Study On Intelligent Fault Diagnosis Method Of Fan Based On Time-frequency Analysis

Failure of the fan during working may cause major property loss and personal safety issues.The congruent relationship between fault and feature extraction method should be seriously considered when signal processing methods are used to extract features of fan faults.And it is difficult to accomplish online diagnosis because of the changeful fault feature extracting from signal in different working scenarios.Deep neural networks can effectively extract the deep features of fan faults.Efficient and stable status recognition and fault diagnosis can be realized with the usage of transfer learning methods.Researchers usually use the time-frequency spectrogram of the vibration signal gaining from running rotating machinery as the input datasets of the deep neural network.In order to reduce the memory usage of the fault diagnosis processor and improve the calculation efficiency,a method for fault identification and diagnosis using the vibration timing signal of the wind turbine is proposed in this paper,which provides a theoretical basis for optimizing the offline deployment of the fault diagnosis systemIn this paper,the time-frequency analysis method is used to perform vibration signal denoise and feature frequency extraction as preprocessing.Two types of deep learning methods are used to extract the deep feature of the preprocessed fan vibration signal data to realize fault diagnosis and classification.The main conclusions obtained from this study are as follows:(1)Based on the fan failure mechanism,the general form of the vibration signal of the fan failure is derived,the fan vibration signal model is constructed combined with the modulation effect during the operation of the fan,which lays the foundation for the subsequent research on the noise reduction and feature extraction methods of the vibration signal.Simulation experiments of the different types of fan faults have been designed and conducted.The single-domain analysis of the vibration signal obtained from a running fan shows that it is difficult to distinguish the type of the fan fault under the condition of strong background noise from merely time-domain or frequency-domain.(2)The preprocessing methods of fan vibration signal noise reduction and feature extraction are studied.The performance indicator system of signal noise reduction is established.The noise reduction performance of simulation signal is evaluated by SNR,correlation and SNR gain,and the noise reduction performance of measured vibration signal is evaluated by class separability.The influence of the number of decomposition layers on the noise reduction effect of the secondgeneration wavelet decomposition is researched.The influence factors of EMD-CIIT noise reduction method are researched.The noise reduction performance of the EMD-CIIT method is mainly affected by the selection of the threshold function and the modification mode of the first IFM.The signal has the best approximation to the real signal after using the hard threshold function to denoise it.The modification mode of the first modal component will affect the smoothness of the noise reduction signal.Contrast research of a variety of wavelet noise reduction methods and other noise reduction methods based on EMD is conducted.The denoised signal obtained after the noise reduction processing of the EMD-CIIT method has the best approximation to the real signal and is relatively smooth.Since the fan vibration signal usually contains a lot of noise,the EMD-CIIT noise reduction method is used to reduce the noise of the fan vibration signal.(3)Contrast research of the EMD and its optimization algorithm is conducted.The phenomenon of "modal aliasing" exists both in the IFMs calculated by the EMD and EEMD methods.And the result of the fan fault classification using the specific IFM is misadventure.And the CEEMDAN method can effectively extract temporal waveform data containing characteristic frequency bands,and the "modal aliasing" phenomenon is eliminated.(4)Two deep learning methods including long and short-term memory model and residual neural network are used to extract deep features of fan vibration signals to realize fan fault classification.The classification scheme that combines EMD-CIIT noise reduction processing and long and short-term memory model fault identification can achieve high-precision real-time fan fault diagnosis under the condition of single-channel data input.Fan fault classification combining noise reduction and CEEMDAN feature modal extraction preprocessing and residual neural network fault identification can perform offline fan fault classification under the condition of multi-channel data input,and the classification accuracy of the sample reaches to 100%.In this paper,signal noise reduction,feature extraction and deep neural network are combined to study fan operating status recognition and fault diagnosis.In order to deploy a fault diagnosis system under low memory conditions,and to achieve high-precision fault diagnosis under high background noise conditions,a fan fault diagnosis framework that directly processes vibration signals is constructed.The method proposed in this paper can be extended to the fault diagnosis of other industrial equipment,and provides technical support for the operation status recognition and fault diagnosis of various industrial equipment.