Blind Signal Processing Theory And Its Application Research On Machine Fault Detection And Analysis

The development of advanced manufacturing technology demands the machining devices must run reliably and safely. This makes that the operation condition monitor and fault diagnosis of machining devices and machining processing becomes more and more important. The new generation monitor and diagnosis system often integrates sensor, image processing, auto control, artificial intelligent and signal processing technologies, etc. Furthermore, the new generation monitoring and diagnosing system often use several sensors at the same time to collect information, which produce a great deal of observation information, some is useful information and some is redundant or useless. How to syncretize the observation information, filter the useless interferential information and extract the useful signal characteristic is a key problem for machinery monitoring and diagnosing. The blind source separation is a novel signal processing technology developed in recent years. It can directly recover the source signals by only using signals'statistic characteristics without knowing the detail signal's information. It can be used to remove the disturbance emitted from nearby parts or machines from sampling signals, thus the interested signals and disturbance can be separated. So the blind source separation technology have becomes the hot topic in fault diagnosing field since the blind separation theory appears. This dissertation first reviewed the theory and application status quo of machine monitor and fault diagnosis. And pointed out that in the past people remove the disturbance from observation by imposed the precondition that they must know many information, including knowing the detail structure, of the being monitored machine or device. Then it introduced the signification and prospects of blind source separation when applied in monitor and diagnosis of manufacturing and material cutting fields. Based on blind source technology, the dissertation advanced the monitor and fault diagnosis scheme, deeply investigated four main problems which people will be confronted with when blind separation technology is applied in monitor and fault diagnosis.By using m-p-p neural network structure, the dissertation deduced two kinds of adaptive principal component analysis algorithms from the view of output component's energy of neural network and information criteria. The two algorithms solved the problem that how to real-time extract the pointed number principal components from the measured mixture signals when people did not know the actual source number. The stability and convergence of the adaptive information criteria algorithm were fully proved. It has a very fast convergence rate, global stability and global optimum value. Furthermore, it has only one global optimum value. The algorithm's performance was validated by experiments.The existed instantaneous mixture blind separation algorithms were investigated and summarized. The difference of the observed signal's kurtosis at operation mode and faulty mode was investigated. The dissertation pointed out that the kurtosis problem must be solved first when the blind source separation technology is applied in machine noise analysis or fault diagnosis. Then a two-stage adaptive blind separation algorithm was presented to solve the trouble. The algorithm also presented the criteria of nonlinear function selecting. By exploiting the neural network structure's principal component analysis as a preprocessing step, the new two stage algorithm is robust and has good performance.The environment factor in machine noise analysis is complex and unstable. So the short sample date can better ensure the precondition of blind source separation model holding that the researched signals need to be stable stochastic signals. The dissertation pointed out by experiments that the FASTICA and other high cumulation algorithms was not suitable for noise separation when the sample data is short, though the FASTICA algorithm was extensively researched in this field. The second order correlation SOBI batch algorithm and the non-parameter-entropy-based batch blind separation algorithm was investigated to deal with the short sample data situation. Considering the blind separation application when the signals dimension is not high, especially when the dimension is two, a new separation algorithm was presented based on smooth filter idea of m-spacing entropy. By experiments, the performance index validated that the new algorithm is superior to FASTICA algorithm.Several traditional blind deconvolution algorithms were introduced first based on convolution mixture noise model. Then two kinds of Fourier transforming technology were investigated. By exploiting sliding Fourier transforming, a new and simple blind deconvolution algorithm was presented. It used only single frequency bin's information to recover sources which lowered the computation complexity greatly compared with the traditional deconvolution algorithms. In order to improve the robustness of signal separation, two kinds of frequency domain blind deconvolution algorithms were presented subsequently by using several limited frequency bins'information. Especially the second frequency domain approach can find the optimum deconvolution result by exploiting non-parameter entropy to evaluate the performance. It only needs to run single frequency bin deconvolution lgorithm several times through appropriate parameter selection. The parameter selection criteria were presented concretely also. The presented blind deconvolution algorithms all overcome the permutation and amplitude indeterminacy which is great trouble for traditional frequency domain blind deconvolution algorithms. They lowered the complexity and improved the practicability, which showed a clear prospect for engineering application.A general structure of fault monitoring and diagnosing based blind source separation was presented. By combining the newest pattern identification technology, a new fault diagnosis approach based on blind source separation and support vector machine was investigated.A four channel simultaneous measuring device was developed based on industrial personal computer. The corresponding software also was programmed at the same time. By experiments of tmotor and encoder vibration signals'separation, the performances of blind source separation were validated.