Research On Typical Fan Fault Diagnosis On CPU+GPU Heterogeneous Computation

This thesis focus on the needs of fan fault diagnosis and state monitoring,identifing typical fan fault in accordance with the project needs based on CPU+GPU heterogeneous model.The characteristic frequency is extracted according to the fan fault state.In the fault diagnosis,the envelope analysis and the model training are time-consuming.The algorithm is parallelly optimized on CUDA,and finally fault diagnosis software based on the CPU+GPU heterogeneous mode is completed.The fan system is simplified and the typical faults manifestations are summarized.According to common fault type s,the time-domain statistical analysis method is used to analyze the signal,and parameters sensitive to the fan faults are identified.The LMD method is used to decompose the nonstationary vibration signal,and the SG-LMD decomposition method is proposed to extract the fault characteristic frequency of noise-containing viberation signal.The criteria for the effective components are determined based on the correlation coefficient analysis,and the vibration signal is reconstructed through the effective components.The fault characteristic frequency of vibration signal is extracted by SG-LMD method.The fan vibration signal parameters can be observed by sensors when the fan is running,and the current device state can be estimated by using the extracted signal parameters.Because this feature is consistent with the Hidden Markov Model(HMM)structure,HMM model is used to estimate the true state of the critical equipment of fan system by observed signal.The HMM model initialization is improved and the scale factor correction met hod is used for data underflow problem solveing.The validity of the HMM model in fault diagnosis is verified by using the bearing data.Due to the high time consumption of FFT for large data size,the FFT algorithm is accelerated based on CUDA.For the HMM model,the classical Forwarding algorithm,Viterbi algorithm and Baum-Welch algorithm are respectively analyzed in parallel.The training time of the mo del can be reduced by about 60%.Finally,a cross-language fault diagnosis software is realized based on the dynamic link library technique.This software system can monitor the operation of the fan and accurately identify the typical fan fault state.