Methodologies For Fault Diagnosis Of Rotary Machine Based On Deep Learning

For modern industrial production,it is crucial to ensure long-term safety and reliable operation of the mechanical equipment.As rotating machinery is the key component in the mechanical system,it is necessary to carry out health monitoring and fault diagnosis on rotating machinery to find out the potential safety hazards in time,and propose corresponding maintenance strategies.The key part of the mechanical fault diagnosis method is the fault feature extraction.The quality of the extracted features has an important impact on the final fault diagnosis result.Traditional fault diagnosis methods usually use artificial feature extraction,which requires certain background knowledge to achieve accurate fault-related feature selection.With the continuous development of artificial intelligence technology,intelligent diagnosis technology based on deep learning has been continuously applied to mechanical fault diagnosis.The deep learning model automatically learns from raw data to help accurate fault classification and reduce the impact of human participation.Aiming at overcoming the limitations of the existing traditional feature-based rotating machinery fault diagnosis,this thesis utilizes deep learning technology to achieve highlyaccurate fault diagnosis.The main research focus of this thesis includes the following four aspects:(1)In order to solve the uncertainty in traditional fault diagnosis methods caused by artificial feature extraction,a feature learning model based on deep belief network is proposed.Deep belief network has powerful feature learning ability that can learn features from the original sensor signal automatically and it is utilized to explore its applicability in the application of rotating machinery fault diagnosis.Feature extraction is combined with fault classification in one deep architecture,which reduces manual participation and realizes intelligent fault diagnosis.At the same time,an improved model is proposed,which has fast model convergence speed and high classification accuracy,and has certain robustness to model depth and hidden layer dimensions.The effectiveness of the system was verified in an induction motor experimental platform.(2)Aiming at realizing accurate fault diagnosis of rotating machinery and overcoming the limitation caused by single sensor signal,a fault diagnosis model using time-frequency distribution image-based convolutional neural network is proposed.The local feature extraction ability of convolutional neural network is used to capture the time-frequency image characteristics of sensor signals,and the fault mode based on time-frequency image is learned to achieve accurate prediction.Furthermore,a deep convolutional neural network-based multisignal fault diagnosis framework is proposed to improve the classification accuracy and the stability of the deep model.The high-level feature fusion model can independently learn features from different types of sensor signals.(3)For the current deep learning-based fault diagnosis methods,the model depth is limited,the feature learning ability is insufficient,and the model training is difficult.In order to improve the model performance,a novel deep transfer learning framework is developed to achieve highly-accurate machine fault diagnosis,which enables and accelerates the training of deep neural network.It gives the fault diagnosis model reasonable initialization and appropriate finetuning to accelerate model training.Compared with the traditional feature-based fault diagnosis model,this developed deep transfer learning model reduces the degree of human participation and improves the intelligence of the system.Compared with the machine learning-based fault diagnosis method,it overcomes the difficulty of training a large scale deep architecture.(4)For the problem of limited data samples in the field of rotating machinery fault diagnosis,in order to realize the effective training of deep model,an auxiliary classifier generative adversarial network-based framework is developed to learn from mechanical sensor signals and generate synthetic one-dimensional data.The generated data are used to augment the limited and unbalanced training data.The model is based on one-dimensional convolution network and such structure is able to learn hierarchical representations through convolution operation and easy to train.Batch normalization is performed within generator to avoid the problem of gradient vanishing during training,and category labels are used as the auxiliary information in this framework to help train the model.The approach is designed to produce realistic synthesized signals with labels and the generated signals can be used as augmented data for further applications in machine fault diagnosis.