Research On Fault Diagnosis Methods For Rotating Machinery Based On Non-Stationary Analysis And Neural Network

To meet requirements of the non-stationary signal processing of modern rotating machinery, improve the accuracy rate of fault diagnosis of complex signal, and get intuitive results of fault diagnosis, this thesis proposes to apply non-stationary signal analysis and neural network in fault diagnosis, which mainly including three ways: Short-Time Fourier Transform, Wigner Distribution and Wavelet Transform.For the characteristics of the results of the non-stationary signal analysis, the thesis presents an effective method of feature extraction, that is, to do some mathematical process for the data of the non-stationary signal analysis result which at the same frequency (or scale) and at different times, so the result can be changed from three-dimensional graphics into two-dimensional spectrum, And then selects the data in characteristic frequency (or scale) as characteristic values. The Feature extractions are identified through the pilot. Test results show that, Short-Time Fourier Transform should use variance program, Wigner distribution should use maximum program, while Wavelet Transform can use the program of sum of squares to find the characteristic scales. At last several eigen values which can Maximum reflect fault signature are successfully picked up, as the input of neural network. The thesis uses MATLAB to develop an object-oriented, general BP neural network software. The software mainly has three innovations: the samples can be automatically read; trained network can be stored and mobile; software generalization.To study the effect of the fault diagnosis method based on non-stationary analysis and neural network, the thesis uses the fault data of Bently Test-bed, and makes contrast with the method based on traditional spectrum analysis and neural network. The results showed that, the diagnosis effects of the three non-stationary analysis methods combined with the neural network are all better than spectral analysis, in particular, Short-Time Fourier Transform and Wavelet Transform combined with the neural network have more satisfactory diagnostic accuracy, which can meet the needs of modern rotating machinery fault diagnosis, and has good application prospect. In addition, the result of the diagnosis is no longer complex three-dimensional graphics but the intuitive output of neural network, so the way of fault diagnosis is rose from judging complex graphics to artificial intelligence.