Research On Intelligent Diagnosis Of Machinery Equipment Based On Deep Transfer Learning

Rotating machines are widely used in petrochemical,aerospace,automobile manufacturing,rail transportation,wind power and other important engineering fields.It is necessary and important to timely and reliable monitor and detect abnormal behaviors and recognize faults of various components of rotating machinery,so as to ensure the stable operation,improve the operating efficiency of machines,and avoid the occurrence of major accidents.Deep Learning refers to deep neural network including multiply non-linear processing layers by simulating the hierarchical structure of human brain.Each layer can be regarded as one transformation at one level.Multiply layers are stacked together to learn hidden discriminative features from raw data.Transfer learning is a new learning technique.It can adapt the learned knowledge of model from multiply tasks,and transfer and apply it to new problems,so that the model has the ability to make inferences.Therefore,the thesis takes deep transfer learning as the core techniques to carry out the intelligent fault diagnosis of rotating machinery.There are some key issues,which are concerned including feature extraction and enhancement,few-shot learning,diagnosis under variable operating conditions,and multisensor information fusion.The research contents include the following four aspects:1)In rotating machines,the vibration signals generated by the defects in key components are weak,and the features are difficult to extract.Therefore,a feature enhancement method integrating Convolutional Neural Network(CNN)and subspace transfer learning technique is proposed.The 2D Cyclic Spectral Coherence(CSCoh)maps of vibration signals are estimated by cyclic spectral analysis to provide discriminative patterns for specific type of faults.In addition,a CNN model is constructed to learn high-level feature representations and conduct fault classification.More specifically,Group Normalization(GN)and CORrelation Alignment(CORAL)unsupervised adaptive layers are designed and employed in CNN to automatically adjust deep features of CNN,which is helpful to improve the classification capability of deep model.Experiments are conducted on different bearing cases.Results demonstrate that the proposed method can achieve better feature representations and higher diagnosis accuracy.2)For a given fault diagnosis task,deep neural network diagnosis model easily overfits on the small training data and leads to poor performance on the testing data.In an effort to deal with such problems,a Transferable CNN(TCNN)method is proposed to improve the performance of few-shot diagnosis issues.In the proposed method,raw time-series signals,collected from source domain datasets are directly used to train the designed TCNN without any hand-crafted feature extraction.Then,based on the model transfer scheme,TCNN obtained from the pre-trained model can be adopted to transfer source domain knowledge to new target domain tasks,improving the diagnosis performance.Two case studies across different operation conditions and test rigs are considered and investigated.Results show that the proposed method exhibits better classification performance and less computational cost.3)To reduce the domain shift problem between training and testing data in rotary machinery,a novel Domain Adversarial Transfer Network(DATN)is developed for handling large distribution discrepancy across domains.Two asymmetric encoder networks integrating deep convolutional neural networks are designed.The weight transfer strategy and domain adversarial training technique are introduced to minimize the difference between source and target distributions,which improves the diagnosis performance.Furthermore,when a large set of source data classes are available and target data only cover a subset of classes,a Weighted Adversarial Transfer Network(WATN)is proposed for partial domain fault diagnosis.Adversarial training is introduced to learn both class discriminative and domain invariant features,and a weighting learning strategy is adopted to weigh the contributions.As such,the irrelevant source examples can be identified,and the distribution discrepancy of shared classes between domains can be reduced.Experiments on bearing and gearbox datasets demonstrate that the proposed DATN and WATN achieve satisfactory performance.4)To effectively utilize multi-sensor information in industrial test systems,a multi-sensor information fusion method based on deep transfer network is proposed.Firstly,the effectiveness is preliminary validated by adopting Sparse Auto Encoder(SAE)for feature-level fusion of data from multiple sensors,and Deep Belief Network(DBN)for classification.Furthermore,by considering that the training and test data may come from multiple sensors,resulting in the data distribution discrepancy,an extension of multi-sensor decision-level fusion using deep transfer network is developed.It can make full use of multiple sensor information,and match feature distribution between target domain and multiple source domain data.Finally,a decision-level fusion is implemented to effectively improve generalization ability of the diagnostic model.The proposed scheme can effectively improve diagnosis accuracy and generalization ability in dealing with issues including feature extraction and enhancement,few-shot learning,diagnosis under variable operating conditions and multi-sensor information fusion.Additional,the proposed scheme also expands the application scope of existing diagnostic methods,providing new solutions and ideas for fault diagnosis of machinery in complex environments.