Application Of ICA In Vibration Signal Processing

IntroductionThe independent component analysis is the developed multi-information source processing method in recent years. The basic meaning is to decompose the multi channel observation signal to certain independent ingredients according to the statistical independence through the optimized algorithm. The goal is to seek a linearity but uncertain orthogonal coordinate system to express the multi-dimensional data. It requests little priori knowledge of the primary signal and the mix system. Based on the ICA method, the statistical dependence between the various ingredients of the analyzed signal is obtained the minimum, thus the source signal essential structure is highlighted. As one multipurpose statistical method, the independent component analysis has widely applied in the speech enhancement, array antenna processing and communication, time series analysis, digital watermark, remote sensing, radar, and so on.The mechanical device vibration signal contains the plenty operation of equipment information. It is the important origin of diagnosis. But the vibration signal is often not single, because it is influenced by installation position of sensor and uncertainty of the breakdown actual vibration direction when measuring the vibration signal. So other disturbance signals are included in the survey signal inevitably. On the other hand, the new vibration can be produced possibly because of the induction vibration of the breakdown. The above reason has caused the present equipment examination and the breakdown processing accuracy improved difficulty. The independent component analysis opens a new research way for solution above question. The independent component analysis applied research aiming at the vibration signal has been made a rapid progress and obtained certain effect currently.After introducing the independent component analysis basic principle, the independent criterion and the commonly used algorithm, the essential two factors are pointed out in independent component analysis algorithm based on the neural network; they are the nonlinear function and the auto-adapted length of step-size selection. In traditional algorithms, the nonlinear function selection is all fixed. The traditional algorithms can separate one kind of non-Gaussian composite signal effectively, but them separation performance becomes bad when separating two kind of type composite signals. Therefore this article takes the Parzen window estimation in the non-parameter estimation to estimate nonlinear function according to different composite signal. The traditional algorithm is improved by taking this method. The improved algorithm not only separates the sole type composite signals, but also two kind of types the composite signals well. The step-size factor in optimized algorithm has also been made the improvement. It not only enhances the restraining speed, but the restraining effect. Each kind of breakdown bearing signal is gathered through the setup test bench. Then the gathered signal is separated and analyzed by utilizing the improvement algorithm. The result indicates that ICA has the very strong ability to separating composite signals in the blind separation aspect. The each separated independent component is not only carried on the breakdown judgment directly, but provided the signal pretreatment work for the characteristic extraction, the information fusion and the data mining and so on.1 The variant step-size algorithm based on Parzen window estimationThe current independent component analysis research mostly concentrates in algorithm based on the artificial neural networks. In this kind of algorithm, the appraisal function is needed as the activation function of network, and the appraisal function is defined by the source signal probability density function. Because in the independent component analysis the source signal is unknown, therefore the real source signal appraisal function is unable to be gained. In general algorithm certain specific nonlinear function is used to substitute of unknown appraisal function, and separation of certain signals can be realized finally. But in order to guarantee algorithm restraining, different nonlinear function must be selected according to the non-Gaussian type of source signal. The separation performance is bad when the difference between real probability density distribution and its estimation is quite big. Especially both Sup-Gaussian signal and Sub-Gaussian signals existing in the source, the kurtosis of signals is needed to be judged to choose the probability density distribution function in each iterative process, so the computation load is big.In order to solve question algorithm based on the artificial neural networks cannot separate composite signals have both sup-Gaussian and sub-Gaussian effectively, the Parzen window estimation is used to estimate the source signal appraisal function directly, and the estimated result is taken as the activation function to improve the representative EASI algorithm of independent component analysis. Two rank nuclear functions of the Parzen window is used to estimate probability density of x, and the method is to smooth and weight each sample point of random variable x through a series of continual function window, then all weighting window is averaged. In the Parzen window estimation the crucial function is the nucleus length parameterσ. In the article optimized method of this parameter is discussed, and the simulation is used to explain influence of the nucleus length to probability density estimation.Independent component analysis algorithm based on the neural network belongs to the LMS algorithm mostly. The convergence rate and the stable state performance of algorithm are affected by selection of step-size parameter. Namely the step-size is small, the stable state performance of algorithm is good, but the convergence rate is slow; On the contrary, convergence speed is fast, but the stable state performance is bad. How to solve the contradiction between convergence rate and stable state performance is always the research hot spot, also an important question. The request of convergence rate and static error cannot be satisfied simultaneously because the fixed step-size is adopted in the original EASI algorithm. In this article, the step-size parameter is adjusted dynamICAlly through the method of tracking change of the appraisal function, and contradiction between the convergence rate and the static error is amended effectively. Combination the above method with the Parzen window estimation, the variant step-size based on the Parzen window estimation is obtained. Simulation experiment is carried on to compare the improvement algorithm in this article algorithm and EASI and the EXT-ICA algorithm. The result indICAtes the improvement algorithm in this article has the fine performance to separate the arbitrary composite signals.2 Vibration signal analysis and processingThe vibration signal processing method is the important method to determine mechanical device breakdown. The rolling bearing most commonly used in the mechanical device is taken as the object of study in this article. Its vibration mechanism is analyzed. Its various parts'vibration frequency caused by breakdown is discussed. Vibration data acquisition system is constructed using the fault simulation test bench and the data acquisition system. Signal analysis of bearing breakdown is divided into three kinds of situations as the addendum circle, the inner loop and the rolling body. Four same model bearings are chosen, one kind of breakdown artificially is made on each three bearings besides one. The vibration signal under the different breakdown can be gathered through replacement of different breakdown bearing. The improvement algorithm in this article is used to analyze the vibration signals under different breakdown. The result indicates the aliasing different breakdown characteristic in the gathered signals can be separated well. After separation, one kind of breakdown characteristic is demonstrated basically by each sensor signal spectrograph. But the rotation frequency signal and the power frequency signal can not be eliminated completely from the separation result.In the end the summary to the full text work and the forecast to the future work are put forward.