An Analysis On Independent Component Of Convolution Mixed Vibration Signals

Industry informatization promotes the continuous development of various signal processing techniques. Compared with the traditional signal processing methods, Blind Source Separation (BSS) analyzes and processes the multi-input and multi-output system of mixed-signals to recover the multi-dimensional source signals under the condition of unknown prior knowledge such as the source signals and transmission channels. Independent Component Analysis (ICA) plays an important role in the Blind Source Separation. The independence criterion among the signal vectors as the objective function, ICA isolates the independent source signals from the observed mixed-signals by optimizing algorithm to adjust the separation matrix. Because of its novel algorithm and excellent performance, it is widely used in hybrid speech recognition, communications processing, seismic monitoring and other fields.Mechanical vibration signals are usually the convolutions of the vibration source signals and the transmission channel impulse responses. ICA based on the instantaneous mixture model can not overcome the influence of the path effect. Taking the rotating machinery vibration signals as the study case, this thesis has researched the ICA of convolution mixed vibration signals. Firstly, the observed time-domain convolution mixed-signals are transformed into the instantaneous mixtures of frequency domain through the Fourier transform. Then the FastICA algorithm is used to separate the mixed-signals. During the above processing, the time-domain signals transformed from Fourier are transformed from the real value into the complex value in the frequency domain. Consequently, the objective function and constraints should be expanded to obtain the complex value FastICA algorithm to do the blind source separation.In the natural environment, there is no absolute"blind". This thesis has also studied the improved algorithm based on the semi-blind separation. Firstly, the reference signals are constructed from the amplitude information of the source signals in the main energy band obtained by the use of the band-pass filter on the basis of the complex value FastICA algorithm. Secondly, the mathematical model based on the improved algorithm of the semi-blind separation is established after the amplitude approach function to construct the new inequality constraints is introduced. Finally, the improved complex value FastICA algorithm is obtained after the Augmented Lagrangian to optimize the constraints is introduced. Simulation results have showed that this improved method can quickly get the target source signals and improve the separation accuracy and computing time of the algorithm without having to repeat the calculation of other signals. In the conclusion part of the thesis, the vibration signals of the rotating machinery are analyzed by the improved complex value FastICA algorithm. The two experiments of the broken tooth fault and the bearing cup raceway fault have demonstrated the results by the use of the convolution mixed model and the improved algorithm based on the semi-blind separation can obtain the clear and abundant fault information, and can basically eliminate the influence of path effect on the measurement results.