Research On Fault Diagnosis Of Rotating Components Based On Deep Learning

With the increasingly large,high speed,integration and intelligence of modern industrial equipment,the requirements for the reliability and safety of the equipment are also increasingly high.As one of the key technologies to maintain the safety and reliability of equipment,fault diagnosis aims to quickly and accurately determine the fault type and location,reducing the cost of equipment operation and maintenance.Data-driven fault diagnosis technology is the mainstream technology in modern times.It can automatically learn fault patterns from historical big data,and has the advantages of simple modeling,high accuracy,and many applicable scenarios.Among them,the emerging deep learning technology in recent years has more powerful feature extraction ability and big data processing ability.Compared with traditional fault diagnosis technology,deep learning technology can realize end-to-end intelligent fault diagnosis in more complex scenarios.Although deep learning has achieved fruitful results in the field of fault diagnosis,there are still several problems to be solved:(1)Deep learning model is a black box model whose internal principles are difficult to explain.(2)It is difficult to predict the transfer learning effect under multiple working conditions,and the essential application and mining depth of vibration signals are not enough.(3)The design of deep learning model requires a lot of professional experience,and the super parameters space is huge,so the model design is time-consuming.All the above problems limit the practicability and performance improvement of deep learning.In order to solve the above problems,this paper mainly studies the following aspects:(1)We explains the convolutional neural network from the perspective of feature visualization and class activation mapping,and preliminarily reveals the working principle of convolutional neural network in fault diagnosis.Through feature visualization,it can be clearly observed that features of the same category gradually gather together with the deepening of the network,features of the different categories diverged gradually.Through the method of class activation diagram,the basis of classification can be marked in time domain and frequency domain,namely abnormal vibration in real time domain and abnormal frequency band in frequency domain.(2)In the supervised transfer learning scenario under multiple working conditions,a fine-tune transfer learning method is proposed to realize homogeneous and heterogeneous transfer learning between two different components.At the same time,the transfer prediction problem is proposed to predict the transfer learning effect under the given base model and target domain data.The correlation between the predicted results and the actual results is up to 0.9927 for the gearbox data set under multiple working conditions.(3)In the unsupervised transfer learning scenario under multiple working conditions,based on the characteristics of vibration signals and the frequency domain characteristics of the convolution kernel,a general convolutional neural network structure,frequency-domain fusion convolutional neural network(FFCNN),is proposed,which can be matched with different domain adaptation loss functions.Compared with the original network structure,the average accuracy of FFCNN on the two bearing datasets can be up to 2.83% higher.(4)The method of neural architecture search is applied to fault diagnosis,and the network architectures suitable for general classification and unsupervised transfer learning are searched in the huge search space respectively.For common classification tasks,a one-shot model containing all search spaces is trained,which can predict the performance of each candidate structure.Then genetic algorithm is used to search for the optimal network structure.Compared with the benchmark model,the accuracy of the model searched on the gearbox dataset improved by 7.33%.For unsupervised transfer learning,differentiable method is used to make the search space continuous,and gradient descent is used to search the network structure and feature mask.Compared with the benchmark model,the average accuracy of the model searched on the two bearing datasets was up to 5.76%.Finally,this paper also developed a fault diagnosis platform based on the Internet of Things,deployed the deep learning model to the edge computing node,communicated with the cloud Web system,and realized the real-time monitoring and management of fault diagnosis.The research shows that the methods proposed in this paper have obvious effects on the interpretation of neural network,the prediction and improvement of the accuracy of transfer learning,and the automatic design of network architectures,etc.,which have guiding significance for the application of deep learning in fault diagnosis...