| In the field of machinery,with the increasing application of computer technologies,sensor technologies and communication technologies,major rotating machinery such as aero-engines,large-scale wind power equipment,and steam turbine generator sets are increasingly moving toward high speed,high efficiency,high precision,integration and intelligence.In order to eliminate safety hazards to the maximum extent and ensure efficient and safe operation of mechanical equipment,a reliable health monitoring system must be established to realize real-time monitoring and safety warning for key components such as bearings,planetary gearboxes in rotating machinery equipment.However,due to the large scale of modern equipment groups,the large number of measuring points required,the high sampling frequency of a single measuring point and the long span of data acquisition,the monitoring data presents the characteristics of "Big data",such as huge volume,rapid generation,various modes,multi-source heterogeneity,high value but low density,etc.,In this background,it has been difficult to meet the requirements for diagnostic accuracy and efficiency for mechanical equipment in practical engineering by relying only on conventional fault diagnosis techniques.The paper mainly is aimed at how to improve the reliability,diagnostic accuracy and efficiency of rotating machinery equipment fault diagnosis under the background of mechanical "Big data",the related research on automatic feature extraction and intelligent fault diagnosis for rotating machinery is carried out from the perspective of deep learning application.Based on the traditional signal analysis,by studying the characteristics of each representation domain for fault signal,it is proved that the time-frequency domain is more helpful for the deep learning model to pick up the effective features.The discrete wavelet transform is used to reveal time-frequency domain information for fault signal,convolutional neural network is used to extract feature efficiently from the time-frequency domain constructed;By studying the fault characteristics of bearing,the WPT-PWVD time-frequency domain information construction method is proposed,and the deep convolutional neural network is further constructed and used to learn the fault characteristics;By combining the information fusion technology with the deep learning theory,the advantages of the residual convolutional neural network are used to extract the deep features of the multi-channel and multi-evidence domain for fault feature fusion,thus a diagnostic method for deep feature fusion based on multi-source information is proposed.The research content of the thesis is as follows:(1)Aiming at the problem about the adaptive and generalization ability of fault features extracted manually is weak and it is difficult to match the specific fault state of rotating machinery,thus a fault diagnosis method for planetary gearboxes based on convolutional neural network and discrete wavelet transform is proposed.Firstly,different signal processing methods are used to characterize the faults from three different representation domains in time domain,frequency domain and time-frequency domain,and analyze the quality difference for revealing the fault information among three representation domains.By analyzing and comparing,the fault signal is processed by discrete wavelet transform to reveal the frequency domain characteristics;then deep learning theory knowledge is implemented to construct Convolutional Neural Network(CNN),and the advantages of convolutional neural networks in feature extraction and pattern recognition is used to pick fault feature distribution from the constructed time-frequency domain;Finally,Softmax multi-classifier is constructed and Backpropagation(BP)algorithm is used to adjust network parameters layer by layer to gradually establish the mapping relationship between fault feature space and fault space,revealing fault internal pattern and generating appropriate Classifier.The effectiveness of the proposed method is verified by using fault data of planetary gearboxes.(2)To solve the problem that the learning effect which the deep learning models learn feature from different time-frequency domain representation is easy to fall into the bottleneck prematurely due to the limitation of different time-frequency domain analysis method in fault signal representation ability,thus a fault diagnosis method for bearing based on deep convolution neural network and WPT-PWVD is proposed.First,the characteristics of the bearing fault signal are analyzed,and wavelet package decomposition(WPT)with the merits of adaptive narrow-band filtering and full-frequency domain resolution is used to decompose the fault condition signal to reveal information in different frequency bands.Meanwhile,multiple high-frequency signal components which the energy is concentrated are extracted,the background noise interference is reduced,and the fault signal impact information is obtained.The several extracted high-frequency signal components are reconstructed separately and the reconstructed signal is demodulated separately by Hilbert algorithm to reveal the low-frequency fault information.Pseudo Wigner-Ville distribution(PWVD)is applied to each demodulated signal components,and the calculation results are accumulated.Then the time-frequency domain representation of fault signal WPT-PWVD without cross-term and high resolution is obtained.The Deep Convolution Neural Network(DCNN)with stronger feature learning ability is further constructed by using the deep learning theory knowledge,and used to learn fault feature from WPT-PWVD time-frequency domain representation.The proposed method is applied to a variety of bearing fault data,and the effectiveness of the proposed method is verified.(3)To address the problem that the traditional information fusion method divides the fault feature extraction and information fusion algorithm,which makes the information fusion matching degree difficult to evaluate,a fault diagnosis method for planetary gearboxes based on multi-source information deep fusion is proposed.In this paper,the vertical and horizontal sensors mounted on rotating machinery are used to obtain the multi-channel signals during the operation of the rotating machinery to describe its overall operating condition.The acquired signal is processed by using WPT time-frequency analysis method and Ensemble Empirical Mode Decomposition(EEMD)to reveal the multi-evidence domain information of the fault signal.Deep learning theory is used to construct multiple Deep Residual Convolution Neural Network(DR-CNN),and used separately to learning feature from each evidence domain to establish a mapping between local evidence space and fault space.The deep fault features of each evidence domain are extracted from each deep residual network,and the evidence set of the feature space is constructed.Using the Random Forest(RF)to fuse the feature evidence set,establishing the relationship between the global feature space and the fault space,and thus making full use of the fault multi-evidence domain information from multiple aspects.Finally,the fusion diagnosis result is obtained.The proposed method is applied to the fault data of planetary gearbox,and a decision criterion consistent with its operation state is obtained.At the end of thesis,this work is summarized and looks forward the follow-up research direction. |
PDF Full Text Request