermore, when the number of observed signals is less than the number of source signals, the effect of sources separation by FastICA algorithm is poor. The method based on EMD and FastICA algorithm can separated signals, do wavelet threshold de-noising and get fault characteristic information of machinery vibratory signal. EMD is used to decompose the vibratory signals with noises to obtain a number of IMF and then instantaneous frequency and amplitude of the signal can be obtained from each IMF after Hilbert transforming and the Hilbert spectrum is obtained which shows the complete distribution of time and frequency. In this way, enough signals are obtained, then the separated source signals are de-noised, and the features of rotating machinery vibration signals are extracted effectively.4) To overcome intrinsic shortcomings of BP neural network, including low learning efficiency, slow convergence rate and easy trapping in local minimum, a new genetic neural network based on FastICA is proposed. First, the mixed signals of rotating machinery by FastICA algorithm are estimated, many individual estimations are obtained from the source signals. Second, the genetic algorithm is used to optimize the weights and thresholds of BP neural network. Third, the normalized power of independent estimation from source signals by FastICA algorithm is obtained, they are put into the genetic neural network to be trained and predicted, and pattern recognition of the rotating machinery fault is shown. The proposed method ensures the global astringency of neural network training and improves the ability and accuracy of fault identification.5) The high evidence conflict is caused by the false evidence in multi-sensor systems, and it can make the D-S combination rule out of action. To reduce the negative impact of the false evidence, a D-S combination rule base on false evidence identification is proposed and it is used as follows:first, the false evidence is identified according to consistent evidence focus of the D-S combination rule and extracted of false evidence. Second, the false evidence and conflicting evidence are restructured as an substitute evidence. Third, the substitute evidence is instead of false evidence to be composed by D-S combination rule and weaken the influence of the false evidence. Comparing the diagnosis result of the new method with results of other methods in the engine fault diagnosis, the diagnosis result of the new method turns out to be more accurate. The new false evidence identification method is proved to be credible and excellent.