Considering that periodic excitation normally exists in running machine the effects of periodic excitation on modal analysis results were studied by using numerical simulation. Through extracting the amplitude curve diagram from correlation function of response of the structure the modal components of harmonic response could be directly identified and the interference of periodic excitation elimilated by digital filtering technique The results of simulation test show that periodic excitation could be identified accurately by this method. This is of value in operational modal identification of linear stationary system.

A comprehensive fault diagnosis method combining Empirical Mode Decomposition (EMD) with Fuzzy Entropy and Support Vector Machine (SVM) for rotor system was proposed. Firstly, the fault signal of the rotor system was decomposed with EMD method into a number of intrinsic mode functions (IMFs), and a method for canceling pseudo mode function in EMD was presented based on energy conservation law. Then as the fact that Fuzzy Entropy can express the complexity of signal and has relative stability, the fuzzy entropies of the first four IMFs were taken as eigenvectors of fault signals. Finally, the eigenvectors were put into SVM to distinguish the faults of the rotor system. The result indicates that this method can effectively extract fault characteristics and diagnose the fault of the rotor system.

To study the effects of the wavelet transform in intensive operational modal parameter identification, a mode system with three degrees of freedom containing two intensive frequencies was established. Applying narrow-band white noise excitation and using the improved Morlet wavelet as the continuous wavelet transform base function, the modal parameter identification of this system was simulated. It was found that reducing bandwidth can improve the frequency resolution effect and decouple the dense modals when applying the continuous wavelet transform method to modal parameter identification. However, it also exacerbated the problem of edge effect which can reduce the accuracy for the dense modal recognition. In order to suppress the edge effect, the length of the useful part of the transformation signal was extended by using the SVM technique. Simulation results show that the better recognition accuracy can be obtained with this method. Finally, through identifying the first two modal parameters of a grinder mill, the feasibility and effectiveness of the method were verified.

An improved BP neural network with the methods of momentum and adaptive learning rate is applied to build the fault diagnosis system of the fans. In the process of training, standard training samples and samples with white noise are employed to train the neural network so that the neural network has some ability of fault tolerance. Results of the simulation and the fault diagnosis of a fan show that the improved BP neural network needs less training times, the learning efficiency is raised, and the phenomenon of trapping in the local minimum for the network is effectively repressed. This method is effective for the fault diagnosis of fans.