| Gearbox is the key part of mechanical equipment to transfer power,which plays an important role in connecting,supporting and transmitting power.Once the fault occurs,it will affect the safety and reliability of the whole mechanical system,and even bring huge economic losses or casualties to enterprises and national production.Therefore,it is of great significance to develop the research of gearbox fault diagnosis and condition monitoring technology for effectively ensuring the safe and reliable operation of the system and avoiding the occurrence of safety accidents in industrial production.In this thesis,the high-accuracy intelligent fault diagnosis of gearbox is taken as the research subject,and the deep learning algorithm is taken as the core technical means,deeply studying several key technical problems existing in the intelligent fault diagnosis of gearbox.The main contents of this thesis includes:(1)Research on semi-supervised intelligent fault diagnosis technique of gearbox.This thesis analyzes the issues existing in the fault diagnosis models based on supervised learning and unsupervised learning,and proposes a semi-supervised intelligent fault diagnosis method for gearbox based on wavelet transform(WT)and generative adversarial nets(GANs).The proposed method involves two parts.In the first part,WT is adopted to transform onedimensional raw vibration signals into two-dimensional time-frequency images.In the second part,the labeled and unlabeled time-frequency images are inputted into the built adversarial learning model to realize fault diagnosis of the gearbox.Compared with fault diagnosis methods based on supervised learning,the proposed approach does not need to label all training samples,and high-accuracy fault diagnosis of gearbox can be achieved by labeling a small number of training samples.Compared with fault diagnosis methods based on unsupervised learning,the accuracy of our proposed method is more excellent.Finally,two case studies are implemented to verify the proposed method.(2)Research on the compound fault diagnosis technique of gearbox.At present,most of the intelligent fault diagnosis technique of gearbox is only applicable to single failure modes,but it is not applicable to the compound fault modes of gearbox with the failure of multiple parts at the same time.Therefore,we propose a novel compound fault diagnosis method of gearbox based on WT and multi-label convolutional neural network(MLCNN).The proposed method involves three parts.In the first part,WT is adopted to transform one-dimensional raw vibration signals into two-dimensional time-frequency images.In the second part,by combining with the multi-label classification,convolutional neural network(CNN)is improved and can be used to deal with the compound faults of gearbox.Finally,the time-frequency images are inputted into the built MLCNN model to realize compound fault diagnosis of the gearbox.This method combines wavelet transform,multi-label classification and convolution neural network,and makes full use of their characteristics and advantages.The effectiveness of the proposed method is verified by two case studies.(3)Research on data augmentation and intelligent fault diagnosis technique of gearbox.Fault diagnosis methods based on deep learning heavily rely on a large amount of training data,but the data in real industrial application is limited and unbalanced.Therefore,we propose a novel fault diagnosis of gearbox based on Stockwell transform(ST),data augmentation generative adversarial nets(DAGANs)and capsule neural network(Caps Net).The proposed method involves three parts.In the first part,ST is adopted for extracting time-frequency image features from one-dimension raw vibration signals of gearbox.In the second part,DAGANs are employed to generate more training image samples.Finally,the original training timefrequency images and the generated fake training time-frequency images are input into the built Caps Net model to accomplish the fault diagnosis of gearbox.Its effectiveness is verified by the experiment.(4)Research on automatic feature extraction and fault diagnosis technique of gearbox.For traditional intelligent fault diagnosis approaches,it is necessary to extract and select features manually,which will heavily rely on expert knowledge in the domain of signal analysis,increasing the cost of fault diagnosis.In addition,fault diagnosis performance can decline greatly if these manually selected features are inadequate.Therefore,we propose an automatic feature extraction and intelligent fault diagnosis method based on two-stream convolutional neural network(TSCNN)and extreme learning machine(ELM).The proposed method involves two parts.In the first part,TSCNN is adopted for extracting fault features of gearbox.In the second part,ELM is employed to realize the classification of fault features.This method can not only achieve higher accuracy of fault diagnosis,but also has adaptive ability.It can automatically extract the fault features of gearbox from the original vibration signal,and reduce the dependence on professional knowledge and engineering experience of technicians.Finally,the effectiveness of this method is verified by three case studies. |
