Research On The Fault Information Mining And Condition Evaluation Of Rotating Machinery

The vibration signals carry rich information that can reveal the running state of equipment.Due to the mechanical system complexity and the effect of noise,it is difficult to guarantee the accuracy of the result if the health conditions of rotating machinery are evaluated by directly analyzing the collected vibration signals.It is an effective way to improve the accuracy of evaluation by mining the useful information from the vibration signals to eliminate the influences of the mechanical system complexity and noise.In terms of the modulation characteristics of the vibrational signals of rotating machinery,this dissertation studies the mining method of the fault information,and evaluates the health conditions of machinery based on the mined fault information.The content includes four parts: the mining method of the fault information in demodulation frequency domain,the mining method of the fault-induced vibration responses,the transferring method of fault features for conditions of data scarcity,and the evaluating method of the fault severity based on feature fusing.(1)The mining method of the fault information in demodulation frequency domain is studied.In the method,the phase function set is developed to demodulate the vibration signals,so the fault information is mined.To solve the problem that the vibrational signals are affected by the random fluctuation of the rotating speed,a new phase function design method is proposed.It can demodulate the frequencies that present the same physical meaning under the fluctuation of different rotating speeds into a same frequency value.The demodulation spectra are obtained via the method.The basic frequency and the allowable demodulation error are defined,based on which the fault information in demodulation frequency domain is searched.The new algorithm can mine the fault information from the demodulation spectra containing noise.The mean experimental recognition rate is higher than 99% and the accuracy is improved.(2)The mining method of the fault-induced vibration responses is studied.The method introduces the Matrix profile algorithm and mines the fault-induced vibration responses.According to the modulation characteristics of the vibration signals(that is,the vibration responses take the resonance frequency as the carrier frequency,and the carrier frequency depends on the mechanical system and is not affected by the working condition;the amplitudes of the vibration responses are affected by the fault and the frequency modulation),the Matrix profile algorithm is constructed using the Znormalized Euclidean distance as the measuring method.It measures the waveforms of the vibration responses.Thus,the mining results are not affected by the signal amplitudes.Based on the mined vibration responses,convolutional neural network(CNN)is adopted to recognize the fault severity.The experiment indicates that the faultinduced vibration responses can be mined from the vibration signals of the experiment table and the planetary gearbox of the wind turbine by the proposed Matrix profile algorithm,and that the fault severity recognition accuracy is improved by using the mined data.(3)The transferring method of fault features for conditions of data scarcity is studied.In the method,the Matrix profile algorithm is used to mine the fault-induced vibration responses from the vibration signals of the experiment table and the planetary gearbox of the wind turbine.A two-stream CNN transfer learning model is constructed to transfer the knowledge learned from the data collected by the experiment setup to the data training model of the wind turbine.The scarcity of the available data of equipment leads to the low accuracy of the network model in the mechanical condition evaluation.To fully learn features from the limited data,a two-stream CNN model is built.Meanwhile,features are also learned from the time-domain waveforms and the timefrequency spectra of the mined data.To optimize the network model under the condition of the limited data,the data collected from the experiment table are firstly used to pretrain the two-stream CNN model,and then the data obtained from the wind turbine are used to fine-tune the model.The experiment shows that the two-stream CNN model can achieve 99.87% accuracy which is higher than the single-stream CNN model in the fault severity recognition,and can converge more quickly and stably.(4)The evaluating method of fault severity based on feature fusing is studied.In the method,self-organizing map(SOM)is firstly utilized to fuse features under normal conditions,obtaining the U-matrix mapping units,and then the fault severity is quantified through calculating the quantitative difference index between the recursive score values of the features under different running conditions and the U-Matrix best matching unit.To remove the redundant features,the Laplacian score of each feature is calculated,and the features are refined according to the scores.Based on the good topology preserving ability and visualization performance of SOM,the U-Matrix mapping units obtained from the refined SOM are used to qualitatively describe the degradation process of the bearing.A method of quantifying the fault severity by calculating the quantitative difference index of recursive score values is proposed.The experiment proves that the U-matrix mapping units obtained by using the mined data can reflect the bearing degradation trend more accurately,and the fused recursive score quantitative difference index can evaluate the fault severity more accurately.